When more is not better: Predicting human emotional states with brain and bodily biosensors

Recent technological advances have made wearable biosensors increasingly accessible, affordable, and sensitive, driving their widespread adoption for physical health applications. Their success has sparked growing interest in leveraging biosensors for monitoring mental health using machine learning techniques. However, many studies, particularly those employing deep learning, prioritize predictive accuracy through multimodal signal fusion, often at the cost of interpretability and a deeper understanding of individual sensor contributions. To address this gap, our study performed a systematic approach to identify which biosensors and features are most informative for emotional state prediction. We recorded eight distinct physiological signals, including underexplored sensors such as electrogastrography (EGG) and trans-radial-ulnar electrical bioimpedance velocimetry(TRUEV), from 39 participants as they listened to music and reported their valence and arousal levels. We first isolated the most informative spectral features from 128-channel electroencephalography (EEG) data. Next, we evaluated each bodily signal independently and then fused the top-performing bodily signals with pre-selected EEG features using a K-nearest neighbor classifier and decision-level fusion. Our results revealed that features from the anterior EEG regions were the most powerful individual predictors of emotion states. Furthermore, incorporating top-performing bodily signals did not significantly improve upon the EEG-only model performance. Our findings provide empirical insights into the crucial biosensors for emotional state prediction, thereby informing the development of parsimonious wearable devices optimized for mental health monitoring in healthcare applications.