Received: 22 November 2022 Revised: 18 May 2023 Accepted: 2 June 2023

DOI: 10.1002/alz.13412

REVIEW ARTICLE

Artificial intelligence for diagnostic and prognostic

neuroimaging in dementia: A systematic review

Robin J. Borchert1,2 Tiago Azevedo3 AmanPreet Badhwar4,5

Jose Bernal6,7,8 Matthew Betts7,8,9 Rose Bruffaerts10,11
Michael C. Burkhart12 Ilse Dewachter11 Helena M. Gellersen8,12
Audrey Low13 Ilianna Lourida14 Luiza Machado15
Christopher R. Madan16 Maura Malpetti1 Jhony Mejia17
Sofia Michopoulou18 Carlos Muñoz-Neira19,20 Jack Pepys1,21
Marion Peres1 Veronica Phillips22 Siddharth Ramanan23
Stefano Tamburin24 Hanz M. Tantiangco25 Lokendra Thakur26,27,28
Alessandro Tomassini23 Ashwati Vipin29 Eugene Tang30
Danielle Newby31 The Deep Dementia Phenotyping (DEMON) Network
Janice M. Ranson14 David J. Llewellyn14,32 Michele Veldsman33
Timothy Rittman1

Correspondence
Robin J. Borchert, Department of Clinical Abstract
Neurosciences, University of Cambridge,
Herchel Smith Building, Forvie Site, Robinson
Introduction: Artificial intelligence (AI) and neuroimaging offer new opportunities for
Way, Cambridge Biomedical Campus, diagnosis and prognosis of dementia.
Cambridge, CB2 0SZ, UK.
Email: rb729@[Link]
Methods: We systematically reviewed studies reporting AI for neuroimaging in
diagnosis and/or prognosis of cognitive neurodegenerative diseases.
Funding information
Results: A total of 255 studies were identified. Most studies relied on the Alzheimer’s
Alzheimer’s Research UK; National Institute
for Health and Care Research (NIHR); National Disease Neuroimaging Initiative dataset. Algorithmic classifiers were the most com-
Institute for Health Research (NIHR);
monly used AI method (48%) and discriminative models performed best for differ-
Alzheimer’s Research UK and the Alan Turing
Institute/Engineering and Physical Sciences entiating Alzheimer’s disease from controls. The accuracy of algorithms varied with
Research Council, Grant/Award Number:
the patient cohort, imaging modalities, and stratifiers used. Few studies performed
EP/N510129/1; Medical Research Council,
Grant/Award Number: MR/X005674/1; validation in an independent cohort.
National Health and Medical Research Council
Discussion: The literature has several methodological limitations including lack
(NHMRC); National Institute on
Aging/National Institutes of Health, of sufficient algorithm development descriptions and standard definitions. We
Grant/Award Number: RF1AG055654
make recommendations to improve model validation including addressing key
clinical questions, providing sufficient description of AI methods and validat-
ing findings in independent datasets. Collaborative approaches between experts

This is an open access article under the terms of the Creative Commons Attribution License, which permits use, distribution and reproduction in any medium, provided
the original work is properly cited.
© 2023 The Authors. Alzheimer’s & Dementia published by Wiley Periodicals LLC on behalf of Alzheimer’s Association.

Alzheimer’s Dement. 2023;19:5885–5904. [Link]/journal/alz 5885

 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5886 BORCHERT ET AL .

in AI and medicine will help achieve the promising potential of AI tools in

practice.

KEYWORDS
artificial intelligence (AI), Alzheimer’s disease, dementia, machine learning (ML), neurodegenera-
tive diseases, neuroimaging

Highlights
∙ There has been a rapid expansion in the use of machine learning for diagnosis and
prognosis in neurodegenerative disease
∙ Most studies (71%) relied on the Alzheimer’s Disease Neuroimaging Initiative
(ADNI) dataset with no other individual dataset used more than five times
∙ There has been a recent rise in the use of more complex discriminative models (e.g.,
neural networks) that performed better than other classifiers for classification of AD
vs healthy controls
∙ We make recommendations to address methodological considerations, addressing
key clinical questions, and validation
∙ We also make recommendations for the field more broadly to standardize
outcome measures, address gaps in the literature, and monitor sources of
bias

1 INTRODUCTION ment using features such as medial temporal lobe atrophy18 and white
matter hyperintensity load.19,20 However, the development of more
There is a pressing need to improve diagnosis and prognosis for people sophisticated approaches and richer data may mean that the most
with dementia. Up to 20% of people may receive the wrong diagnosis,1 informative features are not amenable to human measurement or
and differentiating between early symptoms in dementia based on observation. For example, resting-state functional MRI can be used to
clinical information and neuropsychological testing alone is subjective derive a variety of connectivity metrics between 1000s of nodes that
and prone to error. There is large geographic variability in the likeli- are amenable to machine learning (ML) approaches.21 Deep learning
hood of receiving a diagnosis, even within a single country.2 Diagnostic methods have also demonstrated superiority to human neuroimaging
investigations such as neuroimaging and cerebrospinal fluid (CSF) tests interpretation.22,23
can support clinical diagnosis; however it can take years to receive a ML algorithms facilitate the automation of neuroimaging interpre-
diagnosis from the initial onset of symptoms.3 Receiving a timely and tation and have the potential to reduce bias and improve clinical
accurate diagnosis is critical for people with dementia, their carers, and decision making.24–26 Neuroimaging data are particularly well-suited
families:4,5 it provides the opportunity for forward planning; and with to analysis using ML, particularly deep learning, given its high dimen-
the advent of disease modifying treatments an early accurate diagno- sionality, non-linear nature and high covariance within the data. A
sis will guide treatment selection, working toward a precision medicine large and growing number of ML studies have investigated how neu-
approach.6 roimaging features can be used to predict cognitive diagnoses and
Neuroimaging is a non-invasive investigation used in routine clinical conversion to dementia, fueled by the availability of large datasets,
practice to support the diagnosis of dementia.7,8 A range of neu- such as the Alzheimer’s Disease Neuroimaging Initiative (ADNI).27
roimaging methods are used in dementia and magnetic resonance However, uncertainty remains about which ML approaches have the
imaging (MRI) is one of the most widely used to examine brain greatest potential to inform clinical decision making and how their
structure,9,10 longitudinal patterns of atrophy,11 and changes in brain performance compares to human decision making.
function.12–14 Positron emission tomography (PET) is available in spe- We therefore conducted a systematic review to establish: (1) the
cialist centers and is more expensive; it is used to measure metabolic extent to which ML approaches for neuroimaging have been used for
activity, or using protein-specific ligands to identify underlying the diagnosis and/or prognosis of neurodegenerative diseases; (2) how
pathologies.15–17 this field has progressed over time; (3) methodological challenges; and
Human clinical judgment has traditionally been used to interpret (4) the future directions to facilitate the translation of ML methods for
clinical neuroimaging.9 Visual rating scales may support this assess- patient benefit in dementia.
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5887

This review is part of a Special Issue on “Artificial Intelligence for

Alzheimer’s Disease and Related Dementias” published in Alzheimer’s RESEARCH IN CONTEXT
& Dementia. Together, this series provides a comprehensive overview
1. Systematic Review: We conducted comprehensive
of current applications of artificial intelligence (AI) to dementia, and
searches of MEDLINE, Embase, Cochrane Library, BNI,
future opportunities for innovation to accelerate research. Each
PsycINFO, CINAHL, and Emcare to identify studies that
review focuses on a different area of dementia research, including
examine the potential of artificial intelligence (AI) and
experimental models28 , drug discovery and trials optimization29 ,
machine learning methods applied to neuroimaging to
genetics and omics30 , biomarkers31 , neuroimaging (this article),
inform clinical diagnosis and prognosis in dementia and
prevention32 , applied models and digital health33 , and methods
other neurodegenerative diseases.
optimization34 .
2. Interpretation: The use of AI in neuroimaging is expand-
ing rapidly with the evidence base being dominated by
studies conducted using the ADNI dataset, algorithmic
2 METHODS
classifiers, and structural MRI focusing on Alzheimer’s
disease. Improved diagnostic accuracy was observed
We conducted a systematic review to investigate the use of ML meth-
when a combination of neuroimaging modalities was
ods for diagnosis and/or prognosis in cognitive disorders including
used, e.g., PET and structural MRI. Findings also suggest
Alzheimer’s disease (AD), mild cognitive impairment (MCI), Parkin-
superior performance of discriminative models compared
son’s disease (PD), vascular dementia, Lewy body dementia (LBD),
to algorithmic and generative classifiers for the classifica-
frontotemporal dementia (FTD), progressive supranuclear palsy (PSP),
tion of Alzheimer’s disease vs healthy controls.
Huntington’s disease (HD) and corticobasal degeneration (CBD). The
3. Future Directions: We highlight gaps in knowledge, cur-
review is reported according to PRISMA (Preferred Reporting Items
rent challenges, and issues to be addressed in future
for Systematic Reviews and Meta-Analyses) guidelines,35 and the pro-
research around reproducibility and reporting, relevant
tocol was registered with PROSPERO (ID: CRD42021232249) prior to
clinical questions, and validation of results. We advo-
the screening of abstracts.
cate wider collaboration between clinical, neuroimaging,
and data science teams, and present recommendations to
move toward clinically useful, machine learning methods
2.1 Search strategy
applied to neuroimaging for dementia.

The databases MEDLINE (via Ovid), Embase (via Ovid), Cochrane

Library, BNI (via ProQuest), PsycINFO (via EBSCOhost), CINAHL (via
EBSCOhost), and Emcare (via Ovid) were searched using the title,
abstract, keyword, and MeSH term fields from inception to January 2.3 Inclusion & exclusion criteria
8, 2021, with the support of the Cambridge University Clinical School
Library. Results were limited to English language studies. Full search The inclusion and exclusion criteria used during the screening process
terms for each database can be found in Supplementary Material to determine which studies would be included in the systematic review
1. Studies which were known to the authors and met the inclu- can be found below:
sion/exclusion criteria of the review, but were not initially identified Inclusion criteria:
using the search strategy, were also included.
1. Primary research studies only.
2. Patient population consisting of AD, MCI, PD, vascular dementia,
2.2 PICOS framework LBD, FTD, PSP, CBD, HD, and/or all-cause dementia.
3. Involving at least one of the following neuroimaging or neurophysi-
Outline of the parameters of this systematic review according to the ological modalities: structural or functional MRI, PET, single-photon
PICOS framework: computed tomography (SPECT), electroencephalogram (EEG), mag-
netoencephalography (MEG), or ultrasound.
∙ Participants: Patients with cognitive disorders due to neurodegen- 4. Used ML methods to investigate diagnosis and/or prognosis of
erative diseases. cognitive neurodegenerative disease(s).
∙ Index: Neuroimaging data assessed with ML for diagnosis and/or
prognosis. Exclusion criteria:
∙ Comparator: Traditional manual/subjective diagnostic/prognostic
assessment. 1. Studies which did not include human participants.
∙ Outcome: Accuracy of diagnosis and/or prognosis. 2. Studies published in languages other than English.
∙ Study design: Controlled study. 3. Conference abstracts and book chapters.
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5888 BORCHERT ET AL .

