Channel Selected Stratified Nested Cross Validation for Clinically Relevant EEG Based Parkinsons Disease Detection

TL;DR

This paper addresses the challenge of developing reliable EEG-based Parkinson's disease detectors by confronting common methodological flaws such as patient leakage and dataset heterogeneity. It introduces a model-agnostic nested cross-validation framework with three safeguards: outer-loop patient-level stratification, inner-loop channel selection, and multi-layered windowing to harmonize heterogeneous recordings. Applied to three diverse EEG datasets, a CNN using this framework achieves an accuracy of 0.806 and demonstrates robust generalization in held-out population tests, highlighting the framework’s potential for clinical translation. The approach also emphasizes interpretability through band-aligned representations and Grad-CAM analyses, providing a reproducible blueprint for biomarker discovery in EEG and other biomedical signals.

Abstract

The early detection of Parkinsons disease remains a critical challenge in clinical neuroscience, with electroencephalography offering a noninvasive and scalable pathway toward population level screening. While machine learning has shown promise in this domain, many reported results suffer from methodological flaws, most notably patient level data leakage, inflating performance estimates and limiting clinical translation. To address these modeling pitfalls, we propose a unified evaluation framework grounded in nested cross validation and incorporating three complementary safeguards: (i) patient level stratification to eliminate subject overlap and ensure unbiased generalization, (ii) multi layered windowing to harmonize heterogeneous EEG recordings while preserving temporal dynamics, and (iii) inner loop channel selection to enable principled feature reduction without information leakage. Applied across three independent datasets with a heterogeneous number of channels, a convolutional neural network trained under this framework achieved 80.6% accuracy and demonstrated state of the art performance under held out population block testing, comparable to other methods in the literature. This performance underscores the necessity of nested cross validation as a safeguard against bias and as a principled means of selecting the most relevant information for patient level decisions, providing a reproducible foundation that can extend to other biomedical signal analysis domains.

Figures (3)

• Figure 1: The nested cross‑validation architecture addresses the challenge of high‑dimensional EEG data by starting with all channels and systematically reducing to the most informative. The inner loop selects channels contributing most to inference, while the outer loop validates this choice without bias. This principled “many‑to‑best” reduction is dataset‑, channel‑, and model‑agnostic, yielding results that are more efficient, generalizable, and interpretable.
• Figure 2: Suggested Workflow for EEG Channel Selection Modeling: Subjects are represented by raw EEG time‑series, which are transformed into spectrograms via short‑time Fourier transform. An inner cross‑validation loop trains a model to estimate per‑channel performance, retaining only the most informative electrodes. These selected channels are reintegrated into the outer training and validation framework, where the final model is trained. Predictions are then aggregated into a single patient‑level probability, providing a concise outcome measure.
• Figure 3: Grad‑CAM overlays for Patient 123 highlight consistent band‑level focus. Each row shows one EEG channel (32, 41, 54, 55), progressing from spectrogram (yellow–green) to Grad‑CAM heatmap (red) to overlay. The model emphasizes recurrent theta activity (4–8 Hz), with secondary attention in alpha (8–12 Hz) and occasional beta (12–30 Hz). These theta–alpha patterns broaden and narrow rhythmically, producing a pulsing appearance distinct from noise or artifacts. While edge effects (e.g., aliasing, zero‑padding) sometimes attract attention, the dominant focus remains on physiologically plausible oscillations, underscoring both the strengths and limits of convolutional interpretability