1

Interactive Diabetes Risk Prediction Using Explainable

Machine Learning: A Dash-Based Approach with SHAP,
LIME, and Comorbidity Insights
Udaya Allani, Department of Computer Science, University of Arkansas at Little Rock, Arkansas,
USA
 However, the lack of model transparency and
Abstract— This paper presents an interactive machine interpretability has limited their adoption in real-world
learning-based system for diabetes risk prediction and healthcare settings. To address this, explainable
health assessment using the Behavioral Risk Factor
Surveillance System (BRFSS) dataset. The proposed artificial intelligence (XAI) approaches such as SHAP
framework integrates data preprocessing, feature
Udaya Allani is a Graduate Student in the Department of Computer and
engineering, class imbalance handling, and explainable
Information Sciences, University of Arkansas at Little Rock, Little Rock, AR
AI to generate both predictive and personalized health 72204 USA (e-mail: uallani@[Link]).
insights. Multiple classifiers, including Logistic
Regression, Random Forest, XGBoost, and LightGBM, (SHapley Additive exPlanations) and LIME (Local
were evaluated across three sampling strategies: original, Interpretable Model-Agnostic Explanations) have
SMOTE oversampling, and random undersampling. A been introduced to demystify model outputs and
LightGBM model trained on an undersampled dataset improve trust among stakeholders.
was selected based on recall-optimized cross-validation This paper presents an end-to-end pipeline for
performance. To enhance interpretability, SHAP and
LIME were applied to explain both global and individual diabetes risk prediction using the BRFSS dataset. The
predictions. The system also introduces composite proposed framework includes: (1) a structured
lifestyle and healthcare scores and derives comorbidity preprocessing pipeline with feature engineering and
risk insights using correlation analysis. The final model is class imbalance handling; (2) evaluation of multiple
deployed as a multi-step Dash web application machine learning models across SMOTE and
supporting interactive inputs, risk visualization, and
personalized recommendations. This work demonstrates undersampled datasets, with a recall-optimized
the practical potential of explainable AI to support LightGBM model trained on the undersampled data
public health screening and awareness, especially in selected for final deployment; (3) integration of SHAP
underserved populations. and LIME for dual-level interpretability; and (4) a
deployed interactive Dash application for real-time
Index Terms— Diabetes prediction, predictive
modeling, explainable artificial intelligence, SHAP,
risk analysis. The system also introduces composite
LIME, LightGBM, BRFSS, health risk assessment, lifestyle and healthcare scores to aid understanding
comorbidity analysis, Dash application, machine and identifies comorbidity risks through correlation-
learning, public health analytics. based insights.
In contrast to existing studies focused primarily on
accuracy metrics, this work emphasizes transparency,
I. INTRODUCTION iabetes interpretability, and actionable outputs through an

D
is a growing public health accessible web interface. It is intended to support both
concern, with significantimplications individuals and public health efforts in understanding
for long-term health outcomes and diabetes risk factors and promoting preventive care.
healthcare systems. Early identification of individuals
at risk can enable timely lifestyle II. RELATED WORK
interventions, reduce complications, and Numerous studies have leveraged the Behavioral
support population-level disease prevention. The Risk Factor Surveillance System (BRFSS) dataset to
Behavioral Risk Factor Surveillance develop machine learning (ML) models for diabetes
System (BRFSS) provides a large-scale, prediction. These efforts highlight the potential of
self-reported dataset capturing health- data-driven approaches in public health, but vary
related behaviors, chronic conditions, and access to significantly in methodology, feature selection, model
care making it a valuable resource for data-driven interpretability, and deployment.
health analytics. Chowdhury et al. [1] applied various sampling
Machine learning (ML) techniques have shown techniques such as SMOTE, SMOTE-Tomek, and
promise in analyzing such complex health datasets to SMOTE-EN alongside models like Logistic
predict chronic disease risks, including diabetes.
 2

