Explainable EEG-based prediction of depression therapy outcomes using local Fibonacci pattern analysis

Depression treatment outcomes remain difficult to predict for selective serotonin reuptake inhibitor (SSRI) and repetitive transcranial magnetic stimulation (rTMS) therapies, leading to trial-and-error treatment selection and delayed clinical benefit. This study develops an EEG-based prediction model using local Fibonacci pattern (LFP) features to classify treatment responders before therapy initiation. Pre-treatment EEG signals from the Mumtaz SSRI dataset and the Atieh Hospital rTMS datasets were processed using finite impulse response filtering and multi-scale principal component analysis. The LFP method extracts nonlinear temporal features directly from 19-channel EEG recordings through Fibonacci-indexed local differences and histogram encoding. Features were ranked using neighborhood component analysis and classified with a feedforward neural network under segment-level 10-fold cross-validation. The model achieved accuracies of 99.12% for SSRI, 100.00% for the small rTMS dataset, and 94.51% for the big rTMS dataset under this protocol. To address the risk of overfitting and to provide a stricter estimate of subject-level generalization, subject-wise leave-one-subject-out (LOSO) validation was also performed. Under LOSO validation, the corresponding accuracies were 61.83%, 77.93%, and 71.57%, with balanced accuracy/macro-F1 values of 59.07%/59.16%, 65.49%/66.68%, and 75.43%/71.51% for SSRI, small rTMS, and big rTMS datasets, respectively. Key discriminative channels corresponded to frontal, temporal, and parietal regions involved in emotion regulation and cognitive control. The revised findings support LFP as a computationally efficient EEG representation, but the lower LOSO results indicate that larger independent cohorts are required before any clinical-deployment claim can be made.