EARLY DETECTION OF PARKINSON’S DISEASE USING NKM-SSM MODEL

Abstract

Parkinson's Disease (PD) is the second most common neurodegenerative disease across the globe, impacting more than 10 million people by 2024. At the time when motor symptoms manifest themselves, 50 to 80 percent of dopaminergic cells in the Substantia Nigra pars compacta (SNpc) are irreversibly damaged – "presymptomatic gap" which justifies the necessity for development of biomarker-based artificial intelligence (AI) classifiers. This paper suggests a novel hybrid computation model, NeuroKAN-Mamba, which incorporates Kolmogorov-Arnold Networks (KAN), Mamba State Space Models (SSM) and Random Forest (RF) calibrator to perform a binary classification task on PD vs. Healthy Controls (HC) from multi-modal MRI neuroimaging biomarkers on the PPMI database. On the 4,000 samples from the PPMI database (2,400 PD/1,600 HC) with 20 MRI-based features, NeuroKAN-Mamba reaches 96.25% accuracy, 98.31% AUC-ROC, 96.88% F1-Score, 96.88% Sensitivity and 95.31% Specificity – significantly outperforming all 9 comparison models. Proposed pipeline includes DICOM-based raw neuroimaging data processing, 7-stage pre-processing, learning and evaluation of hybrid model and generation of explanations by means of extracting symbolic formula and calculating SHAP values. The presented model demonstrates clinically plausible biologically-driven results.