4. Articles which did not include primary research, for example, 3. ML methods, extracted neuroimaging features.
reviews. 4. Receiver-operator curve (ROC) analysis results from the ML algo-
5. Studies where access to the full text was not available despite rithm used to predict diagnosis/prognosis in the patient population,
attempts from multiple individuals involved in the screening pro- including accuracy (ACC), sensitivity (SEN), specificity (SPE), area
cess. under the curve (AUC), positive predictive value (PPV), and/or
6. Studies which did not use ML methods or only used simple logistic negative predictive value (NPV).
or linear regression methods for classification.
7. Studies which combined neuroimaging with other biomarkers,
including CSF markers and/or genetics data, in the ML algorithms 2.6 Risk of bias assessment
without reporting of model performance for neuroimaging features
without these additional biomarkers. Following the second stage of screening, all included studies were
8. Studies which focused on automated segmentation techniques assessed for risk of bias by one reviewer using a hybrid version of
which did not directly relate to diagnosis/prognosis of neurodegen- the Joanna Briggs Institute (JBI) Critical Appraisal checklist covering
erative diseases. the areas we deemed most relevant to this area of research.36 The
9. Studies which used AI methods for feature extraction but not specific questions used for risk of bias assessment and their outcome
classification. for each study can be found in Supplementary Material 2. We only
excluded studies exhibiting clear methodological concerns, such as lack
of reporting of basic participant demographics, in order to accurately
2.4 Study selection depict and identify current barriers in the literature limiting translation
to clinical practice.
The initial records were identified using the search criteria. These
records underwent de-duplication using a Zotero ([Link]
automation tool, which flagged possible duplicate studies, and were 2.7 Data synthesis and approaches to
manually screened by a reviewer to merge genuine duplicates. Fol- classification
lowing de-duplication, all studies were screened across two stages.
During the first stage, each abstract was independently reviewed by We used descriptive statistics to determine the following character-
two reviewers to determine their eligibility for inclusion based on the istics of the extracted dataset: source of neuroimaging data, type of
outlined criteria using the screening tool Rayyan ([Link] neuroimaging used, ML methods, focus on diagnosis and/or prognosis,
ai/). Once both reviewers screened their allocated abstracts, inclu- accuracy of diagnostic/prognostic classifications, and global distribu-
sion/exclusion decisions were unblinded. For abstracts where there tion of first authors’ institutions. Studies using MRI were labeled
was disagreement between screeners, a third independent reviewer according to the types of features used for the classification task
assessed the abstract and made the final decision as to (1) progression including volumetric structural, non-volumetric structural, and func-
to full-text screening stage or (2) exclusion. tional MRI. Volumetric structural imaging was defined as MRI methods
The second stage involved full-text screening of all included studies measuring the volume of specific regions using voxel-based segmen-
by one reviewer per paper. For studies where the reviewer was unsure tation techniques. Studies were classified as using non-volumetric
if the study met the outlined criteria, a second opinion was sought and structural MRI if the features used for classification were related to
a joint decision made after discussion with the second reviewer. cortical thickness, texture, or surface area using T1- or T2-weighted
images and/or diffusion tensor imaging (DTI) data. The type of AI
algorithm used for the diagnostic/prognostic classification task was
2.5 Data extraction extracted. Studies which used AI methods for feature extraction but
not classification were excluded.
One reviewer per paper manually collected data from each report inde- Given a training set of labeled features, there are multiple ways to
pendently into an Excel spreadsheet without the use of automation learn a classifier that can then be used to predict class membership for
tools. The following data were extracted from the included studies: new, unlabeled instances. We categorized classifiers according to the
object they seek to learn or model.
1. Article information: First author, year, journal, country of first
author’s affiliated institution. 1. Generative classifiers learn the joint distribution of the features
2. Study method: Patient population(s), neuroimaging modality, and labels.37 Examples include naïve Bayes and linear/quadratic dis-
source of data. For studies using different datasets relating to a criminant analysis. After training, it is possible to generate (hence
study, information regarding which specific dataset was extracted the name) new pairs of features and labels by sampling from the
where possible. For example, for ADNI studies, the specific dataset learned joint distribution.
used (ADNI-1, ADNI-2, ADNI-GO, J-ADNI) was identified and 2. Discriminative classifiers learn the conditional distribution of
recorded where available. the labels given the features.38 Examples include logistic and
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5889

Gaussian process regression with potential regularization, k- 3.1 Datasets

nearest neighbors, and most ensemble methods (such as random
forests). Few studies used more than a single dataset, with 233 studies using one
3. Non-probabilistic, algorithmic classifiers directly learn the decision dataset, 18 used two datasets, and the remainder used three or more
boundary in feature space.39 Examples include maximum margin datasets. The most commonly used dataset was ADNI (see Figure 2). In
classifiers and support vector machines. the majority of the studies using data from ADNI, the specific cohort
used (ADNI-1, ADNI-2, ADNI-GO, J-ADNI) was not stated (129 of 181)
We note that some non-probabilistic classifiers can be reframed in (Table 2 in Supplementary Material 1). Where the cohort was available
a probabilistic light.40,41 For this reason, some authors consider these (n = 52), 36 (69.2%) studies used a single cohort, 8 (15.4%) used two
methods to be discriminative in nature and draw less of a distinction cohorts, and 8 (15.4%) used three cohorts. Of those that used ADNI-
between our types (2) and (3). 2 and ADNI-GO (n = 11), a majority (n = 9) also used ADNI-1. Apart
In order to determine how well a classifier generalizes to new data, from using the ADNI dataset alone, 19 studies used data from ADNI
models are typically evaluated using a validation set consisting of combined with other datasets including the UK Biobank and AIBL. The
labeled data withheld from the training process. The model’s predic- majority (n = 11) of these combination studies used a local dataset in
tions in the validation data can be compared to known labels using addition to the ADNI dataset.
a variety of different metrics; precision, recall, accuracy, AUC, and F-
scores are all estimated in this way. If a classifier performs much better
on training data than on validation data, this can indicate overfitting. 3.2 AI methods
In such a case, the model may be refitted with regularization terms or
priors that penalize model complexity. The classifier type most frequently used was a non-probabilistic
Following data extraction, we conducted a meta-analysis. Consid- algorithmic approach (48%), an example of which is support vec-
ering the large number of studies from a single cohort and significant tor machines (SVM), followed by discriminative classifiers (32%)
overlap of datasets, there is a risk of identifying spurious associa- which includes most neural networks. Generative classifiers and
tions and false-positive findings when running a comprehensive meta- “other” methods, mainly consisting of studies which combined
analysis.42–44 We attempted to overcome these barriers by running multiple AI algorithms to generate novel or complex classifica-
a focused evaluation of the performance of ML algorithms, mea- tion tools were difficult to categorize; each constituted 10% of
sured with AUC values, for a specific task: classification of AD versus the literature. Most of these studies focused heavily on com-
healthy controls. This was achieved using a Stratified Weighted Ran- putational methods which are not easily accessible to a clinical
dom Method (SWRM) approach by assigning weights to the datasets audience.
and features (see further methodological details in Supplementary The number of studies which used algorithmic classifiers (mainly
Material 1). SVM) increased considerably between 2013 and 2015, after which its
use stabilized. In contrast, there was a sharp rise in the number of stud-
ies using discriminative approaches (e.g., neural networks) starting in
3 RESULTS 2017, with discriminative studies outnumbering algorithmic studies for
the first time in 2019 (Figure 3).
The initial search strategy yielded 2709 studies, which underwent In order to unveil potential differences in performance between ML
abstract screening following de-duplication. Three additional studies methods, we examined AUC values for classifying AD versus healthy
which were not picked up in the initial search strategy but met the controls across studies (Figure 4). Of note is that only 13% (11 of 84)
inclusion criteria were identified by experts in the field and underwent of these studies reported a confidence interval for the AUC value. Of
full-text screening. The studies were consolidated to 255 studies after these 11 studies, 5 did not report the range of the confidence interval
full-text screening (full list of references in Supplementary Material (e.g., 90% or 95%).
3). A flow chart of the screening process reported according to the We employed a meta-analytic approach using the stratified
PRISMA 2020 guidelines35 is shown in Figure 1. The publication time weighted random method (SWRM) to weigh results based on the
period ranged from 2005 to 2021. The included studies were classified dataset, imaging modality, and type of ML method used (method-
by country based on the institutional affiliation of the first author. The ological details in the Supplementary Material 1). We found that for
most common countries included China (26%), USA (17%), Italy (7%), classification of AD versus healthy controls (i) discriminative models
France (6%), and South Korea (6%). (SWRM = 3.39, RSD = 0.948, Heterogeneity = Considerable) per-
Risk of bias assessment resulted in exclusion of three studies which formed better compared to algorithmic (SWRM = 2.42, RSD = 0.758,
exhibited clear methodological concerns, such as lack of reporting of Heterogeneity = Substantial) and generative (SWRM = 2.14,
basic participant demographics (supplementary material 2). The major- RSD = 0.784, Heterogeneity = Moderate) classifiers; and (ii) each
ity of studies used clearly defined inclusion criteria (95%) with detailed R table expected to have 49 rows but has in the range of 6-8, which
descriptions of participants and settings (91%). Only 41% of studies indicates that most of the literature was limited to only few datasets
explicitly identified potential confounding factors. and imaging modalities.
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5890 BORCHERT ET AL .

F I G U R E 1 PRISMA (Preferred Reporting

Items for Systematic Reviews and
Meta-Analyses) 2020 flow diagram for
systematic review outlining the number of
studies identified and excluded at each stage.

F I G U R E 2 Datasets used across included studies. The majority of studies (n = 181, 71.0%) used the ADNI dataset alone or in combination with
another dataset. Local data were used in 69 (27.1%) studies. Multiple studies used a combination of two datasets or more resulting in an overlap
between the categories listed here. ADNI = Alzheimer’s Disease Neuroimaging Initiative, OASIS = Open Access Series of Imaging Studies,
AIBL = Australian Imaging, Biomarker & Lifestyle study of ageing, Bdx-3C = Bordeaux 3 Cities study, BLSA = Baltimore Longitudinal Study of
Aging, CADDementia = Computer-Aided Diagnosis of Dementia challenge, NACC = National Alzheimer’s Coordinating Center.

We identified four studies which used transfer learning for ing was typically used for fine tuning neural networks, particu-
classification45–48 which were trained on ImageNet45 ADNI (nor- larly when the authors felt the dataset was not sufficiently large
mal controls and AD),46 Human Connectome Project (HCP),47 and enough to properly train the neural network algorithm. Accuracy var-
generic images,48 and were transferred to ADNI,45 ADNI (stable ied between these studies, including for the following classification
and progressive MCI),46 ADNI,47 and ADNI (sMRI).48 Transfer learn- tasks: AD versus healthy controls (90.4–99.1), MCI versus healthy
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5891

F I G U R E 3 Changes in classification methods over time. This figure shows the rise in the use of discriminative classifiers in the last 4 years. The
use of algorithmic classifiers increased up to 2015 and has remained steady since. The use of generative models has stayed relatively stable since
its first use in 2005.

controls (83.2–99.2), and MCI converters versus non-converters line measures alone for the diagnosis of AD,62 and were par-
(70.6–81.6). ticularly useful when applied to the prediction of MCI to AD
conversion.69,75,77 Of interest, longitudinal changes in volumetric MRI
may need to be considered in the context of baseline volumetry to be
3.3 MRI meaningful.74
Twenty-eight studies investigated the use of non-volumetric
The number of imaging modalities used across the included studies structural imaging features for diagnosis (n = 24) and/or progno-
can be found in Figure 5. Structural MRI and PET/SPECT were the sis/conversion (n = 7). The input consisted of T1- or T2-weighted
most frequently used imaging modalities for diagnosis and prognosis of images, DTI data, or a combination thereof, to estimate non-volumetric
dementia, being used in approximately 71% and 25% of studies respec- features such as cortical thickness, texture, and surface area. These
tively. Around half of studies leveraged structural MRI alone (134 of studies focused on (i) optimization of image pre-processing techniques,
255) and those making use of multiple modalities (49 of 255) often (ii) investigation of feature selection methods, and (iii) optimization
used sMRI and PET (35 of 49) together. It is only since 2020 that stud- of classifiers and subsequent validation of the developed method.
ies incorporating three or more different modalities have begun to The accuracy for differentiating between AD patients and healthy
appear.49–51 controls ranged from 79.2% to 99.1%. Promising developments were
In total, 68.6% (175 of 255) of studies relied on volumetric struc- noted for differential diagnosis (e.g., vascular dementia vs. AD)78
tural MRI measurements. In the few studies that tested traditional and early diagnosis distinguishing MCI and healthy controls.79–82 As
and AI approaches head-to-head, AI methods outperformed raw vol- expected, differentiating MCI subtypes and between MCI and AD
umetric measurements, for example, against hippocampal volume for cohorts was a more difficult task, which is also often the case in clinical
diagnosis52,53 and for predicting conversion of MCI to AD.54 The practice. We found that performance was lower when predicting
reported accuracy of AI methods for the diagnosis of AD varied MCI conversion to AD, or conversion of stable MCI to progressive
between 60.2% and 99.3%. Of note, estimates in the lower range MCI.83–85
were found when using a multi-class classifier (i.e., AD vs. MCI vs. Twenty-six studies (the first published in 2012) used resting-state
healthy controls, rather than AD vs. healthy controls)55,56 or where an MRI (rsMRI); we did not identify any studies using task-based MRI.
independent validation group was used.57 All but 4 studies51,52,86,87 focused on diagnosis and the majority
Contributing to heterogeneity, the aim of “diagnosis” differed (20 of 26) used ADNI data, either as the primary dataset or as
between studies using structural MRI. For example, there were 17 a replication dataset. Graph measures were often used to summa-
studies specifically targeting early diagnosis in which “early” disease rize network characteristics. Overall, the accuracy of discriminat-
was variably defined by: MMSE score < 2458–60 ; CDR 0.5-148,61–63 ; ing between AD and controls ranged between 85% and 97%, but
progression from MCI to AD within 18 months,64,65 2 years,66 3 dropped when discriminating between MCI and controls (70-88%).
years67,68 ; conversion more than 12 months after imaging69 ; or was Most studies reported the nodes which contribute most to discrimi-
not clearly defined.70–72 nation between AD and controls: there was some heterogeneity, but
Studies using longitudinal structural MRI measures (n = 6)69,73–77 most often components of the default mode network (DMN) were
suggest that multiple timepoints may be more accurate than base- identified.88–91
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5892 BORCHERT ET AL .

F I G U R E 4 Forest plot depicting AUC values for classifications of AD patients versus healthy controls. Confidence intervals are shown where
this was reported. Studies were stratified according to the type of machine learning method used including algorithmic (orange), discriminative
(blue), generative (green) and other (red). Unweighted average AUC values for each type of machine learning method is depicted with a diamond.

