Um modelo de classificação híbrido para identificação de estudantes universitários com depressão

Abstract

Depression is a serious mental health condition that is more prevalent among college students, compared to the general population. Moreover, it is a consensus that early identification is crucial for effective mental health treatment. Recent advances in Machine Learning (ML) have explored Digital Phenotyping (DP) data for depression detection, and Natural Language Processing (NLP) applied to social network data have been used to detect depressive symptoms, achieving accuracy rates of 70–80%. Despite these advances, each data modality presents limitations. Sensor data captures physiological behaviors but is less effective at identifying depressive mood, while textual data better reflects depression related thoughts. Although achieving promising results, there are still few studies investigating the combination of these modalities. In this context, this study developed a hybrid model integrating mobile sensor data and NLP-extracted textual features to identify Possible Depressive Profile (PDP) in college students. To leverage the experiments on the combination of these data modalities, two approaches were tested: feature-level integration and late fusion of predictions. The hybrid model did not surpass the baseline model, achieving 73% accuracy, slightly below the sensor-only baseline (79%), while late fusion reached 70% accuracy and 77% recall for true positives identification. The identified limitations relate to the scarcity of textual data and challenges in NLP feature extraction, which motivated further exploration of the fusion strategy.