Um estudo sobre aprendizado de máquina multimodal para identificar a depressão em adolescentes brasileiros usando dados coletados via smartphones

Abstract

Depression affects an estimated 322 million people worldwide, according to the WHO. The first symptoms of depression usually appear between late adolescence and early adulthood. However, its highly heterogeneous clinical presentation makes diagnosis difficult in the early stages, especially in non-specialized settings. In addition, there are no laboratory tests to diagnose depression. Clinical interviews and self-reported questionnaires are the gold standard for diagnosing depression, despite the subjective bias of the assessment. In this context, mobile technologies, such as smartphones and wearable sensors, have provided new ways of collecting data on depressed behavior in domestic environments, which are not clinically observable, in addition to traditional approaches. In this study, we investigated multimodal and multitasking machine learning models to predict the current and future depressed mood state of adolescents from the Identifying Depression Early in Adolescence Risk Stratified Cohort (IDEA-RiSCo) database. We performed extensive experiments using 81 features extracted from acoustic data from the adolescents’ speech and the environment, accelerometer, and GPS collected through smartphones. Also, we compared the performance of nomothetic (generalized) and idiographic (personalized) predictive models and two data aggregations. The experiments considered the sMFQ questionnaire scores as the ground truth. We established a machine learning model that combines activity data, reported audios, and mobility using cumulative data aggregation as an important alternative to complement the diagnosis of depression in Brazilian adolescents. In addition, we highlight the attributes that most contributed to optimizing the performance of predictive models related to activity: average daily walking time; reported audios: MFCC coefficients, spectrograms, and spectral contrasts; and mobility: maximum distance from home, and location variance.