Regression, AdaBoost, and Gradient Boost to address support diabetes risk prediction using machine
class imbalance in BRFSS 2021 data. While the study learning techniques.
emphasized recall improvement, it lacked B. Data Preprocessing
explainability components, making it difficult to The BRFSS dataset underwent a structured
understand model decisions. preprocessing pipeline to ensure data quality and
Liu et al. [2] conducted a comparative analysis of consistency prior to modeling. No missing values
Logistic Regression, Random Forest, and XGBoost were present in the dataset, and although some
using BRFSS 2015 data. They employed SMOTE for extreme BMI values were observed, they were
balancing and used SHAP to identify key predictors retained to reflect realistic variation in self-reported
like general health and blood pressure. However, their health metrics. Duplicate records were identified and
study did not explore LIME, nor did it integrate removed to reduce redundancy. Categorical variables
statistical evaluation across multiple sampling such as income were encoded using ordinal encoding,
methods. and all columns were converted to appropriate data
Ahmed et al. [3] proposed an explainable ML types to ensure compatibility with machine learning
framework using Logistic Regression and Random frameworks.
Forest, integrating SHAP and LIME to provide local The original target variable, Diabetes_012, included
and global model explanations. Although they three classes: non-diabetic (0), prediabetic (1), and
achieved 86% accuracy and presented compelling diabetic (2). To frame the task as binary classification,
visual interpretations, they did not address the class prediabetic entries were excluded. The remaining
imbalance problem in the dataset. Furthermore, their diabetic (2) entries were re-coded as 1, and non-
framework lacked real-time application, composite diabetic (0) entries remained unchanged. This yielded
scoring, and comorbidity risk analysis. a class distribution of approximately 84.41% non-
Nguyen and Zhang [4] utilized Decision Tree, K- diabetic and 15.59% diabetic. The resulting dataset
Nearest Neighbors, and Logistic Regression on was clean, structured, and suitable for training and
BRFSS data to build interpretable models. Their focus evaluation of machine learning models.
remained on basic performance metrics without using C. Feature Engineering
advanced sampling techniques or XAI methods. To improve model performance and interpretability,
In contrast to these studies, the present work several engineered features were introduced:
integrates advanced preprocessing, undersampling, • Lifestyle Score: A composite score
and a LightGBM classifier selected through cross- combining behavioral indicators such as
validation based on recall. Additionally, it combines physical activity, fruit and vegetable intake,
SHAP and LIME for both global and local smoking, alcohol consumption, and days of
interpretability, introduces lifestyle and healthcare poor mental and physical health. The score
scores for personalized risk feedback, and deploys the was normalized and scaled from 1 to 5.
model in a web-based Dash application. A further • Healthcare Access Score: Aggregated from
contribution includes correlation-based comorbidity responses related to insurance coverage,
insights, which have not been addressed in prior ability to afford care, and general access to
BRFSS-based research. healthcare services.
• Risk Factor Count: Created by summing
III. METHODOLOGY
binary indicators of high-risk conditions:
A. Dataset Description HighBP, HighChol, Stroke, and
This study uses a subset of the 2015 Behavioral HeartDiseaseorAttack. This count provided a
Risk Factor Surveillance System (BRFSS) dataset, a cumulative measure of an individual’s
health-related telephone survey conducted by the chronic health burden and emerged as the
Centers for Disease Control and Prevention (CDC). most influential predictor in SHAP analysis.
After preprocessing and feature selection, the working These derived features improved the model’s ability
dataset included approximately 250,000 self-reported to capture health behavior patterns and clinical risk
responses and 22 relevant features covering profiles beyond individual variables.
demographic, lifestyle, medical, and healthcare- D. Handling Class Imbalance
related variables. The original target variable in the Due to the natural imbalance between diabetic and
dataset included three classes: diabetic, non-diabetic, non-diabetic cases, three sampling strategies were
and prediabetic. These features were selected to employed for comparison:
 3