3.4 Neurophysiological imaging diseases including PD and FTD.92–94 The majority of the studies
(n = 21) used quantitative EEG, while the remaining used either MEG,95
We identified 24 studies which used neurophysiological imaging meth- event-related potential EEG96 , or combined EEG with SPECT.97
ods, only three of which investigated non-AD neurodegenerative Although half (n = 12) of these studies have been published since
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5893

AD (min accuracy: 56% for PET alone vs. 72% for PET and other
modalities).109,115–117,120,133–135 An additional approach used PET
and structural MRI data in combination with other markers (i.e.,
apolipoprotein E4 [APOE4] status and cognitive scores) to train a clas-
sifier, then selected neuroimaging features for classification, showing
better performance when neuroimaging data (gray matter density,
amyloid burden, APOE4 status; r = −0.68) were used to predict
individualized rate of cognitive decline in MCI, compared to cogni-
tive predictors (depression, memory and executive function scores;
r = −0.4).136 Similarly, three studies showed that SPECT is able to
classify MCI and AD, but its predictive value for MCI conversion
improved when combined with other imaging modalities or cognitive
F I G U R E 5 Imaging modalities used across included studies.
assessments.97,137,138
fMRI = functional MRI, DTI = diffusion tensor imaging,
DWI = Diffusion weighted imaging, EEG = Electroencephalography,
MEG = magnetoencephalography MT = magnetization transfer,
PET = positron emission tomography, sMRI = structural MRI, 3.6 Approaches to prognosis in AD
SPECT = single photon emission computed tomography.
Fifty-four studies investigated either prognosis or a combination of
2018, this cohort of publications also included some of the earliest diagnosis and prognosis. The majority were retrospective designs
studies identified in this review starting in 2005.98,99 All neurophys- (51 of 54). Of 54 studies, 47 (87%) looked at prognosis in terms
iological studies used data from their local institution, the largest of of MCI to AD conversion. Of these studies, two approaches were
which included EEG recordings from 272 participants,100 although used to evaluate the performance of prognostic predictions; some
most studies (n = 13) included less than 50 participants. In a man- exclusively used baseline data (fixed), while others used multi-
ner similar to other imaging modalities, SVM was the most common ple imaging time points (continuous) and related these to time to
(n = 12) ML tool used and no other algorithm was used in more than conversion.
three studies. Accuracy of discrimination between AD and healthy con- MRI alone was the main imaging modality used (36 of 54 studies)
trols varied from 69% in the single MEG study95 up to 100% in one with an additional six studies combining MRI and PET. Nine studies
study using four EEG features.101 used only PET data,102,111–113,116,122,126,139 one used SPECT,137 and
two used EEG data.93,95 The main outcome measure for these studies
was conversion to AD from MCI over a prespecified period of time (47
3.5 PET/SPECT imaging of 54 studies). A smaller proportion of studies (n = 4) used cognitive
decline as an outcome measure. Similar to the diagnostic studies dis-
Sixty-five studies were identified using PET imaging, aiming to improve cussed in this review, the majority of the neuroimaging data came from
early diagnosis (n = 46), prognosis (n = 13), or both (n = 6) using ML the ADNI database (78%, 42 of 54 studies). An additional three studies
approaches. The most commonly used approach was SVM (n = 27), combined local datasets with ADNI.
which when applied to FDG PET, demonstrated an accuracy of over Thirty-eight studies used only baseline imaging data to predict a
85% in studies for detecting AD hypometabolic patterns102–104 and future diagnosis with a range of accuracy between 65% and 96% (mean
outperformed structural MRI when compared head-to-head.105,106 AUC 0.79, standard deviation 0.09). Seven used multiple imaging time-
Using SVM with FDG PET data distinguished AD (>86% accuracy) points to make predictions with accuracies between 73% and 92%
and MCI (>78.8% accuracy) from controls and predicted MCI conver- (mean AUC 0.81, standard deviation 0.10). One paper found a sub-
sion within 12 months and up to 5 years with accuracies ranging from stantial improvement with longitudinal data (AUC 0.93) compared to
72% to 80%.107–118 The same approach applied to amyloid PET also baseline data alone (AUC 0.54),111 and a second paper achieved a
demonstrated accuracies of >85% for predicting MCI conversion and high level of accuracy using baseline neuroimaging information with
diagnosing AD.115,117,119–121 Non-SVM approaches, such as convolu- longitudinal cognitive scores (AUC = 0.90).70
tional neural networks and deep learning, on FDG PET and amyloid Time to conversion was divided into two categories: conver-
PET showed variable performance in predicting a final diagnosis of AD, sion within a fixed timeframe (42 of 47), or a continuous mea-
cognitive decline, or MCI conversion,46,122–132 with accuracy between sure of time of conversion (5 of 47). Of those that used a fixed
75% and 100%. Model accuracy in multicenter studies (>70% accu- timeframe, 5 studies considered conversion within 1 year (AUC
racy) was lower than that of those relying on local datasets (>78% range: 0.72-0.90), 8 studies within 18 months (AUC range: 0.68-
accuracy). 0.79), 5 studies within 2 years (AUC range: 0.74-0.96), 17 stud-
Compared to ML methods which used PET alone, those which ies within 3 years (AUC range: 0.65-0.93), and 7 studies pre-
combined imaging modalities (i.e., FDG PET, amyloid PET, and/or dicted conversion over 3 years with a maximum of within 10 years
MRI) were more accurate in terms of diagnosis of both MCI and (AUC range: 0.54-0.91).
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5894 BORCHERT ET AL .

The main outcome of the remaining studies that did not focus on tural MRI alone or in combination with another MRI modality or PET,
MCI to AD progression (7 of 54) varied; 2 of 7 predicted cognitive almost all of which focused on AD. The size of the ADNI data has led
scores (Alzheimer’s Disease Assessment Scale—Cognitive Subscale to a rapid rise since 2017 in the use of more complex discriminative AI
[ADAS-Cog]) over time using longitudinal MRI,140,141 while 2 other methodologies, including deep learning models. These more complex
studies predicted both cognitive scores (Mini Mental State Exami- models have in general outperformed simpler algorithmic and gen-
nation [MMSE]) and MCI to AD conversion within 24 months.77,142 erative models, although comparison between studies is challenging
Additionally, two of seven studies predicted conversion from cogni- given differences in diagnostic criteria and outcome measures. Most
tively normal to AD in 759 and 2 years.64,143 Finally, only one paper studies of diagnosis published ROC curve analysis results; however,
examined prognosis in non-AD neurodegenerative diseases, namely there were marked differences between studies in definitions such
PD and DLB93 with an AUC of 0.87. as “early” dementia, and in the outcome measures used in prognos-
tic studies. There remain significant gaps in the literature including
non-Alzheimer’s neurodegenerative diseases (most strikingly vascular
3.7 Non-Alzheimer’s dementias dementia with only two studies), the limited application of promis-
ing neurophysiology methods, and validation in clinically relevant
The majority of studies that included patients with non-Alzheimer’s populations.
dementia used neuroimaging features to improve the differential ML methods have been successfully applied to almost every aspect
diagnosis between different dementia diagnoses. In total, 17 studies of neurodegenerative disease.155 A previous review of ML for neu-
included a non-AD dementia group, 14 featured a non-AD dementia roimaging in dementia included studies up to 2016,42 since when
as the diagnosis of interest, with the remaining 3 using the non-AD the field has expanded rapidly. Approximately 60% of the studies we
groups as a control group. FTD or behavioral variant FTD (bvFTD) was included (n = 152) have been published since 2016. Some progress has
the most commonly investigated non-AD dementia, with seven stud- been made on the concerns raised by Pellegrini and colleagues, includ-
ies having FTD or bvFTD as their main focus.92,93,144–148 These studies ing the overreliance on SVM classifiers and MRI. SVM was still the most
attempted differential diagnosis of FTD (from AD and/or LBD) most frequently used classifier in our cohort which is unsurprising given that
often using neuropsychological data and structural imaging (four of it was one of the first widely adopted methods. However, the overre-
seven studies), with two studies using EEG92,94 and one using struc- liance on SVM classifiers has reduced, reflecting the rapid growth of
tural MRI for classification based on post-mortem pathology.147 Five this field and moving toward the use of a range of ML methodologies,
studies used data routinely collected in clinics (for example, from mem- as well as PET and/or multimodal approaches. However, despite this
ory clinics) to attempt differential diagnosis between patient groups surge in studies, several barriers prevent the integration of these novel
based on imaging features and typically included FTD, LBD, PSP, CBD, methods into everyday clinical practice. Below we discuss three critical
PD dementia, and vascular dementia.53,149–152 issues identified from this systematic review: (1) reporting and repro-
Structural MRI was the most frequently used imaging modality ducibility of methodology, (2) addressing clinically relevant questions,
(11 of 17 studies). Two studies focused on the differential diagnosis (3) validation of results.
between PD and LBD,93,153 and only two on vascular dementia.78,154
The majority of studies used data from local hospitals or memory clin-
ics (14 of 17 studies); one paper used local data combined with ADNI,57 4.1 Methodological considerations
and three studies used multi-center or cohort data.144,148,150 Since
the majority of studies utilized prospective or retrospective data from While it is encouraging to see a wide range of methods applied to neu-
local clinics, datasets were relatively small compared to multi-center roimaging data, the multiplicity of approaches creates a challenge in
studies like ADNI with most studies including 60 to 100 patients and assessing the validity of each method, comparing between differing
some as low as 15 patients in a single diagnostic category.78 The stud- models, and independently reproducing the results. Although we did
ies with larger patient numbers tended to come from multi-center not systematically review reproducibility, in general we found limited
studies144,150 or used retrospective data over a long period of time.147 descriptions of many models, and only a minority of studies reported
the availability of code to enable replication.
Reproducibility and transparency in neuroimaging research is an
4 DISCUSSION increasingly prominent issue, most clearly outlined by Poldrack and
colleagues.156 The neuroimaging field has led the way in open science
In this systematic review, we examined 255 published studies using efforts, such as large data sharing platforms pioneered by the Human
neuroimaging alone for the diagnosis or prognosis of neurodegenera- Connectome Project,157 and introducing best practice for analysis and
tive disease. The vast majority of studies (71%) used the ADNI dataset data sharing through the COBIDAS guidelines.158,159 To increase the
which primarily uses MRI and focuses on the conversion from MCI reliability of results, pre-registering analysis through platforms such
to AD. The dominance of ADNI means that this emphasis is reflected as the Open Science Framework160 has been advocated for in both
in the published literature, with the majority of studies using struc- neuroimaging studies161 and ML methodologies.162 More generally,
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5895

