Predicting Positive Psychological States using Machine Learning and Digital Biomarkers from Everyday Wearable Data

Wearable devices offer continuous physiological data collection, presenting new opportunities for real-world mental health monitoring. Previous research has primarily emphasized detecting stress and psychological states associated with mental illnesses, whereas predicting positive psychological states, such as self-esteem, positive affect, and meaning in life, remains underexplored. In this study, 34 participants wore research-grade smartwatches over an eight-day period, resulting in 6,528 hours of physiological data--including heart rate variability (HRV), electrodermal activity (EDA), and accelerometer-derived movement features. Psychological states were concurrently assessed using Ecological Momentary Assessment (EMA), yielding 247 observations with a total of 4,446 self-reported labels across 18 psychological states related to positive and negative affect, self-esteem, sense of meaning in life, and personal relationships. Machine learning models--including Random Forest, Long Short-Term Memory networks (LSTM), Convolutional Neural Networks (CNNs), and Transformers--were trained to predict these EMA-reported psychological states. Explainable AI techniques (SHAP and LIME) identified physiological markers of importance, while conformal prediction methods assessed model uncertainty. Results indicate that absolute prediction accuracy remains challenging; however, CNN models achieved threshold accuracies of up to 62%, with accelerometer-based features emerging as most predictive. Self-esteem-related states demonstrated the clearest physiological signatures, and among negative affect states, anger was more reliably detectable than anxiety or sadness. These findings highlight the potential of wearable-derived biomarkers for monitoring positive psychological states, despite current limitations in predictive accuracy. Future research should focus on refining feature extraction methods, enhancing model generalizability, and integrating multimodal data sources to improve real-world applicability in mental health monitoring.