Abstract

The pressing need for Artificial Intelligence (AI) applications in healthcare is evident, particularly in the context of depression prediction. Literature underscores the significance of visual cues as crucial indicators of depression. Primary objective of this work is to design a complete machine learning pipeline for more accurate depression prediction, which includes several stages like: data collection stage, feature extraction stage, feature selection stage, classification stage, and performance evaluation stage. Data collection involved video recording of participants (n = 219) while conducting emotion elicitation (triggering emotions by showing photos/videos) to depressed and non-depressed subjects. Then, numerous visual features like geometrical features and facial action unit features were extracted. Filter and Wrapper Feature Selection (FS) methods were used to extract the optimal feature set from high-dimensional visual features. In the Filter method, experiments are conducted using three strategies: quasi-constant strategy, mutual information gain, and linear discriminant analysis. In the wrapper method, experiments are conducted using three strategies: forward selection, backward elimination, and recursive feature elimination. Accuracy for the classification of non-depressed or depressed subjects was used as the performance metric. Obtained results with an accuracy of 85.6% show that the backward elimination approach (even though only ten features were selected) outperformed other experiments conducted in current work and also with the state-of-the-art methods. In addition to this, our method is also applied to publicly available benchmarking dataset to show its effectiveness on diverse dataset. These findings demonstrate the applicability of visual features using filter and wrapper feature selection method is reliable in depression prediction. Hence implications extend to a potential application in mental health assessment.