Artificial Intelligence In Medicine 150 (2024) 102821

Contents lists available at ScienceDirect

Artificial Intelligence In Medicine

journal homepage: [Link]/locate/artmed

Research paper

Analyzing entropy features in time-series data for pattern recognition in

neurological conditions
Yushan Huang a,f , Yuchen Zhao b , Alexander Capstick d,f , Francesca Palermo d,f , Hamed Haddadi c ,
Payam Barnaghi d,e,f,g ,∗
a
Dyson School of Design Engineering, Imperial College London, London, UK
b Department of Computer Science, University of York, York, UK
c Department of Computing, Imperial College London, London, UK
d Department of Brain Sciences, Imperial College London, London, UK
e
The Great Ormond Street Institute of Child Health, University College London, London, UK
f
Great Ormond Street Hospital for Children, London, UK
g
Care Research and Technology Centre, The UK Dementia Research Institute, London, UK

ARTICLE INFO ABSTRACT

Keywords: In the field of medical diagnosis and patient monitoring, effective pattern recognition in neurological time-
Entropy series data is essential. Traditional methods predominantly based on statistical or probabilistic learning and
Time-series data inference often struggle with multivariate, multi-source, state-varying, and noisy data while also posing
Pattern recognition
privacy risks due to excessive information collection and modeling. Furthermore, these methods often overlook
Neurological conditions
critical statistical information, such as the distribution of data points and inherent uncertainties. To address
these challenges, we introduce an information theory-based pipeline that leverages specialized features to
identify patterns in neurological time-series data while minimizing privacy risks. We incorporate various
entropy methods based on the characteristics of different scenarios and entropy. For stochastic state transition
applications, we incorporate Shannon’s entropy, entropy rates, entropy production, and the von Neumann
entropy of Markov chains. When state modeling is impractical, we select and employ approximate entropy,
increment entropy, dispersion entropy, phase entropy, and slope entropy. The pipeline’s effectiveness and
scalability are demonstrated through pattern analysis in a dementia care dataset and also an epileptic and
a myocardial infarction dataset. The results indicate that our information theory-based pipeline can achieve
average performance improvements across various models on the recall rate, F1 score, and accuracy by up to
13.08 percentage points, while enhancing inference efficiency by reducing the number of model parameters
by an average of 3.10 times. Thus, our approach opens a promising avenue for improved, efficient, and critical
statistical information-considered pattern recognition in medical time-series data.

1. Introduction obstacles become even more critical in areas such as dementia, heart
disease, and epilepsy analysis, where swift and accurate interpretation
With the progress in digital health technologies, an ever-increasing of time-series data can greatly influence the quality of patient care.
amount of health-related data is being produced, creating unparal- The analysis of neural time-series data occupies a central role in
leled opportunities for biomedical research and the improvement of fields such as neuroscience, medicine, and bioinformatics. This data
healthcare services [1,2]. The amalgamation of real-world patient data is typically sourced from electroencephalography (EEG) [5], functional
with these technologies allows for the analysis and extraction of crit- magnetic resonance imaging (fMRI) [6], or other neural recording tech-
ical health information, which in turn facilitates more precise clinical nologies. It is essential for diagnostic decision-support and healthcare
decision-making and enhances patient outcomes. However, the analysis applications. Moreover, there are more generalized forms of time-series
of large, multi-sourced, and often noisy datasets introduces signifi- data, such as activity data, which are crucial for analyzing neurologi-
cant challenges, particularly regarding privacy issues and the effective cal conditions like dementia and traumatic brain injury [7–9]. These
recognition of patterns in neurological time-series data [3,4]. These time-series datasets are characterized by their high dimensionality,

∗ Corresponding author at: Department of Brain Sciences, Imperial College London, London, UK.
E-mail address: [Link]@[Link] (P. Barnaghi).

[Link]
Received 2 June 2023; Received in revised form 14 February 2024; Accepted 21 February 2024
Available online 22 February 2024
0933-3657/© 2024 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license ([Link]
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 1. An overview of the healthcare monitoring IoT platform.

non-linearity, and non-stationarity, which make their analysis and in- (1) We present an information theory-based method for analyzing
terpretation quite challenging. Traditional linear methods, such as the neural time-series data, tailored for use within a decentralized health-
Fourier transform [10] and autoregressive models [11], can be effective care monitoring IoT platform (as depicted in Fig. 1). This method is en-
in certain scenarios but often struggle to capture the complex dynamics gineered to consistently perform well across multivariate, multi-source,
and non-linear patterns inherent in the data. Consequently, more so- state-varying, and noisy neural time-series datasets. We specifically
phisticated methods, including neural network-based and information differentiate neurological time series data into two categories: state-
theory-based approaches, have become increasingly popular for analyz- transition and non-state-transition data. Depending on the scenario and
ing neural time-series data. For instance, Recurrent Neural Networks data characteristics, we apply and choose different entropy methods.
(RNNs) [12] and Long Short-Term Memory networks (LSTMs) [13] In our decentralized healthcare monitoring IoT platform, data prepro-
have been utilized to handle the data’s long-term dependencies and cessing and the extraction of entropy features are performed locally,
dynamic changes. Additionally, information theory-based methods like whereas model execution and medical decision-making are centralized
entropy and mutual information have been applied to quantify the on cloud servers.
complexity and uncertainty of the data, uncovering valuable insights (2) Compared to traditional neural network-based methods, the pro-
hidden within [14,15]. posed approach can select and extract the most informative, relevant,
Despite the progress made by information theory-based and neu- and representative high-level features from the data. These features are
ral network methods to analyze neural time series data, they still based on information theory, consider critical statistical information,
face significant limitations and challenges when dealing with complex and are easier to understand and interpret, thereby enhancing the inter-
and dynamic data. Most of the current studies are task-specific and pretability of neural networks. This also mitigates privacy and security
lack generalization, requiring specific analysis and algorithm/model concerns because feature extraction is performed locally, eliminating
design when faced with new tasks or scenarios. On the other hand, the need to upload raw data to cloud servers. These entropy features
some of the methods are highly sensitive to raw data and perform enhance prediction accuracy.
poorly when handling multivariate, multi-source, state-varying, and (3) Our work showcases broad applicability and generalizability
noisy data. In addition, the existing methods often overlook critical sta- across a variety of datasets. We implement the proposed pipeline on a
tistical information, such as the distribution of data points and inherent dementia care dataset derived from our clinical study on remote health-
uncertainties in the raw data. The neural network-based methods often care monitoring. To further validate its scalability, we test the pipeline
embed unnecessary non-task-related information, increasing privacy on two publicly available datasets to demonstrate the effectiveness
risks [16]. For example, in Federated Learning (FL), uploading raw across different healthcare contexts: the Epileptic Seizure Recognition
time-series data to cloud servers can result in privacy breaches and Dataset [17] and the PTB Diagnostic ECG Database [18,19].
impose significant transmission bandwidth constraints. Therefore, there (4) Our study highlights the performance and efficiency gains
is an unmet need to develop more general, robust, and efficient ap- achievable with various models through the use of our method. Specif-
proaches to extract high-level features from multivariate, multi-source, ically, we evaluate the effectiveness of the extracted features using
state-varying, and noisy neural time series data. This is crucial to en- different models such as Logistic Regression (LR), Support Vector
hance model performance and accuracy, meeting the high requirements Machines (SVM), Multi-Layer Perceptron (MLP), Convolutional Neural
for data analysis precision and interpretability in clinical and research Network (CNN), and Long Short-Term Memory (LSTM). Our exper-
settings. imental results show that, for the three datasets, compared to the
In previous work, we conducted a preliminary analysis of three baseline methods, the information theory-based pipeline can signifi-
Markov chain-based entropy features via heat maps [7], highlighting cantly improve the accuracy, recall, and F1 scores of the models by
the potential of entropy in analyzing multivariate, multi-source, rapidly an average of up to 13.08 percentage points (pp), and can simplify the
state-varying, and noisy neurological time-series data. However, it model structure and enhance efficiency, reducing the number of model
stopped short of presenting a complete pipeline for analyzing such parameters by an average of 3.10 times.
data and did not validate the effectiveness of these methods with The remainder of this paper is structured as follows: Section 2 delves
machine learning (ML) models. In this paper, we introduce an in- into related works and examines the challenges within the scope of
formation theory-based method for neural time-series analysis within this research. Section 3 outlines the technical aspects of our proposed
non-centralized medical decision systems, designed to maintain stable method, including an introduction to several entropy methods related
performance across multivariate, multi-source, state-varying, and noisy to Markov chains and our feature selection methods. In Section 4, we
data contexts. The primary contributions of this paper are outlined as detail the implementation and evaluation framework for our experi-
follows: ments, covering datasets, models, and performance metrics. Section 5

2
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 2. Method overview.

provides a thorough analysis of the experimental outcomes. Discussion data through the extraction of key features, the complexity of the data
of the method and its limitations is presented in Section 6. The pa- and the model’s ‘‘black-box’’ nature are reduced. This reduction not
per concludes in Section 7, summarizing our findings and suggesting only streamlines model input but also addresses privacy and security
directions for future research. concerns by minimizing the amount of information processed by the
The source code, constructed models, and links to the public network, thereby focusing the model on specific goals. In decentralized
datasets are made available on a GitHub repository [20]. machine learning systems, this aspect of privacy protection is accentu-
ated, as only the essential extracted features, rather than raw data, are
2. Related works transmitted to the server side. The concept of entropy, introduced by
Shannon to quantify information uncertainty, lays the foundational sci-
Among the solutions developed to tackle these challenges, deep entific theory for modern information theory [26]. Following Shannon’s
neural networks (DNNs) have gained popularity for their proficiency in entropy, various entropy variants like spectral entropy [27] and sample
learning the spatio-temporal characteristics of data and autonomously entropy [28] have been developed. Studies such as those by Nurwulan
extracting features for pattern recognition and prediction outcomes et al., comparing traditional features with multi-scale entropy (MSE)
[21]. These models often employ feature extraction techniques such features extracted from 3-axis acceleration data, have demonstrated
as convolutional neural networks (CNN) and long short-term memory MSE’s superior performance in KNN and random forest (RF) classi-
(LSTM) units. Notable innovations include the CNN-LSTM architecture fications [29]. Similarly, Bao et al. have shown the effectiveness of
analyzed by Hussain et al. [22], the CNN-LSTM framework augmented frequency-domain entropy features combined with statistical measures
with a self-attention mechanism by Park et al. [23], and the inte- like mean, energy, and correlation for building predictive models [30].
gration of transformers and generative adversarial networks (GANs) These approaches not only enhance learning model performance and
by Shankar et al. [24]. However, despite their benefits, DNNs can safeguard privacy but also often outperform traditional deep neural
be cumbersome, inefficient, and may not fully leverage critical sta- network models, which is crucial for informed clinical decision-making.
tistical data insights, such as point distributions and the inherent However, many existing studies either focus solely on a single entropy
uncertainties within these distributions, potentially rich in useful, high- feature or use entropy features alongside traditional ones without tai-
level features [25]. Additionally, these models risk privacy breaches loring the selection to specific data characteristics, such as Markovian
by often requiring the input of all raw data for processing, which systems and stochastic state transitions, or justify the choice of entropy
collects more information than necessary. A further issue is the opacity features. Moreover, these methods tend to be task-specific and lack a
and limited interpretability of neural network-based methods, hin- comprehensive, scalable entropy-based pipeline applicable to various
dering scientists’ and engineers’ deep comprehension of the models tasks.
and restricting their application in crucial decision-making scenarios, A study closely related to ours is the entropy measurement model
particularly in medicine and healthcare. In these fields, model inter- developed by Howedi et al. [31], which employs approximate entropy
pretability is essential, as it significantly influences the confidence (ApEn), sample entropy (SampEn), and fuzzy entropy (FuzzyEn) for
and reliance that healthcare professionals and patients have on the detecting visitors in a home environment. However, this study does
model’s predictions. This underscores the necessity for methodologies not tailor the selection of entropy features to the data characteristics,
that can adeptly manage complex data, safeguard privacy, and eluci- such as Markovian systems and stochastic state transitions, nor does it
date high-level features, thus facilitating improved decision-making in provide a rationale for the chosen entropy features.
health-related contexts.
Information theory-based methods are increasingly utilized in ana- 3. Methodology
lyzing neural time-series data, particularly for feature extraction using
mathematical principles combined with machine learning models for The process proposed in this paper consists of four primary com-
data analysis. These approaches enable the identification and extraction ponents: data collection, data pre-processing, feature construction, and
of the most informative, relevant, and representative features, which modeling, as illustrated in Fig. 2. The initial three stages—data collec-
are often more interpretable due to their direct relationship with the tion, pre-processing, and feature construction are executed locally on
data’s inherent properties and structure. By simplifying the original an edge device (such as a gateway), whereas the modeling phase takes

