Multimodal Fusion of Environmental and Physiological Data for Real-World Personalised Comfort Modelling

People spend the majority of their lives within environments shaped by multiple interacting exposures, including thermal conditions, acoustic noise, lighting, and air quality, yet remain largely unaware of how these settings affect their comfort. Existing comfort research treats domains in isolation under controlled laboratory conditions, leaving real-world multi-domain effects on personal comfort underexplored. This paper proposes a unified Comfort Framework that fuses three practical data layers: macro-environmental conditions retrieved via location-based APIs, kinematic and micro-environmental context captured through smartphone sensors, and physiological responses recorded by a chest-worn ECG sensor. Binary comfort states are labelled in real time using a minimal-disruption lap-button protocol on a consumer smartwatch. We validate the pipeline through a single-subject pilot of 18 free-living sessions. Random Forest classification across 10 valid leave-one-session-out folds achieved an F1 macro of 0.456 ± 0.151, indicating that consumer wearable comfort prediction in unconstrained free-living conditions is more challenging than controlled chamber studies suggest. Descriptive statistics showed dataset-level differences between comfort states in wrist skin temperature (31.9 vs. 33.3 °C), heart rate (70.7 vs. 77.1 bpm), and RMSSD (42.1 vs. 34.3 ms), with overlap between classes consistent with the modest classification performance. SHAP analysis identified acoustic features, HRV features, and wrist temperature as the strongest comfort signals. The framework is architecturally designed to address all four IEQ domains, though this pilot empirically validated only thermal and acoustic signals.