Stress is reflected in measurable physiological changes, which can affect decision making in mentally demanding activities. The goal was to determine whether physiological parameters can be used to predict chess performance and events, and whether they improve predictions based on chess features. Data collected at an official chess tournament was used to build a dataset combining each player's physiological and chess features for every move. In five experiments, machine learning models were trained and evaluated using physiological and chess features, and their combination. SHAP was used to explain predictions. When identifying the player with the lower ELO rating, physiological features achieved an accuracy of 62.4%, compared with 59.1% using chess features. Physiological features alone identified a substantial advantage with an accuracy of 67.1%. Adding them to a model using chess features increased accuracy from 72.5% to 75.5%. Using chess features, we detected 54.8% of mistakes with a precision of 26.9%, while adding physiological features did not improve predictive performance. Across all experiments, models using physiological features noticeably outperformed a random classifier, though to a lesser extent when predicting mistakes.
|