The Deep Learning Method Differentiates Patients with Bipolar Disorder from Controls with High Accuracy Using EEG...

Area of Science:

• Neuroscience
• Computational Psychiatry
• Medical Technology

Background:

• Bipolar disorder (BD) diagnosis is challenging due to overlapping symptoms with other mood disorders.
• Advanced techniques like deep learning (DL) are being explored for precise BD detection.
• Previous studies have not utilized DL techniques with electroencephalogram (EEG) signals for BD analysis.

Purpose of the Study:

• To investigate the efficacy of deep learning algorithms applied to EEG signals for diagnosing bipolar disorder.
• To compare the performance of 1D-CNN+LSTM and 2D-CNN models in classifying BD patients and controls.
• To identify specific EEG electrode activities and brain regions associated with bipolar disorder.

Main Methods:

• Collected EEG signals from 169 individuals with BD and 45 controls.
• Applied artifact cleaning and processed EEG data using 1D-CNN combined with LSTM and 2D-CNN models.
• Utilized Class Activation Maps (CAMs) to visualize and identify discriminative brain regions.

Main Results:

• The 2D-CNN model achieved an overall accuracy of 95.91% in identifying BD patients.
• The 1D-CNN+LSTM model demonstrated an overall accuracy of 93%.
• Analysis revealed that activities at F4, C3, F7, and F8 electrodes are predominant features for BD detection, with CAMs indicating prefrontal changes.

Conclusions:

• This study represents the first use of EEG-based DL analysis for bipolar disorder.
• The findings suggest that DL algorithms analyzing raw EEG data can effectively differentiate individuals with BD from healthy controls.
• CAM analysis highlights the significance of prefrontal cortex activity in the EEG data of BD patients.