💻 computer science

MO-AIWOA-XAI: A Multi-Objective Explainable Diabetes Risk Prediction integrating Feature Optimization and Cross-Dataset Validation

This paper proposes the MO-AIWOA-XAI framework, which integrates multi-objective adaptive improved whale optimization for feature selection, an Optuna-optimized Extra Trees classifier, and explainable AI techniques to deliver a robust, accurate, and clinically interpretable diabetes risk prediction model validated across multiple datasets.

Original authors: Sumalatha U, krishnaprakasha k

Published 2026-11-12

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Sumalatha U, krishnaprakasha k

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Diabetes is a condition where the body struggles to manage sugar, leading to serious health problems if not caught early. Doctors have long relied on blood tests and physical exams to spot the warning signs, but these methods often require a visit to a clinic and a lab. In recent years, scientists have turned to computers to help. By feeding medical records into machine learning programs, researchers can teach computers to spot patterns that suggest a person is at risk of developing diabetes. However, these computer programs often act like black boxes. They might give a correct answer, but they cannot explain why they reached that conclusion. For a doctor to trust a computer's advice, they need to understand the reasoning behind it. They need to know which specific factors, like age or blood pressure, tipped the scale. Furthermore, many computer models are trained on just one type of patient data, making it unclear if they would work for a different group of people.

A team of researchers has developed a new system designed to solve these problems. Their approach, which they call MO-AIWOA-XAI, is built to do three things at once: find the most important health clues, predict the risk of diabetes with high accuracy, and explain exactly how it made that prediction. The system starts by sifting through a large list of patient details to find the few that matter most. It then uses a smart learning program to make its prediction. Finally, it generates a clear report for the doctor, showing not just the risk level, but also what changes a patient could make to lower that risk. The researchers tested this system on two very different sets of patient data to ensure it works reliably.

The researchers began by gathering two distinct collections of patient information. The first set came from a group of women of Pima Indian heritage and included detailed physiological measurements like glucose levels, blood pressure, and body mass index. The second set was different; it focused on early warning symptoms, such as frequent thirst, sudden weight loss, and blurred vision, collected from a mixed population. These two datasets represented different ways doctors might encounter the disease: one through lab results and the other through patient complaints. Before the computer could learn from this data, the researchers had to clean it up. They filled in missing numbers, adjusted for differences in how the data was recorded, and balanced the groups so the computer would not be biased toward one outcome over another.

To find the best predictors, the team used a method inspired by the hunting behavior of humpback whales. In nature, these whales work together to trap fish in a bubble net. The computer algorithm mimics this by sending out a group of virtual "whales" to explore different combinations of health factors. Each whale tests a specific set of features, like checking if age and glucose together are better predictors than age and blood pressure. The algorithm improves over time, learning which combinations work best. Crucially, this system was designed to look for two goals simultaneously: it wanted the highest possible accuracy in predicting diabetes, but it also wanted to use as few features as possible. This is vital for real-world use, as doctors prefer simple, quick checks over complex, expensive panels of tests.

After running this optimization process, the system identified a small, powerful group of four factors that carried the most weight: the number of pregnancies a woman has had, her glucose level, her blood pressure, and her body mass index. By focusing only on these four, the system reduced the amount of data it needed to process by about 60 percent, yet it maintained its ability to spot the disease. The researchers then trained a final prediction model using these four key factors. To ensure the model was tuned perfectly, they used an automated tool that tested thousands of different settings to find the one that worked best.

The results of this new framework were impressive. When tested on the Pima Indian dataset, the system correctly identified diabetes cases 83.76 percent of the time. On the second dataset, which focused on early symptoms, it achieved an even higher accuracy of 95.19 percent. More importantly, the system proved it was not just guessing; it was stable. When the researchers tested it on different slices of the data, the results remained consistent. They also checked to see if the probabilities the computer gave matched reality, and they found that the predictions were well-calibrated, meaning a 90 percent risk score truly meant a high likelihood of the disease.

Perhaps the most significant part of this work is the explanation. The system does not just say "high risk"; it tells the doctor why. Using a technique called SHAP, it showed that glucose and body mass index were the most influential factors across the entire group of patients. For an individual patient, the system uses LIME to explain that specific person's risk, highlighting exactly which of their measurements pushed the score up. It even goes a step further with a method called DiCE, which generates counterfactual explanations. This means the system can tell a patient what would need to change to lower their risk. For example, it might show that if a patient's glucose level were slightly lower or their weight slightly reduced, the computer's prediction would shift from "diabetes" to "no diabetes." This turns a static prediction into a dynamic tool for prevention.

The researchers validated their work by comparing it against older methods and by testing it across the two different datasets. They found that their approach was more robust and reliable than standard techniques. They also confirmed that the explanations the system gave were stable; if they ran the test again, the system would point to the same important factors. This consistency is essential for building trust in medical settings. The study concludes that by combining smart feature selection, a powerful prediction model, and clear explanations, it is possible to create a tool that is both accurate and understandable. This framework offers a path forward for clinical decision support, where computers can assist doctors not just by calculating numbers, but by providing the context and reasoning needed to make informed choices about patient care.