3
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 3. An example of the state-transition data and Markov chains.

place in the cloud. In detail, sensors gather raw data, which is then construct entropy features based on the Markov chain. For non-state-
forwarded to the edge device. One advantage of this approach is that transition data, we first construct various entropy features (such as
the processing of the raw data can be conducted locally at the edge, dispersion entropy and phase entropy) and then proceed with feature
and only the extracted features are transmitted to the cloud, enhancing selection. Next, we provide a detailed introduction to the entropy
the privacy protection of medical information. This also maximizes features based on the Markov chain and the feature selection methods.
the computational capabilities of both the edge and cloud, improving The construction methods for other entropy features can be found in
system efficiency. the Appendix.
A key benefit of this methodology is its local processing of raw data
at the edge, permitting only the distilled features to be uploaded to 3.2.1. State-transition data and Markov chains
the cloud. This strategy significantly bolsters the privacy protection of State-transition data, exemplified by the output from PIR sensors,
medical data. Furthermore, it leverages the computational strengths of encapsulates both the time and location of sensor activation, thereby
both edge and cloud platforms, thereby enhancing the overall efficiency providing rich spatial and temporal insights. Spatially, each activation
of the system. location is considered an individual node, and nodes are sequentially
linked according to the order of sensor triggers to construct a pathway,
3.1. Data collection and pre-processing representing a first-order Markov chain, as depicted in Fig. 3. This
arrangement allows for the computation of how frequently and how
Data and activity collection are achieved through various sensors many times a specific location or path is activated. Temporally, the
installed in the home. We classify the collected time-series signals analysis extends to calculating the duration of occupancy at a given
into state-transition and non-state-transition based on whether the
location and the transit time between two points, location A to location
time-series signals can constitute a state transition. State-transition
B. Extracting these high-level features, which encapsulate both spatial
time-series signals can form state transitions, for example, PIR sensors,
and temporal dimensions, presents a challenge for conventional deep
as shown in Fig. 2.A, which are placed at different locations in the
learning models due to their complexity. Hence, our feature extraction
home and are triggered as someone passes by, record both the location
process leverages the unique properties of state-transition data through
and time simultaneously. As Fig. 3 shows, these points are discrete.
the application of Markov chains and entropy techniques.
Through data pre-processing, the discrete points can be connected to
represent transition signals. Therefore, this kind of time-series data is
referred to as state-transition signals. In contrast, non-state-transition 3.2.2. Shannon’s entropy of a Markov chain
signals cannot form signals representing state transitions, such as EEG We assume that a specific human activity, such as a sequence of
signals. These signals are continuous and challenging to transform into locations, could be modeled as a Markov chain. In this model, the
signals that depict state transitions. occurrences of these activities are treated as random events, and their
In the data pre-processing phase, addressing missing values is our frequency and pattern could be quantitatively measured. Shannon’s
initial step. To handle these, we utilize forward-fill or backward- entropy is used to effectively represent the complexity and variability
fill techniques, contingent on their contextual placement and dis- in human activity patterns. This choice is motivated by Shannon’s
tribution within the dataset. Predominantly, forward-fill is applied, entropy’s ability to quantify the uncertainty and complexity of informa-
but backward-fill comes into play for missing values situated at the tion in a data-driven manner, eliminating the need for pre-established
dataset’s commencement. This strategy is adopted to maintain the con- assumptions and models. Suppose that there are 𝑛 locations 𝑋 =
tinuity and integrity of the data. Further, we implement label encoding 𝑥1 , 𝑥2 , … , 𝑥𝑛 in a participant’s activity, then Shannon’s entropy of a
to convert categorical data into a machine-readable format and apply Markov chain 𝐻(𝑥) can be described as:
Z-score standardization. The latter process normalizes the data, setting ∑
𝑛
( ) ( )
the mean to 0 and standard deviation to 1, thereby optimizing the 𝐻(𝑋) = − 𝑃 𝑥𝑖 log 𝑃 𝑥𝑖 (1)
data for improved model performance. The final step involves aligning 𝑖=1

the data to ensure uniformity in scale across all features, facilitating In which 𝑃 (𝑋𝑖 ) is the probability of activity 𝑥𝑖 . When the frequency
smoother subsequent analysis. of a participant’s activity changes, 𝐻(𝑥) will change accordingly to
represent the change in activity pattern. A higher value indicates
3.2. Feature construction increased uncertainty and complexity, reflecting a more diverse and
unpredictable human activity pattern. Conversely, a lower value signi-
We utilize entropy and its variants to extract features from the raw fies reduced uncertainty, indicative of a more regular and predictable
data. Specifically, for state-transition data, after pre-processing it, we pattern of behavior.

4
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 4. An example of the Entropy rate of a Markov chain. In ⃝ 1 , the rectangular boxes represent the locations (states) in the Markov chain, and the arrows represent the routes
between locations in the house. Different colors represent different start locations (blue: bathroom, orange: bedroom, green: kitchen, black: lounge, and red: hallway). The numbers
next to the lines represent route probabilities which correspond to the table ⃝ 2 and Transition Matrix 𝑇 ⃝3 . (For interpretation of the references to color in this figure legend, the
reader is referred to the web version of this article.)

Algorithm 1 Entropy rate of a Markov chain respectively. Then the probability 𝑃𝑖𝑗 of the route from 𝑥𝑖 to 𝑥𝑗 can be
1: Define: 𝑆 = {𝑠1 , 𝑠2 , ..., 𝑠𝐿 } is a Markov chain trajectory, where 𝐿 is the represented as:
length of the trajectory, and 𝑠 ∈ 𝑋, 𝑋 = {𝑥1 , 𝑥2 , ..., 𝑥𝑛 }, 𝑛 is the number ( )
𝑃𝑖𝑗 = 𝑃 𝑥𝑗 ∣ 𝑥𝑖 (2)
of states in the Markov chain. 𝑇 𝑊1 is the time window required for the
stationary distribution, 𝑇 𝑊1 ≤ 𝐿. 𝑇 𝑊2 is the time window required for the Where 𝑥𝑖 and 𝑥𝑗 ∈ 𝑋. Suppose that there are 𝑛 states in a Markov
target task, 𝑇 𝑊2 ≤ 𝐿. 𝑃𝑖𝑗 is the probability from state 𝑥𝑖 to state 𝑥𝑗 . 𝑆𝑃 is chain, then the Markov chain can be represented as 𝑛×𝑛 matrix 𝑃𝑖𝑗 𝑖,𝑗∈𝑋 ,
the start point;
called Transition Matrix 𝑇 , an example is shown in Fig. 4. From Markov
Input: Markov chain trajectory 𝑆;
chains, stationary distributions 𝜋 can be calculated, which represents:
Output: Entropy rate 𝜉 of the Markov chain;
2: Set 𝑇 𝑊1 and 𝑇 𝑊2 ; 𝜋 = 𝜋𝑇 (3)
3: // Stationary Distribution Function
4: Function StationaryDistribution(𝑆, 𝑇 𝑊1 ) In which, 𝜋 is an n-dimension vector associated with a Markov chain
5: 𝑆𝑇 𝑊1 = 𝑆[0 ∶ 𝑇 𝑊1 ]; with 𝑛 states. Using this, the entropy rate of a Markov chain can be
( )
6: 𝑇 = 𝑃 𝑙𝑏′ = 𝑥′𝑗 ∣ 𝑙𝑎′ = 𝑥′𝑖 , where 𝑙𝑎′ , 𝑙𝑏′ ∈ 𝑋, represent the previous state expressed as [33]:
and the current state, 𝑎′ ∈ [2, 𝑇 𝑊1 ], 𝑏′ ∈ [1, 𝑇 𝑊1 − 1], 𝑥′𝑖 ∈ 𝑋, 𝑥′𝑗 ∈ 𝑋; ∑
𝑛
7: 𝜋 = 𝜋𝑇 ; 𝜉=− 𝜋𝑖 𝑃𝑖𝑗 log 𝑃𝑖𝑗 (4)
8: return 𝜋; 𝑖𝑗
9: // Entropy Rate Function
In which, 𝜋𝑖 is the probability in the stationary distribution asso-
10: Function EntropyRate(𝜋, 𝑆, 𝑇 𝑊2 )
11: for 𝑆𝑃 = 0; 𝑆𝑃 + 𝑇 𝑊2 ≤ 𝐿; 𝑆𝑃 = 𝑆𝑃 + 𝑇 𝑊2 do
ciated with activity 𝑥𝑖 ∈ 𝑋 in a Markov chain with the stationary
12: 𝑆𝑇 𝑊2 = 𝑆[𝑆𝑃 ∶ 𝑇 𝑊2 ]; distribution. When calculating the entropy rate of a Markov chain,
( )
13: 𝑃𝑖𝑗 = 𝑃 𝑙𝑏 = 𝑥𝑗 ∣ 𝑙𝑎 = 𝑥𝑖 , where 𝑙𝑎 , 𝑙𝑏 ∈ 𝑋, represent the previous there are two time-windows that need to be set, one time-window is
state and the current state, 𝑎 ∈ [2, 𝑇 𝑊2 ], 𝑏 ∈ [1, 𝑇 𝑊2 − 1], 𝑥𝑖 ∈ 𝑋, 𝑥𝑗 ∈ 𝑋; used to calculate 𝑃𝑖𝑗 for target time-series data, and the other is used
to calculate 𝜋𝑖 to represent the characteristics of time-series data. The
∑𝑛
14: 𝜉𝑚 = − 𝑥 ,𝑥 ∈𝑋 𝜋𝑖 𝑃𝑖𝑗 log 𝑃𝑖𝑗 ; time window to calculate 𝑃𝑖𝑗 is set by the task objective. It should be
𝑖 𝑗
15: end for noted that the time window to calculate the stationary distribution 𝜋𝑖 is
16: return 𝜉 = {𝜉1 , 𝜉2 , ..., 𝜉𝑚 }; important, as it should reflect the stationary pattern of the participant.
17: end
For example, participants’ routines may be affected by the seasons,
thus we need to avoid the possible effects of the seasons when setting
up the time windows to calculate the stationary distribution, such as
3.2.3. Entropy rate of a Markov chain setting the time windows to override the seasonal variations. The dual
The adoption of the entropy rate of a first-order Markov chain in time-window approach for estimating ensures that the entropy rate is
our study stems from the inherent limitations associated with Shan- sensitive to both immediate and long-term behavioral patterns, offering
non’s entropy when applied to human activity data. While Shannon’s a robust metric for analyzing and interpreting the multifaceted nature
entropy is instrumental in quantifying the uncertainty or randomness of human activities. The complete procedure for calculating the Entropy
of individual events, it falls short of capturing the sequential and Rate of a Markov Chain is shown in the Algorithm 1.
dependent nature of human activities. In scenarios where activities
are interconnected, and one event potentially influences the next, a
3.2.4. Entropy production of a Markov chain
more detailed measure is required to encapsulate the complexity and
variability inherent in such data. This motivates our approach to the In dealing with complex dynamic time-series data, Shannon’s en-
entropy rate of a Markov chain, a metric that not only accounts for tropy and entropy rate of Markov chains still have limitations. Shan-
the uncertainty of individual states but also integrates the probabilistic non’s entropy, while effective in quantifying the uncertainty of informa-
dependencies between consecutive states, offering a holistic view of the tion, often overlooks the connectivity and associativity of information.
dynamism in human activity patterns [32]. On the other hand, although the entropy rate of Markov chains con-
Suppose that 𝑋 = {𝑥1 , 𝑥2 , … , 𝑥𝑛 } represents 𝑛 states in a Markov siders the probability of state transitions, it primarily describes the
chain. Let 𝑥𝑖 , 𝑥𝑗 ∈ 𝑋 represent the previous state and the current state, system’s uncertainty and complexity at a statistical level, potentially

