Abstract
Maintaining optimal thermal comfort in buildings is critical for occupant well-being and energy efficiency. This study introduces a permutation-importance-based feature selection method for multiclass thermal comfort classification using ensemble algorithms. Data comprising environmental (air temperature, humidity, CO₂ concentration) and physiological (SpO₂, blood pressure, BMI, HRV metrics) variables were collected from 1536 samples and labeled on the ASHRAE 7-point scale. Decision Tree, AdaBoost, and CatBoost models were trained on the full feature set, then re-evaluated using only the ten most informative predictors identified via permutation importance. Results show that reduced-feature models match or slightly outperform full-feature counterparts: CatBoost accuracy improved from 0.961 to 0.971, AdaBoost from 0.841 to 0.857, and Decision Tree from 0.688 to 0.825, while dramatically lowering sensor and computational requirements. Paired t-tests confirmed no significant performance loss. This streamlined approach enables cost-effective, real-time thermal comfort monitoring and supports the deployment of intelligent HVAC systems with minimal hardware.