tsta

Home / سی و دومین کنفرانس ملی و دهمین کنفرانس بین المللی مهندسی زیست پزشکی ایران

Deep Learning and Fuzzy Entropy in Parkinson's Diagnosis: a Framework Based on Task-Based EEG Signals

Authors :

Amir Hossein Tajarrod¹ Tania Hossein Khani² َAsghar Zarei³ Mousa Shamsi⁴

1- دانشکده مهندسی پزشکی، دانشگاه صنعتی سهند، تبریز، ایران 2- دانشکده مهندسی پزشکی، دانشگاه صنعتی سهند، تبریز، ایران 3- دانشکده مهندسی پزشکی، دانشگاه صنعتی سهند، تبریز، ایران 4- دانشکده مهندسی پزشکی، دانشگاه صنعتی سهند، تبریز، ایران

Keywords :

Deep learning،Parkinson’s disease،EEG،Fuzzy entropy،LSTMFCN

Abstract :

Parkinson's disease (PD) is the second most common neurodegenerative disorder worldwide, characterized by reduced dopamine levels in the central nervous system. Electroencephalography (EEG) signals have emerged as a promising tool for diagnosing PD due to their non-invasive nature, low cost, and high temporal resolution. This paper proposes a framework for diagnosing PD in healthy individuals. The proposed framework involves the extraction of fuzzy entropy from sub-bands of wavelets, combined with deep learning networks to classify EEG signals obtained under an auditory oddball paradigm. The deep learning networks used in this study include the EEG Network (EEGNet), Residual Networks within EEG (ResNetEEG), EEG Transformer, and Long Short-Term Memory Fully Convolutional Network (LSTMFCN). Four classification scenarios were explored: healthy control (CTRL) vs. PD patients off medication (PD-OFF), CTRL vs. PD patients on medication (PD-ON), PD-ON vs. PD-OFF, and a multi-class. The results indicated that the ResNetEEG network achieved the best average accuracy of 99.78% for the CTRL vs. PD-OFF classification. In contrast, the LSTMFCN network demonstrated optimal performance for the other classifications, with average accuracies of 99.81% for CTRL vs. PD-ON, 99.38% for PD-ON vs. PD-OFF, and 99.85% for the multi-class scenario. Both the EEGNet and EEG Transformer networks also showed comparable performance. Even the ROC curves for these networks showed AUC values of 1.0, further confirming the effectiveness of the implemented networks. These results emphasize the significant potential of utilizing EEG-derived features and deep learning techniques for the accurate detection of PD across various clinical scenarios.