Early Detection Depression Based On Action Unit and Eye Gaze Features Using a Multi-Input CNN-WoPL Framework

Abstract

Depression is a common mental disorder with significant life impact, including a high risk of suicide. Patients with depression attempt suicide five times more often than the general population. Self-reporting, subjective judgement and clinician expertise influence conventional diagnostic methods. For timely intervention and effective treatment, early and accurate diagnosis of depression is essential. This study proposes a framework called Multi-Input CNN-WoPL, a CNN-based method without a pooling layer that combines two features - action units and gaze - to improve accuracy and robustness in automatic depression detection. Pooling layer reduces spatial dimension of feature map, resulting in loss of information related to expression data, affecting depression detection result. The performance of the proposed method results in an accuracy of 0.994 and F1 score = 0.993, the F1 score value close to 1.0 indicates that the proposed method has good precision, recall and performance.