staged approaches to model validation in ML are available to improve 4.3 Validation of results
confidence in model performance.25
We found that the combination of multiple imaging modalities, We found that studies using an independent dataset for validation, as
such as MRI and PET, improved the performance of ML models for opposed to cross-validation or other similar methods, reported much
classification tasks related to AD. We speculate that using features lower accuracy, particularly when a community-based population was
from multiple modalities enables the models to train on several dif- used. For instance, applying an SVM classifier trained on ADNI and
ferent biomarkers which provide a more holistic representation of the applied to memory clinics found markedly reduced accuracy in the clin-
underlying disease mechanisms, such as changes in structure (volu- ical setting (AUC = 0.76 for AD diagnosis) compared to that in the
metric MRI), network-connectivity metrics (resting-state fMRI), and training dataset (AUC = 0.96).57 A few recent studies have addressed
metabolic physiology (PET). Although the results suggest this approach the risk of overfitting by assessing generalizability in unseen inde-
may be beneficial, the limited number of studies identified here using pendent research datasets,104,165,166 collectively demonstrating the
this method means that it is difficult to suggest which combinations of value of this approach in identifying methodological issues relevant to
modalities will be best at improving the performance of ML models. the overall model performance. Therefore, validation studies are crit-
ical, particularly those in a memory clinic setting where the tools are
ultimately to be used.
4.2 Addressing key clinical questions The over-reliance on a single dataset such as ADNI introduces
potential ethnic and socio-economic biases to models that may ham-
Relevant clinical questions can be split into early diagnosis, differential per generalization, an issue that has been specifically raised in the
diagnosis, prognosis and predicting response to treatment. There were ADNI dataset.167 Concerns have been raised more generally about bias
no studies investigating the response to treatment, perhaps unsurpris- in ML models,168 including in the context of health applications.169
ingly given that the currently widely available treatments for dementia This is of particular concern in marginalized ethnic groups who
are symptomatic rather than disease modifying. The majority of stud- have poorer health indicators in general,170 and who may miss out
ies considered the diagnosis of AD, or the prognostic prediction of on access to health services due to socio-demographic, cultural, or
MCI conversion to early AD. However, variability in definitions such as religious beliefs,171 including dementia services.172,173 More repre-
“early Alzheimer’s disease” limited comparison between studies. This sentative datasets are critical for models to translate reliably to all
partly reflects the wider field where, for example, a clear definition parts of the population, to inform risk prediction models, and work
of MCI has remained elusive despite recent efforts to reach such a toward closing gaps in health inequality related to dementia. Address-
consensus.163 ing bias in these collected datasets, and differences between genetic
We found no studies that assessed the common clinical challenge or ethnic groups in model performance, or applicability to different
of differential diagnosis from among multiple (>2) possible diagnoses. socio-economic populations, will be critical to address in ongoing data
This is a much harder problem to solve for ML algorithms because it collection. It is unlikely that a single study or a single dataset can
requires a multi-class classifier which is computationally more chal- properly address these challenging issues, so collaboration between
lenging and typically yields lower accuracy than a binary classifier. studies and between countries is required. This is happening to some
The lack of appropriate multiclass data is a major limitation, partic- extent in initiatives such as J-ADNI in Japan which is almost identi-
ularly given the reliance on the ADNI dataset that consists almost cal to the North American protocol and has been used to compare
exclusively of amnestic MCI or AD patients. The National Alzheimer’s diagnosis and progression in dementia between both cohorts.174
Coordinating Center dataset has Alzheimer’s and non-Alzheimer’s Other examples include the Longitudinal Aging Study in India (LASI-
dementia patients from a real-world setting,164 but is much more vari- DAD)175 and through initiatives such as the Genetic Frontotemporal
able in scanning sequences (including MRI field strengths), and reports dementia Initiative (GeNFI),176 which recruits multi-nationally. Feder-
clinically defined diagnoses rather than research diagnostic criteria. ated learning may also help address this issue by providing broader
ROC curve analysis was widely used to characterize diagnostic accessibility to datasets from diverse backgrounds and international
classification performance. In particular, we found the AUC is often sources.
reported as the main measure of classification between groups, usually A number of methodological approaches are available for mea-
accompanied by the PPV and NPV. The PPV and NPV are more rele- suring or mitigating bias.177 Examples include the geometric solution
vant to clinical practice, providing interpretation of the proportion of to learn fair representations (He et al. 2020),178 which removes
correct positive and negative results for a classification. The outcome correlations between the data and specified protected features, as
measure for prognostic studies is more challenging. We found that well as IBM’s AI Fairness 360 toolkit (Bellamy et al. 2019),179
studies predicting prognosis usually grouped outcomes and applied which provides an accessible set of fairness metrics for a model and
ROC curve analysis. This is particularly relevant for predicting MCI to accompanying explanations to help mitigate bias. We did not find
AD conversion; however, it is not applicable to other situations, such as the issue of bias to be discussed or addressed in the studies we
predicting the rate of cognitive decline in established dementia. reviewed.
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5896 BORCHERT ET AL .

In addition to overcoming these barriers related to transparency,

BOX 1: Recommendations to move toward clinically use- establishing large, diverse datasets, external validation and consensus

ful, machine learning methods applied to neuroimaging for definitions, we will also need to address translational challenges more

dementia broadly to implement AI into real-world clinical settings.180 Over-

coming the technical obstacles of integrating AI will be required for
Recommendations for machine learning studies
different types of bias/artifacts when data are conglomerated from
Methodological considerations
various sources/institutions181 while ensuring the security and privacy
∙ Provide sufficient description of the methods, with avail-
of sensitive health records for storage and sharing.182 Several factors
able code, to enable independent replication
currently limit the adoption of AI tools by clinicians including iden-
∙ Use a staged approach to model validation
tity threat,183,184 disruption of clinical workflow, and the uncertainty
∙ Pre-register analysis
surrounding the basis of “black box” algorithms, particularly when the
∙ Consider using multiple modalities
output disagrees with their own clinical judgement.185 By improving
Addressing key clinical questions
interpretability, explainable AI may be the most amenable approach
∙ Clearly state the diagnostic criteria used
to building trust and understanding in the medical profession.186 Fur-
∙ For diagnosis, report performance in terms of ROC curve
thermore, social and legal issues will require significant attention if
analysis, including PPV and NPV, and confidence intervals
implementation of AI into clinical practice is to be successful. For exam-
∙ Clearly define measures of prognosis, and consider the use
ple, there remains uncertainty about which party is responsible when
of odds ratios and survival analysis
the use of AI tools result in harm from both legal187 and patient188 per-
Validation
spectives, while patients in general may prefer human supervision over
∙ Independently validate models in at least one independent
AI.189
dataset
∙ Validate findings in a real-world dataset (e.g., memory
clinics)
4.5 Limitations
Recommendations for the field more broadly
∙ Work toward consensus on outcome measures for diagno-
This systematic review has three main limitations. First, although
sis and prognosis
we aimed to provide an informed and broad overview of the exist-
∙ Establish large datasets of non-AD and/or multiple types
ing literature on this subject, our exclusion of reports not written in
of dementia
English and those where the full text was not available meant that
∙ Establish open datasets for EEG comparable to those with
some studies which would have otherwise met the inclusion crite-
MRI
ria may not have been covered in this review. Two key additional
∙ Monitor ethnic and sociodemographic bias in data collec-
exclusion criteria were the decisions not to include studies using lin-
tion and encourage cross-study collaboration to address
ear regression for classification, and studies combining neuroimaging
these biases
with other biomarkers without reporting the model performance for
the neuroimaging features in isolation. Our motivation was to focus
specifically on neuroimaging, and specifically on recognized ML meth-
ods, but it is possible we excluded studies with high clinical value and
translational potential.
4.4 Challenges for the field Second, the heterogeneity in classification tasks, ML methods used
and statistical reporting across studies may have introduced bias when
Some of the issues we have highlighted can be addressed by individual trying to decipher which tasks and results to extract. More specifically,
researchers, but others require engagement from the neuroimaging, this was an issue with the more technical studies which compared mul-
ML, and clinical communities more generally. This kind of collaboration tiple (often > 5) ML methods across three or more classification groups
has proven successful in initiatives such as ADNI. Although ADNI is a introducing a large number of comparisons and results to consolidate
powerful dataset and has facilitated the use of more complex method- and extract. For this reason, we decided to run our meta-analysis on
ologies, similar collaborations for data collection and curation are a very specific task from which we could extract the AUC value for
required to help address ML for non-Alzheimer’s neurodegenerative classifying AD versus healthy controls. This heterogeneity in AI meth-
disease, and for EEG data. ods, imaging modalities, and patient cohorts also meant that we were
Given the challenges of comparisons between studies using differ- unable to provide insight into which features performed best for spe-
ent methodologies and definitions, we suggest the field move toward cific classification tasks. We do not address significant ethical issues
consensus on outcome measures. Diagnostic criteria exist for the in big data analysis of data security, consent to data sharing, and the
major neurodegenerative disorders, but better definitions of ‘early’ dis- acceptability of AI methods to clinicians and the general public.
ease, and standard methods to assess prognosis would facilitate model Third, we employed a risk of bias screening tool that depended on a
selection. We outline our recommendations in Box 1. subjective judgment for each paper’s inclusion or exclusion, and there
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5897

18
may have been heterogeneity in this assessment between screeners. Imaging Physics, University Hospital Southampton NHS Foundation Trust,
We chose a low threshold for inclusion based on study quality in Southampton, UK
19
order to accurately depict and identify current barriers in the litera- Research into Memory, Brain sciences and dementia Group (ReMemBr Group),
Translational Health Sciences, Bristol Medical School, University of Bristol,
ture limiting translation to clinical practice. We only excluded studies
Bristol, UK
exhibiting clear methodological concerns, such as lack of reporting 20
Artificial Intelligence & Computational Neuroscience Group (AICN Group),
of basic participant demographics. The screening tool had a binary Sheffield Institute for Translational Neuroscience (SITraN), Department of Neu-
outcome (inclusion/exclusion), and we were unable to investigate the roscience, University of Sheffield, Sheffield, UK
potential relationship between study quality and ML performance. 21
Department of Biomedical Sciences, Humanitas University, Pieve Emanuele,
Italy
22
University of Cambridge Medical Library, Cambridge, UK

5 CONCLUSIONS 23
Medical Research Council Cognition and Brain Sciences Unit, University of
Cambridge, Cambridge, UK
24
In this systematic review, we generate a number of recommendations Department of Neurosciences, Biomedicine and Movement Sciences, Univer-
sity of Verona, Verona, Italy
to facilitate translation of ML methods for patient benefit in the diag-
25
Information School, University of Sheffield, Sheffield, UK
nosis and prognosis of dementia. We highlight issues of methodological
26
heterogeneity, clinical relevance of results, and validation/replication Division of Genetics and Genomics, Boston Children’s Hospital, Harvard
Medical School, Boston, Massachusetts, USA
of findings. We offer a set of recommendations to address key gaps
27
Broad Institute of MIT and Harvard, Cambridge, UK
in the literature including the importance of addressing key clinical
28
Department of Neurology, Massachusetts General Hospital, Harvard Medical
questions, providing sufficient details of AI methods, and validating
School, Boston, Massachusetts, USA
findings in independent datasets which are clinically relevant. Look-
29
Nanyang Technological University, Singapore
ing forward, the field is likely to move toward the establishment
30
Population Health Sciences Institute, Newcastle University, Newcastle upon
of real-world datasets, multi-model imaging methods, and complex
Tyne, UK
ML algorithms emphasizing the importance of providing sufficient 31
Department of Psychiatry, University of Oxford, Oxford, UK
methodological details to enable independent replication. We are opti-
32
Alan Turing Institute, London, UK
mistic that addressing these concerns will accelerate the translation of
33
Department of Experimental Psychology, University of Oxford, Oxford, UK
ML methods for patient benefit in neurodegenerative disease.

AUTHOR CONTRIBUTIONS
AFFILIATIONS
1
Department of Clinical Neurosciences, University of Cambridge, Cambridge, Robin J. Borchert, Michele Veldsman, Timothy Rittman contributed to
UK the conception of the work, drafting and revision of the manuscript
2
Department of Radiology, University of Cambridge, Cambridge, UK for intellectual content. Robin J. Borchert, Michele Veldsman, Tim-
3
Department of Computer Science and Technology, University of Cambridge, othy Rittman, Jose Bernal, Eugene Tang contributed to the devel-
Cambridge, UK opment of the protocol. Veronica Phillips conducted the literature
4
Department of Pharmacology and Physiology, University of Montreal, Montreal, search. Robin J. Borchert coordinated the screening process. Robin J.
Canada
Borchert, Michele Veldsman, Timothy Rittman, Tiago Azevedo, Aman-
5
Centre de recherche de l’Institut Universitaire de Gériatrie (CRIUGM), Mon- Preet Badhwar, Jose Bernal, Matthew Betts, Rose Bruffaerts, Helena
treal, Canada
M. Gellersen, Audrey Low, Christopher R. Madan, Maura Malpetti,
6
Centre for Clinical Brain Sciences, The University of Edinburgh, Edinburgh, UK
Jhony Mejia, Sofia Michopoulou, Carlos Muñoz-Neira, Marion Peres,
7
Institute of Cognitive Neurology and Dementia Research, Otto-von-Guericke
Siddharth Ramanan, Stefano Tamburin, Hanz M. Tantiangco, Lokendra
University Magdeburg, Magdeburg, Germany
Thakur, Alessandro Tomassini, Ashwati Vipin, Eugene Tang, Danielle
8
German Center for Neurodegenerative Diseases (DZNE), Magdeburg, Germany
Newby screened papers for inclusion in the review. Robin J. Borchert,
9
Center for Behavioral Brain Sciences, University of Magdeburg, Magdeburg,
Jose Bernal, Helena M. Gellersen, Audrey Low, Jhony Mejia, Carlos
Germany
10
Muñoz-Neira, Marion Peres, Hanz M. Tantiangco extracted data from
Computational Neurology, Experimental Neurobiology Unit, Department of
Biomedical Sciences, University of Antwerp, Antwerp, Belgium eligible papers. Robin J. Borchert, Michele Veldsman, Timothy Rittman,
11
Biomedical Research Institute, Hasselt University, Diepenbeek, Belgium
Jose Bernal, Lokendra Thakur contributed to analysis and interpreta-
12 tion of the data. Lokendra Thakur contributed to the meta-analytic
Department of Psychology, University of Cambridge, Cambridge, UK
13
approach. Robin J. Borchert, Michele Veldsman, Timothy Rittman,
Department of Psychiatry, University of Cambridge, Cambridge, UK
AmanPreet Badhwar, Jose Bernal, Matthew Betts, Rose Bruffaerts,
14
University of Exeter Medical School, Exeter, UK
Michael C. Burkhart, Ilse Dewachter, Audrey Low, Luiza Machado,
15
Department of Biochemistry, Universidade Federal do Rio Grande do Sul, Porto
Maura Malpetti, Jhony Mejia, Sofia Michopoulou, Jack Pepys, Stefano
Alegre, Brazil
16
Tamburin, Lokendra Thakur, Ashwati Vipin contributed to the writing
School of Psychology, University of Nottingham, Nottingham, UK
of the manuscript. Michele Veldsman and Timothy Rittman provided
17
Department of Biomedical Engineering, Universidad de Los Andes, Bogotá,
Colombia
study supervision. Janice M. Ranson and David J. Llewellyn conceived
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5898 BORCHERT ET AL .

and organized the symposium from which this paper and others in Research and Treatment (ISTAART), through the Artificial Intelligence
the series originated, obtained funding, contributed to the concep- for Precision Dementia Medicine professional interest area. The views
tion of the work, revised the manuscript for intellectual content, and and opinions expressed by authors in this publication represent those
harmonized the manuscript with other papers in the series. Ilianna of the authors and do not necessarily reflect those of the PIA mem-
Lourida revised the manuscript for intellectual content and harmo- bership, ISTAART or the Alzheimer’s Association. [Correction added on
nized the manuscript with other papers in the series. All authors read 01 September 2023, after first online publication: The preceding two
and approved the final manuscript. sentences were added.]

