Abstract
Many diseases display heterogeneity in clinical features and their progression, indicative of the existence of disease subtypes. Extracting patterns of disease variable progression for subtypes has tremendous application in medicine, for example, in early prognosis and personalized medical therapy. This work presents a novel, data-driven, network-based Trajectory Clustering (TC) algorithm for identifying Parkinson's subtypes based on disease trajectory. Modeling patient-variable interactions as a bipartite network, TC first extracts communities of co-expressing disease variables at different stages of progression. Then, it identifies Parkinson's subtypes by clustering similar patient trajectories that are characterized by severity of disease variables through a multi-layer network. Determination of trajectory similarity accounts for direct overlaps between trajectories as well as second-order similarities, i.e., common overlap with a third set of trajectories. This work clusters trajectories across two types of layers: (a) temporal, and (b) ranges of independent outcome variable (representative of disease severity), both of which yield four distinct subtypes. The former subtypes exhibit differences in progression of disease domains (Cognitive, Mental Health etc.), whereas the latter subtypes exhibit different degrees of progression, i.e., some remain mild, whereas others show significant deterioration after 5 years. The TC approach is validated through statistical analyses and consistency of the identified subtypes with medical literature. This generalizable and robust method can easily be extended to other progressive multi-variate disease datasets, and can effectively assist in targeted subtype-specific treatment in the field of personalized medicine.
Introduction: Addressing Disease Heterogeneity
Parkinson’s Disease (PD) is notoriously heterogeneous; two patients can have vastly different symptoms and rates of progression. Identifying "subtypes" is critical for personalized medicine and early prognosis. However, most existing methods rely on "baseline" data—a single snapshot in time—which ignores the crucial temporal evolution of the disease.
While deep learning has been applied to this problem, these models often lack interpretability, making it hard for clinicians to understand why a patient was categorized into a specific group. The challenge is to create a data-driven method that captures the "trajectory" of multiple symptoms simultaneously while remaining transparent and grounded in clinical reality.
Methodology: The Stacked Multi-layer Network
The proposed Stacked Multi-layer Network approach consists of three main phases:
- Bipartite Variable-Community Detection: For every "layer" (either a time point or a range of disease severity), the algorithm builds a bipartite network where one set of nodes represents patients and the other represents clinical variables (symptoms). Using Louvain community detection, the algorithm identifies "variable-communities"—groups of symptoms that tend to be severe in the same patients at that specific stage.
- Trajectory Definition: Each patient’s journey is mapped as a path through these variable-communities across the layers.
- Second-Order Trajectory Clustering (TC): This is the core innovation. Instead of just looking at direct overlaps between patient paths, the TC algorithm calculates a "node closeness" metric. It accounts for second-order similarities: if Patient A and Patient B don't overlap directly but both overlap significantly with Patient C, they are considered more similar. This makes the clustering more robust to noise and missing data.
The algorithm was tested using two different layering strategies: Temporal Layers (stacking data by year from Baseline to Year 4) and Outcome-Based Layers (stacking data by ranges of a specific severity score, such as MDS-UPDRS3).
Datasets and Clinical Variables
The study utilized the Parkinson’s Progression Markers Initiative (PPMI) database, focusing on a cohort of 194 patients with complete longitudinal data over 5 time points (Baseline to Year 4). The analysis tracked 17 clinical variables categorized into domains: Cognitive (e.g., memory, processing speed), Behavioral/Mental Health (depression, anxiety), Sleep, Autonomic, Disability, and General PD Severity.
Key Results: Identifying Four Distinct Subtypes
The TC algorithm identified four distinct subtypes in both the temporal and outcome-based analyses. These subtypes are characterized not just by their symptoms at one point, but by their entire progression path.
The identified subtypes include:
- Subtype $S_0$ (Mild/Cognitive): Older population, relatively mild motor symptoms, but showing specific cognitive impairment (visuospatial skills) that worsens over time.
- Subtype $S_1$ (Cognitive-Heavy): High severity across most cognitive variables, but relatively low motor disability.
- Subtype $S_2$ (Motor/Mental Health): Characterized by severe motor symptoms and significant deterioration in mental health and daily living activities over the 5-year period.
- Subtype $S_3$ (Autonomic/Sleep): Severe autonomic and sleep issues, but with limited cognitive or motor severity in the early years.
Statistical validation confirmed that these clusters were significantly different in their variable distributions and progression rates, aligning with existing medical literature.
Discussion and Future Directions
A key insight of this work is the variable-centric perspective. Unlike standard clustering, this method first identifies which symptoms "travel together." For example, the study found that in early stages, mental health symptoms form their own cluster, but as the disease progresses, they begin to co-express with motor disability. This dynamic subtyping allows clinicians to view patients as following specific "paths" through the disease space.
While the current study is limited by sample size (194 patients) and the use of variable communities rather than individual variables, the framework is highly flexible. Future work could extend this to include genetic markers or fluid biomarkers and adapt the algorithm to account for the effects of medical treatments. This generalizable method can effectively assist in targeted subtype-specific treatment in the field of personalized medicine.
Summary updated