Strategies for Generalisable Machine Learning with Small Data for Exercise Fatigue Detection

This thesis contributes to the field of human-centred computing by exploring strategies and pitfalls for developing generalisable machine learning models for sensor-based exercise fatigue detection with small data. Machine learning faces several challenges in exercise fatigue detection due to the scarcity of available data, as fatigue research typically relies on sample data from a limited number of subjects due to the time and effort required to design, conduct, and analyse studies with human subjects. Although machine learning models can provide robust predictions, even when trained on small data sets, careful consideration of variability and data distribution is required to improve their generalisability. A step-by-step framework for exercise fatigue detection with machine learning and small data is introduced in this thesis. The framework is implemented in a case study of 48 subjects performing squat exercises, using inertial measurement units and 2D pose estimation to capture movement patterns and correlate these patterns with ratings of perceived exertion. The results are analysed in terms of generalisability, including different numbers of classes and subjects, class imbalances, k-fold cross-validation, oversampling, inter-individual variability, evaluation metrics, and evaluation types. Based on a comparison between inertial units and 2D pose estimation, it is concluded that 2D pose estimation can be used for fatigue detection. From the literature survey and case study, it is also concluded that most exercise fatigue detection models trained on small data sets may not perform well in a real-world application.