ACKNOWLEDGMENTS CONFLICT OF INTEREST STATEMENT

With thanks to the Deep Dementia Phenotyping (DEMON) Network The authors declare that they have no conflicts of interest. Author
State of the Science symposium participants (in alphabetical order): disclosures are available in the supporting information.
Peter Bagshaw, Robin Borchert, Magda Bucholc, James Duce, Char-
lotte James, David Llewellyn, Donald Lyall, Sarah Marzi, Danielle ORCID
Newby, Neil Oxtoby, Janice Ranson, Tim Rittman, Nathan Skene, Robin J. Borchert [Link]
Eugene Tang, Michele Veldsman, Laura Winchester, Zhi Yao. This Tiago Azevedo [Link]
paper was the product of a DEMON Network state of the science AmanPreet Badhwar [Link]
symposium entitled “Harnessing Data Science and AI in Dementia Jose Bernal [Link]
Research” funded by Alzheimer’s Research UK. Race against Demen- Matthew Betts [Link]
tia Alzheimer’s Research UK (ARUK-RADF2021A-010). Jose Bernal Rose Bruffaerts [Link]
is supported by the MRC Doctoral Training Programme in Precision Michael C. Burkhart [Link]
Medicine (Award Reference No. 2096671). Amanpreet Badhwar is Ilse Dewachter [Link]
supported by Fonds de recherche du Québec Santé—Chercheur bour- Helena M. Gellersen [Link]
siers Junior 1 and Fondation Courtois. Matthew Betts is supported by Audrey Low [Link]
the Deutsche Forschungsgemeinschaft (DFG, German Research Foun- Ilianna Lourida [Link]
dation) – Project-ID 425899996 – SFB 1436 Project A08 and by the Luiza Machado [Link]
German Federal Ministry of Education and Research (BMBF, fund- Christopher R. Madan [Link]
ing code 01ED2102B) under the aegis of JPND. Eugene Tang, NIHR Maura Malpetti [Link]
Clinical Lecturer, is funded by the National Institute for Health and Jhony Mejia [Link]
Care Research (NIHR). The views expressed in this publication are Sofia Michopoulou [Link]
those of the author(s) and not necessarily those of the NIHR, NHS Carlos Muñoz-Neira [Link]
or the UK Department of Health and Social Care. Sofia Michopoulou, Jack Pepys [Link]
NIHR Clinical Lecturer, was funded by the National Institute for Marion Peres [Link]
Health Research (NIHR), the NIHR Applied Research Collaboration Veronica Phillips [Link]
ARC Wessex, the Southampton Academy of Research and the Health Siddharth Ramanan [Link]
Education England Topol Fellowship program. The views expressed Stefano Tamburin [Link]
in this publication are those of the author and not necessarily those Hanz M. Tantiangco [Link]
of the funding bodies. Carlos Muñoz-Neira was supported by the Lokendra Thakur [Link]
Government of Chile through ‘Becas Chile’ and CONICYT—National Alessandro Tomassini [Link]
Commission for Scientific and Technological Research [CONICYT— Ashwati Vipin [Link]
Comisión Nacional de Investigación Científica y Tecnológica], the Eugene Tang [Link]
University of Bristol (Grant Code G100030-150), and its Postdoc- Danielle Newby [Link]
toral Research Associate position at the University of Sheffield. Janice M. Ranson [Link]
Janice Ranson and David Llewellyn are supported by Alzheimer’s David J. Llewellyn [Link]
Research UK and the Alan Turing Institute/Engineering and Physical Michele Veldsman [Link]
Sciences Research Council (EP/N510129/1). DJL also receives fund- Timothy Rittman [Link]
ing from the Medical Research Council (MR/X005674/1), National
Institute for Health Research (NIHR) Applied Research Collabora- REFERENCES
tion South West Peninsula, National Health and Medical Research 1. Fischer CE, Qian W, Schweizer TA, et al. Determining the impact
Council (NHMRC), and National Institute on Aging/National Insti- of psychosis on rates of false-positive and false-negative diagno-
tutes of Health (RF1AG055654). Timothy Rittman is supported by the sis in Alzheimer’s disease. Alzheimers Dement Transl Res Clin Interv.
2017;3:385-392. doi:10.1016/[Link].2017.06.001
Cambridge Centre for Parkinson’s Plus Disorders and the Cambridge
2. Cook LD, Nichol KE, Isaacs JD. The London memory service audit and
Biomedical Research Centre. This manuscript was facilitated by the quality improvement programme. BJPsych Bull. 2019;43:215-220.
Alzheimer’s Association International Society to Advance Alzheimer’s doi:10.1192/bjb.2019.18
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5899

3. Nedelec T, Couvy-Duchesne B, Monnet F, et al. Identifying health 22. Davatzikos C. Machine learning in neuroimaging: progress and chal-
conditions associated with Alzheimer’s disease up to 15 years before lenges. NeuroImage. 2019;197:652-656. doi:10.1016/[Link].
diagnosis: an agnostic study of French and British health records. 2018.10.003
Lancet Digit Health. 2022;4:e169-78. doi:10.1016/S2589-7500(21) 23. Li X, Xiong H, Li X, et al. Interpretable deep learning: interpreta-
00275-2 tion, interpretability, trustworthiness, and beyond. Knowl Inf Syst.
4. de Vugt ME, Verhey FRJ. The impact of early dementia diagnosis and 2022;64:3197-3234. doi:10.1007/s10115-022-01756-8
intervention on informal caregivers. Prog Neurobiol. 2013;110:54-62. 24. Hainc N, Federau C, Stieltjes B, Blatow M, Bink A, Stippich C.
doi:10.1016/[Link].2013.04.005 The bright, artificial intelligence-augmented future of neuroimaging
5. Robinson L, Tang E, Taylor J-P. Dementia: timely diagnosis and early reading. Front Neurol. 2017;8:489. doi:10.3389/fneur.2017.00489
intervention. BMJ. 2015;350:h3029. doi:10.1136/bmj.h3029 25. Kohoutová L, Heo J, Cha S, et al. Toward a unified framework for
6. Meco AD, Vassar R. Early detection and personalized medicine: interpreting machine-learning models in neuroimaging. Nat Protoc.
future strategies against Alzheimer’s disease. Prog Mol Biol Transl Sci. 2020;15:1399-1435. doi:10.1038/s41596-019-0289-5
2021;177:157-173. doi:10.1016/[Link].2020.10.002 26. Nielsen AN, Barch DM, Petersen SE, Schlaggar BL, Greene DJ.
7. Rittman T. Neurological update: neuroimaging in dementia. J Neurol. Machine learning with neuroimaging: evaluating its applications in
2020;267:3429-3435. doi:10.1007/s00415-020-10040-0 psychiatry. Biol Psychiatry Cogn Neurosci Neuroimaging. 2020;5:791-
8. Filippi M, Agosta F, Barkhof F, et al. EFNS task force: the use of neu- 798. doi:10.1016/[Link].2019.11.007
roimaging in the diagnosis of dementia. Eur J Neurol. 2012;19:1487- 27. Mueller SG, Weiner MW, Thal LJ, et al. Ways toward an early diagno-
1501. doi:10.1111/j.1468-1331.2012.03859.x sis in Alzheimer’s disease: the Alzheimer’s Disease Neuroimaging Ini-
9. Harper L, Barkhof F, Scheltens P, Schott JM, Fox NC. An algo- tiative (ADNI). Alzheimers Dement J Alzheimers Assoc. 2005;1:55-66.
rithmic approach to structural imaging in dementia. J Neurol doi:10.1016/[Link].2005.06.003
Neurosurg Psychiatry. 2014;85:692-698. doi:10.1136/jnnp-2013- 28. Marzi SJ, Nott A, Sala Frigerio C, et al. Artificial intelligence for
306285 neurodegenerative experimental models. Alzheimer’s Dementia. Sub-
10. Karas GB, Burton EJ, Rombouts SA, et al. A comprehensive study of mitted.
gray matter loss in patients with Alzheimer’s disease using optimized 29. Doherty T, Yao Z, Al Khleifat A, et al. Artificial intelligence for demen-
voxel-based morphometry. NeuroImage. 2003;18:895-907. doi:10. tia drug discovery and trials optimization. Alzheimer’s Dementia.
1016/s1053-8119(03)00041-7 30. Bettencourt C, Skene N, Bandres-Ciga S, et al. Artificial intelligence
11. Young AL, Marinescu RV, Oxtoby NP, et al. Uncovering the het- for dementia genetics and omics. Alzheimer’s Dementia. Submitted.
erogeneity and temporal complexity of neurodegenerative diseases 31. Winchester LM, Harshfield EL, Shi L, et al. Artificial intelligence for
with Subtype and Stage Inference. Nat Commun. 2018;9:4273. doi:10. alzheimer’s disease and associated dementia biomarkers. Alzheimer’s
1038/s41467-018-05892-0 Dementia. Submitted.
12. Pievani M, de Haan W, Wu T, Seeley WW, Frisoni GB. Functional 32. Newby D, Orgeta V, Marshall CR, et al. Artificial intelligence for
network disruption in the degenerative dementias. Lancet Neurol. dementia prevention. Alzheimer’s Dementia. Submitted.
2011;10:829-843. doi:10.1016/S1474-4422(11)70158-2 33. Lyall DM, Kormilitzin A, Lancaster C, et al. Artificial intelligence for
13. Greicius MD, Srivastava G, Reiss AL, Menon V. Default-mode net- dementia applied models and digital health. Alzheimer’s Dementia.
work activity distinguishes Alzheimer’s disease from healthy aging: Submitted.
evidence from functional MRI. Proc Natl Acad Sci. 2004;101:4637- 34. Bucholc M, James C, Al Khleifat A, et al. Artificial intelligence
4642. doi:10.1073/pnas.0308627101 for dementia research methods optimization. Alzheimer’s Dementia.
14. Badhwar A, Tam A, Dansereau C, Orban P, Hoffstaedter F, Bellec Submitted.
P. Resting-state network dysfunction in Alzheimer’s disease: a sys- 35. Page MJ, McKenzie JE, Bossuyt PM, et al. The PRISMA 2020 state-
tematic review and meta-analysis. Alzheimers Dement Amst Neth. ment: an updated guideline for reporting systematic reviews. BMJ.
2017;8:73-85. doi:10.1016/[Link].2017.03.007 2021;372:n71. doi:10.1136/bmj.n71
15. Klunk WE, Engler H, Nordberg A, et al. Imaging brain amyloid 36. Moola S, Munn Z, Tufanaru C, et al. Chapter 7: systematic reviews of
in Alzheimer’s disease with Pittsburgh Compound-B. Ann Neurol. etiology and risk. 2019. doi:10.46658/JBIRM-17-06
2004;55:306-319. doi:10.1002/ana.20009 37. Bishop CM. Probabilistic Generative Models (section 4.2). Pattern Recog-
16. Lowe VJ, Curran G, Fang P, et al. An autoradiographic evaluation of nit. Mach. Learn. Newer. Springer-Verlag New York Inc.; 2007.
AV-1451 Tau PET in dementia. Acta Neuropathol Commun. 2016;4. 38. Bishop CM. Probabilistic Discriminative Models (section 4.3). Pat-
doi:10.1186/s40478-016-0315-6 tern Recognit. Mach. Learn. Newer. Springer-Verlag New York Inc.;
17. Chételat G, Arbizu J, Barthel H, et al. Amyloid-PET and 18 F-FDG- 2007.
PET in the diagnostic investigation of Alzheimer’s disease and other 39. Bishop CM. Sparse Kernel Machines Pattern Recognition and Machine
dementias. Lancet Neurol. 2020;19:951-962. doi:10.1016/S1474- Learning (Chapter 7). Pattern Recognit. Mach. Learn. Newer. Springer-
4422(20)30314-8 Verlag New York Inc.; 2007.
18. Scheltens P, Launer LJ, Barkhof F, Weinstein HC, van Gool WA. 40. Franc V, Zien A, Schölkopf B, Support Vector Machines as Probabilis-
Visual assessment of medial temporal lobe atrophy on magnetic reso- tic Models. 2011.
nance imaging: interobserver reliability. J Neurol. 1995;242:557-560. 41. Murphy K. A probabilistic interpretation of SVMs. Mach. Learn. Proba-
doi:10.1007/BF00868807 bilistic Perspect. MIT Press; 2012.
19. Wahlund LO, Barkhof F, Fazekas F, et al. A new rating scale for 42. Pellegrini E, Ballerini L, Hernandez M del CV, et al. Machine learn-
age-related white matter changes applicable to MRI and CT. Stroke. ing of neuroimaging for assisted diagnosis of cognitive impairment
2001;32:1318-1322. doi:10.1161/[Link].32.6.1318 and dementia: a systematic review. Alzheimers Dement Diagn Assess
20. Fazekas F, Chawluk JB, Alavi A, Hurtig HI, Zimmerman RA. MR sig- Dis Monit. 2018;10:519-535. doi:10.1016/[Link].2018.07.004
nal abnormalities at 1.5 T in Alzheimer’s dementia and normal aging. 43. LeBlanc M, Zuber V, Thompson WK, et al. A correction for sample
AJR Am J Roentgenol. 1987;149:351-356. doi:10.2214/ajr.149.2. overlap in genome-wide association studies in a polygenic pleiotropy-
351 informed framework. BMC Genomics. 2018;19:494. doi:10.1186/
21. Wang J, Zuo X, He Y. Graph-based network analysis of resting-state s12864-018-4859-7
functional MRI. Front Syst Neurosci. 2010;4. [Link] 44. Han B, Duong D, Sul JH, de Bakker PIW, Eskin E, Raychaudhuri
[Link]/20589099/ S. A general framework for meta-analyzing dependent studies
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5900 BORCHERT ET AL .