1. Original dataset without any sampling To extend the system’s utility, a correlation-based
adjustment. analysis was conducted to identify comorbidities most
2. SMOTE (Synthetic Minority associated with diabetes. Pearson correlation
Oversampling Technique) to synthetically coefficients were calculated between diabetes and
balance classes by generating new minority related variables such as high blood pressure, high
samples. cholesterol, stroke, and heart disease. These insights
3. Random Undersampling, which reduced were integrated into the application’s output as risk
the majority class to match the minority flags and visual summaries to support further health
class. risk assessment.
Each approach was applied independently, and H. Dash Application Deployment
model performance was evaluated using cross- The entire pipeline was deployed as a multi-step
validation to determine the most effective technique. interactive web application using Dash (Plotly). The
Ultimately, undersampling combined with LightGBM user interface guides individuals through sequential
yielded the highest recall and was selected as the final stages basic information, lifestyle habits, healthcare
strategy. access, and medical history. Upon submission, the
E. Model Training and Evaluation system generates real-time predictions, personalized
Several machine learning models were tested, health recommendations, comorbidity insights and
including Logistic Regression, Decision Tree, alerts, and LIME/SHAP-based explainability
Random Forest, XGBoost, and LightGBM. To address visualizations. The deployment prioritizes usability,
class imbalance, each model was evaluated across accessibility, and interpretability.
three sampling strategies: original distribution,
SMOTE oversampling, and random undersampling. IV. RESULTS AND DISCUSSION
The evaluation focused on recall as the primary A. Model Performance Across Sampling Strategies
metric, aiming to minimize false negatives in a The performance of six classifiers KNN, Logistic
healthcare setting. Regression, Random Forest, XGBoost, Neural
Hyperparameter tuning was performed using Network, and LightGBM was evaluated under three
GridSearchCV to optimize each model's performance class balancing scenarios: original dataset, SMOTE
within each sampling strategy. To assess the statistical oversampling, and random undersampling. Figures 1-
significance of performance differences, a one-way 3 summarize the results for each strategy using five
ANOVA was conducted on recall scores obtained via evaluation metrics: accuracy, precision, recall, F1-
stratified k-fold cross-validation, followed by Tukey's score, and
HSD post-hoc test to identify specific model ROC-AUC.
differences. As shown in Figure 1, performance on the original
LightGBM combined with random undersampling dataset favored accuracy but suffered in recall,
demonstrated the highest and most consistent recall, indicating the models failed to detect many diabetic
with statistically significant improvements over other cases. SMOTE improved recall significantly (Figure
models. Due to its balance of performance and 2), but led to reduced precision and accuracy.
explainability, it was selected as the final deployed Undersampling (Figure 3) achieved the best recall
model. across all models, especially with Logistic Regression,
F. Explainable AI Integration LightGBM, and XGBoost, making it the preferred
To enhance transparency and support decision- approach for minimizing false negatives.
making, both SHAP (SHapley Additive exPlanations)
and LIME (Local Interpretable Model-Agnostic
Explanations) were integrated. SHAP was used for
global feature importance visualization and
interpretation, revealing that the Risk Factor Count,
General Health, and BMI were among the top
predictors. LIME was used to explain individual Figure 1. Final Model Performance on Original
predictions by approximating the local behavior of the Dataset
model around specific instances. Together, these tools
provided a comprehensive view of how predictions
were generated and why. G. Comorbidity
Correlation Analysis
 4

