Prediction of antidepressant responses to non-invasive brain stimulation using frontal electroencephalogram signals: Cross-dataset comparisons and validation

Cheng-Ta Li,Chi-Sheng Chen,Chih-Ming Cheng,Chung-Ping Chen,Jen-Ping Chen,Mu-Hong Chen,Ya-Mei Bai,Shih-Jen Tsai
DOI: https://doi.org/10.1016/j.jad.2023.08.059
2023-12-15
Abstract:Background: 10-Hz repetitive transcranial magnetic stimulation(rTMS) and intermittent theta-burst stimulation(iTBS) over left prefrontal cortex are FDA-approved, effective options for treatment-resistant depression (TRD). Optimal prediction models for iTBS and rTMS remain elusive. Therefore, our primary objective was to compare prediction accuracy between classification by frontal theta activity alone and machine learning(ML) models by linear and non-linear frontal signals. The second objective was to study an optimal ML model for predicting responses to rTMS and iTBS. Methods: Two rTMS and iTBS datasets (n = 163) were used: one randomized controlled trial dataset (RCTD; n = 96) and one outpatient dataset (OPD; n = 67). Frontal theta and non-linear EEG features that reflect trend, stability, and complexity were extracted. Pretreatment frontal EEG and ML algorithms, including classical support vector machine(SVM), random forest(RF), XGBoost, and CatBoost, were analyzed. Responses were defined as ≥50 % depression improvement after treatment. Response rates between those with and without pretreatment prediction in another independent outpatient cohort (n = 208) were compared. Results: Prediction accuracy using combined EEG features by SVM was better than frontal theta by logistic regression. The accuracy for OPD patients significantly dropped using the RCTD-trained SVM model. Modern ML models, especially RF (rTMS = 83.3 %, iTBS = 88.9 %, p-value(ACC > NIR) < 0.05 for iTBS), performed significantly above chance and had higher accuracy than SVM using both selected features (p < 0.05, FDR corrected for multiple comparisons) or all EEG features. Response rates among those receiving prediction before treatment were significantly higher than those without prediction (p = 0.035). Conclusion: The first study combining linear and non-linear EEG features could accurately predict responses to left PFC iTBS. The bootstraps-based ML model (i.e., RF) had the best predictive accuracy for rTMS and iTBS.
What problem does this paper attempt to address?