5
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 5. The model structure of NEEP of Markovian systems. 𝐻 is the size of the embedding dimension.

failing to unveil the system’s internal dynamic evolution and com- Algorithm 2 Training process of NEEP
plex interactions in depth. Considering the activity patterns obtained 1: Define: 𝑆 = {𝑠1 , 𝑠2 , ..., 𝑠𝐿 } is a Markov chain trajectory, where 𝐿 is the
from PIR sensors as an example, dynamic evolution refers to how the length of the trajectory, and 𝑠 ∈ 𝑋, 𝑋 = {𝑥1 , 𝑥2 , ..., 𝑥𝑛 }, 𝑛 is the number of
behavior and activity patterns change over time and under different states in the Markov chain.
conditions. For example, an individual might exhibit different activity Input: Markov chain trajectory 𝑆;
patterns during various health conditions. Output: The values calculated by the loss function 𝐽̂(𝜃);
To overcome these limitations, we consider using entropy produc- 2: while termination condition not met do
tion rate (EP) [34]. Originating in the field of physics, particularly 3: Embedding layer;
thermodynamics and statistical physics, EP not only quantifies the 4: Objective function
system’s uncertainty but also delves into the system’s instability and ∑[ ( ) ]
𝐽̂(𝜃) = 𝛥𝑆𝜃 𝑠𝑡 , 𝑠𝑡+1 − 𝑒−𝛥𝑆𝜃 (𝑠𝑡 ,𝑠𝑡+1 ) (7)
dynamic changes. This makes entropy production a powerful tool ca- 𝑡∈𝐿
pable of revealing the intrinsic dynamics and complex interactions of
5: Compute gradients ∇𝜃 𝐽̂(𝜃);
complicated dynamic systems. Although entropy production has its
6: Update parameters 𝜃;
roots in physics, its core idea focuses on quantifying and analyzing the
7: end while
instability and dynamic changes of a system. In pattern recognition,
each person’s activity patterns can be viewed as a complex dynamic
system [35,36]. By calculating the entropy production of these systems,
we can unveil the complex dynamic patterns hidden in spatiotemporal entropy measures (such as Shannon’s entropy), von Neumann entropy
data and deeply understand the driving factors behind these patterns. is based on a matrix form. This means that it can not only capture the
EP can be estimated by machine learning models such as the Neu- characteristics of a single variable in the time series but also deeply
ral Estimator for Entropy Production (NEEP), which can estimate EP grasp the interactions and relationships between different variables.
of Markovian systems [37]. Given a Markov chain trajectory 𝑆 = This integrative nature displays its strength in capturing complex data
{𝑠1 , 𝑠2 , … , 𝑠𝐿 } and a function ℎ𝜃 acting over previous state 𝑠𝑡 and the structures.
current state 𝑠𝑡+1 in the Markov chain, where 𝜃 denotes the trainable
Moreover, VNE possesses robustness, especially when dealing with
neural network parameters, then the output of NEEP can be defined
continuous values and limited data samples. The robust nature of von
as [37]:
∑[ ( ) ] Neumann entropy primarily stems from the density matrix in its com-
𝐽̂(𝜃) = 𝛥𝑆𝜃 𝑠𝑡 , 𝑠𝑡+1 − 𝑒−𝛥𝑆𝜃 (𝑠𝑡 ,𝑠𝑡+1 ) (5) putational methodology. By building an appropriate density matrix, we
𝑡∈𝐿
can capture the relationships and interactions between these states.
Where 𝛥𝑆𝜃 is: This allows VNE to consider the continuity and overall characteristics
( ) ( ) ( ) of the data. Even if the data is noisy or incomplete, it can provide
𝛥𝑆𝜃 𝑠𝑡 , 𝑠𝑡+1 ≡ ℎ𝜃 𝑠𝑡 , 𝑠𝑡+1 − ℎ𝜃 𝑠𝑡+1 , 𝑠𝑡 (6)
a suitable approximation of the system’s overall state. This approach
The model structure of NEEP is shown in Fig. 5 and the procedure ensures that, even when faced with minor perturbations or incomplete
for training NEEP is shown in Algorithm 2. In NEEP, an embedding data, VNE can deliver stable and consistent results.
layer is used to transform the discrete state into a trainable continuous
As previously mentioned, we can regard each person’s activity
vector [37], and then the embedded data is input into a hidden MLP
patterns as a complex dynamic system. Therefore, VNE offers us a new
layer. It has to be noted that, the length of the time series data is
perspective to analyze time-series data. Taking PIR sensors as an exam-
very important when training NEEP, as we need to ensure that the
ple, for the complex time-series data collected in household settings,
data for this period of time is sufficient for training and can reflect the
participant’s characteristics. we can approach from both the perspectives of activity frequency and
activity duration, as shown in Fig. 6. This allows us to analyze time-
3.2.5. Von Neumann entropy of a Markov chain series data from both spatial and temporal perspectives. In addition,
The von Neumann entropy (VNE) originates from quantum me- due to the presence of the density matrix, VNE can also conduct a
chanics. In quantum mechanics, VNE is employed to quantitatively comprehensive analysis of the data from both horizontal and vertical
depict the uncertainty of a quantum state [38]. Compared to other perspectives.

6
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 6. An example for the von Neumann entropy of a Markov chain. Suppose that there are five locations (states) in a Markov chain, and we plan to calculate the von Neumann
entropy of one week. From the perspective of spatial, ⃝ 1 : count the frequency 𝑎𝑑𝑖,𝑗 of different routes in a Markov chain for each day of the week, in which 𝑖 and 𝑗 represent
the previous location and the current location, respectively; ⃝2 : aggregate weekly activity frequency; ⃝ 3 : transfer the weekly activity frequency to the activity frequency matrix
𝐴; ⃝4 : calculate the Pearson correlation Matrix 𝑅 between each day; ⃝ 5 : calculate the density matrix 𝜌 by 𝑓 (∙); ⃝6 : calculate the von Neumann entropy by 𝑓 ′ (∙). And from the
perspective of temporal, the only difference is changing the activity frequency to activity duration.

Given a density operator 𝜌 with 𝑁 eigenvalues 𝜆1,…,𝑛 , VNE is defined Algorithm 3 von Neumann entropy of a Markov chain
as follows: 1: Define: 𝑆 = {𝑠1 , 𝑠2 , ..., 𝑠𝐿 } is a Markov chain trajectory, where 𝐿 is the
∑
𝑁 length of the trajectory, and 𝑠 ∈ 𝑋, 𝑋 = {𝑥1 , 𝑥2 , ..., 𝑥𝑛 }, 𝑛 is the number
𝑆(𝝆) = − tr(𝝆 log 𝝆) = − 𝜆𝑗 log 𝜆𝑗 (8) of states in the Markov chain. 𝑉 𝑁 is the von Neumann entropy of a
𝑗=1 Markov chain. 𝑇 𝑊2 is the time window required for the target task, where
One of the key points to calculate VNE is to obtain the density 𝑇 𝑊2 <= 𝐿. 𝑆𝑃 is the start point;
Input: Markov chain trajectory 𝑆;
operator 𝜌, which must satisfy (i) be Hermitian, (ii) have unit trace, and
Output: The 𝑉 𝑁;
(iii) be positive semi-definite. Given 𝑹 ∈ R𝑁 , an N-dimension Pearson
2: Set 𝑇 𝑊2 ;
correlation matrix of the human activity data, then the density operator 3: for 𝑆𝑃 = 0; 𝑆𝑃 + 𝑇 𝑊2 <= 𝐿; 𝑆𝑃 = 𝑆𝑃 + 𝑇 𝑊2 do
𝝆 can be defined as [39]: 4: Calculate original matrix 𝐴 (e.g., activity frequency matrix);
5: 𝑁 × 𝑁 Pearson correlation Matrix 𝑅 of 𝐴;
𝝆 = 𝑹∕𝑁 (9)
6: Density operator 𝜌 ← 𝜌 = 𝑅∕𝑁;
The density operator 𝜌, calculated by Eq. (9) satisfies all the re- 7: von Neumann entropy 𝑉 𝑁(𝜌) ← 𝑉 𝑁(𝜌) = Tr(𝜌 log 𝜌), log 𝜌 =
∑∞ 𝑘
(−1)𝑘+1 (𝐵−𝐼) ;
quirements. However, it has to be noted that the density operator 𝜌, 𝑘=1 𝑘
8: end for
which is calculated from real-world data, may be sparse, and thus, there
may be anomalies in the calculation of 𝑙𝑜𝑔𝜌 using standard classical
mathematical methods. Therefore, we calculate log 𝜌 by Mercator’s Algorithm 4 Feature Selection Using MI and PCC
Series. Suppose 𝐵 is a matrix and sufficiently close to the identity 1: Define: Dataset 𝐷 with features 𝐹 and target variable 𝑇 . Mutual infor-
matrix 𝐼, and ‖𝐵 − 𝐼‖ < 1, then a logarithm of 𝐵 can be computed mation 𝐼, Pearson correlation coefficient 𝜌, and selection threshold 𝜃 =
by means of the following k-power series [40]: 0.8.
Input: Dataset 𝐷, Number of features to select 𝑁;
∑
∞
(𝐵 − 𝐼)𝑘
log(𝐵) = (−1)𝑘+1 (10) Output: Set of selected features 𝐹𝑠 ;
𝑘 2: Calculate MI for each feature 𝑓(𝑖 in 𝐹 with the target 𝑇 using:
𝑘=1
∑ ∑ )
𝑝(𝑓𝑖 ,𝑡)
This means we can obtain log 𝜌 by: 3: 𝐼(𝑓𝑖 ; 𝑇 ) = 𝑡∈𝑇 𝑓𝑖 ∈𝐹 𝑝(𝑓𝑖 , 𝑡) log 𝑝(𝑓 )𝑝(𝑡)
𝑖
4: Rank features in 𝐹 based on calculated MI, resulting in ranked list 𝐹𝑟𝑎𝑛𝑘𝑒𝑑
∑
∞
(𝝆 − 𝐼)𝑘 5: 𝐹𝑠 ← ∅, 𝐹𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 ← first 𝑁 features from 𝐹𝑟𝑎𝑛𝑘𝑒𝑑
log(𝝆) = (−1)𝑘+1 (11)
𝑘 6: while the number of features in 𝐹𝑠 is less than 𝑁 do
𝑘=1
7: 𝐹𝑡𝑒𝑚𝑝 ← 𝐹𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑
Integrating Eq. (8), Eqs. (9) and (10), the VNE can be obtained. The 8: for each pair (𝑓𝑖 , 𝑓𝑗 ) in 𝐹𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 where 𝑖 ≠ 𝑗 do
complete procedure for calculating VNE is shown in the Algorithm 3. 9: Calculate Pearson correlation coefficient 𝜌 for 𝑓𝑖 and 𝑓𝑗 using:
∑
(𝑓𝑖𝑘 −𝑓̄𝑖 )(𝑓𝑗𝑘 −𝑓̄𝑗 )
10: 𝜌(𝑓𝑖 , 𝑓𝑗 ) = √∑ √∑
(𝑓𝑖𝑘 −𝑓̄𝑖 )2 (𝑓𝑗𝑘 −𝑓̄𝑗 )2
3.3. Feature selection
11: if 𝜌(𝑓𝑖 , 𝑓𝑗 ) > 𝜃 then
12: Remove the feature with lower MI from 𝐹𝑡𝑒𝑚𝑝
Feature selection has a crucial role in modeling by identifying 13: end if
relevant variables that contribute to the predictive power of the model 14: end for
while reducing dimensionality and computational cost. When the 15: 𝐹𝑠 ← 𝐹𝑡𝑒𝑚𝑝
dataset allows for state-transition modeling, we prioritize the entropy 16: 𝐹𝑠𝑒𝑙𝑒𝑐𝑡𝑒𝑑 ← next (𝑁 − number of features in 𝐹𝑠 ) features from 𝐹𝑟𝑎𝑛𝑘𝑒𝑑
features associated with Markov chains, including Shannon’s entropy, 17: end while
18: Return 𝐹𝑠 ;
entropy rate, entropy production, and von Neumann entropy of Markov
chains. Conversely, for non-state-transition modeling, we deploy a
feature selection methodology utilizing mutual information (MI) and
the Pearson correlation coefficient (PCC) to select appropriate entropy of variables, which is essential for unraveling complex and non-linear
features from several entropy methods. The detailed steps of the feature relationships. The MI for each feature is computed as:
selection process are shown in Algorithm 4.
( )
Initially, features are ranked based on the MI between each feature ∑ ∑ 𝑝(𝑓𝑖 , 𝑡)
𝐼(𝑓𝑖 ; 𝑇 ) = 𝑝(𝑓𝑖 , 𝑡) log , (12)
𝑓𝑖 and the target label 𝑇 . This ranking reflects the mutual dependency 𝑡∈𝑇 𝑓 ∈𝐹
𝑝(𝑓𝑖 )𝑝(𝑡)
𝑖