Under both sampling conditions, KNN, Logistic

Regression, and Random Forest exhibited the most
pronounced performance gaps compared to
LightGBM, XGBoost, and Neural Network models.
The Tukey HSD results (Figures 5 and 7) confirm
these differences with statistically significant rejection
Figure 2. Final Model Performance with flags in the majority of pairwise comparisons
SMOTE Oversampling involving underperforming models.
The prioritization of recall as the evaluation metric
is critical in this health-focused context. False

gain
insights into feature contributions, SHAP (SHapley Additive
exPlanations) was employed. SHAP is based on cooperative game
Figure 3. Final Model Performance with theory and attributes model output to each feature by
Undersampling negatives i.e., predicting a non-diabetic status for a
truly diabetic individual can result in Figure 4. Cross-
B. Statistical Validation of Model Differences
Validation Recall Scores with SMOTE (Mean ± SD)
To verify the significance of performance
differences between models across sampling
strategies, a one-way ANOVA test was conducted on
recall scores obtained from stratified cross-validation.
For both SMOTE and random undersampling, the
resulting p-values were well below the 0.05 threshold,
indicating that the choice of model significantly
influenced recall outcomes.
Specifically, for SMOTE, the ANOVA test yielded a
p-value of 2.13 × 10⁻²⁴, and for random
undersampling, the p-value was 5.39 × 10⁻¹⁰,
confirming that the observed recall variations are
statistically significant and not due to random Figure 5. Tukey HSD Post-Hoc Comparison of
fluctuations. These findings validate that differences Model Recall Scores (SMOTE)
among model performances are reliable and merit
further exploration.
To identify which specific model pairs differed
significantly, a Tukey HSD (Honestly Significant
Difference) post-hoc test was applied. This method
controls the family-wise error rate and provides
pairwise comparisons of model recall means. The
results are visualized using compact summary tables
that include mean differences, confidence intervals,
and rejection flags indicating statistical significance
(True/False).
Visual inspection of recall scores under SMOTE
and undersampling (Figures. 4 and 6) highlights Figure 6. Cross-Validation Recall Scores with
noticeable differences between classifiers. While some Undersampling (Mean ± SD)
models like Random Forest and KNN perform
reasonably well in isolated conditions, their variability
and sensitivity to sampling techniques reduce
reliability. In contrast, LightGBM consistently
achieved high recall scores, particularly under
undersampling, and displayed minimal variance
across folds.
 C. SHAP-Based Global and Local Interpretability
To ensure transparency in model predictions and
5

calculating their marginal contribution across all

possible feature combinations. It enables both global
(dataset-level) and local (individual-level)
interpretability.
1) Global Feature Importance
A SHAP summary plot was generated to identify
which features most influenced diabetes predictions
across the dataset. As shown in Figure 8, the most
impactful features were:
• Risk Factor Count: An engineered variable
aggregating chronic conditions like HighBP,
HighChol, Stroke, and
HeartDiseaseorAttack.
• Physical Health (PhysHlth): Denotes the
number of days with poor physical health.
• HighBP and General Health (GenHlth):
Indicate chronic disease presence and
perceived well-being.
• Difficulty Walking (DiffWalk) and
Cholesterol Check (CholCheck): Reflect
mobility and engagement with preventive
care.
The SHAP summary plot ranks features based on
their mean absolute SHAP values, reflecting their
overall influence on model predictions. Features
positioned higher in the plot contribute more Figure 8. SHAP Summary Plot of Global Feature
significantly to the prediction outcome. The horizontal Importance
spread of each feature indicates whether its impact
pushes the prediction toward the diabetic (positive 2) Local Explanation with SHAP Waterfall Plot
SHAP value) or non-diabetic class (negative SHAP Figure 9 displays a SHAP waterfall plot that
value). explains how individual features influenced the
For example, individuals with high Risk Factor model’s prediction for a specific case. The base value
Count or poor physical health (PhysHlth) typically (1.393), which represents the average model output in
exhibit strong positive contributions toward diabetes log-odds, is adjusted by the SHAP values of each
classification. The color gradient encodes the actual feature to reach the final model prediction of 2.27 log-
feature values, with red representing higher values and odds for this individual.
blue indicating lower ones. This visualization provides Each numeric value next to a feature (e.g., +0.24 or
a comprehensive and interpretable overview of model –0.03) represents the feature's additive contribution to
behavior, reinforcing the relevance of selected the prediction. A positive SHAP value indicates that
predictors in a clinically meaningful way. the feature increased the predicted risk, while a
negative value decreased it. For example, high values
for PhysHlth (15 days), Risk Factor Count (2), and the
presence of HighBP (1) and DiffWalk (1) had the
strongest positive impact, pushing the prediction well
above the average. In contrast, Smoker (1) showed a
mild negative contribution, slightly lowering the
output.
The final log-odds score is internally converted to a
probability for classification, but this intermediate
representation helps to understand which features had
the greatest effect. Such local explanations not only
improve transparency but also support clinically
 6

