A Clinical AI-Based System (MoodMon) for Affective Disorders: Algorithm Development and Validation

Vollständiger Abstract

Worum geht es in dieser Arbeit?

Abstract Background Psychiatry needs objective technological tools to address global staffing shortages, stigma, and other systemic challenges. An AI-based system for mood monitoring (MoodMon) was developed along with a mobile app for smartphones to detect changes in the mental state of individuals with major depressive disorder (MDD) and bipolar disorder (BD) based on acoustic features derived from speech signals. A long-term, naturalistic study of the MoodMon system represents a breakthrough in biomarker validation. Objective The aim of the study was to determine whether acoustic features are effective biomarkers of mental status changes in individuals with affective disorders and whether they are useful in the remote clinical monitoring of patients by psychiatrists. Methods To evaluate the effectiveness of AI algorithms in detecting changes in mental state based on acoustic features, data from 75 patients with BD and 25 patients with MDD over a period of 944 days were used. This makes this the longest analysis in the world covering two of the most common mental disorder diagnoses. A wealth of clinical, behavioral, and technical data were collected and used to train the MoodMon machine learning models under the supervision of human experts—experienced psychiatrists. In the first stage, the AI was trained using objective data and clinical assessments conducted by psychiatrists, including 17-item version of the Hamilton Depression Rating Scale and the Young Mania Rating Scale, as well as the Clinical Global Impression Scale. The second stage involved further refinement of the AI models using individual and population data, generating alerts when subtle changes in mental state were detected. Results In total, 243 acoustic features were extracted from the speech signals of patients and considered as input for the AI models aimed at detecting changes in mental state. The system demonstrated high performance, achieving the following sensitivity (true positive rate [TPR]) and specificity (true negative rate [TNR]) values: for both diagnoses, a TPR of 89.5% (1151/1286) and a TNR of 98.8% (30,224/30,579); for BD only, a TPR of 89.6% (866/966) and a TNR of 98.9% (22,664/22,905); and for MDD only, a TPR of 89.1% (285/320) and a TNR of 98.5% (7560/7674). Voice alerts in the MoodMon system are a key tool supporting clinical decision-making. They increase the probability of a clinical visit and exert a significant influence on the likelihood of treatment modification. Conclusions The system confirmed the presence of parameters that may serve as biomarkers of mental state changes in individuals with BD and MDD. A key clinical implication is the increased probability of prompt treatment modification following an alert, thereby supporting the primary objective underlying the development of the AI-based MoodMon system.