Personality‑Based Digital Phenotyping of Psychosomatic Health Using Machine Learning

Abstract

This study aimed to investigate whether integrating personality traits with smartphone‑derived digital phenotyping indicators can predict psychosomatic health using machine learning models. A cross‑sectional study was conducted with 524 adult participants recruited from Germany. Personality traits were assessed using the Big Five Inventory (BFI‑44), psychosomatic symptoms were measured using the Patient Health Questionnaire (PHQ‑15), and psychological distress was evaluated using the Depression Anxiety Stress Scales (DASS‑21). Participants also installed a custom smartphone application that passively collected behavioral data for four weeks, including daily phone usage duration, nighttime phone inactivity, communication frequency, response latency, and mobility variability. Data analysis involved descriptive statistics, correlation analyses, and machine learning modeling using Random Forest, Support Vector Machine, Logistic Regression, and Gradient Boosting algorithms implemented in Python with scikit‑learn. Model performance was evaluated using accuracy, precision, recall, F1‑score, and AUC‑ROC metrics. The sample consisted of 524 participants (mean age = 29.4 ± 7.8 years), including 56% females and 44% males. Moderate to severe psychosomatic symptoms were identified in 31.5% of participants based on PHQ‑15 scores. Neuroticism showed significant positive correlations with psychosomatic symptoms (r = 0.46), depression (r = 0.49), anxiety (r = 0.52), and stress (r = 0.47), while conscientiousness and agreeableness showed negative correlations with psychosomatic symptoms (r = −0.28 and r = −0.21, respectively). Participants with higher psychosomatic symptom scores demonstrated longer daily smartphone usage (mean = 5.2 hours vs. 3.7 hours), reduced nighttime inactivity periods (mean = 5.8 hours vs. 7.1 hours), and lower mobility variability. Among the predictive models, Gradient Boosting achieved the best performance with an accuracy of 0.82, F1‑score of 0.80, and AUC‑ROC of 0.87, outperforming Random Forest (accuracy = 0.79), Support Vector Machine (accuracy = 0.76), and Logistic Regression (accuracy = 0.73). The findings suggest that combining personality assessments with smartphone‑based digital phenotyping can effectively predict psychosomatic health and may support early identification of psychosomatic risk through machine learning approaches.