with overlapping subjects in association mapping. Hum Mol Genet. 61. Dai Z, Yan C, Wang Z, et al. Discriminative analysis of early
2016;25:1857-1866. doi:10.1093/hmg/ddw049 Alzheimer’s disease using multi-modal imaging and multi-level char-
45. Jain R, Jain N, Aggarwal A, Hemanth DJ. Convolutional neural acterization with multi-classifier (M3). NeuroImage. 2012;59:2187-
network based Alzheimer’s disease classification from magnetic res- 2195. doi:10.1016/[Link].2011.10.003
onance brain images. Cogn Syst Res. 2019;57:147-159. doi:10.1016/j. 62. Hojjati SH, Ebrahimzadeh A, Babajani-Feremi A. Identification of
cogsys.2018.12.015 the early stage of Alzheimer’s disease using structural MRI and
46. Lu D, Popuri K, Ding GW, Balachandar R, Beg MF. Alzheimer’s Dis- resting-state fMRI. Front Neurol. 2019;10:904. doi:10.3389/fneur.
ease Neuroimaging I. Multiscale deep neural network based analysis 2019.00904
of FDG-PET images for the early diagnosis of Alzheimer’s disease. 63. Khedher L, Ramírez J, Górriz JM, Brahim A, Segovia F. Early diagnosis
Med Image Anal. 2018;46:26-34. of Alzheimer‘s disease based on partial least squares, principal com-
47. Li W, Zhang L, Qiao L, Shen D. Toward a better estimation of func- ponent analysis and support vector machine using segmented MRI
tional brain network for mild cognitive impairment identification: a images. Neurocomputing. 2015;151:139-150. doi:10.1016/[Link].
transfer learning view. IEEE J Biomed Health Inform. 2020;24:1160- 2014.09.072
1168. doi:10.1109/JBHI.2019.2934230 64. Pan D, Zeng A, Jia L, Huang Y, Frizzell T, Song X. Early detection
48. Nanni L, Interlenghi M, Brahnam S, et al. Comparison of transfer of Alzheimer’s disease using magnetic resonance imaging: a novel
learning and conventional machine learning applied to structural approach combining convolutional neural networks and ensemble
brain MRI for the early diagnosis and prognosis of Alzheimer’s learning. Front Neurosci. 2020;14.
Disease. Front Neurol. 2020;11. 65. Salvatore C, Cerasa A, Battista P, et al. Magnetic resonance imaging
49. Li T-R, Wu Y, Jiang J-J, et al. Radiomics analysis of magnetic reso- biomarkers for the early diagnosis of Alzheimer’s disease: a machine
nance imaging facilitates the identification of preclinical Alzheimer’s learning approach. Front Neurosci. 2015;9:307. doi:10.3389/fnins.
Disease: an exploratory study. Front Cell Dev Biol. 2020;0. doi:10. 2015.00307
3389/fcell.2020.605734 66. Chincarini A, Bosco P, Calvini P, et al. Local MRI analysis approach in
50. de Vos F, Schouten TM, Koini M, et al. Pre-trained MRI-based the diagnosis of early and prodromal Alzheimer’s disease. NeuroIm-
Alzheimer’s disease classification models to classify memory clinic age. 2011;58:469-480. doi:10.1016/[Link].2011.05.083
patients. NeuroImage Clin. 2020;27:102303. doi:10.1016/[Link].2020. 67. Hu K, Wang Y, Chen K, Hou L, Zhang X. Multi-scale features extrac-
102303 tion from baseline structure MRI for MCI patient classification
51. Rabin JS, Neal TE, Nierle HE, et al. Multiple markers contribute to risk and AD early diagnosis. Neurocomputing. 2016;175:132-145. doi:10.
of progression from normal to mild cognitive impairment. NeuroImage 1016/[Link].2015.10.043
Clin. 2020;28:102400. doi:10.1016/[Link].2020.102400 68. Moradi E, Pepe A, Gaser C, Huttunen H, Tohka J. Alzheimer’s
52. Li F, Liu M. Alzheimer’s disease neuroimaging initiative. A hybrid con- Disease Neuroimaging Initiative. Machine learning framework for
volutional and recurrent neural network for hippocampus analysis in early MRI-based Alzheimer’s conversion prediction in MCI subjects.
Alzheimer’s disease. J Neurosci Methods. 2019;323:108-118. doi:10. NeuroImage. 2015;104:398-412. doi:10.1016/[Link].2014.10.
1016/[Link].2019.05.006 002
53. Morin A, Samper-Gonzalez J, Bertrand A, et al. Accuracy of MRI 69. Zhu Y, Kim M, Zhu X, Kaufer D, Wu G. Alzheimer’s disease
classification algorithms in a tertiary memory center clinical routine neuroimaging initiative. Long range early diagnosis of Alzheimer’s
cohort. J Alzheimers Dis JAD. 2020;74:1157-1166. doi:10.3233/JAD- disease using longitudinal MR imaging data. Med Image Anal.
190594 2021;67:101825. doi:10.1016/[Link].2020.101825
54. Costafreda SG, Dinov ID, Tu Z, et al. Automated hippocampal shape 70. Li H, Fan Y. Early prediction of alzheimer’s disease dementia based on
analysis predicts the onset of dementia in mild cognitive impairment. baseline hippocampal mri and 1-year follow-up cognitive measures
NeuroImage. 2011;56:212-219. doi:10.1016/[Link].2011.01. using deep recurrent neural networks. Proc IEEE Int Symp Biomed
050 Imaging. 2019;2019:368-371. doi:10.1109/ISBI.2019.8759397
55. Guo Y, Zhang Z, Zhou B, et al. Grey-matter volume as a potential fea- 71. Lisowska A, Rekik I. Joint pairing and structured mapping of convolu-
ture for the classification of Alzheimer’s disease and mild cognitive tional brain morphological multiplexes for early dementia diagnosis.
impairment: an exploratory study. Neurosci Bull. 2014;30:477-489. Brain Connect. 2019;9:22-36. doi:10.1089/brain.2018.0578
doi:10.1007/s12264-013-1432-x 72. Singanamalli A, Wang H, Madabhushi A. Cascaded multi-view canon-
56. Cárdenas-Peña D, Collazos-Huertas D, Castellanos-Dominguez G. ical correlation (CaMCCo) for early diagnosis of Alzheimer’s disease
Enhanced data representation by kernel metric learning for demen- via fusion of clinical, imaging and omic features. Sci Rep. 2017;7:8137.
tia diagnosis. Front Neurosci. 2017;11:413. doi:10.3389/fnins.2017. doi:10.1038/s41598-017-03925-0
00413 73. Davatzikos C, Xu F, An Y, Fan Y, Resnick SM. Longitudinal progression
57. Klöppel S, Peter J, Ludl A, et al. Applying automated MR-based of Alzheimer’s-like patterns of atrophy in normal older adults: the
diagnostic methods to the memory clinic: a prospective study. J SPARE-AD index. Brain J Neurol. 2009;132:2026-2035. doi:10.1093/
Alzheimers Dis JAD. 2015;47:939-954. doi:10.3233/JAD-150334 brain/awp091
58. Cheng B, Liu M, Shen D, Li Z, Zhang D. Alzheimer’s disease neu- 74. Chincarini A, Sensi F, Rei L, et al. Integrating longitudinal informa-
roimaging initiative. multi-domain transfer learning for early diag- tion in hippocampal volume measurements for the early detection of
nosis of Alzheimer’s disease. Neuroinformatics. 2017;15:115-132. Alzheimer’s disease. NeuroImage. 2016;125:834-847. doi:10.1016/j.
doi:10.1007/s12021-016-9318-5 neuroimage.2015.10.065
59. Coupé P, Fonov VS, Bernard C, et al. Detection of Alzheimer’s 75. Cui R, Liu M. Alzheimer’s Disease Neuroimaging Initiative. RNN-
disease signature in MR images seven years before conversion to based longitudinal analysis for diagnosis of Alzheimer’s disease. Com-
dementia: toward an early individual prognosis. Hum Brain Mapp. put Med Imaging Graph Off J Comput Med Imaging Soc. 2019;73:1-10.
2015;36:4758-4770. doi:10.1002/hbm.22926 doi:10.1016/[Link].2019.01.005
60. Gorji HT, Haddadnia J. A novel method for early diagno- 76. Farzan A, Mashohor S, Ramli R, Mahmud R. Discriminant analy-
sis of Alzheimer’s disease based on pseudo Zernike moment sis of intermediate brain atrophy rates in longitudinal diagnosis of
from structural MRI. Neuroscience. 2015;305:361-371. alzheimer’s disease. Diagn Pathol. 2011;6:105. doi:10.1186/1746-
doi:10.1016/[Link].2015.08.013 1596-6-105
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5901

