Analysis, Identification and Prediction of Parkinson’s disease sub-types and progression through Machine Learning - Ashwin Ram

Aug 2, 2026

Upload by Asher

Parkinson’s disease is a common neurodegenerative disorder whose causes remain largely unknown, and whose patient symptom trajectories show substantial variability. Using the Parkinson’s Progression Markers Initiative (PPMI) longitudinal dataset, the work reviews prior machine learning efforts and motivates progression-focused modeling beyond standard diagnosis. It analyzes the distribution of supervised versus unsupervised methods, highlights limited integration, and proposes a combined clustering-and-prediction strategy using extended longitudinal data.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages9

Reading Time23 min

AccessFree to download

Introduction

Machine learning study landscape using PPMI

Supervised vs unsupervised approaches for subtyping

Motivation for combined supervised–unsupervised modeling

Methodology and study objective

Analysis, Identification and Prediction of Parkinson’s disease sub-types and progression through Machine Learning Ashwin Ram ([ashwin.ram@utexas.edu](ashwin.ram@utexas.edu)) I. INTRODUCTION Parkinson’s disease (PD) is the second most common type of neurodegenerative disorder, affecting up to 4% of individuals aged 80 and above (Pringsheim et al., 2014) . While damage to midbrain dopaminergic neurons is a well-known cause of PD, most cases are of unknown origin, and very little is currently understood about variations in patient trajectories. To address this, the Parkinson's Progression Markers Initiative (PPMI; Marek et al., 2011, 2018) has been collecting longitudinal data from various patient cohorts and data modalities since 2010 . The data collected includes clinical measures, brain imaging, gene expression levels, protein abundance, genomic variant status, as well as data from sensors and wearable devices, with the aim of identifying biomarkers to aid in the development of new interventions for PD. The PPMI initiative provides a comprehensive and well-annotated dataset for analyzing the progression of several biological variables in conjunction with clinical measures of disease severity. There have been over 110 machine learning studies that used the PPMI database. And, out of the 110 studies on machine learning reviewed, almost 90%(97 studies) used supervised models, while only 19 studies utilized unsupervised methods. Of the studies that reported supervised methods, 55 studies attempted to predict Parkinson’s diagnosis, which is not only the majority of supervised studies but also of all machine learning studies reviewed. Early diagnosis of Parkinson's disease is important, but there are already established clinical tools for diagnosing PD. Thus, machine learning algorithms focused on diagnosis will have little impact on the main goal ofthe PPMI study, which is to understand the variability of patient symptoms and their trajectories over time. Only 26 studies made use of the longitudinal structure of PPMI data, predicting future symptoms from a baseline, which is known as “progression prediction.” Since understanding the biological variables associated with heterogeneity in patient trajectories is a core goal of the PPMI project, these progression prediction papers present much more significance. Thirteen additional studies focused on predicting symptoms measured at the same time asthe predictive features, while five studies focused on predicting neuroimaging results or medication state, rather than symptoms or diagnosis. One study reported predictive accuracy for both diagnosis and symptom levels. The most commonly used supervised models were linear regression, support vector machines, and random forests and/or gradient boosting methods. A smaller number of studies used unsupervised learning to generate latent variables or clusters to capture patient variability. Of the 19 studies that used unsupervised methods, 11 were concerned with subtyping Parkinson’s patients using clustering models. Eleven used latent variable or dimensionality reduction methods with continuous latent factors, and three used both subtyping and continuous latent variables. Surprisingly, only six papers combined supervised and unsupervised methods, despite the stated focus of much PD research on finding subtypes that can predict differential progression across groups of Parkinson’s patients. To discover latent sub-groups of patients and find predictors of future sub-group membership, it is likely that supervised and unsupervised models will need to be integrated. Notably, three papers combined clustering of patients into subtypes with prediction of current or future symptoms. For instance, Faghri et al. (2018) used a combination of Nonnegative Matrix Factorization (NMF) and Gaussian Mixture Models (GMMs) to cluster patients into subtypes and random forests for supervised prediction of symptom levels four years later. Valmarska et a

FAQs

Why is progression prediction more important than diagnosis prediction for Parkinson’s disease in this context?

Diagnosis-focused models add limited value because clinical tools for diagnosing Parkinson’s disease already exist. The PPMI effort targets understanding variability in symptoms and their trajectories over time, which progression prediction addresses.

What does the PPMI dataset provide for Parkinson’s progression research?

PPMI provides longitudinal data including clinical measures, brain imaging, gene expression, protein abundance, genomic variant status, and sensor or wearable device signals. The goal is to identify biomarkers that support new interventions.

How do prior studies typically use supervised and unsupervised machine learning for Parkinson’s subtyping?

Most reviewed studies use supervised models, often to predict diagnosis or symptoms. Fewer studies use unsupervised methods for latent variables or patient clusters, and only a small number integrate both approaches to improve subtype discovery and prediction.

Aug 2, 2026

Upload by Asher