4 - (SYM 111) Leveraging Artificial Intelligence for the Assessment of Severity of Depressive Symptoms

Background: Major Depressive Disorder (MDD) remains a significant global health concern, impacting millions worldwide. While measurement-based care is considered the gold standard, it faces challenges due to limited resources and a shortage of mental health specialists, delaying optimal treatment. With approximately 3.4 billion smartphones globally and over two-thirds of the U.S. population owning at least one, these devices have become integral to daily life. Similarly, one in five Americans uses a fitness tracker, and it is estimated that by 2028 there will be 87.2 million users of commercial wristbands or health trackers worldwide. Smartphones and wristband sensors can record biological and behavioral markers of depression including activity level, physiological health, and socialization level. Moreover, artificial intelligence-based models integrating multimodal sensor data can be developed to detect mood changes and deliver digital health interventions at the most opportune time. Despite promising preliminary data, current algorithms need to be further improved before being deployed in clinical practice. This study aims to develop a machine-learning (ML) model assessing depression severity changes by leveraging data from smartphone and wristband sensors, including electrodermal activity (EDA) measurements.

Method: Participants include 98 individuals with MDD who were asked to install an app on their phones tracking socialization and activity levels, wear two wristband sensors daily, complete daily surveys, and undergo weekly clinician-based assessments of depression for 12 weeks.

Results: At the time of this submission, 61 (65%) participants have completed the study, and 5 are active. Participants wore the wristbands for 75% of the days and completed 88% of the daily surveys, and 98% of the weekly clinician-based assessments. Several machine learning methods have been evaluated on the data, with the best performance currently obtained by mixed-effects random forest models. These outperform group-level baselines while currently matching personalized baselines. Physiological features including EDA, and location signals were the most important for the estimation of depression severity.

Conclusion: Monitoring patients with MDD through the combined use of smartphones and wrist sensors may be feasible. This approach holds significant promise for estimating changes in depressive symptom severity over time, offering potential breakthroughs in depression management and delivering just-in-time interventions.