3DCNN-SL framework for the diagnosis of Parkinson's disease using frequency-preserving 3D EEG tensors

Parkinson's disease (PD) still requires scalable and non-invasive diagnostic biomarkers that support the decoding of medication-responsive neural activity. This study introduces 3DCNN-SL, a hybrid 3D CNN + Stacked LSTM framework designed to assess frequency-preserving 3D EEG tensors generated from Morlet wavelet transforms and azimuthal equidistant projections of auditory oddball responses, while preserving band-specific resolution for intercomparison at single-band and multi-band scales either separately or together. The model was applied and validated on the PRED + CT dataset of 25 PD patients measured under ON and OFF medication conditions and 25 age-matched controls, achieving trial-level accuracy of 99.1% and subject-level accuracy of 94.0% under subject-independent group cross-validation with permutation testing that confirmed statistical significance (p < 0.001). Analysis of learned convolutional features revealed frontoparietal power asymmetries in beta and gamma domains in PD patients in the OFF state, which partially normalized with levodopa administration. These findings suggest that the model captures medication-responsive neural patterns which were supported by permutation testing and were unlikely to arise from random label associations. The multi-band 3D EEG tensor analysis, combined with task-evoked EEG and learned-feature statistical analysis, provides a data-driven framework for exploring EEG-derived features associated with Parkinson's disease within the studied dataset and medication-related neural modulation.