Identifying network state-based Parkinson's disease subtypes using clustering and support vector machine models.

Area of Science:

• Neuroscience
• Genetics
• Machine Learning

Background:

• Parkinson's disease (PD) heterogeneity complicates the development of effective therapeutic targets.
• Identifying distinct PD subtypes is crucial for advancing personalized medicine.

Purpose of the Study:

• To identify network-specific patterns characterizing PD subtypes using clustering algorithms.
• To evaluate the diagnostic power of brain features and network patterns in differentiating PD subtypes.
• To investigate the association between PD subtypes and APOE genotype.

Main Methods:

• Applied K-means and hierarchical clustering to Parkinson's Progression Markers Initiative (PPMI) data.
• Utilized gray matter volume and dopaminergic features of the neostriatum (caudate, putamen, anterior putamen).
• Employed machine learning (ML) algorithms (Random Forest, Logistic Regression, SVM) for classification and biomarker evaluation.

Main Results:

• Identified three network states: one for healthy controls (HC) and two distinct PD subtypes.
• Found significant dopaminergic deficit (DAT) in PD patients, accelerated by APOE ε2/ε4.
• ML models, particularly SVM, achieved high accuracy (99.3%) in classifying PD subtypes using brain features and network patterns.

Conclusions:

• PD exhibits intrinsic heterogeneity influenced by genetic factors, particularly APOE genotype.
• Network states and ML models can characterize PD subtypes, offering insights for personalized drug development.
• Distinct PD subtypes show varying levels of gray matter volume and DAT deficits.