Identifying Network State-Based Parkinson's Disease Subtypes Using Clustering and Support Vector Machine Models

BENEDICTOR ALEXANDER NGUCHU ^1,2,3 Han Yifei ^2,3 Yanming Wang ¹ Peter Shaw ^2*

• ¹ University of Science and Technology of China, Hefei, China
• ² Oujiang Lab, Wenzhou, Zhejiang Province, China
• ³ Wenzhou Medical University, Wenzhou, Zhejiang Province, China

Parkinson's disease (PD) heterogeneity poses challenges to the development of personalized treatment. Here, we employ K-means and hierarchical clustering algorithms on data from the Parkinson's Progression Markers Initiative (PPMI) to identify network-specific patterns that describe PD subtypes using the optimal number of brain features. Specifically, the features were the gray matter volume and dopaminergic features of the neostriatum, i.e., the caudate, putamen, and anterior putamen. We use machine learning (ML) algorithms, including Random Forest, Logistic Regression, and Support Vector Machine, to evaluate the diagnostic power of the brain features and network patterns in differentiating the PD subtypes and distinguishing PD from HC. Finally, we assessed whether PD subtypes described through network-specific patterns are APOE genotype-dependent. Using data from 2396 subjects, we show that PD (n=2037) is highly associated with APOE ε2/ε4. Our findings reveal a significant DAT deficit in the left/right structures of the caudate, putamen, and anterior putamen in subjects with PD compared to subjects with SWEDD(n=137) or HC(n=222), and that APOE ε2/ε4 may accelerate DAT deficits and brain alterations in both PD and SWEDD. Furthermore, clinical symptoms of PD in SWEDD subjects, hardly validated by DAT scan data, can be explained by APOE-genotype variations and other brain features beyond DAT. We show the existence of three networks states, with the first network state describing the HC subjects, while the remaining two network states describing the two PD subtypesone network state typified by a mildly sparsely connected network and the other network state characterized by a more intensified sparsity. We show that the two subtypes of PD are characterized by distinctly different levels of total gray matter volume and DAT deficit. ML models show that brain features and network patterns can serve as reliable biomarkers for PD and its subtypes, with fine-tuned SVM model showing better performance (100% AUC, 99.3% accuracy, 0.993 F1). Our findings suggest that, while PD is generally associated with DAT deficit, it exhibits intrinsic heterogeneity across individuals, possibly stemming from genetic factors. Such heterogeneity can be characterized by ML models and optimally mapped into network states, providing new insights to consider when developing personalized drugs.