7
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 7. An example of a PLWD with clear routine activities (a) and another PLWD with fewer routine activities (b). The participant with more routine activities tends to have a
more consistent daily activity pattern at the same time each day, as shown in the red boxes. The 𝑥-axis shows the time of the day, the 𝑦-axis shows different days, and the different
colors represent different locations in the house. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Where 𝑓𝑖 represents the 𝑖th feature and 𝑇 is the target label. This 4.1. Datasets
ranking process is fundamental as it guides the subsequent selection of
features by highlighting their respective importance. The datasets we used in our experiment are: the Minder dataset,
Upon ranking the features, we select the top 𝑁 features for further collecting activity data of the people living with dementia (PLWD) [33];
analysis. To address potential multicollinearity, we calculate the pair- the epileptic seizure recognition dataset (ESRD), collecting EEG data
wise PCC among these features. If the correlation coefficient between [17]; and the PTB diagnostic ECG database (PTBDB), collecting ECG
any two features 𝑓𝑖 and 𝑓𝑗 exceeds the empirically determined thresh- data [18,19].
old, we retain only the feature with the higher MI score and select the
next feature in the rank until the top 𝑁 features are distinct and less 4.1.1. Minder dataset
correlated. The PCC between two features 𝑓𝑖 and 𝑓𝑗 can be calculated We use data from an in-home monitoring study (illustrated in
as: Fig. 1), called Minder, to support PLWD [33]. The Minder platform
∑
(𝑓𝑖 − 𝑓̄𝑖 )(𝑓𝑗 − 𝑓̄𝑗 ) collects various digital markers, including activity data, home device
𝜌(𝑓𝑖 , 𝑓𝑗 ) = √ √∑ ,𝑖 ≠ 𝑗 (13) usage, and clinical information. The Minder study protocol received
∑
(𝑓𝑖 − 𝑓̄𝑖 )2 (𝑓𝑗 − 𝑓̄𝑗 )2 ethical approval from the London-Surrey Borders Research Ethics Com-
Where 𝜌 denotes the PCC between features 𝑓𝑖 and 𝑓𝑗 . mittee and South West London Ethics Committee (linkhere) and we
In this paper, the default threshold is 0.8, following the convention obtained informed written consent from all study participants.
established in prior research where it is often considered a cutoff point The dataset is labeled by our monitoring team in response to alerts
for high correlation [41,42]. This value is widely recognized as a generated by the Minder platform, which operates 24/7. These alerts
practical balance between including informative features and excluding are confirmed with either the person living with dementia (PLWD)
redundant ones. Our approach ensures that the retained features are not or their caregivers, ensuring the accuracy of information on potential
only relevant to the target but also provide unique information, thereby health-related events such as falls, abnormal motor behavior, hos-
mitigating the risk of multicollinearity. pital admissions, urinary tract infections, and symptoms of anxiety,
depression, agitation, confusion, and disturbed sleep patterns. Data
4. Implementation & evaluation setup corresponding to participants who have experienced such events are
specifically labeled to indicate these adverse health occurrences.
We have implemented our entropy-based analysis framework on In our research, we concentrate on analyzing Minder’s activity data,
three distinct datasets: a dementia care dataset from the ongoing UK which falls under the category of state-transition data. This dataset
Dementia Research Institute’s Minder study, and two publicly available encompasses 3,762 person-weeks of information, gathered from Decem-
datasets on epileptic seizure (ESRD) and heart disease (PTBDB) focusing ber 2020 to March 2022. The average age of the study participants is
on EEG and ECG signals, respectively. In terms of model selection, 79 years. It is important to note that all data used in this study has been
our analysis encompasses Logistic Regression (LR), Support Vector anonymized to protect participant privacy.
Machines (SVM), Multi-Layer Perceptron (MLP), Convolutional Neural The activity data in the Minder platform is collected using PIR
Network (CNN), Long Short-Term Memory (LSTM), and a combined sensors installed in various locations, including the kitchen, bathroom,
CNN-LSTM architecture (further details are provided in Section 4.2). bedroom, lounge, and hallway, as shown in Fig. 2.A. The PIR sensor
The datasets are partitioned into training, validation, and test sets, logs an event with seconds precision when a person passes by. The
following a 70%, 15%, 15% split ratio, respectively. To ensure the recorded data shows the household’s life patterns over time; an example
reliability of our findings, all experiments are conducted 30 times, with of raw data is shown in Fig. 7, which compares the routine activities
both the mean and standard deviation of the results being computed. of two PLWDs.

8
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 8. Visualization of ESRD. The 𝑥-axis represents sampling points, and the 𝑦-axis represents EEG signals (𝜇𝑉 ).

Fig. 9. The histogram color maps for PTBDB are marked as abnormal (a) and normal (b). The 𝑥-axis represents sampling points, and the 𝑦-axis represents the normalized value
of the heartbeat.

4.1.2. Epileptic seizure recognition dataset we specifically extract heartbeat signals from ECG lead 2 [43], concen-
The Epileptic Seizure Recognition Dataset (ESRD) comprises 11,500 trating on the myocardial infarction and healthy control groups. The
time-series EEG signal samples from 500 subjects, designed for ana- dataset comprises a total of 14,552 samples. Color maps in histograms
lyzing EEG signal variations during seizures, categorizing it as non- depicting the PTB data as abnormal and normal are illustrated in Fig. 9.
state-transition data [17]. Each sample is divided into 23 segments,
with each segment containing 178 data points spanning one second. 4.2. Models and performance metrics
The UCI has processed the original dataset, shuffling the segments
to create the 11,500 time-series EEG signal samples. An illustrative In this section, we introduce models of the state-transition data,
visualization of ESRD is presented in Fig. 8. This dataset encapsu- models of the non-state-transition data, and the performance metrics.
lates five distinct health states, one of which is specific to epileptic
seizures, while the remaining four represent normal conditions with- 4.2.1. Models of the state-transition data
out epilepsy symptoms. Despite the dataset’s potential for multi-class For the model of the state-transition data, after feature extrac-
tion, we utilize LR, SVM, MLP, and LSTM as classifiers, following the
analysis, many researchers opt for binary classification, focusing on dif-
standard procedure in the domain of medical feature extraction [44–
ferentiating class 1 (epileptic seizures) from the other classes. Similarly,
46].
our research aims to distinguish between individuals with and without
To evaluate the performance of our method on both linear and
epileptic seizures.
non-linear classifiers, we utilize LR and SVM (with a linear kernel)
as linear classifiers and MLP and LSTM as non-linear classifiers. To
4.1.3. PTB diagnostic ECG database optimize the parameter configurations for the LR and SVM models, we
To further assess the applicability and scalability of our analysis employ a grid search strategy. This strategy is implemented using the
pipeline, we evaluated it using the PTB Diagnostic ECG Database 𝐺𝑟𝑖𝑑𝑆𝑒𝑎𝑟𝑐ℎ𝐶𝑉 tool from the 𝑠𝑐𝑖𝑘𝑖𝑡 − 𝑙𝑒𝑎𝑟𝑛 library in 𝑃 𝑦𝑡ℎ𝑜𝑛. For LR,
(PTBDB), which consists of 549 records from 290 subjects (n = 209 the grid search covers multiple candidate values for the regularization
male, and 81 female), classified as non-state-transition data [18,19]. strength parameter C (0.001, 0.01, 0.1, 1, 10, 100), regularization
The participants’ ages range from 17 to 87 years, with an average age methods (L1 and L2), and solver (liblinear, newton-cg, lbfgs, sag, saga).
of 57.2 years. The data is recorded at a sampling frequency of 125 Hz. For SVM, we consider different C values (0.001, 0.01, 0.1, 1, 10, 100).
The Diagnostic class encompasses a variety of conditions including To optimize MLP and LSTM, we used a strategy that combined
myocardial infarction, cardiomyopathy/heart failure, bundle branch temperature warm-up and cosine annealing to optimize the learning
block, dysrhythmia, myocardial hypertrophy, valvular heart disease, rate. For the network structures of MLP and LSTM, to ensure a fair
myocarditis, among others, as well as healthy controls. For this study, comparison, we set the same number of hidden layers (two hidden

