A Selective RAG-Enhanced Hybrid ML-LLM Framework for Efficient and Explainable Fatigue Prediction Using Wearable...

Area of Science:

• Computational science
• Occupational health
• Artificial intelligence

Background:

• Fatigue significantly impacts performance in demanding professions.
• Wearable sensors offer continuous monitoring, but traditional machine learning (ML) models lack reliability and interpretability in real-world scenarios.
• Existing fatigue prediction models struggle with unstable, poorly calibrated, and opaque outputs.

Purpose of the Study:

• To develop a robust and interpretable fatigue classification framework for high-stress occupations.
• To integrate the efficiency of ML with the reasoning capabilities of large language models (LLMs) for enhanced prediction.
• To improve the reliability and transparency of fatigue monitoring using wearable data.

Main Methods:

• A selective Retrieval-Augmented Generation (RAG)-enhanced hybrid ML-LLM framework was developed.
• ML models (logistic regression, XGBoost, LSTM) were trained on wearable and ecological momentary assessment data from 297 emergency responders.
• LLMs were selectively activated for borderline ML predictions (0.45 ≤ p ≤ 0.55) using symbolic rules and retrieved examples.

Main Results:

• The hybrid framework improved classification performance in the uncertainty region (accuracy: 0.556 to 0.617, precision: 0.684 to 0.703, recall: 0.635 to 0.748, F1: 0.659 to 0.725).
• Overall test set performance also improved (accuracy: 0.707 to 0.718, precision: 0.739 to 0.741, recall: 0.918 to 0.937, F1: 0.819 to 0.827).
• SHAP and LLM analyses identified sleep duration and heart-rate variability as key predictors, enhancing model transparency.

Conclusions:

• The RAG-enhanced hybrid ML-LLM framework significantly boosts the robustness, interpretability, and efficiency of fatigue classification.
• This scalable solution offers a more reliable method for real-world fatigue monitoring in high-stress occupations.
• The integration of ML and LLMs provides transparent explanations for fatigue prediction, addressing limitations of conventional models.