relevant, personalized interpretation—highlighting elevated cholesterol, all of which increased

risk factors specific to an individual case. the likelihood of a positive classification.
This type of local explanation is particularly • Features such as income ≤ 5, moderate
valuable in healthcare applications, where physical health issues, and age between 8
understanding why a prediction was made is as and 10 (coded bin) also reinforced the
important as the prediction itself. By quantifying the diabetic prediction.
contribution of each feature, SHAP enables clinicians, On the right side of the figure, the corresponding
public health professionals, or end-users to trace the feature values used by the model are displayed,
reasoning behind a high-risk or low-risk classification. confirming the individual had elevated values for
In this case, the prediction can be clearly attributed to BMI (40.0), GenHlth (5), and multiple chronic
measurable health conditions such as elevated conditions.
physical health concerns and comorbid risks. These This visual explanation not only confirms the
insights can guide individualized interventions, decision rationale but also supports user trust and
promote informed decision-making, and build trust in transparency in a healthcare context. By clearly
machine learning systems deployed in sensitive identifying the top risk-driving factors, LIME
domains like health risk assessment. empowers patients and clinicians to focus on the most
influential health areas.
In healthcare applications, the ability to generate
such individualized explanations is essential for
fostering transparency and patient engagement.
LIME's local approximations allow users to explore
how specific health factors influence model output in
a manner that aligns with clinical reasoning. This
makes the tool particularly effective for patient
education, shared decision-making, and early
intervention planning. By surfacing interpretable,
instance-level justifications, LIME enables both
clinicians and patients to move beyond black-box
predictions toward actionable, personalized healthcare
Figure 9. SHAP Waterfall Plot for an Individual insights.
Prediction

D. LIME-Based Local Explanations

To enhance model interpretability at the individual
level, LIME (Local Interpretable Model-Agnostic
Explanations) was applied. LIME generates
Figure 10. LIME Explanation for an Individual
explanations by approximating the model locally
Prediction
around a given prediction using a simple interpretable
E. Comorbidity Insights
model (typically linear regression). This allows for
Understanding how diabetes interacts with other
clear visualization of which features contributed most
chronic health conditions is critical for assessing
to a specific classification.
overall patient risk. To explore these associations, a
In this study, a representative prediction was
Pearson correlation analysis was conducted between
selected for analysis. As shown in Figure 10, the
the diabetes outcome variable and other risk factors
model predicted a 92% probability of diabetes for
such as High Blood Pressure, High Cholesterol,
the chosen individual. The bar chart highlights
Stroke, and Heart Disease.
features that contributed to this prediction:
As illustrated in Figure 11, the strongest
• BMI > 32, GenHlth > 3, and Risk Factor
comorbidity was observed between diabetes and high
Count = 2 were the strongest contributors
blood pressure (r = 0.26), followed by high cholesterol
driving the prediction toward the diabetic
(r = 0.20), heart disease or heart attack (r = 0.17), and
class.
stroke (r = 0.10). These moderate but meaningful
• Other contributing factors included high correlations align with established clinical research,
blood pressure, difficulty walking, and which identifies these conditions as common
 7