9
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

layers). For the number of nodes, we scale them proportionally to the 4.2.3. Performance metrics
input size to ensure a fair comparison. For example, if the input size To evaluate the performance of these models comprehensively, we
of the baseline is 𝑎, the hidden size of the baseline is 𝑏, the input size utilize accuracy, F1 score, recall, and Area Under the Curve (AUC)
of the entropy-based method is 𝑐, and the hidden size of the entropy- as the evaluation metrics. Accuracy intuitively reflects the overall
based method is 𝑑, we ensure that the ratio of 𝑎∕𝑐 is consistent with classification capability of the models, while F1 score and recall provide
𝑏∕𝑑. We utilize Binary Cross-Entropy loss function, and SGD optimizer. a deeper understanding of the model performance in binary classifi-
Considering the effect of the month, we set the time step of LSTM to 4 cation, especially important in medical scenarios. Additionally, AUC,
weeks of data. as an important metric, can reflect the model’s ability to differenti-
ate between positive and negative classes at different thresholds. It
We divide the entire dataset into three parts: a training set, a
takes into account the model’s sensitivity and specificity and serves
validation set, and a test set. The specific split ratio is 70% for the
as an effective tool for assessing the accuracy of model predictions in
training set, 15% for the validation set, and 15% for the test set. The
terms of probabilities. The combined use of these metrics allows us to
training set is used for model training, the validation set is used for
comprehensively assess the performance of the models from multiple
tuning model parameters, and the test set is used to evaluate the final dimensions, ensuring a comprehensive understanding of model effec-
performance of the models. tiveness. In addition, for the evaluation of the non-state-transition data,
∙ Baseline. For the state-transition data (Minder dataset), there since we compare the entropy-based method to the deep models, we
is limited research on feature extraction for PIR sensors. We follow also calculate the model parameter count by the 𝑠𝑢𝑚𝑚𝑎𝑟𝑦() function
the research of Chimamiwa et al. [47], utilize the average frequency from 𝑇 𝑒𝑛𝑠𝑜𝑟𝐹 𝑙𝑜𝑤 or by iterating through the model in 𝑃 𝑦𝑇 𝑜𝑟𝑐ℎ.
and average duration of different activities as the baseline, and also
split one day into daytime (06:00–18:00) and nighttime (18:00–24:00 5. Evaluation results
and 00:00–06:00). The extracted feature is then fed into the classifier
including LR, SVM, MLP, and LSTM. We evaluate our method on three datasets: Minder, ESRD, and
∙ Entropy-Based. For the state-transition data, we prioritize the PTBDB. Minder is a state-transition dataset, ESRD and PTBDB are
entropy features associated with Markov chains, including Shannon’s non-state-transition datasets. We utilize accuracy, F1 score, recall, and
ROC-AUC as the evaluation metrics.
entropy of Markov chains, entropy rate of Markov chains, EP of Markov
chains, VNE of Markov chains (activity frequency), VNE of Markov
5.1. Minder database
chains (activity duration), and activity duration difference of Markov
chains in each week (daytime and nighttime).
Minder data represents a state-transition model in which each lo-
cation is a node/state, and movements between locations are repre-
4.2.2. Models of the non-state-transition data sented as transitions. We prioritize the entropy features associated with
For non-state transition data (ESRD and PTBDB dataset), since Markov chains. Additionally, we consider the effect of sundowning
deep model-based feature extraction methods are popular recently, and circadian rhythms in people living with dementia (PLWD) [50]
we utilize CNN-based feature, LSTM-based feature, and CNN-LSTM- by dividing one day into two time periods: daytime (06:00–18:00) and
based feature as the baseline. Considering that these three approaches night (18:00–24:00 and 00:00–6:00). Following these configurations,
the baseline features are the average frequency and average duration
utilize deep learning models and to ensure experimental consistency,
of different activities in each week (daytime and nighttime) [47]. The
fairness, and effective comparisons, our entropy-based method employs
output of the models is healthcare-related events (True or False).
MLP as the classifier. Additionally, MLP offers advantages in terms
The entropy features are Shannon’s entropy of Markov chains,
of computational efficiency and model simplification. All these deep
Entropy rate of Markov chains, EP of Markov chains, VNE of Markov
learning models utilize the strategy of combining temperature warm-up chains (activity frequency), VNE of Markov chains (activity duration),
and cosine annealing to optimize the learning rate. and activity duration difference of Markov chains in each week (day-
∙ Baseline. For the CNN-based features, when dealing with the ESRD time and nighttime). Since the frequency of our dataset labels is one
and PTBDB datasets, we employ a 1D-CNN model. This model design week (indicating whether there are any anomalies within the week),
is followed by Khalilpour et al.’s research [48], and we optimize its we calculate a series of entropy values by the week. For Shannon’s
structure to suit our data better. The raw data is used as input to entropy, we first calculate the entropy of each day (distinguishing
this model. Similarly, for LSTM-based features, we reference the LSTM between daytime and nighttime) and then compute the weekly average
model proposed by Farisi et al. [49]. This model also takes the raw value. For EP, as it is calculated by machine learning models, for each
data as input, and we make adjustments to it to better align with participant, we first train the model using the training data and then
our data. For the CNN-LSTM-based feature analysis, we follow the calculate the EP for each week. For the entropy rate, to represent the
model proposed by Hussain et al. [22]. Specifically, they propose a stationary pattern and avoid the impact of seasonal changes, we set the
CNN-LSTM model, in which the CNN is used for front-end feature time window to four months to calculate the stationary distribution
extraction, while the LSTM is employed for learning temporal patterns 𝜋, with the target time window being one day, and then calculate
in the back end. To better extract features and capture non-stationary the weekly average value. For VNE, we calculate the frequency (or
duration) of various activities within a day and then compute the VNE
and time-varying information, they transform the original signal into
for each week. The output of the models is healthcare-related events
time–frequency domain signals through a five-level decomposition and
(True or False).
a one-level approximate discrete wavelet transform (DWT). The loss
Fig. 10, Fig. 11, and Table 1 show the evaluation results of Minder.
function and optimizer of the above models are kept consistent with
The entropy-based method exhibits effectiveness across different mod-
the state-transition model.
els. Through comprehensive analysis, we find that the entropy-based
∙ Entropy-Based (Ours). For non-state-transition data, we first calcu- method can significantly improve performance for both linear models,
late several entropy features such as increment entropy, approximate such as LR and SVM (linear kernel), as well as non-linear models like
entropy, and slope entropy; we then utilize MI and PCC for feature MLP and LSTM. Specifically, we observe an average increase of 13.08pp
selection (as illustrated in Section. 3.3). After feature extraction, we in recall rate, 10.80pp in F1 score, and 7.88pp in accuracy.
add an MLP classifier. The model structure of the MLP is designed to This performance improvement can be attributed to the enhanced
be as simple as possible while maintaining model performance. The loss ability of the entropy-based method to capture the complexity of pat-
function and optimizer are kept consistent with the baseline model. terns and temporal dependencies. Traditional baseline features, such

10
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 10. The evaluation results of the Minder database. The 𝑥-axis represents different evaluation metrics. We can find that, for four different models, compared with the baseline
features, modeling by the entropy features can improve the recall rate, F1 score, and accuracy.

Fig. 11. The ROC-AUC curves of the Minder database.

as average activity frequency and duration, may have limitations in Table 1

The average performance of the models for the Minder.
revealing complex data structures with temporal dependencies. In con-
trast, the entropy-based method provides a powerful way to deepen the Evaluation Baseline Entropy Improvement

understanding of intrinsic dynamic variations in data by quantifying LR Recall rate 50.13 ± 4.68% 58.97 ± 4.61% 8.84pp
F1 score 52.63 ± 4.11% 58.46 ± 4.42% 5.83pp
the uncertainty and randomness of state transitions in Markov chains.
Accuracy 55.52 ± 4.13% 58.03 ± 4.71% 2.51pp
Measures of entropy such as Shannon’s entropy, entropy rate, EP, and SVM Recall rate 56.05 ± 4.60% 62.72 ± 6.56% 6.67pp
VNE not only capture the probability distribution of activities but also F1 score 52.82 ± 4.75% 61.63 ± 4.30% 8.81pp
reflect the diversity and uncertainty of system states, helping the model Accuracy 50.12 ± 5.75% 60.94 ± 4.03% 10.82pp
to uncover nonlinear relationships and complex interactions in the data MLP Recall rate 66.14 ± 5.70% 84.16 ± 5.21% 18.02pp
F1 score 71.55 ± 5.61% 84.97 ± 4.97% 13.42pp
more effectively. Accuracy 78.37 ± 8.11% 85.88 ± 5.51% 7.51pp
The entropy-based method performed better on non-linear models LSTM Recall rate 71.51 ± 4.04% 90.29 ± 4.41% 18.78pp
such as LSTM and MLP. For example, after adopting the entropy-based F1 score 76.17 ± 4.27% 91.29 ± 3.72% 15.12pp
method, the LSTM model shows an 18.78pp increase in recall rate, Accuracy 81.72 ± 6.41% 92.41 ± 4.18% 10.69pp
15.12pp in F1 score, and 10.69pp in accuracy. Similarly, the MLP Average Recall rate – – 13.08pp
model also exhibits an 18.02pp increase in recall rate, 13.42pp in F1 score – – 10.80pp
Accuracy – – 7.88pp
F1 score, and 7.51pp in accuracy. This is because these non-linear
models have multi-layer processing and memory capabilities, signif-
icantly improving the performance of complex pattern recognition
and the utilization of high-dimensional features. This enhancement is models as baselines. All of these baseline models are derived from state-
particularly important when dealing with multi-variate, multi-source, of-the-art (SOTA) research [22,48,49] and are adjusted to better fit the
state-varying, and noisy time-series data, highlighting the potential of characteristics of our dataset. We perform Z-score normalization and
the entropy-based method in time-series analysis. alignment on the raw data. For CNN-based and LSTM-based models,
Furthermore, we evaluate the model performance using ROC-AUC the input to the models is the pre-processed data. For the CNN-LSTM-
curves. As shown in Fig. 11, all models exhibit an improvement in based model, in addition to data pre-processing, we also conduct DWT
AUC values after utilizing entropy-based features. In the LR model, transformations and utilize the transformed signals as model inputs.
the introduction of entropy-based features increases the AUC value The output of the models is participants with epileptic seizures (True
from 0.56 to 0.63, while the SVM model’s AUC value improves from or False).
0.59 to 0.70. In the MLP model, the entropy method improves the For the entropy-based method, we first calculate a series of entropy
AUC value from 0.78 to 0.93, indicating its sufficient performance in features, then employ feature selection by MI and PCC. The feature
distinguishing between positive and negative classes. The LSTM model selection result is shown in Fig. 12. Based on the MI, we first select the
also achieves an AUC value of 0.98 after adopting entropy features. top four features, including increment entropy, approximate entropy,
These improvements underscore the effectiveness of the entropy-based slope entropy, and sample entropy. Subsequently, to avoid feature
method in processing complex time-series data. Consistent with our collinearity and redundancy, we conduct a secondary selection of the
previous analysis, the entropy-based method significantly enhances the selected features based on the PCC. We observe that the sample entropy
model’s ability to analyze time-series data. and approximate entropy exhibit a relatively high PCC (higher than
0.8), with approximate entropy having a higher MI score. Therefore, we
5.2. Epileptic seizure recognition dataset (ESRD) retain approximate entropy and, in descending order according to MI
ranking, select phase entropy as the next feature to keep. Afterwards,
ESRD is a non-state-transition dataset; we aim to differentiate be- by checking the PCC again, we find that all the selected features have
tween the normal participants and those with epileptic seizures. Due no significant collinearity or redundancy and have sufficient MI scores.
to the popularity of deep model-based feature extraction methods in Therefore, the selected entropy-based features for ESRD are increment
the analysis of EEG and ECG data, we use CNN, LSTM, and CNN-LSTM entropy, approximate entropy, slope entropy, and phase entropy. Since

11
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 12. The feature selection result of ESRD, including Pearson relationship matrices (the left) and mutual information (the right). In the right figure, the selected features are
in red while the unselected features are in blue. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Fig. 13. The results of ESRD. (a) is the evaluation results, including recall rate, F1 score, accuracy, and ROC-AUC. (b) is the comparison of the model structure between the
baseline and entropy models. As we utilize pre-processed data with less data noise, the AUC-ROC performances of all the models are ideal.

Table 2 average improvement of 3.12pp in recall, 3.25pp in F1 score, and

Comparison of ESRD classification results.
3.35pp in accuracy. The ROC-AUC curve also shows a sufficient model
Recall rate F1 score Accuracy
performance of the entropy-based model. Furthermore, the model’s
Baseline-CNN 91.18 ± 1.04% 92.51 ± 0.98% 93.90 ± 1.07%
parameter count is significantly reduced, with an average reduction of
Baseline-LSTM 95.74 ± 1.09% 95.12 ± 0.54% 94.52 ± 0.93%
Baseline-CNN-LSTM 96.30 ± 1.35% 96.08 ± 0.77% 95.88 ± 0.53% 3.05 times. This improvement shows the effectiveness and efficiency of
Entropy-MLP 97.53 ± 0.74% 97.82 ± 0.43% 98.12 ± 0.67% the entropy-based method.
Avg Improvement 3.12pp 3.25pp 3.35pp The effectiveness of the entropy-based method can be explained
from several perspectives. Firstly, entropy can measure signal uncer-
tainty and complexity, which is particularly crucial in the analysis of
the baseline utilizes deep learning models, we employ MLP as the clas- EEG and ECG signals. By combining and integrating series entropy
sifier for the entropy-based method, which can also offer advantages in features, we can capture the nonlinear characteristics and underlying
terms of computational efficiency and model simplification. dynamic variations from different aspects, which are vital for recog-
The experimental results for ESRD are shown in Fig. 13 and Table 2, nizing atypical pattern changes such as epileptic seizures. Secondly,
revealing improvements in both model performance and efficiency with feature selection by MI and PCC can effectively remove redundant
the entropy-based method. In the task of epileptic seizure recognition, features while retaining the most informative ones. This reduces model
compared to the baselines, the entropy-based MLP method shows an complexity, prevents overfitting, and enhances model generalization.

