Enhancing Depression Prediction via Entropy-Based Behavioral Irregularity and Feature Engineering on Smartphone Sensor Data

Depression has emerged as a major global mental health issue, yet current diagnostic methods largely rely on self-reported questionnaires (such as the PHQ-9), which has the limitation of being based on subjective responses. Against this backdrop, research on objective depression prediction using smartphone sensor data has been actively conducted. Previous research [1] revealed that place diversity entropy showed a significant correlation with depressive symptoms, suggesting that unpredictability(place-based entropy) can be utilized as a mental health indicator. However, the study[1] has limitations in that it only addressed entropy in a single dimension place diversity and did not comprehensively address reflecting the irregularity across the diverse behavioral domains that constitute daily life. Unlike the previous study that was limited to place diversity entropy alone, this study conducts a multidimensional analysis by computing entropy from multiple behavioral indicators, including time-of-day based entropy, screen usage time entropy, physical activity entropy, and light intensity change entropy. Through this, our study seeks to more comprehensively capture patterns of behavioral irregularities associated with depression and to improve prediction accuracy by simultaneously incorporating multiple entropy variables into the model.