complications or co-existing risks in diabetic After all inputs are submitted, users are shown the
individuals. prediction result, lifestyle and healthcare scores,
This comorbidity analysis not only reinforces the and explainable insights derived from SHAP and
predictive importance of these variables in the model LIME (Figure 18–21). The final screen also provides:
but also enhances the real-world relevance of the • Top SHAP features influencing the
deployed tool. In the application, users flagged as prediction
diabetic can be simultaneously alerted to increased  LIME interpretation for the specific
risks of cardiovascular disease, thereby promoting individual  Personalized health
preventive awareness and early clinical engagement. improvement suggestions  Comorbidity
risks and recommendations
This user-friendly interface bridges technical output
with public usability, enabling non-technical users to
understand their diabetes risk and contributing factors.
The platform emphasizes accessibility and
interpretability, combining modern UI design with
explainable AI outputs. Visual elements such as SHAP
and LIME plots, lifestyle scoring, and comorbidity
insights help users understand both their risk and
contributing factors. This integration of prediction and
explanation supports informed decision-making in
both clinical and public health contexts.

Figure 11. Correlation Heatmap Between Diabetes

and
Related Chronic Conditions
D. Interactive Web Interface and User Experience
To make the predictive model accessible and user-
friendly, an interactive web application was developed
using the Dash framework. The application features a
multi-step form that collects demographic, lifestyle,
medical history, and healthcare-related information to
generate personalized diabetes risk predictions. The
interface was designed with glassmorphism styling
for modern aesthetics and clarity.
Users are guided through the following sequential
steps:
 Step 1: Basic Information (Age, Sex, BMI –
Figure 12 . Step 1: Basic Information Input
Figure
12)
• Step 2: Lifestyle Habits (Smoking, alcohol
consumption, fruit/vegetable intake, physical
activity – Figure 13)
• Step 3: Medical History (High BP, High
Chol,
Stroke, Heart Disease – Figure 14)
 Step 4: Healthcare Access & Physical
Condition
(General health, mental/physical health days,
walking difficulty – Figure 15)
• Step 5: Socioeconomic Details (Education
and Income levels – Figure 16)
 8

Figure 13. Step 2: Lifestyle Habits Form

Figure 14. Step 3: Medical History Form

Figure 15. Step 4: Healthcare & Physical

Condition Form
 9

Figure 16. Step 5: Education & Income Level Form Figure 19. Prediction Result with Healthcare
Score

Figure 17. Prediction Submission Interface

Figure 20. Top SHAP Features and LIME
Explanations

Figure 18. Prediction Result with Lifestyle

Score
Figure 21. Personalized Suggestions and
Comorbidity Insights
V. CONCLUSION AND FUTURE WORK
This study developed an explainable, user-centered
diabetes risk prediction system using the 2015 BRFSS
dataset and machine learning. The final LightGBM
model, selected through extensive cross-validation
and statistical testing, demonstrated superior recall
performance under an undersampling strategy. To
promote model transparency, SHAP and LIME were
employed to provide both global and local
explanations, offering insights into the key drivers of
diabetes predictions at both the population and
individual levels.
A key contribution of this work is the deployment
of a web-based Health Risk Calculator that transforms
static predictions into actionable health insights. The
 10

