Generalisable machine learning models trained on heart rate variability data to predict mental fatigue.

Area of Science:

• Physiology
• Machine Learning
• Cognitive Science

Background:

• Mental fatigue from prolonged cognitive tasks increases accident risk.
• Previous Heart Rate Variability (HRV) studies for fatigue detection lack generalizability due to single-task training.
• Developing generalized fatigue detection requires diverse task-based data.

Purpose of the Study:

• To develop and validate machine learning algorithms for detecting and predicting mental fatigue severity.
• To improve the generalizability of fatigue detection by training on heterogeneous, multi-task datasets.
• To optimize algorithm performance by evaluating different time window lengths and data types (resting vs. task-related).

Main Methods:

• Combined datasets from three experiments with different cognitive fatigue induction tasks.
• Trained and compared machine learning algorithms (Support Vector Classifier, CatBoost regression) for fatigue detection and severity prediction.
• Evaluated performance using various time window lengths and compared resting vs. task-related HRV data.

Main Results:

• Support Vector Classifier achieved best detection performance (AUC=0.843, accuracy=0.761) using task-related HRV in a 5-min window.
• CatBoost regression showed best severity prediction performance (R²=0.248, RMSE=17.058) using 3-min HRV and self-reported measures.
• Heterogeneous, multi-task training significantly improved model generalizability.

Conclusions:

• Machine learning models trained on diverse cognitive tasks effectively detect and predict mental fatigue using HRV.
• Optimized time window lengths and data types enhance the accuracy of fatigue detection and severity prediction.
• This approach offers a promising method for preventing fatigue-related injuries and accidents.