A Multimodal Deep Learning Based Framework for Early and Accurate Diagnosis of Depression Using Electroencephalography and Event‐Related Potential Signals
Machine learning study demonstrates up to 91% diagnostic accuracy in patients with depression, highlighting the promise of multimodal neural biomarkers.
Key Points
To develop and evaluate a multimodal deep learning framework for early, objective, and accurate diagnosis of major depressive disorder using electroencephalography and event-related potential signals.
Analyzed 3-electrode and 128-electrode EEG and ERP recordings from the public Multi-modal Open Dataset for Mental-Disorder Analysis (MODMA).
Extracted spatial features using Node2vec on electrode graphs, time-frequency features via cross wavelet transformation with AlexNet, and spectral features via power spectral density with a temporal-separable convolutional network.
Classified combined multimodal feature representations using Gated Recurrent Unit (GRU) and Transformer architectures evaluated with k-fold cross-validation.
The GRU model outperformed Transformer models across evaluation benchmarks, achieving k-fold cross-validation accuracies of 70.91%, 81.33%, 75.33%, and 91.00% across dataset configurations.
Multimodal integration of spatial, spectral, and time-frequency features achieved a peak classification accuracy of 91% for detecting major depressive disorder.