multi-step interface not only guides users through The author would like to sincerely thank Dr.
demographic, lifestyle, and clinical input collection Mariofanna Milanova, Professor of Computer
but also delivers tailored predictions, interpretability Science at the University of Arkansas at Little Rock
visuals, and comorbidity alerts. This integration of and IEEE Senior Member, for her invaluable
predictive analytics and explainable AI makes the mentorship and technical guidance throughout the
system practical and accessible for real-world use. development of this research project. The author also
Future Work will focus on expanding the model’s wishes to thank Dr. Francesco Cavarretta, Assistant
scope beyond diabetes to include multi-disease risk Professor in the Department of Computer Science and
prediction, using longitudinal or more recent datasets Research Fellow at the Emerging Analytics Center at
for improved temporal relevance. Integration with the University of Arkansas at Little Rock, for his
electronic health records (EHRs), inclusion of real- insightful feedback and support. Their expertise in
time wearable data, and implementation of multi- artificial intelligence, machine learning, and
language support are also planned to improve computational modeling played a crucial role in
usability and accuracy. Additionally, incorporating shaping both the predictive framework and the
user feedback loops and external clinical validation interpretability components of the system. This work
will further strengthen the model's reliability and was conducted as part of a graduate research project in
applicability in healthcare settings. the Department of Computer and Information
The explainable nature of the system ensures that Sciences at the University of Arkansas at Little Rock.
predictions are not only accurate but also
interpretable, helping bridge the gap between black- REFERENCES
box machine learning models and clinical trust. By [1] M. M. Chowdhury, R. S. Ayon, and M. S. Hossain, “Diabetes
surfacing clear reasoning for each prediction, the diagnosis through machine learning: Investigating algorithms
and data augmentation for class imbalanced BRFSS dataset,”
system empowers both patients and healthcare Preprint, Dept. of Mathematics and Statistics, Texas Tech
providers to make informed decisions. This Univ., Lubbock, TX, USA, and Dept. of Electronics and
transparency is critical for fostering the adoption of Telecommunication Eng., Rajshahi Univ. of Engineering and
Technology, Rajshahi, Bangladesh, 2023.
AI-driven tools in healthcare environments,
[2] Z. Liu, Q. Zhang, H. Zheng, S. Chen, and Y. Gong, “A
particularly when applied to chronic disease comparative study of machine learning approaches for
prevention and early detection. diabetes risk prediction: Insights from SHAP and feature
Moreover, this research contributes to the growing importance,” Preprints, Nov. 19, 2024. [Online]. Available:
[Link]
field of interpretable machine learning by [3] S. Ahmed, M. S. Kaiser, M. S. Hossain, and K. Andersson, “A
demonstrating a complete pipeline from data comparative analysis of LIME and SHAP interpreters with
preprocessing and model selection to post hoc explainable ML-based diabetes predictions,” IEEE Access,
vol. 12, pp. 87912–87925, 2025. doi:
explanation and real-time deployment. The modular 10.1109/ACCESS.2024.3422319.
design of the framework allows for easy adaptation to [4] B. Nguyen and Y. Zhang, “A comparative study of diabetes
other health conditions and datasets. As health data prediction based on lifestyle factors using machine learning,”
unpublished manuscript, 2025.
becomes more complex and high-dimensional, the [5] H. Nguyen, H. Cao, V. Nguyen, and D. Pham, "Evaluation of
need for interpretable, interactive, and scalable Explainable Artificial Intelligence: SHAP, LIME, and CAM,"
prediction systems will only increase, positioning this *ResearchGate preprint*, 2021. [Online]. Available:
[Link]
work as a strong foundation for future advances in
[6] A. L. R. Agahan, M. S. A. Magboo, and V. P. C. Magboo,
patient-centered predictive analytics. "Predicting the Risk of Diabetes Using Explainable Artificial
In summary, this study not only highlights the Intelligence," in *Proc. 2023 Int. Conf. on Electrical,
predictive capabilities of modern machine learning Computer and Energy Technologies (ICECET)*, 2023, pp. 1–
6. doi: 10.1109/ICECET58911.2023.10389419.
techniques but also emphasizes the importance of [7] N. Gandhi and S. Mishra, “Explainable AI for healthcare: A
usability, interpretability, and personalization in digital study for interpreting diabetes prediction,” in Proc. Int. Conf.
health applications. By integrating statistical rigor on Machine Learning and Big Data Analytics (ICMLBDA),
Cham, Switzerland: Springer, 2022, pp. 95–105.
with user-focused design and explainable AI, the [8] Y. Ramon, D. Martens, F. Provost, and T. Evgeniou, “A
system demonstrates a balanced approach to comparison of instance-level counterfactual explanation
technology-driven health risk assessment. As AI algorithms for behavioral and textual data: SEDC, LIME-C
and SHAP-C,” Advances in Data Analysis and Classification,
continues to shape the future of healthcare, solutions vol. 14, no. 4, pp. 801–819, Dec. 2020.
like this provide a template for building transparent, [9] Y. Zhao, J. K. Chaw, M. C. Ang, M. M. Daud, and L. Liu, “A
inclusive, and impactful tools that align with both diabetes prediction model with visualized explainable
artificial intelligence (XAI) technology,” in *Advances in
clinical standards and public health needs. Visual Informatics (Lecture Notes in Computer Science)*,
VI. ACKNOWLEDGEMENT vol. 14322, H. B. Zaman, Ed., Singapore: Springer, 2024. doi:
10.1007/978-981-99-7339-2_52.
 11

