Explainable artificial intelligence for multimodal visual comfort prediction: Integrating environmental, physiological, and perceptual data

Visual comfort plays an essential role in shaping occupants’ visual experience and performance in daylighted sports environments. To advance its quantitative assessment, this study proposes a multimodal and interpretable modelling approach for predicting visual comfort by integrating luminance-based environmental metrics, physiological measurements, and subjective evaluations. A field experiment was conducted in a naturally daylit university gymnasium, where heart rate variability and electrodermal activity were recorded, high dynamic range imaging captured luminance distributions, and structured questionnaires collected subjective comfort responses. Six ensemble classifiers, including three tree-based models and three boosting-based models, were evaluated using a leave-one-subject-out cross-validation scheme to ensure participant-independent generalization. Among them, LightGBM achieved the highest classification accuracy of 87.4% for three-level visual comfort states. To enhance interpretability, Shapley additive explanations were applied to quantify feature contributions. The results indicate that luminance structure and subjective perceptual factors jointly shape comfort prediction, while physiological indicators provide complementary information on inter-individual variability. Model-sensitive luminance ranges were identified, including L w ≈ 2000 cd/m2, CR w ≈ 18, and L FOV ≈ 140 cd/m 2 . Overall, the proposed framework provides a data-driven means to interpret complex human–environment interactions and supports occupant-centred daylighting evaluation in large-span indoor spaces.