Classifying Depression in Imbalanced Datasets Using an Autoencoder- Based Anomaly Detection Approach

Depression is the most prevalent mental health ailment in the United States, affecting 15% of the population. Untreated depression can significantly decrease quality of life, physical health, and has significant economic and societal costs. The traditional method of diagnosing depression requires the patient to respond to medical questionnaires and is subjective. Passive methods to autonomously detect depression are desirable. Prior work on smartphone sensing of depression has utilized machine learning classification of smartphone sensor data. However, as with many ailments, the percentage of afflicted users in most populations is small compared with those unaffected, leading to severe class imbalance. In this work, we explore anomaly detection methods as a method for mitigating class imbalance for depression detection. Our approach adopts a multi-stage machine learning pipeline. First, using autoen-coders, we project the mobility features of the majority class (undepressed users). Thereafter, the trained autoencoder then classifies a test set of users as either depressed (anomalous) or not depressed (inliers) using a One Class SVM algorithm. Our method, when applied to the real-world StudentLife data set shows that even with an extremely imbalanced dataset, our method is able to detect individuals with depression symptoms with an AUC-ROC of 0.92, significantly outperforming traditional machine learning classification approaches.