LEVERAGING MACHINE LEARNING FOR ENHANCED DETECTION OF BIPOLAR DISORDER: A NEW FRONTIER IN MENTAL HEALTH DIAGNOSTICS

Abstract

Mental health disorders have become a critical global concern, affecting individuals across all demographics. Early diagnosis and accurate detection of these conditions are essential for effective intervention, as delayed identification can result in severe consequences, including self-harm and fatal outcomes. This study introduces a novel approach for analyzing facial expressions using the AffectNet and 2013 Facial Emotion Recognition (FER) datasets. Unlike conventional diagnostic methods, this research develops an advanced system that compiles an extensive mental health dataset and predicts mental health conditions based on facial emotional cues.

A distinctive hybrid framework is presented, incorporating the state-of-the-art YOLOv10 object detection algorithm to recognize and classify visual indicators linked to specific mental disorders. To enhance predictive performance, the system employs an ensemble learning model that combines Convolutional Neural Networks (CNNs) with Visual Transformers (ViT), achieving an accuracy of approximately 81% in detecting mental health conditions such as depression and anxiety.

To ensure interpretability and reliability, this study integrates Gradient-weighted Class Activation Mapping (Grad-CAM) and saliency maps to identify significant  facial regions influencing model predictions. These insights enhance transparency, allowing healthcare professionals to understand the reasoning behind the system's outputs, thereby fostering confidence in the decision-making process.

A flowchart outlines the mental disorder detection pipeline, beginning with data collection and preprocessing, which then branches into training and testing datasets. Both datasets are processed through a YOLOv10 model, followed by a logic module that links to a dedicated mental disorder dataset. This dataset feeds into an ensemble model consisting of MDNet, ViT, and ResNet50. The final output is refined through Explainable AI (XAI) techniques, ensuring a transparent and interpretable prediction process.