77. Zhang D, Shen D, Initiative ADN. Predicting future clinical changes 94. Dottori M, Sedeño L, Martorell Caro M, et al. Towards affordable
of MCI patients using longitudinal and multimodal biomarkers. PLOS biomarkers of frontotemporal dementia: a classification study via
ONE. 2012;7:e33182. doi:10.1371/[Link].0033182 network’s information sharing. Sci Rep. 2017;7:3822. doi:10.1038/
78. Castellazzi G, Cuzzoni MG, Cotta Ramusino M, et al. A machine s41598-017-04204-8
learning approach for the differential diagnosis of alzheimer and vas- 95. Furutani N, Nariya Y, Takahashi T, et al. Decomposed temporal com-
cular dementia fed by MRI selected features. Front Neuroinformatics. plexity analysis of neural oscillations and machine learning applied to
2020;0. doi:10.3389/fninf.2020.00025 Alzheimer’s disease diagnosis. Front Psychiatry. 2020;11.
79. Li Q, Wu X, Xu L, Chen K, Yao L. Classification of Alzheimer’s disease, 96. Chapman RM, McCrary JW, Gardner MN, et al. Brain ERP compo-
mild cognitive impairment, and cognitively unimpaired individuals nents predict which individuals progress to Alzheimer’s disease and
using multi-feature kernel discriminant dictionary learning. Front which do not. Neurobiol Aging. 2011;32:1742-1755. doi:10.1016/j.
Comput Neurosci. 2018;11:no pagination. neurobiolaging.2009.11.010
80. Jung WB, Lee YM, Kim YH, Mun CW. Automated classification to 97. Holler Y, Bathke AC, Uhl A, et al. Combining SPECT and quantitative
predict the progression of alzheimer’s disease using whole-brain EEG analysis for the automated differential diagnosis of disorders
volumetry and DTI. Psychiatry Investig. 2015;12:92-102. with amnestic symptoms. Front Aging Neurosci. 2017;9.
81. Ebadi A, Dalboni da Rocha JL, Nagaraju DB, et al. Ensemble classifica- 98. Dauwels J, Vialatte F, Musha T, Cichocki A. A comparative study
tion of Alzheimer’s disease and mild cognitive impairment based on of synchrony measures for the early diagnosis of Alzheimer’s dis-
complex graph measures from diffusion tensor images. Front Neurosci. ease based on EEG. NeuroImage. 2010;49:668-693. doi:10.1016/j.
2017;11. neuroimage.2009.06.056
82. Kruthika KR, Rajeswari, Maheshappa HD. CBIR system using Capsule 99. Cichocki A, Shishkin SL, Musha T, Leonowicz Z, Asada T, Kurachi
Networks and 3D CNN for Alzheimer’s disease diagnosis. Inform Med T. EEG filtering based on blind source separation (BSS) for early
Unlocked. 2019;14:59-68. detection of Alzheimer’s disease. Clin Neurophysiol Off J Int Fed Clin
83. Gao F, Yoon H, Xu Y, et al. AD-NET: age-adjust neural network Neurophysiol. 2005;116:729-737. doi:10.1016/[Link].2004.09.017
for improved MCI to AD conversion prediction. NeuroImage Clin. 100. Buscema M, Vernieri F, Massini G, et al. An improved I-FAST system
2020;27:no pagination. for the diagnosis of Alzheimer’s disease from unprocessed electroen-
84. Hett K, Ta V-T, Manjón JV, Coupé P. Adaptive fusion of texture-based cephalograms by using robust invariant features. Artif Intell Med.
grading for Alzheimer’s disease classification. Comput Med Imaging 2015;64:59-74. doi:10.1016/[Link].2015.03.003
Graph Off J Comput Med Imaging Soc. 2018;70:8-16. doi:10.1016/j. 101. Gallego-Jutglà E, Solé-Casals J, Vialatte F-B, Elgendi M, Cichocki A,
compmedimag.2018.08.002 Dauwels J. A hybrid feature selection approach for the early diag-
85. Eskildsen SF, Coupe P, Garcia-Lorenzo D, et al. Prediction of nosis of Alzheimer’s disease. J Neural Eng. 2015;12:016018. doi:10.
Alzheimer’s disease in subjects with mild cognitive impairment from 1088/1741-2560/12/1/016018
the ADNI cohort using patterns of cortical thinning. Neuroimage. 102. Toussaint P-J, Perlbarg V, Bellec P, et al. Resting state FDG-PET
2013;65:511-521. functional connectivity as an early biomarker of Alzheimer’s dis-
86. Hojjati SH, Ebrahimzadeh A, Khazaee A, Babajani-Feremi A. ease using conjoint univariate and independent component analyses.
Alzheimer’s Disease Neuroimaging I. Predicting conversion from NeuroImage. 2012;63:936-946. doi:10.1016/[Link].2012.03.
MCI to AD using resting-state fMRI, graph theoretical approach and 091
SVM. J Neurosci Methods. 2017;282:69-80. 103. Gray K, Wolz R, Heckemann R, Rueckert D, Hammers A. Structural
87. Hojjati SH, Ebrahimzadeh A, Khazaee A, Babajani-Feremi A. Predict- differences in cognitively normal elderly individuals with abnormal
ing conversion from MCI to AD by integrating rs-fMRI and structural amyloid biomarkers: detection using volumetric MRI in ADNI and
MRI. Comput Biol Med. 2018;102:30-39. doi:10.1016/[Link]. AIBL. Alzheimers Dement. 2012(1):P337-8.
2018.09.004 104. De Carli F, Nobili F, Pagani M, et al. Accuracy and generalization
88. Li Y, Yang H, Lei B, Liu J, Wee CY. Novel effective connectivity infer- capability of an automatic method for the detection of typical brain
ence using ultra-group constrained orthogonal forward regression hypometabolism in prodromal Alzheimer disease. Eur J Nucl Med Mol
and elastic multilayer perceptron classifier for MCI identification. Imaging. 2019;46:334-347. doi:10.1007/s00259-018-4197-7
IEEE Trans Med Imaging. 2019;38:1227-1239. 105. Ferreira LK, Rondina JM, Kubo R, et al. Support vector machine-based
89. Nguyen DT, Ryu S, Qureshi MNI, Choi M, Lee KH, Lee B. Hybrid multi- classification of neuroimages in Alzheimer’s disease: direct compari-
variate pattern analysis combined with extreme learning machine for son of FDG-PET, rCBF-SPECT and MRI data acquired from the same
Alzheimer’s dementia diagnosis using multi-measure rs-fMRI spatial individuals. Rev Bras Psiquiatr. 2018;40:181-191.
patterns. PLoS ONE Electron Resour. 2019;14:e0212582. 106. Fan Y, Resnick SM, Wu X, Davatzikos C. Structural and functional
90. Jin D, Wang P, Zalesky A, et al. Grab-AD: generalizability and repro- biomarkers of prodromal Alzheimer’s disease: a high-dimensional
ducibility of altered brain activity and diagnostic classification in pattern classification study. NeuroImage. 2008;41:277-285. doi:10.
Alzheimer’s Disease. Hum Brain Mapp. 2020;41:3379-3391. 1016/[Link].2008.02.043
91. Wang M, Lian C, Yao D, Zhang D, Liu M, Shen D. Spatial-temporal 107. Pardo JV, Lee JT, Kuskowski MA, et al. Fluorodeoxyglucose positron
dependency modeling and network hub detection for functional MRI emission tomography of mild cognitive impairment with clinical
analysis via convolutional-recurrent network. IEEE Trans Biomed Eng. follow-up at 3 years. Alzheimers Dement. 2010;6:326-333.
2020;67:2241-2252. 108. Pan X, Adel M, Fossati C, Gaidon T, Wojak J, Guedj E. Multiscale spa-
92. Garn H, Coronel C, Waser M, Caravias G, Ransmayr G. Differ- tial gradient features for 18F-FDG PET image-guided diagnosis of
ential diagnosis between patients with probable Alzheimer’s dis- Alzheimer’s disease. Comput Methods Programs Biomed. 2019;180:no
ease, Parkinson’s disease dementia, or dementia with Lewy bodies pagination.
and frontotemporal dementia, behavioral variant, using quantita- 109. Ortiz A, Munilla J, Alvarez-Illan I, Gorriz JM, Ramirez J. Exploratory
tive electroencephalographic features. J Neural Transm Vienna Austria. graphical models of functional and structural connectivity patterns
1996 2017;124:569-581. doi:10.1007/s00702-017-1699-6 for Alzheimer’s disease diagnosis. Front Comput Neurosci. 2015;9:1-
93. Ruffini G, Ibañez D, Castellano M, et al. Deep learning with EEG 18.
spectrograms in rapid eye movement behavior disorder. Front Neurol. 110. Li Y, Lu J, Jiang J, Zhang H, Zuo C. Radiomics: a novel feature extrac-
2019;10. tion method for brain neuron degeneration disease using 18F-FDG
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5902 BORCHERT ET AL .

PET imaging and its implementation for Alzheimer’s disease and mild 130. Liu M, Cheng D, Yan W. Alzheimer’s disease neuroimaging initiative.
cognitive impairment. Ther Adv Neurol Disord. 2019;12. classification of Alzheimer’s disease by combination of convolu-
111. Teng L, Li Y, Zhao Y, et al. Predicting MCI progression with FDG-PET tional and recurrent neural networks using FDG-PET images. Front
and cognitive scores: a longitudinal study. BMC Neurol. 2020;20:148. Neuroinformatics. 2018;12.
112. Cabral C, Morgado PM, Campos Costa D, Silveira M. Alzheimer’s Dis- 131. Suk H-I, Lee S-W, Shen D. Alzheimer’s Disease Neuroimaging Ini-
ease Neuroimaging I. Predicting conversion from MCI to AD with tiative. Latent feature representation with stacked auto-encoder
FDG-PET brain images at different prodromal stages. Comput Biol for AD/MCI diagnosis. Brain Struct Funct. 2015;220:841-859. doi:10.
Med. 2015;58:101-109. 1007/s00429-013-0687-3
113. Shen T, Jiang J, Lu J, et al. Predicting Alzheimer disease from mild cog- 132. Zhang F, Li Z, Zhang B, Du H, Wang B, Zhang X. Multi-modal deep
nitive impairment with a deep belief network based on 18F-FDG-PET learning model for auxiliary diagnosis of Alzheimer’s disease. Neu-
images. Mol Imaging. 2019;18:1536012119877285. rocomputing. 2019;361:185-195. doi:10.1016/[Link].2019.04.
114. Ota K, Oishi N, Ito K, Fukuyama H, SEAD-J Study Group. Alzheimer’s 093
Disease Neuroimaging Initiative. Effects of imaging modalities, brain 133. Shao W, Peng Y, Zu C, Wang M, Zhang D. Hypergraph based multi-
atlases and feature selection on prediction of Alzheimer’s disease. task feature selection for multimodal classification of Alzheimer’s
J Neurosci Methods. 2015;256:168-183. doi:10.1016/[Link]. disease. Comput Med Imaging Graph. 2020;80:no pagination.
2015.08.020 134. Zu C, Jie B, Liu M, et al. Label-aligned multi-task feature learning for
115. Zhan Y, Chen K, Wu X, et al. Identification of conversion from normal multimodal classification of Alzheimer’s disease and mild cognitive
elderly cognition to Alzheimer’s disease using multimodal support impairment. Brain Imaging Behav. 2016;10:1148-1159.
vector machine. J Alzheimers Dis. 2015;47:1057-1067. 135. Liu L, Fu L, Zhang X, et al. Combination of dynamic (11)C-PIB PET and
116. Ben Bouallègue F, Mariano-Goulart D, Payoux P, Alzheimer’s Disease structural MRI improves diagnosis of Alzheimer’s disease. Psychiatry
Neuroimaging Initiative (ADNI). Joint Assessment of Quantitative Res. 2015;233:131-140. doi:10.1016/[Link].2015.05.014
18F-Florbetapir and 18F-FDG Regional Uptake Using Baseline Data 136. Giorgio J, Landau SM, Jagust WJ, Tino P, Kourtzi Z. Modelling
from the ADNI. J Alzheimers Dis JAD. 2018;62:399-408. doi:10.3233/ prognostic trajectories of cognitive decline due to Alzheimer’s
JAD-170833 disease. NeuroImage Clin. 2020;26:102199. doi:10.1016/[Link].2020
117. Yang BH, Chen JC, Chou WH, et al. Classification of Alzheimer’s .102199
Disease from 18F-FDG and 11C-PiB PET imaging biomarkers using 137. Borroni B, Anchisi D, Paghera B, et al. Combined 99mTc-ECD
support vector machine. J Med Biol Eng. 2020;40:545-554. SPECT and neuropsychological studies in MCI for the assessment
118. Liu F, Wee C-Y, Chen H, Shen D. Inter-modality relationship con- of conversion to AD. Neurobiol Aging. 2006;27:24-31. doi:10.1016/j.
strained multi-modality multi-task feature selection for Alzheimer’s neurobiolaging.2004.12.010
Disease and mild cognitive impairment identification. NeuroImage. 138. Habert M-O, Horn J-F, Sarazin M, et al. Brain perfusion SPECT with
2014;84:466-475. doi:10.1016/[Link].2013.09.015 an automated quantitative tool can identify prodromal Alzheimer’s
119. El-Gamal FEA, Elmogy MM, Ghazal M, et al. A novel early diagnosis disease among patients with mild cognitive impairment. Neurobiol
system for mild cognitive impairment based on local region analysis: Aging. 2011;32:15-23. doi:10.1016/[Link].2009.01.013
a pilot study. Front Hum Neurosci. 2018;11:no pagination. 139. Segovia F, Bastin C, Salmon E, Górriz JM, Ramírez J, Phillips C. Com-
120. Xu L, Wu X, Chen K, Yao L. Multi-modality sparse representation- bining PET images and neuropsychological test data for automatic
based classification for Alzheimer’s disease and mild cognitive diagnosis of Alzheimer’s disease. PloS One. 2014;9:e88687. doi:10.
impairment. Comput Methods Programs Biomed. 2015;122:182-190. 1371/[Link].0088687
121. Nozadi SH, Kadoury S. Classification of Alzheimer’s and MCI patients 140. Bhagwat N, Pipitone J, Voineskos AN, Chakravarty MM. Alzheimer’s
from semantically parcelled PET images: a comparison between Disease Neuroimaging Initiative. An artificial neural network model
AV45 and FDG-PET. Int J Biomed Imaging. 2018;2018:no pagination. for clinical score prediction in Alzheimer disease using structural
122. Choi H, Jin KH. Alzheimer’s Disease Neuroimaging I. Predicting cog- neuroimaging measures. J Psychiatry Neurosci JPN. 2019;44:246-260.
nitive decline with deep learning of brain metabolism and amyloid doi:10.1503/jpn.180016
imaging. Behav Brain Res. 2018;344:103-109. 141. Zhou J, Liu J, Narayan VA, Ye J. Modeling disease progression via
123. Ding Y, Sohn JH, Kawczynski MG, et al. A deep learning model to pre- multi-task learning. NeuroImage. 2013;78:233-248. doi:10.1016/j.
dict a diagnosis of Alzheimer disease by using 18F-FDG PET of the neuroimage.2013.03.073
Brain. Radiology. 2019;290:456-464. 142. Lei B, Hou W, Zou W, Li X, Zhang C, Wang T. Longitudinal score pre-
124. Huang Y, Xu J, Zhou Y, Tong T, Zhuang X. Diagnosis of Alzheimer’s diction for Alzheimer’s disease based on ensemble correntropy and
disease via multi-modality 3D convolutional neural network. Front spatial-temporal constraint. Brain Imaging Behav. 2019;13:126-137.
Neurosci. 2019;13. doi:10.1007/s11682-018-9834-z
125. Son HJ, Oh JS, Oh M, et al. The clinical feasibility of deep learning- 143. Battineni G, Chintalapudi N, Amenta F, Traini E. A comprehen-
based classification of amyloid PET images in visually equivocal cases. sive machine-learning model applied to magnetic resonance imaging
Eur J Nucl Med Mol Imaging. 2020;47:332-341. (MRI) to predict Alzheimer’s disease (AD) in older subjects. J Clin Med.
126. Blazhenets G, Ma Y, Sorensen A, et al. Principal components analysis 2020;9:2146. doi:10.3390/jcm9072146
of brain metabolism predicts development of Alzheimer dementia. J 144. Bachli MB, Sedeño L, Ochab JK, et al. Evaluating the reliability of neu-
Nucl Med. 2019;60:837-843. rocognitive biomarkers of neurodegenerative diseases across coun-
127. Morgado P, Silveira M, Marques JS. Diagnosis of Alzheimer’s disease tries: a machine learning approach. NeuroImage. 2020;208:116456.
using 3D local binary patterns. Comput Methods Biomech Biomed Eng doi:10.1016/[Link].2019.116456
Imaging Vis. 2013;1:2-12. 145. Cajanus A, Hall A, Koikkalainen J, et al. Automatic MRI quantifying
128. Popuri K, Balachandar R, Alpert K, et al. Development and valida- methods in behavioral-variant frontotemporal dementia diagno-
tion of a novel dementia of Alzheimer’s type (DAT) score based on sis. Dement Geriatr Cogn Disord Extra. 2018;8:51-59. doi:10.1159/
metabolism FDG-PET imaging. NeuroImage Clin. 2018;18:802-813. 000486849
129. Lu D, Popuri K, Ding GW, Balachandar R, Beg MF. Alzheimer’s disease 146. Möller C, Pijnenburg YAL, van der Flier WM, et al. Alzheimer dis-
neuroimaging I. Multimodal and multiscale deep neural networks for ease and behavioral variant frontotemporal dementia: automatic
the early diagnosis of Alzheimer’s disease using structural MR and classification based on cortical atrophy for single-subject diagnosis.
FDG-PET images. Sci Rep. 2018;8:5697. Radiology. 2016;279:838-848. doi:10.1148/radiol.2015150220
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
BORCHERT ET AL . 5903

