Skip to main content

Subtyping Parkinson’s Disease with Deep Learning Models (2016 PPMI Data Challenge Winner)

Study Rationale:
As Parkinson's disease (PD) is highly heterogeneous, identifying coherent PD subtypes is crucial for understanding the underlying mechanism of the disease and designing and testing appropriate therapies. Existing studies mostly focus on discovering static patterns among PD patients with vector-based models. However, the disease progression patterns of PD patients could also be very important as they can tell us how disease status will evolve over time. The rationale behind this study is to utilize the comprehensive data provided by the Parkinson's Progression Markers Iniatitive (PPMI) to discover the cohesive progressive subtypes of PD patients.

Summary of Results:
We identified three progressive subtypes for idiopathic PD patients. The first is characterized by function decay of both cognitive and motor abilities. Patients in the second subtype have significant function decay of cognitive abilities (e.g., cannot memorize words well), but their motor abilities are normal. The third subtype has normal cognitive abilities but severe motor problems, especially right-hand tremor.

Description of Methods Used:
Each subtype is a coherent group of patients with similar PD progression patterns. From computational perspective, this is a clustering problem, where the key is to evaluate the progression similarities between pairwise patients. We first linked the patient records in all 35 data tables and then established a progression path for each patient using a neural network model called long short-term memory. We then adopted a method called dynamic time warping to evaluate the pairwise similarities among those learned progression paths. With the similarity measure, we further embedded every patient into a two-dimensional space with an algorithm called t-SNE so that the pairwise patient similarities can be preserved. We then used a simple clustering algorithm called k-means to get the patient groups.

Impact of Results:
This work identifies the progressive subtypes of PD patients. The detected subtypes characterized the three major PD progression pathways, which can help the community to understand how PD progresses over time. We hope that these subtypes will help us develop effective early detection methods and interventions.

Next Steps for Further Analysis/Development:
The next step would be to pair with clinical PD experts to 1) finetune the process to make the discovered subtypes more precise and reliable; and 2) perform more in-depth and complete associated feature analysis across clinical, biological, imaging and genetic assessments to characterize each subtype more comprehensively.


  • Fei Wang, PhD

    New York, NY United States

  • Jian Liang, MS/PhD candidate

    Beijing China

  • Cao Xiao, MS/PhD candidate

    Seattle, WA United States

Discover More Grants

Within the Same Program

Within the Same Funding Year

We use cookies to ensure that you get the best experience. By continuing to use this website, you indicate that you have read our Terms of Service and Privacy Policy.