Weight Change Prediction for Automated Depression Diagnosis

Early detection of depression through self-diagnosis is challenging. Especially survey-based self-diagnosis can lead to incorrect results because the answer contains the subjective opinion of the respondent. To overcome this problem, the multi sensor-based method using a smartphone has been proposed as an alternative for self-diagnosis of depression. These methods diagnose depression by using various sensor data that represented typical depression symptoms from the smartphone. However, sensor data have difficulty estimating weight changes, a representative symptom of depression. In this paper, we propose a facial image-based weight change estimation method for use in diagnosing depression. In particular, we show i) simple feature engineering, designed to fast and robustly compare between two facial images using 3D face landmarks, ii) face alignment processing that makes the model accurately compare facial images which were taken from various viewpoints, and iii) a novel preprocessing with monocular depth estimation to solve scale variance according to variation of distance between the camera and face. This code is available at https://github.com/sejong-rcv/Weight-Change-Prediction.