Enhanced Diagnostic and Interpretable Model for Febrile Diseases Using Explainable AI and Large Language Model

Résumé

Tropical febrile diseases pose significant public health challenges in tropical regions, particularly in resource-limited settings, where timely and accurate diagnoses are critical for effective disease management. However, existing diagnostic systems are black-box and often lack interpretability, making it difficult for healthcare practitioners to understand and trust machine learning (ML) predictions. Additionally, the existing system excludes younger patients, limiting its applicability. This study aimed to develop an integrated diagnostic system that enhances interpretability by leveraging machine learning (ML), Explainable AI (XAI), and a large language model (LLM). A dataset comprising 3,914 patient records with 32 symptoms was obtained from a New Frontiers in Research Fund-sponsored project and preprocessed for analysis. Random Forest (RF) was trained with hyperparameter tuning using GridSearchCV and evaluated with 5-fold cross-validation to optimize its performance. The tuned model achieved high diagnostic performance, demonstrating its effectiveness in diagnosing six febrile diseases. To enhance interpretability, a Model Interpretability Framework (MIF) was incorporated, integrating Local Interpretable Model-Agnostic Explanations (LIME) for visual insights and ChatGPT for textual explanations. This combination provided users with a clear understanding of the diagnostic process, addressing the limitations of black-box ML models. The system was implemented using Python in a development environment that included Google Colaboratory, Visual Studio Code, Flet, and MySQL for database management. Agile development methodology and Unified Modeling Language tools were employed to create a user-friendly design for clinicians and patients. The diagnostic system was evaluated using precision, recall, F1-score, and AUC-ROC metrics. The RF model achieved outstanding precision for most diseases, high recall (≥0.97), and an AUC-ROC of 0.99, indicating nearly flawless training performance with few misclassifications. On the test dataset, the model performed better in diagnosing malaria (F1-score = 88%, precision = 85%), urinary tract infection (F1-score = 72%, precision = 80%), and respiratory tract infection (F1-score = 72%, precision = 77%). The evaluated system demonstrated strong predictive performance, correctly identifying 72 out of 99 tested patient cases, with 26 cases misclassified and one case completely missed. It achieved the highest detection rates for malaria and HIV/AIDS, accurately identifying all cases without false positives. This work enhances diagnostic solutions for tropical healthcare settings by improving explainability and ensuring inclusivity for younger patients, addressing key limitations in existing systems. The integration of ML, XAI, and LLM enhances diagnostic performance and lays the groundwork for future AI-driven healthcare innovations, ultimately improving health outcomes in resource-scarce regions. Healthcare providers should adopt this system to enhance the efficiency of disease diagnosis and extend the system to diagnose additional medical conditions.

Citer ce document

Accès au document

Ce document est hébergé sur son dépôt institutionnel d'origine.

Auteur(s)

Statistiques

Consultations : 3

Téléchargements : 0