Detection of Parkinson's disease via EEG using machine learning and spectral analysis techniques DOI: 10.29188/2712-9217-2025-11-2-24-28

6

This paper addresses the critical challenge of early instrumental diagnosis of Parkinson's disease (PD). Existing clinical assessment methods are frequently subjective, necessitating the implementation of automated biomarker analysis systems. The aim of this study was to develop and validate a classification system for PD patients and healthy controls based on quantitative EEG (qEEG) analysis.

The study utilized data from the open-source UCSD Resting State EEG dataset, comprising 16 PD patients and 16 healthy controls. The efficacy of two feature extraction methods was compared: Fast Fourier Transform (FFT) and wavelet transform (using Daubechies and Morlet wavelets). A Convolutional Neural Network (CNN) was employed as the classifier. Results demonstrated a significant advantage of the FFT method for analyzing resting-state recordings, achieving a classification accuracy of 97%, with a sensitivity of 91%, specificity of 95%, and an AUC of 0.97. Wavelet analysis yielded lower accuracy (78%), which can be attributed to the quasi-stationary nature of resting-state signals, where temporal feature localization is redundant.

It was determined that the key spectral markers of the pathology include EEG slowing (increased power in theta and delta bands) and a reduction in alpha rhythm. The practical significance of this work lies in the development of a Python-based microservice (utilizing FastAPI and ONNX), enabling the integration of the model into clinical practice as a Clinical Decision Support System (CDSS).