The Sound of Mental Health: Audio Features as Indicators of Depression and Anxiety Symptoms in Behavioral Treatment

Measurement-based care (MBC) is critical for assessing clients' mental health in therapy and guiding evidence-based treatment. However, self-report assessments have low completion rates, necessitating less intrusive methods to obtain objective data. Audio features present a promising avenue for passive prediction of clients' mental health status during therapy sessions. This study investigated the use of audio features extracted from therapy sessions to predict depression and anxiety symptoms in a diverse dataset of 2,348 therapy sessions delivered in real-world behavioral health programs. Leveraging machine learning models and a range of audio features, the algorithms predicted whether clients had either mild or moderate-severe depression and anxiety. Our models achieved high accuracy, precision, and recall rates (for anxiety: accuracy of 75.1%, precision of 81.3%, and recall of 61.3% ; for depression, prediction accuracy of 73.2%, precision of 71.4% , and recall of 68.1%). Despite limitations such as limited demographic data and agnosticism of contextual factors, our approach represents a significant step towards improving MBC in mental health settings. Audio analysis emerges as a cost-effective and non-intrusive method for passive prediction, thereby improving therapy monitoring and enhancing patient outcomes.