Quantitative Assessment of rTMS Therapeutic Efficacy in Treatment-Resistant Depression Using Multi-Band EEG Analysis and Machine Learning Classifications

Background and Objectives: Major depressive disorder (MDD) affects about 350 million people worldwide and is the fourth leading cause of disease burden. Repetitive transcranial magnetic stimulation (rTMS) has become a non-invasive treatment for depression that does not respond to typical therapies. This study aimed to assess the effectiveness of rTMS for patients with treatment-resistant MDD by analyzing electroencephalogram (EEG) signals and using machine learning methods to classify neural patterns before and after treatment. Methods: A randomized clinical trial was carried out with 89 patients diagnosed with treatment-resistant MDD who received rTMS at the Zima Cognition and Brain Therapy Center over six weeks. EEG recordings were taken from 60 channels before treatment began and right after the last treatment session. Two preprocessing methods were used: frequency-domain filtering with finite impulse response filters (0.4-100 Hz) and time-frequency filtering using wavelet transform with Daubechies-10 mother wavelet. Frequency sub-bands (alpha, beta, and gamma) were extracted, and 13 features were calculated from each sub-band, including temporal features (mean absolute value, waveform length), frequency features (mean frequency, median frequency), and nonlinear features (Lyapunov exponent). Principal component analysis was used to reduce dimensions while keeping 95% of the variance. Classification was done with three algorithms: support vector machine (SVM) with multiple kernels, k-nearest neighbor (KNN), and decision tree. Results: Frequency-domain filtering showed better performance than time-frequency filtering in maintaining spectral components. Statistical t-tests revealed significant differences (p<0.05) between pre- and post-treatment classes across temporal and nonlinear features in all frequency bands. The gamma band had the highest classification accuracy. SVM with nonlinear kernels and weighted KNN achieved the best results, with accuracy exceeding 99% for gamma band features using frequency-domain filtering. Beta and alpha bands also showed the ability to distinguish effects, indicating that rTMS influences more than just gamma-band activity. The combination of principal component analysis and statistical feature selection improved both classification speed and performance. Conclusion: This study provides quantitative evidence of rTMS effectiveness for treatment-resistant depression through EEG signal analysis and machine learning classification. The method successfully identified neural patterns before and after treatment with high accuracy, suggesting possible uses for predicting treatment response and optimizing personalized therapy in clinical practice.