Optimizing Biomedical Sensor Selection Using Artificial Intelligence and Game Theory: A Shapley Value Approach

This thesis proposes a new route toward biomedical sensor optimization by integrating Game Theory and Artificial Intelligence (AI), specifically, using the Shapley value method. The prime objective is to identify significant sensors and choose an optimal sub-collection that can deliver high predictive accuracy with lower numbers of sensors, thereby enhancing resource optimization. Two datasets were utilized: the HuGaDB dataset for human activity recognition, and a synthetically generated dataset covering a broad range of biomedical sensor data simulating stress responses. Machine learning techniques such as Random Forests and Support Vector Machines (SVM), supported by Shapley value-based explainability (SHAP), were employed to compare the individual and joint contributions of sensors to predictions in models. The game-theoretic Shapley value solution methodically quantifies sensor contributions with provision of unambiguous and effective sensor selection. Experimental results demonstrated the suggested AI-based Shapley model efficiently chose close-to-optimal predictive accuracy offering subsets of sensors, significantly reducing computational complexity and energy consumption. Thus, this thesis introduces a new approach to sensor optimization, highlighting complementarity between AI explainability and game-theoretic fairness with immense potential towards biomedical monitoring system design for high-performance and efficient systems.