Construction and comparison of machine learning models for driving comfort assessment using multimodal physiological signals

To achieve an objective assessment of driving comfort, this study developed and systematically compared multiple machine learning models based on multimodal physiological signals. 64 individuals had their electrocardiograms (ECGs), surface electromyograms (sEMGs), muscle oxygen saturation (SmO₂), and arterial blood flow signals recorded while they drove. Using comfort scores as labels, six types of models, including support vector machine (SVM), random forest (RF), and adaptive Boosting (AdaBoost), were constructed and optimized with principal component analysis, synthetic minority oversampling technique (SMOTE), and grid search. Results showed that multimodal signal fusion significantly improved classification performance. The SVM model achieved optimal performance with the feature set containing arterial blood flow parameters, reaching an average area under the curve (AUC) of 0.97 and an accuracy of 0.90. RF models consistently exceeded 0.95 AUC with fused signals. Practical signal combinations for portable monitoring were identified: "sEMG + ECG" dual-signal combination (average AUC is 0.90) and "SmO₂ + ECG + sEMG" triple-signal combination (average AUC is 0.92). This study offers empirical support for model selection and signal setup in objective driving comfort evaluation, providing insightful information for the development of in-car health monitoring systems.