Technical Summary: MO-AIWOA-XAI Framework for Diabetes Risk Prediction

Problem Statement
The paper addresses critical challenges in applying machine learning to clinical decision-making for diabetes management. While existing models often achieve high predictive accuracy, their adoption in healthcare is hindered by a lack of transparency ("black-box" nature) and the presence of feature redundancy in high-dimensional medical datasets. Furthermore, many existing studies rely on single datasets, lack comprehensive validation (such as cross-dataset testing or calibration analysis), and fail to integrate multi-objective optimization with explainable artificial intelligence (XAI) to provide actionable clinical insights.

Methodology
The authors propose MO-AIWOA-XAI, a comprehensive framework integrating multi-objective feature optimization, ensemble learning, and explainable AI. The methodology consists of the following core components:

Data Preprocessing:
- Imputation: K-Nearest Neighbor (KNN) imputation was used for missing numerical values (e.g., glucose, insulin) to preserve data distribution, while mode imputation was used for categorical data.
- Feature Engineering: Interaction features (e.g., Age × BMI, Glucose × BMI) were created to capture non-linear relationships.
- Normalization: Min-Max normalization scaled numerical features to the [0, 1] interval.
- Class Balancing: The SMOTE-Tomek technique was applied to address class imbalance, resulting in a balanced dataset of 491 positive and 491 negative cases.
Feature Selection (MO-AIWOA):
- The study introduces a Multi-Objective Adaptive Improved Whale Optimization Algorithm (MO-AIWOA). This is a modification of the standard Whale Optimization Algorithm (WOA) designed to overcome premature convergence and poor diversity.
- Improvements: The algorithm incorporates adaptive exploration-exploitation control, dynamic position updating, and a random perturbation strategy to maintain population diversity.
- Objective: It simultaneously optimizes two objectives: maximizing predictive performance (AUC) and minimizing feature dimensionality. The fitness function is defined as $0.7 \times \text{AUC} + 0.3 \times \text{FeatureReduction}$ .
- Outcome: The algorithm reduced the feature space by approximately 60%, selecting a subset of four clinically relevant features: Pregnancies, Glucose, Blood Pressure, and BMI.
Classification Model:
- An Extra Trees (ET) classifier was employed for risk prediction.
- Hyperparameter Tuning: The Optuna library, utilizing Bayesian optimization with a Tree-structured Parzen Estimator (TPE), was used to optimize the ET hyperparameters (number of trees, max depth, min samples split/leaf).
Explainable AI (XAI) Integration:
- SHAP (SHapley Additive exPlanations): Provided both global and local explanations, quantifying feature contributions based on cooperative game theory.
- LIME (Local Interpretable Model-agnostic Explanations): Generated personalized, local explanations for individual patient predictions.
- DiCE (Diverse Counterfactual Explanations): Produced counterfactual instances to identify minimal changes required in input features to alter the prediction outcome (e.g., changing a prediction from diabetic to non-diabetic).
Validation Strategy:
- The framework was evaluated on two distinct datasets: the PIMA Indians Diabetes Dataset (physiological parameters) and the Early Stage Diabetes Risk Prediction Dataset (symptoms and demographics).
- Metrics: Evaluation included Accuracy, AUC, F1-score, MCC, PR-AUC, and Cohen's Kappa.
- Robustness Checks: The study employed 10-fold stratified cross-validation, bootstrap confidence interval analysis (1,000 iterations), calibration analysis, and Decision Curve Analysis (DCA) to assess clinical utility.

Key Results

Performance on PIMA Dataset: The framework achieved an accuracy of 83.76%, an AUC of 0.9004, an F1-score of 0.84, an MCC of 0.6757, and a PR-AUC of 0.9068.
Performance on Early Stage Dataset: The model achieved an accuracy of 95.19% and an AUC of 0.992.
Stability: The 10-fold cross-validation yielded a mean accuracy of 80.12% ± 3.78%, indicating stable performance across data partitions. The bootstrap 95% confidence interval for accuracy was [77.16%, 87.31%].
Explainability: The SHAP stability score was 1.0, indicating perfect consistency in feature rankings across resampling experiments. The framework successfully generated stable global, local, and counterfactual explanations.
Clinical Utility: Decision Curve Analysis demonstrated a positive net benefit across a wide range of probability thresholds, suggesting the model provides meaningful support for clinical decision-making.

Significance and Claims
The paper claims that the MO-AIWOA-XAI framework effectively bridges the gap between high predictive performance and model interpretability in diabetes risk assessment. By combining multi-objective optimization with ensemble learning and a triad of XAI techniques (SHAP, LIME, DiCE), the framework offers a solution that is not only accurate but also transparent and clinically actionable.

The authors position the work as a step toward reliable clinical decision support systems that can handle feature redundancy and provide the necessary transparency for healthcare practitioners to trust AI-driven diagnoses. The framework is presented as a computationally efficient candidate for future clinical applications, balancing the trade-off between feature reduction and diagnostic capability.

Limitations and Future Directions
The authors acknowledge several limitations:

The experiments were conducted on public benchmark datasets, which may not fully capture the heterogeneity of real-world clinical populations.
The optimization process requires more computational resources than simple filter-based methods.
The XAI techniques used are post-hoc and may not fully reflect the internal logic of the classifier.

Future work is proposed to include validation on large-scale, multi-center clinical datasets, the integration of longitudinal patient records and wearable sensor data, and the development of federated learning frameworks to ensure patient privacy while improving predictive power.

Technical Summary: MO-AIWOA-XAI Framework for Diabetes Risk Prediction

More like this