AI-driven Model for Detection and Classification of Major Depressive Disorder using EEG data

Abstract

Problem: Major Depressive Disorder (MDD) is a serious mental health condition affecting millions worldwide, which leads to reduced quality of life, disability, and an increased risk of suicide. Traditional MDD diagnostic methods mainly rely on clinical interviews, medical history, family history, and self-reported symptoms, which are subjective and inconsistent across based on the healthcare provider. These limitations emphasise the need for objective diagnostic tools. Recent studies have shown that electroencephalography (EEG) signals have the potential to detect MDD based on their signal patterns. This research proposes an AI-driven model to improve MDD diagnosis accuracy using EEG data, addressing existing gaps in current detection methodologies. Methodology: This study proposes an AI-driven approach to improve MDD diagnosis using EEG data. It utilizes deep learning techniques and advanced feature engineering techniques to analyze spatial and temporal features of EEG signals. The pipeline includes data preprocessing, feature extraction, model training, and performance evaluation. Various EEG biomarkers are extracted, including spectral entropy, fractal dimension, and power spectral density. Deep learning models such as CNN, LSTM, and BiLSTM are implemented to classify EEG data. Strategies like data augmentation and regularization are applied to enhance model robustness and mitigate overfitting. Initial Results: Promising results were obtained from the first prototype evaluation, with the CNN model demonstrating the highest performance across key performance metrics. The CNN model achieved a training accuracy of 93.44% and an ROC-AUC score of 0.9850, highlighting its strong ability to differentiate between MDD and healthy EEG signals. These results confirm the feasibility of an AI-driven approach for improving MDD diagnosis through EEG analysis. Moving forward, the final model development will focus on utilizing transformers while refining CNN architectures, incorporating larger datasets, and further optimizing feature extraction techniques to enhance generalizability and clinical applicability.