12
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 14. The feature selection result of PTBDB, including Pearson relationship matrices (the left) and mutual information (the right). In the right figure, the selected features are
in red and the unselected features are in blue. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

The efficiency improvement of the entropy-based method stems Table 3

Comparison of PTBDB classification results.
from the fewer model parameters compared to the baselines, reducing
computational complexity. This not only lowers the risk of overfitting Recall rate F1 score Accuracy

but also makes the model more suitable for real-time analysis and large- Baseline-CNN 92.16 ± 1.83% 92.92 ± 1.22% 93.70 ± 0.88%
Baseline-LSTM 93.11 ± 2.02% 93.61 ± 1.27% 94.14 ± 0.60%
scale data processing. Additionally, parameter reduction implies that
Baseline-CNN-LSTM 94.07 ± 1.57% 94.32 ± 1.28% 94.59 ± 1.58%
the model is more efficient during both training and inference stages, Entropy-MLP 95.02 ± 0.76% 95.21 ± 0.64% 95.42 ± 0.99%
which is particularly important for medical applications requiring rapid
Avg Improvement 1.91pp 1.59pp 1.28pp
responses.

5.3. PTBDB
enhance the models’ efficiency, allowing for the achievement of desired
PTBDB is a non-state-transition dataset. We aim to distinguish the performance levels without resorting to overly complex models. This
ordinary participants and the participants with any heart disease. Fol- aspect is particularly beneficial in healthcare settings, where quick and
lowing the configurations of the experiments on the ESRD dataset, we accurate feedback is paramount.
also employ CNN-based, LSTM-based, and CNN-LSTM-based models as
baselines [22,48,49]. The preprocessing steps are the same as the ESRD 6. Discussion and future work
experiments. The output of the models is the participants with any heart
disease (True or False).
To evaluate our method, we conduct experiments on three different
The feature selection result is shown in Fig. 14. Based on the MI, we
datasets. The results show that, compared to the baseline, our method
first select the top four features, including phase entropy, dispersion
can improve performance on various evaluation metrics and different
entropy, approximate entropy, and slope entropy. Then we conduct
types of datasets. This can help to develop more robust decision-support
a secondary selection of the selected features based on the PCC, and
tools for applications that use neural time-series data or applications
all the selected features exhibit a relatively low Pearson correlation
in other fields that use similar data. The first dataset is the daily
coefficient (lower than 0.8). Therefore, we select the above entropy
activity data of PLWD, which is state-transition data collected by
features as the entropy-based features. Similar to the experiment con-
PIR sensors. Information theory-based methods can effectively identify
figuration with the ESRD dataset, we employ MLP as the classifier for
unusual patterns in this type of data. The second and third datasets are
the entropy-based method.
the seizure dataset and the electrocardiography dataset, which are non-
The experimental results for PTBDB are shown in Fig. 15 Table 3,
state-transition data. Information theory-based methods also accurately
showing improvements in both model performance and efficiency with
identify patients with epilepsy and heart disease, achieving an effective
the entropy-based method. The entropy-based method achieves similar
model performance to the SOTA, showing an average improvement diagnosis. We attribute the improvements mainly to:
of 1.91pp in recall rate, 1.59pp in F1 score, and 1.28pp in accuracy. ∙ Effective high-level entropy features. Entropy features consider criti-
The ROC-AUC curves also demonstrate that we achieve similar model cal statistical information, aiding in the extraction of advanced features
performance to the SOTA methods. For the model’s parameter count, from raw data that traditional neural networks struggle to automati-
compared to the baselines, the entropy-based model can achieve an cally extract. We design, apply, and select different entropy features
average reduction of 3.14 times. for various time-series data. For state-transition data, besides using
The entropy-based approach has shown consistent effectiveness Shannon entropy, we also preprocess the original data into a first-order
across both state-transition and non-state-transition datasets, marked Markov chain and apply the entropy rate, entropy production, and von
by both robust model performance and operational efficiency. This Neumann entropy. For non-state-transition data, we employ different
success is largely due to the utilization of various entropy features variants of entropy, such as increment entropy and slope entropy.
for feature extraction and the implementation of strategic feature ∙ Multi-dimensional Entropy Feature Analysis. Our methodology in-
selection processes. Incorporating multiple entropy features facilitates tegrates a variety of entropy features, each designed to uncover the
a comprehensive analysis of time-series data by harnessing a broader inherent complexity of the data from distinct viewpoints [51,52]. By
spectrum of information, which traditional methods might overlook. adopting this multi-dimensional strategy, we achieve a deeper and
Simultaneously, the process of feature selection meticulously eliminates more nuanced comprehension of the dynamics and structures present
features that exhibit high redundancy or collinearity, thereby stream- in neural time-series data. Whereas conventional analysis techniques
lining the dataset and improving the efficiency and accuracy of the might concentrate on a singular data characteristic or statistic, thus
model. The strategic choice and use of entropy features significantly risking the omission of other crucial data attributes, our approach to

13
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. 15. The results of PTBDB database. (a) is the evaluation results. (b) is the comparison of the model structure between the baseline and entropy models. As we utilize
pre-processed data with less data noise, the AUC-ROC performances of all the models are ideal.

multi-dimensional entropy feature analysis ensures a broad capture of our method’s effectiveness across diverse medical data contexts. This
the data’s facets, providing a more detailed and precise representation enhancement will significantly extend the applicability and robustness
∙ Efficient feature selection. To mitigate the potential decline in model of our approach in various medical scenarios.
performance caused by redundant and highly collinear features, we ∙ Model Interpretation. While incorporating diverse entropy features
implement a feature selection process grounded in information theory. has improved both the performance and stability of our models, it has
This process employs both mutual information and the Pearson corre- also somewhat complicated the interpretation of these models. In our
lation coefficient matrix to discern and select the optimal 𝑘 entropy forthcoming research endeavors, we intend to investigate and devise
features for our modeling efforts. more streamlined approaches and tools specifically aimed at facilitating
∙ Model simplicity enhanced by efficient entropy features. Traditional the interpretation and comprehension of complex entropy features, as
deep learning models often necessitate extensive computational re- well as elucidating their impact on model performance.
sources and time to analyze complex time-series data effectively. Yet, ∙Integration of Multimodal Datasets. We also recognize the complexity
by leveraging entropy features, our approach enables the attainment and diversity of neural time series data and plan to explore how
of favorable outcomes using simpler model architectures, like LR and to effectively incorporate data from different sources and modalities
MLP. This advancement not only boosts the efficiency of the model but into our analysis framework. This will involve developing data fusion
also aids in minimizing model complexity and enhancing interpretabil- techniques and algorithms to enhance the model’s generalizability and
ity.
accuracy.
However, there are also some limitations in the proposed solutions.
∙ Expansion of Application Scope. The proposed method has the
We plan to investigate and address the limitations of the solution in
potential to be extended to other types of time series data. In future
future work. The key limitations are listed below.
work, we plan to explore how this methodology can be applied to
∙ Dynamic Feature Selection. We introduce a feature selection
fields such as finance, meteorology, and energy, aiming to unlock new
methodology centered around Mutual Information (MI) and Pearson
insights and enhance analytical precision in these areas.
Correlation Coefficient (PCC), designed for static and comprehensive
∙ Personalization and Customization. We will continue to explore how
medical time-series data. This selection approach is initially formulated
to customize and optimize our method to meet the needs of specific
in the Cloud and subsequently implemented locally. Nonetheless, the
applications and groups. This may include developing personalized
dynamic nature and frequent incompleteness of medical data pose
significant challenges in real-world settings [53–55]. Liu et al. [56] de- models that take into account individual differences and specifications.
veloped a technique for dynamic mutual information feature selection
that caters to evolving datasets over time, proving adept at real-time 7. Conclusions
adjustments and responsiveness to data fluctuations. Similarly, Luo
et al. [57] have concentrated on feature selection for datasets with in- We propose a highly general pipeline that uses information theory
complete information, addressing the critical issue of handling missing and entropy to extract high-level features and analyze pattern recog-
or partial data. In our subsequent efforts, we aim to incorporate these nition from inherent low-level neurological time-series data, which
dynamic and incomplete data-focused techniques to broaden and refine can reduce privacy risks, improve model performance and enhance

14
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

efficiency. This pipeline’s effectiveness and scalability have been val- Appendix
idated through pattern analysis in datasets concerning dementia care, A.1. Entropy and entropy variants
epilepsy, and myocardial infarction. For instances involving stochastic
state transitions, our approach utilizes features derived from Shannon’s A.1.1. Approximate entropy
entropy, entropy rates, entropy production rate, and von Neumann For non-Markovian chain systems, Approximate Entropy (ApEn) can
entropy of Markov chains. In situations where state transition modeling be used to quantify the complexity of the system. Given a time series
does not apply, we employ approximate entropy, increment entropy, dataset {𝑢(𝑖) ∶ 1 ≤ 𝑖 ≤ 𝑁} with 𝑁 samples, form the sequence in order
dispersion entropy, phase entropy, and slope entropy. Furthermore, to generate an m-dimension vector:
we introduce an entropy feature selection method based on mutual 𝑢′ (𝑖) = [𝑢(𝑖), 𝑢(𝑖 + 1), … , 𝑢(𝑖 + 𝑚 − 1)] 𝑖 = 1, 𝑁 − 𝑚 + 1 (A.1)
information and the Pearson correlation matrix. The results show that,
compared with the baseline, the entropy-based method improves the Define the distance between the vectors and to be the𝑢′ (𝑖) 𝑢′ (𝑗)
recall rate, F1 score, and accuracy on average by up to 13.08pp. We maximum of the differences between the corresponding elements of the
two vectors:
also compared the pipeline with deep learning models on ESRD and
PTBDB. The results show that the pipeline can also enhance efficiency 𝑑[𝑢′ (𝑖), 𝑢′ (𝑗)] = max [|𝑢′ (𝑖 + 𝑘) − 𝑢′ (𝑗 + 𝑘)|] (A.2)
𝑘=0,𝑚−1
with an average reduction of 3.10 times in the number of model
parameters. The proposed pipeline offers a versatile, high-precision, Given a threshold 𝑝, count the number of 𝑑[𝑢′ (𝑖), 𝑢′ (𝑗)] <= 𝑝, denoted
privacy-preserved, and efficient solution for analyzing time-series data as 𝐴𝑚
𝑁
(𝑝), and calculate the ratio of 𝐴𝑚
𝑁
(𝑝) to 𝑁 −𝑚+1, denoted as 𝐵𝑁 𝑚 (𝑝):

