Predicting Psychological States using Machine Learning and Digital Biomarkers from wearable data

Background: Prior research has shown that digital biomarkers can predict mental states associated with depression and anxiety. However, their ability to detect positive psychological states—such as positive emotions, meaning in life, self-esteem, and accomplishment—remains less explored. This pilot study aims to address this gap by investigating whether smartwatch-derived digital biomarkers can predict daily psychological states aligned with the PERMA framework (Positive Emotion, Engagement, Relationships, Meaning, and Accomplishment), providing a more comprehensive understanding of mental well-being.

Methods: Thirty-four healthy volunteers wore research-grade smartwatches for eight consecutive days, enabling the passive collection of physiological signals, including heart rate variability (HRV), electrodermal activity (EDA), and actigraphy. Participants also completed end-of-day ecological momentary assessments measuring 18 PERMA-related psychological states. After data preprocessing, we extracted 177 features from actigraphy, HRV, and EDA. Principal component analysis (PCA) was applied to identify relevant feature subsets. Machine learning models—including Random Forest, Long Short-Term Memory (LSTM) networks, and Convolutional Neural Networks (CNNs)—were employed to predict psychological states from the complete feature set and the reduced subset. Classification performance was evaluated using three metrics: (1) prediction accuracy, (2) threshold accuracy (correct if within ±1 of the true label), and (3) top3 accuracy (true label among the top three predictions). Additionally, interpretable AI techniques identified
the most influential features, while conformal prediction assessed prediction uncertainties.

Findings: While overall classification accuracy for PERMA-related psychological states was low across all models, the CNN-based model achieved a threshold accuracy of 62%, indicating that all predictions fell within ±1 of the true label. The LSTM model, when applied to PCA-extracted features, achieved the highest threshold accuracy of 60%. Performance analysis of individual psychological states revealed that self-esteem constructs had the highest prediction accuracy for both positive and negative emotions. Anger-related negative emotions also demonstrated strong accuracy, likely due to well-documented physiological responses associated with anger in prior research. Constructs related to meaning in life and personal relationships exhibited moderate accuracy, reflecting their complex cognitive and emotional components. Our findings demonstrated moderate explainability of daily psychological states, with actigraphy consistently emerging as the strongest predictor across different algorithms. These patterns held true even after accounting for physiological signals from HRV and EDA, highlighting the contribution of movement data. Movement patterns explained substantial variations in participants’ positive mental states, emphasizing the role of physical activity.

Discussion: These findings suggest that wearable-derived biomarkers can effectively track fluctuations in positive psychological states in real-world settings. By integrating interpretable AI techniques and uncertainty quantification, wearable technology holds promise for personalized mental health monitoring and timely interventions. Future research should refine feature extraction methods and validate these results in larger, more diverse populations to enhance generalizability.