147. Perry DC, Brown JA, Possin KL, et al. Clinicopathological correla- using structural MRI and mini-mental state examination. J Neurosci
tions in behavioural variant frontotemporal dementia. Brain J Neurol. Methods. 2018;302:66-74. doi:10.1016/[Link].2018.01.003
2017;140:3329-3345. doi:10.1093/brain/awx254 166. Qiu S, Joshi PS, Miller MI, et al. Development and validation of an
148. Wang J, Redmond SJ, Bertoux M, Hodges JR, Hornberger M. A interpretable deep learning framework for Alzheimer’s disease clas-
comparison of magnetic resonance imaging and neuropsychological sification. Brain J Neurol. 2020;143:1920-1933. doi:10.1093/brain/
examination in the diagnostic distinction of Alzheimer’s disease and awaa137
behavioral variant frontotemporal dementia. Front Aging Neurosci. 167. Mendelson AF, Zuluaga MA, Lorenzi M, Hutton BF, Ourselin S. Selec-
2016;8:119. doi:10.3389/fnagi.2016.00119 tion bias in the reported performances of AD classification pipelines.
149. Kloppel S. Brain morphometry and functional imaging techniques NeuroImage Clin. 2017;14:400-416. doi:10.1016/[Link].2016.12.
in dementia: methods, findings and relevance in forensic neurology. 018
Curr Opin Neurol. 2009;22:612-616. 168. Sun W, Nasraoui O, Shafto P. Evolution and impact of bias in
150. Bruun M, Rhodius-Meester HFM, Koikkalainen J, et al. Evaluating human and machine learning algorithm interaction. PLOS ONE.
combinations of diagnostic tests to discriminate different dementia 2020;15:e0235502. doi:10.1371/[Link].0235502
types. Alzheimers Dement Amst Neth. 2018;10:509-518. doi:10.1016/ 169. Parikh RB, Teeple S, Navathe AS. Addressing bias in artificial intelli-
[Link].2018.07.003 gence in health care. JAMA. 2019;322:2377-2378. doi:10.1001/jama.
151. Houmani N, Vialatte F, Gallego-Jutglà E, et al. Diagnosis of 2019.18058
Alzheimer’s disease with electroencephalography in a differential 170. Williams DR. Miles to go before we sleep: racial inequities
framework. PloS One. 2018;13:e0193607. doi:10.1371/[Link]. in health. J Health Soc Behav. 2012;53:279-295. doi:10.1177/
0193607 0022146512455804
152. Koikkalainen J, Rhodius-Meester H, Tolonen A, et al. Differential 171. Obermeyer Z, Powers B, Vogeli C, Mullainathan S. Dissecting racial
diagnosis of neurodegenerative diseases using structural MRI data. bias in an algorithm used to manage the health of populations. Science.
NeuroImage Clin. 2016;11:435-449. doi:10.1016/[Link].2016.02.019 2019;366:447-453. doi:10.1126/science.aax2342
153. Oppedal K, Engan K, Eftestøl T, Beyer M, Aarsland D. Classifying 172. Razai MS, Kankam HKN, Majeed A, Esmail A, Williams DR. Mitigat-
Alzheimer’s disease, Lewy body dementia, and normal controls using ing ethnic disparities in covid-19 and beyond. BMJ. 2021;372:m4921.
3D texture analysis in magnetic resonance images. Biomed Signal doi:10.1136/bmj.m4921
Process Control. 2017;33:19-29. doi:10.1016/[Link].2016.10.007 173. Mukadam N, Cooper C, Livingston G. A systematic review of eth-
154. Ritter K, Lange C, Weygandt M, et al. Combination of structural MRI nicity and pathways to care in dementia. Int J Geriatr Psychiatry.
and FDG-PET of the brain improves diagnostic accuracy in newly 2011;26:12-20. doi:10.1002/gps.2484
manifested cognitive impairment in geriatric inpatients. J Alzheimers 174. Iwatsubo T, Iwata A, Suzuki K, et al. Japanese and North Ameri-
Dis JAD. 2016;54:1319-1331. doi:10.3233/JAD-160380 can Alzheimer’s disease neuroimaging initiative studies: harmoniza-
155. Myszczynska MA, Ojamies PN, Lacoste AMB, et al. Applications of tion for international trials. Alzheimers Dement. 2018;14:1077-1087.
machine learning to diagnosis and treatment of neurodegenerative doi:10.1016/[Link].2018.03.009
diseases. Nat Rev Neurol. 2020;16:440-456. doi:10.1038/s41582- 175. Lee J, Banerjee J, Khobragade PY, Angrisani M, Dey AB. LASI-DAD
020-0377-8 study: a protocol for a prospective cohort study of late-life cogni-
156. Poldrack RA, Baker CI, Durnez J, et al. Scanning the horizon: towards tion and dementia in India. BMJ Open. 2019;9:e030300. doi:10.1136/
transparent and reproducible neuroimaging research. Nat Rev Neu- bmjopen-2019-030300
rosci. 2017;18:115-126. doi:10.1038/nrn.2016.167 176. Rohrer JD, Nicholas JM, Cash DM, et al. Presymptomatic cognitive
157. Van Essen DC, Smith SM, Barch DM, et al. The WU-Minn Human Con- and neuroanatomical changes in genetic frontotemporal dementia
nectome Project: an overview. NeuroImage. 2013;80:62-79. doi:10. in the Genetic Frontotemporal dementia Initiative (GENFI) study:
1016/[Link].2013.05.041 a cross-sectional analysis. Lancet Neurol. 2015;14:253-262. doi:10.
158. Nichols TE, Das S, Eickhoff SB, et al. Best practices in data analysis 1016/S1474-4422(14)70324-2
and sharing in neuroimaging using MRI. Nat Neurosci. 2017;20:299- 177. Mehrabi N, Morstatter F, Saxena N, Lerman K, Galstyan A. A sur-
303. doi:10.1038/nn.4500 vey on bias and fairness in machine learning. ACM Comput Surv.
159. Pernet C, Garrido MI, Gramfort A, et al. Issues and recommendations 2021;54:115:1-115:35. doi:10.1145/3457607
from the OHBM COBIDAS MEEG committee for reproducible EEG 178. A Geometric Solution to Fair Representations | Proceedings of the
and MEG research. Nat Neurosci. 2020;23:1473-1483. doi:10.1038/ AAAI/ACM Conference on AI, Ethics, and Society n.d. (accessed April
s41593-020-00709-0 4, 2023). [Link]
160. Sullivan I, DeHaven A, Mellor D. Open and reproducible research on 179. Bellamy RKE, Dey K, Hind M, et al. AI Fairness 360: an extensible
open science framework. Curr Protoc Essent Lab Tech. 2019;18:e32. toolkit for detecting and mitigating algorithmic bias. IBM J Res Dev.
doi:10.1002/cpet.32 2019;63:4:1-4:15. doi:10.1147/JRD.2019.2942287
161. Gentili C, Cecchetti L, Handjaras G, Lettieri G, Cristea IA. The case for 180. Bohr A, Memarzadeh K. The rise of artificial intelligence in health-
preregistering all region of interest (ROI) analyses in neuroimaging care applications. Artif Intell Healthc. 2020:25-60. doi:10.1016/B978-
research. Eur J Neurosci. 2021;53:357-361. doi:10.1111/ejn.14954 0-12-818438-7.00002-2
162. Hildebrandt M. Preregistration of machine learning research design. 181. Obermeyer Z, Emanuel EJ. Predicting the future — Big Data, machine
Against P-Hacking. PROFILEDCOGITAS SUM COGITAS SUM 10 Years learning, and clinical medicine. N Engl J Med. 2016;375:1216-1219.
Profiling Eur. Citiz. Amsterdam University Press; 2018:102-105. doi:10.1056/NEJMp1606181
doi:10.1515/9789048550180-019 182. Yu K-H, Beam AL, Kohane IS. Artificial intelligence in healthcare. Nat
163. Dunne RA, Aarsland D, O’Brien JT, et al. Mild cognitive impairment: Biomed Eng. 2018;2:719-731. doi:10.1038/s41551-018-0305-z
the manchester consensus. Age Ageing. 2021;50:72-80. doi:10.1093/ 183. Lapointe L, Rivard S. A multilevel model of resistance to information
ageing/afaa228 technology implementation. MIS Q. 2005;29:461-491. doi:10.2307/
164. Beekly DL, Ramos EM, van Belle G, et al. The national Alzheimer’s 25148692
coordinating center (NACC) database: an Alzheimer disease 184. Liberati EG, Ruggiero F, Galuppo L, et al. What hinders the uptake
database. Alzheimer Dis Assoc Disord. 2004;18:270-277. of computerized decision support systems in hospitals? A quali-
165. Sørensen L, Nielsen M, Alzheimer’s Disease Neuroimaging Initia- tative study and framework for implementation. Implement Sci IS.
tive. Ensemble support vector machine classification of dementia 2017;12:113. doi:10.1186/s13012-017-0644-2
 15525279, 2023, 12, Downloaded from [Link] by NICE, National Institute for Health and Care Excellence, Wiley Online Library on [22/10/2025]. See the Terms and Conditions ([Link] on Wiley Online Library for rules of use; OA articles are governed by the applicable Creative Commons License
5904 BORCHERT ET AL .

185. Antoniadi AM, Du Y, Guendouz Y, et al. Current challenges and future methods systematic review. Lancet Digit Health. 2021;3:e599-611.
opportunities for XAI in machine learning-based clinical decision sup- doi:10.1016/S2589-7500(21)00132-1
port systems: a systematic review. Appl Sci. 2021;11:5088. doi:10.
3390/app11115088
186. McDermid JA, Jia Y, Porter Z, Habli I. Artificial intelligence explain- SUPPORTING INFORMATION
ability: the technical and ethical dimensions. Philos Transact A Additional supporting information can be found online in the Support-
Math Phys Eng Sci. 2021;379:20200363. doi:10.1098/rsta.2020
ing Information section at the end of this article.
.0363
187. Gerke S, Minssen T, Cohen G. Ethical and legal challenges of artifi-
cial intelligence-driven healthcare. Artif Intell Healthc. 2020:295-336.
doi:10.1016/B978-0-12-818438-7.00012-5 How to cite this article: Borchert RJ, Azevedo T, Badhwar AP,
188. Jamjoom AAB, Jamjoom AMA, Thomas JP, et al. Autonomous
et al. Artificial intelligence for diagnostic and prognostic
surgical robotic systems and the liability dilemma. Front Surg.
neuroimaging in dementia: A systematic review. Alzheimer’s
2022;9:1015367. doi:10.3389/fsurg.2022.1015367
189. Young AT, Amara D, Bhattacharya A, Wei ML. Patient and gen- Dement. 2023;19:5885–5904.
eral public attitudes towards clinical artificial intelligence: a mixed [Link]