for pattern recognition in neurological conditions, which can be applied

to various scenarios in healthcare. 𝐴𝑚
𝑙
(𝑝)
𝐵𝑙𝑚 (𝑝) = (A.3)
𝑁 −𝑚+1
𝑚 (𝑝):
Calculate the average value of 𝐵𝑁
CRediT authorship contribution statement
1 ∑
𝑁−𝑚+1
𝐵 𝑚 (𝑝) = 𝐵 𝑚 (𝑝) (A.4)
𝑁 − 𝑚 + 1 𝑁=1 𝑁
Yushan Huang: Writing – review & editing, Writing – original
draft, Visualization, Validation, Software, Resources, Project adminis- Increase the dimension from 𝑚 to 𝑚 + 1, and repeat the above steps.
tration, Methodology, Investigation, Funding acquisition, Formal anal- For sequences of finite length, an estimate of the sample entropy can
ysis, Data curation, Conceptualization. Yuchen Zhao: Writing – review be obtained as [58]:
& editing, Supervision, Methodology, Conceptualization. Alexander
ApEn(𝑚, 𝑟, 𝑁) = 𝐵 𝑚 (𝑝) − 𝐵 𝑚+1 (𝑝) (A.5)
Capstick: Writing – review & editing, Validation, Methodology, Data
curation. Francesca Palermo: Writing – review & editing, Valida- A.1.2. Increment entropy
tion, Methodology, Data curation. Hamed Haddadi: Writing – review The Incremental Entropy (IncrEn) algorithm is a method for calcu-
& editing, Supervision, Project administration, Methodology, Fund- lating the entropy of a sequence of data points incrementally rather
ing acquisition, Conceptualization. Payam Barnaghi: Writing – re- than computing the entropy of the entire sequence all at once. Given
view & editing, Validation, Supervision, Resources, Project administra- a time series dataset {𝑢(𝑖) ∶ 1 ≤ 𝑖 ≤ 𝑁} with 𝑁 samples. Construct
tion, Methodology, Funding acquisition, Formal analysis, Data curation, an increment time series {𝑣(𝑖), 1 ≤ 𝑖 ≤ 𝑁 − 1] by 𝑣(𝑖) = 𝑥(𝑖 + 1) − 𝑥(𝑖)
Conceptualization. from 𝑢(𝑖). Hence, for a positive integer 𝑚, 𝑁 − 𝑚 vectors of dimension
𝑚 are derived from an incremental time series. These vectors, denoted
as 𝑉 (𝑘) = [𝑣(𝑘), 𝑣(𝑘 + 1), … , 𝑣(𝑘 + 𝑚 − 1)], 1 ≤ 𝑘 ≤ 𝑁 − 𝑚, represent
Declaration of competing interest contiguous segments of the time series. Each element in a vector 𝑉 (𝑘)
is mapped onto a word of two letters. The sign of each component is
represented by 𝑣′𝑘+𝑗 = sgn(𝑣(𝑘 + 𝑗)), 𝑗 = 1 ⋯ , 𝑚 − 1, and the magnitude of
The authors declare the following financial interests/personal rela-
each component in relation to the other components within the vector
tionships which may be considered as potential competing interests:
is represented by 𝑞𝑘+𝑗 , 𝑗 = 1, … , 𝑚 − 1 for a quantifying resolution 𝑟.
Payam Barnaghi reports financial support was provided by EPSRC
As a result, 𝑁 − 𝑚 words, 𝑤𝑘 , 1 ≤ 𝑘 ≤ 𝑁 − 𝑚, are generated. Each word,
PROTECT Project. Hamed Haddadi reports financial support was pro- consisting of 2 × 𝑚 letters, can have (2𝑟 + 1)𝑚 variations, depending on
vided by EPSRC OpenPlus Fellowship. Yushan Huang reports financial the values of 𝑚 and 𝑟. The frequency of occurrence of each unique word
support was provided by China Scholarship Council. Payam Barnaghi 𝑤𝑛 is defined as:
reports financial support was provided by MRC and Alzheimer’s Soci- ( )
( ) 𝑄 𝑤𝑛
ety. Payam Barnaghi reports was provided by the Great Ormond Street 𝑝 𝑤𝑛 = (A.6)
Hospital Children’s Charity Award. 𝑁 −𝑚
( )
where
{ } 𝑄 𝑤𝑛 signifies the count of the unique word 𝑤𝑛 within the
𝑤𝑘 . The Increment Entropy (IncrEn) of order m (where 𝑚 is equal
Acknowledgments to or greater than 2) and resolution 𝑅 is defined as:
𝑚
∑
(2𝑅+1)
( ) ( )
This project is supported by the EPSRC PROTECT Project, United 𝐼𝑛𝑐𝑟𝐸𝑛(𝑚) = − 𝑝 𝑤𝑛 log 𝑝 𝑤𝑛 (A.7)
𝑛=1
Kingdom (grant number: EP/W031892/1), EPSRC OpenPlus Fellowship
(EP/W005271/1), and the UK DRI Care Research and Technology A.1.3. Dispersion entropy
Centre funded by MRC and Alzheimer’s Society (grant number: UKDRI- Dispersion entropy (DE) can be used to describe the complexity of
7002). The raw data from the Minder dataset was accessed using the time series data. For time series with low regularity, DE can reflect
DCARTE library developed by Dr. Eyal Soreq at the UK Dementia the degree of disorder of the series [59]. Given a time series dataset
Research Institute’s Care Research and Technology Centre. Yushan {𝑢(𝑖) ∶ 1 ≤ 𝑖 ≤ 𝑁} with 𝑁 samples. Map 𝑢(𝑖) to 𝑦(𝑖) between 0 and 1 by
Huang is funded by the China Scholarship Council. Payam Barnaghi’s normal cumulative distribution function (NCDF):
research is also partially supported by an award from the Great Ormond 𝑢𝑗 ( )
1 (𝑡−𝜇)2 ∕2𝜎 2 𝑑𝑡
𝑦𝑗 = √ 𝑒− (A.8)
Street Hospital (GOSH), United Kingdom (grant number 21PP30). 2𝜋𝜎 ∫−∞

15
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Fig. A.16. The phase space representation of an HRV signal.

In which the parameter 𝜇 is the expectation of 𝑢(𝑖), while the A.1.5. Slope entropy
parameter 𝜎 is its standard deviation. Map 𝑦 to the range of [1, 2, … , 𝑐], Slope entropy (SlopEn) is a method for measuring the complex-
and obtain a new sequence 𝑧(𝑐)𝑗 : ity of time series data, which is primarily based on transferring the
( ) original time series data to a series of single-threshold and symbolic
𝑧(𝑐)
𝑗 = int 𝑐𝑦𝑗 + 0.5 (A.9)
patterns [62,63]. SlopEn is initially applied in the fields of medicine
In which, 𝑐 is the number of categories, and 𝑖𝑛𝑡 is the rounding and biological signal processing, particularly in the analysis of electro-
function. Then construct the embedding vector 𝑧(𝑚,𝑐)
𝑖 by: cardiograms (ECG) and electroencephalograms (EEG). Its core concept
( ) is based on estimating the uncertainty or complexity of data by exam-
(𝑚,𝑐) (𝑐) (𝑐) (𝑐)
𝑧𝑖 = 𝑧𝑖 , 𝑧𝑖+𝑑 , … , 𝑧𝑖+(𝑚−1)𝑑 , ining changes in the slopes of time-series data. This means that slope
(A.10)
𝑖 = 1, 2, … , 𝑁 − (𝑚 − 1)𝑑 entropy not only considers the absolute values of data points but also
delves into the relative changes between data points.
In which, 𝑚 is the embedding dimension, 𝑐 is the number of classes,
Compared to other traditional linear metrics, SlopEn demonstrates
and 𝑑 is the time delay. Then each 𝑧(𝑚,𝑐)
𝑗 is mapped to dispersion
significant advantages in several aspects. First, it provides a tool to cap-
pattern 𝜋𝑣0 𝑣1 ⋯𝑣𝑚−1 (𝑣 = 1, 2, … , 𝑐), in which 𝑧(𝑐) (𝑐)
𝑖 = 𝑣0 , 𝑧𝑖+𝑑 = 𝑣1 , ..., and ture the inherent dynamics and nonlinear characteristics of time series,
𝑧(𝑐) = 𝑣 𝑚−1 . The number of possible dispersion of each 𝑧(𝑚,𝑐) is 𝑐 𝑚 .
𝑖+(𝑚−1)𝑑 𝑗 making it an effective measure of complexity. Second, it focuses on
Calculate the relative frequency for each potential dispersion pat-
changes in slopes, enabling an intuitive understanding of the dynamics
tern:
( ) of data through it, which is crucial for interpreting and comprehend-
( ) num 𝜋𝑣0 𝑣1 ,…,𝑣𝑚−1 ing patterns and trends in time series. Additionally, SlopeEn exhibits
𝑃 𝜋𝑣0 𝑣1 ,…,𝑣𝑚−1 = (A.11) good robustness, maintaining the stability of its measurement even
𝑁 − (𝑚 − 1)𝑑
in the presence of noise or interference from external factors. When
Finally, based on Shannon’s entropy, DE can be obtained by [60]:
considering time-series signals and their nonlinear and complex nature,
𝑐𝑚
∑ ( ) ( ( )) these signals often contain rich information and intrinsic dynamics that
𝐷𝐸(𝑢, 𝑚, 𝑐, 𝑑) = − 𝑝 𝜋𝑣0 ,…,𝑣𝑚−1 ln 𝑝 𝜋𝑣0 ,…,𝑣𝑚−1 (A.12) can be captured through changes in slopes. Slope entropy allows for a
𝜋=1
deeper understanding of this information, offering a comprehensive and
A.1.4. Phase entropy in-depth insight into the complexity and intrinsic dynamics of the data.
Phase entropy (PhEn) is developed to detect the complexity of Given a time series dataset {𝑢(𝑖) ∶ 1 ≤ 𝑖 ≤ 𝑁} with 𝑁 samples. De-
physiological signals. For example, given a time series dataset {𝑢(𝑖) ∶ compose 𝑢 into 𝑗 subsequences according to the embedded dimension
1 ≤ 𝑖 ≤ 𝑁} with 𝑁 samples, we can represent the data by the Lorenz 𝑚:
plot, as Fig. A.16(a) shows. In the Poincaré plot, if we replace the { }
sequence 𝑢𝑖 by 𝑢𝑖+1 − 𝑢𝑖 , then we can get SODP plot, as Fig. A.16(b) 𝑢𝑚𝑖 = 𝑢𝑖 , 𝑢𝑖+1 , … , 𝑢𝑖+𝑚−1 (A.17)
shows. Specifically, from a given time series 𝑢𝑖 , we can obtain 𝑌𝑖 and
In which, 𝑖 = {1, 2, … , 𝑗}, 𝑗 = 𝑁 − 𝑚 + 1. Define two soft threshold
𝑋𝑖 by [61]:
parameters 𝛿 and 𝛾 to calculate the symbolic patterns of 𝑢𝑚 𝑖 , where
𝑌𝑖 = 𝑢𝑖+2 − 𝑢𝑖+1 0 < 𝛿 < 𝛾.
(A.13)
𝑋𝑖 = 𝑢𝑖+1 − 𝑢𝑖 Define 𝑑 = 𝑢𝑖+1 − 𝑢𝑖 , and compare 𝑑 with the two soft threshold
Then compute the slope angle of each scatter point as shown in parameters 𝛿 and 𝛾, then five patterns can be obtained:
Fig. A.16(b). ⎧ pattern = 2, 𝛾 < 𝑑,
𝑌𝑖 ⎪
𝜃𝑖 = tan −1
(A.14) ⎪ pattern = 1, 𝛿 < 𝑑 ≤ 𝛾,
𝑋𝑖 ⎪
⎨ pattern = 0, |𝑑| ≤ 𝛿, (A.18)
Then the probability distribution 𝑝𝑖 can be calculated by: ⎪
⎪ pattern = −1, −𝛾 ≤ 𝑑 < −𝛿,
𝑆𝜃𝑖 ⎪ pattern = −2, 𝑑 < −𝛾.
𝑝𝑖 = ∑𝑘 (A.15) ⎩
𝑖=1 𝑆𝜃𝑖
Based on the five patterns, we can get 5𝑚−1 sequence combinations.
Finally, based on Shannon’s entropy, the PhEn can be calculated
The relative frequency 𝑝𝑛 of the combination can be calculated by the
as [61]:
number of occurrences 𝑓𝑛 of each combination:
−1 ∑
𝑘
PhEn = 𝑝(𝑖) log 𝑝(𝑖) (A.16) 𝑓𝑛
log 𝑁 𝑖=1 𝑝𝑛 = , 𝑛 = 1, 2, … , 5𝑚−1 (A.19)
𝑗