[10] A. Priyadarshini and J. Aravinth, “Correlation based breast

cancer detection using machine learning,” in Proc. Int. Conf.
on Recent Trends in Electronics, Information &
Communication Technology (RTEICT), 2021, pp. 499–504.
[11] A. Yahyaoui, A. Jamil, J. Rasheed, and M. Yesiltepe, “A
decision support system for diabetes prediction using machine
learning and deep learning techniques,” in Proc. 1st Int.
Informat. Softw. Eng. Conf. (UBMYK), 2019, pp. 1–4.
[12] A. Mujumdar and V. Vaidehi, “Diabetes prediction using
machine learning algorithms,” Procedia Computer Science,
vol. 165, pp. 292–299, 2019.
[13] N. Fazakis, O. Kocsis, E. Dritsas, S. Alexiou, N. Fakotakis,
and K. Moustakas, “Machine learning tools for long-term type
2 diabetes risk prediction,” IEEE Access, vol. 9, pp. 103737–
103757, 2021.
[14] M. A. Sarwar, N. Kamal, W. Hamid, and M. A. Shah,
“Prediction of diabetes using machine learning algorithms in
healthcare,” in Proc. 24th Int. Conf. on Automation and
Computing (ICAC), Sep. 2018, pp. 1–6.
[15] M. U. Emon, M. S. Keya, M. S. Kaiser, M. A. Islam, T.
Tanha, and M. S. Zulfiker, “Primary stage of diabetes
prediction using machine learning approaches,” in Proc. Int.
Conf. on Artificial Intelligence and Smart Systems (ICAIS),
Mar. 2021, pp. 364–367.
[16] Z. Q. Lin, M. J. Shafiee, S. Bochkarev, M. St. Jules, X. Y.
Wang, and A. Wong, “Do explanations reflect decisions? A
machine-centric strategy to quantify the performance of
explainability algorithms,” arXiv preprint arXiv:1910.07387,
2019.
[17] J. J. Bigna and J. J. Noubiap, “The rising burden of non-
communicable diseases in sub-Saharan Africa,” The Lancet
Global Health, vol. 7, no. 10, pp. e1295–e1296, 2019.
[18] L. Breiman, “Random forests,” Machine Learning, vol. 45,
pp. 5–32, 2001.
[19] T. A. Buchanan, A. H. Xiang, et al., “Gestational diabetes
mellitus,” *The Journal of Clinical Investigation*, vol. 115,
no. 3, pp. 485–491, 2005.
[20] A. Budreviciute, S. Damiati, D. K.
Sabir, K. Onder, P.
Schuller-Goetzburg, G. Plakys, A. Katileviciute, S. Khoja, and
R. Kodzius, “Management and prevention strategies for non-
communicable diseases (NCDs) and their risk factors,”
*Frontiers in Public Health*, vol. 8, p. 788, 2020.
[21] J. Burez and D. Van den Poel, “Handling class imbalance in
customer churn prediction,” Expert Systems with
Applications, vol. 36, no. 3, pp. 4626–4636, 2009.
[22] C. J. Caspersen, G. D. Thomas, L. A. Boseman, G. L.
Beckles, and A. L. Albright, “Aging, diabetes, and the public
health system in the United States,” *American Journal of
Public Health*, vol. 102, no. 8, pp. 1482–1497, 2012.