16
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

Finally, SlopEn can be calculated based on the Shannon’s entropy: [26] Shannon CE. A mathematical theory of communication. Bell Syst Tech J
1948;27(3):379–423.
𝑚−1
5∑
[27] Powell G, Percival I. A spectral entropy method for distinguishing regular and
SE(𝑚, 𝛾, 𝛿) = − 𝑝𝑛 ln 𝑝𝑛 (A.20) irregular motion of Hamiltonian systems. J Phys A 1979;12(11).
𝑛=1 [28] Richman JS, Lake DE, Moorman JR. Sample entropy. In: Methods in enzymology,
vol. 384, Elsevier; 2004, p. 172–84.
References [29] Nurwulan NR, Jiang BC. Multiscale entropy for physical activity recognition. In:
Proceedings of the 2020 2nd Asia Pacific information technology conference.
[1] Andreu-Perez J, Poon CC, Merrifield RD, Wong ST, Yang G-Z. Big data for health. 2020, p. 73–7.
IEEE J Biomed Health Inf 2015;19(4):1193–208. [30] Bao L, Intille SS. Activity recognition from user-annotated acceleration data. In:
[2] Motwani A, Shukla PK, Pawar M. Ubiquitous and smart healthcare monitoring International conference on pervasive comp.. Springer; 2004, p. 1–17.
frameworks based on machine learning: A comprehensive review. Artif Intell [31] Howedi A, Lotfi A, Pourabdollah A. Exploring entropy measurements to identify
Med 2022;102431. multi-occupancy in activities of daily living. Entropy 2019;21(4):416.
[3] Zhang C, Song D, Chen Y, Feng X, Lumezanu C, Cheng W, et al. A deep neural [32] Lampard D. A stochastic process whose successive intervals between events form
network for unsupervised anomaly detection and diagnosis in multivariate time a first order Markov chain—I. J Appl Probab 1968;5(3):648–68.
series data. In: Proceedings of the AAAI conference on artificial intelligence. [33] Enshaeifar S, Zoha A, Markides A, Skillman S, Acton ST, Elsaleh T, et al.
2019, p. 1409–16. Health management and pattern analysis of daily living activities of people with
[4] Piccialli F, Giampaolo F, Prezioso E, Camacho D, Acampora G. Artificial intel- dementia using in-home sensors and machine learning techniques. PLoS One
ligence and healthcare: Forecasting of medical bookings through multi-source 2018;13(5).
time-series fusion. Inf Fusion 2021;74:1–16. [34] Dewar R. Information theory explanation of the fluctuation theorem, maximum
[5] Rafiei MH, Gauthier LV, Adeli H, Takabi D. Self-supervised learning for entropy production and self-organized criticality in non-equilibrium stationary
electroencephalography. IEEE Trans Neural Netw Learn Syst 2022. states. J Phys A 2003;36(3):631.
[6] Shah HA, Saeed F, Yun S, Park J-H, Paul A, Kang J-M. A robust approach for [35] Butner J, Amazeen PG, Mulvey GM. Multilevel modeling of two cyclical pro-
brain tumor detection in magnetic resonance images using finetuned efficientnet. cesses: extending differential structural equation modeling to nonlinear coupled
IEEE Access 2022;10:65426–38. systems.. Psychol Methods 2005;10(2):159.
[7] Huang Y, Zhao Y, Haddadi H, Barnaghi P. Using entropy measures for monitoring [36] Vallacher RR, Nowak A. The emergence of dynamical social psychology. Psychol
the evolution of activity patterns. 2022, arXiv preprint arXiv:2210.01736. Inq 1997;8(2):73–99.
[8] Palermo F, Chen Y, Capstick A, Fletcher-Loyd N, Walsh C, Kouchaki S, et al. [37] Kim D-K, Bae Y, Lee S, Jeong H. Learning entropy production via neural
Tihm: An open dataset for remote healthcare monitoring in dementia. Sci Data networks. Phys Rev Lett 2020;125(14).
2023;10(1):606. [38] Bengtsson I, Życzkowski K. Geometry of quantum states: an introduction to
[9] Parkinson ME, Doherty R, Curtis F, Soreq E, Lai HH, Serban A-I, et al. Using quantum entanglement. Cambridge University Press; 2017.
home monitoring technology to study the effects of traumatic brain injury in [39] Felippe H, Viol A, de Araujo D, da Luz M, Palhano-Fontes F, Onias H, et al. The
older multimorbid adults. Ann Clinical Transl Neurol 2023;10(9):1688–94. von Neumann entropy for the pearson correlation matrix: A test of the entropic
[10] Murugappan M, Murugappan S. Human emotion recognition through short time brain hypothesis. 2021, arXiv preprint arXiv:2106.05379.
electroencephalogram (EEG) signals using fast Fourier transform (FFT). In: 2013
[40] MacDuffee CC. The theory of matrices, vol. 5, Springer Science & Business Media;
IEEE 9th international colloquium on signal processing and its applications. IEEE;
2012.
2013, p. 289–94.
[41] Mei K, Tan M, Yang Z, Shi S. Modeling of feature selection based on ran-
[11] Lawhern V, Hairston WD, McDowell K, Westerfield M, Robbins K. Detection and
dom forest algorithm and pearson correlation coefficient. J Phys Conf Series
classification of subject-generated artifacts in eeg signals using autoregressive
2022;2219(1):012046.
models. J Neurosci Methods 2012;208(2):181–9.
[42] Liu Y, Mu Y, Chen K, Li Y, Guo J. Daily activity feature selection in smart homes
[12] Xu S, Wang Z, Sun J, Zhang Z, Wu Z, Yang T, et al. Using a deep recurrent
based on pearson correlation coefficient. Neural Process Lett 2020;51:1771–87.
neural network with EEG signal to detect parkinson’s disease. Ann Transl Med
[43] Kachuee M, Fazeli S, Sarrafzadeh M. Ecg heartbeat classification: A deep transfer-
2020;8(14).
able representation. In: IEEE international conference on healthcare informatics.
[13] Michielli N, Acharya UR, Molinari F. Cascaded LSTM recurrent neural network
IEEE; 2018, p. 443–4.
for automated sleep stage classification using single-channel EEG signals. Comput
[44] Sun S, Chen H, Luo G, Yan C, Dong Q, Shao X, et al. Clustering-fusion feature
Biol Med 2019;106:71–81.
selection method in identifying major depressive disorder based on resting state
[14] Ghosh SM, Bandyopadhyay S, Mitra D. Nonlinear classification of emotion
EEG signals. IEEE J Biomed Health Inf 2023.
from EEG signal based on maximized mutual information. Expert Syst Appl
[45] Abdellatef E, Emara HM, Shoaib MR, Ibrahim FE, Elwekeil M, El-Shafai W,
2021;185:115605.
et al. Automated diagnosis of EEG abnormalities with different classification
[15] Krishnan PT, Raj ANJ, Balasubramanian P, Chen Y. Schizophrenia detection
techniques. Med Biol Eng Comput 2023;1–23.
using MultivariateEmpirical mode decomposition and entropy measures from
multichannel EEG signal. Biocybern Biomed Eng 2020;40(3):1124–39. [46] Chawla P, Rana SB, Kaur H, Singh K, Yuvaraj R, Murugappan M. A decision
[16] McMahan B, Moore E, Ramage D, Hampson S, y Arcas BA. Communication- support system for automated diagnosis of Parkinson’s disease from EEG using
efficient learning of deep networks from decentralized data. In: Artificial FAWT and entropy features. Biomed Signal Process Control 2023;79:104116.
intelligence and statistics. PMLR; 2017, p. 1273–82. [47] Chimamiwa G, Alirezaie M, Banaee H, Köckemann U, Loutfi A. Towards habit
[17] Andrzejak RG, Lehnertz K, Mormann F, Rieke C, David P, Elger CE. Indications of recognition in smart homes for people with dementia. In: Ambient intelligence:
nonlinear deterministic and finite-dimensional structures in time series of brain 15th European conference, amI 2019, rome, Italy, November 13–15, 2019,
electrical activity: Dependence on recording region and brain state. Phys Rev E proceedings 15. Springer; 2019, p. 363–9.
2001. [48] Khalilpour S, Ranjbar A, Menhaj MB, Sandooghdar A. Application of 1-D CNN to
[18] Bousseljot R-D. Nutzung der EKG-signaldatenbank CARDIODAT der PTB über das predict epileptic seizures using eeg records. In: 2020 6th international conference
internet. Biomed Tech/Biomed Eng (Biomed Eng) 1995;40:317–8. on web research. IEEE; 2020, p. 314–8.
[19] Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et [49] Farsi L, Siuly S, Kabir E, Wang H. Classification of alcoholic EEG signals using
al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research a deep learning method. IEEE Sens J 2020;21(3):3552–60.
resource for complex physiologic signals. Circulation 2000;101(23):e215–20. [50] Volicer L, Harper DG, Manning BC, Goldstein R, Satlin A. Sundowning and
[20] Huang Y. EntropyPipeline. 2023, [Link] circadian rhythms in alzheimer’s disease. Am J Psychiatry 2001;158(5):704–11.
EntropyPipeline. [51] Jui SJJ, Deo RC, Barua PD, Devi A, Soar J, Acharya UR. Application of entropy
[21] Niu M, Zhao Y, Haddadi H. Effective abnormal activity detection on multivariate for automated detection of neurological disorders with electroencephalogram
time series healthcare data. In: Proceedings of the 29th annual international signals: A review of the last decade (2012–2022). IEEE Access 2023.
conference on mobile computing and networking. 2023, p. 1–3. [52] Li W, Zhao Y, Wang Q, Zhou J. Twenty years of entropy research: A bibliometric
[22] Hussain W, Sadiq MT, Siuly S, Rehman AU. Epileptic seizure detection us- overview. Entropy 2019;21(7):694.
ing 1 D-convolutional long short-term memory neural networks. Appl Acoust [53] Ke C, Jin Y, Evans H, Lober B, Qian X, Liu J, et al. Prognostics of surgical site
2021;177:107941. infections using dynamic health data. J Biomed Informat 2017;65:22–33.
[23] Park J, Lee K, Park N, You SC, Ko J. Self-attention LSTM-FCN model [54] Tan Q, Ye M, Ma AJ, Yip TC-F, Wong GL-H, Yuen PC. Importance-aware
for arrhythmia classification and uncertainty assessment. Artif Intell Med personalized learning for early risk prediction using static and dynamic health
2023;102570. data. J Am Med Inf Assoc 2021;28(4):713–26.
[24] Shankar V, Yousefi E, Manashty A, Blair D, Teegapuram D. Clinical-gan: Trajec- [55] Ranjbari S, Arslanturk S. Integration of incomplete multi-omics data using knowl-
tory forecasting of clinical events using transformer and generative adversarial edge distillation and supervised variational autoencoders for disease progression
networks. Artif Intell Med 2023;138:102507. prediction. J Biomed Inform 2023;147:104512.
[25] Ding S, Zhu H, Jia W, Su C. A survey on feature extraction for pattern [56] Liu H, Sun J, Liu L, Zhang H. Feature selection with dynamic mutual information.
recognition. Artif Intell Rev 2012;37(3):169–80. Pattern Recognit 2009;42(7):1330–9.

17
Y. Huang et al. Artificial Intelligence In Medicine 150 (2024) 102821

[57] Luo C, Li T, Chen H, Lv J, Yi Z. Fusing entropy measures for dy- [61] Rohila A, Sharma A. Phase entropy: A new complexity measure for heart rate
namic feature selection in incomplete approximation spaces. Knowl-Based Syst variability. Physiol Measur 2019;40(10).
2022;252:109329. [62] Li Y, Mu L, Gao P. Particle swarm optimization fractional slope entropy: a new
[58] Chen X, Solomon IC, Chon KH. Comparison of the use of approximate entropy time series complexity indicator for bearing fault diagnosis. Fractal Fractional
and sample entropy: applications to neural respiratory signal. In: 2005 IEEE 2022;6(7).
engineering in medicine and biology 27th annual conference. IEEE; 2006, p. [63] Li Y, Tang B, Yi Y. A novel complexity-based mode feature representation for
4212–5. feature extraction of ship-radiated noise using VMD and slope entropy. Appl
[59] Rostaghi M, Azami H. Dispersion entropy: A measure for time-series analysis. Acoust 2022;196:108899.
IEEE Signal Process Lett 2016;23(5):610–4.
[60] Chakraborty M, Mitra D, et al. Automated detection of epileptic seizures using
multiscale and refined composite multiscale dispersion entropy. Chaos Solut
Fractals 2021;146.

18