Disease Progression Modelling and Stratification for detecting sub-trajectories in the natural history of pathologies: application to Parkinson's Disease trajectory modelling

Alessandro Viani, Boris A. Gutman, Emile d’Angremont, Marco Lorenzi

TL;DR

DP-MoSt is a new probabilistic framework that models disease progression as continuous trajectories rather than a sequence of discrete events. By identifying distinct "sub-trajectories" for specific biomarkers, it successfully stratifies Parkinson’s disease patients into clinically meaningful subgroups with higher temporal precision and better interpretability than previous state-of-the-art models.

What Problem Does This Paper Address?

Degenerative diseases like Parkinson’s are notoriously "noisy" and heterogeneous. Two patients might share a diagnosis but follow completely different paths: one might suffer primarily from tremors, while another struggles with gait and balance.

Existing tools like SuStaIn (Subtype and Stage Inference) have tried to map these differences, but they often treat disease progression as a series of discrete "events" (e.g., a biomarker crossing a specific threshold). This approach lacks temporal granularity—it can tell you the order of events, but not exactly how the biomarkers evolve over a continuous timeline. Furthermore, existing models often struggle to quantify uncertainty at the biomarker level or fail to account for the fact that some biomarkers might be identical across all patients even if others diverge.

Methods

The authors introduce DP-MoSt (Disease Progression Modelling and Stratification), which shifts the paradigm from discrete events to continuous functions.

  1. Continuous Trajectories: Instead of steps, DP-MoSt uses sigmoidal (S-shaped) functions to model how biomarkers change over a long-term disease time-axis.
  2. Two-Level Mixture Model: The model operates on two levels simultaneously. The first level determines if a specific biomarker actually exhibits different "sub-trajectories" (clusters). The second level calculates the probability of an individual patient belonging to one of those sub-populations.
  3. Time-Shifting: Since clinical data usually consists of short-term "snapshots," the model uses Gaussian Process theory to estimate an absolute "disease age" for each patient, mapping their short-term observations onto a long-term timeline.
  4. Optimization: The parameters are estimated using a two-level Expectation-Maximization (EM) algorithm, which balances the discovery of new sub-types against a penalty for overfitting (using Laplace and Inverse-Gamma priors).

Datasets

  • Synthetic Benchmarks: The authors generated 100 synthetic datasets with varying "Signal-to-Noise Ratios" (SNR) and numbers of biomarkers (2, 5, and 10) to test the model's robustness in identifying known clusters.
  • PPMI (Parkinson’s Progression Markers Initiative): A massive real-world dataset. The study focused on 1,954 patients with longitudinal data (averaging 5 time points each). They specifically looked at five key measures: PIGD (gait) scores, TD (tremor) scores, cognitive assessments (MCATOT), and motor scores (NP2PTOT/NP3PTOT).

Key Insights and Ideas

  • Biomarker-Specific Splitting: One of the "aha" moments is that DP-MoSt doesn't force every biomarker to have multiple sub-types. In the Parkinson's data, the model found that the TD_score (Tremor Dominant) was the primary driver of the split, while other markers like cognitive scores remained relatively consistent across groups.
  • Continuous vs. Discrete: By using continuous curves, the model captures the rate of change. For example, it identified that one sub-population sees a much faster increase in tremor severity after the 6-year mark on the disease timeline.
  • Uncertainty Quantification: Unlike previous methods, DP-MoSt estimates the confidence ($\xi_b$) that a sub-trajectory actually exists for a given biomarker, preventing the model from "hallucinating" subtypes where there is only noise.

Key Results

  • Clinical Accuracy: When tested on real PPMI data, DP-MoSt's unsupervised clusters aligned remarkably well with existing clinical labels. It assigned 76% of Tremor-Dominant (TD) patients to one group and 62% of Postural Instability and Gait Disorder (PIGD) patients to the other.
  • Superior Stratification: Compared to SuStaIn, which produced highly unbalanced groups (72% of all subjects in one group), DP-MoSt provided a more balanced and clinically interpretable 56/44 split.
  • Robustness: In synthetic tests, the trajectory approximation error (measured via the OSPA metric) remained low and stable even as the data complexity and noise increased.

Limitations and Future Work

  • Fixed Number of Sub-populations: Currently, DP-MoSt requires the user to specify the number of sub-populations to look for. The authors note that SuStaIn has an advantage here as it can estimate the optimal number of clusters automatically.
  • Computational Complexity: Expanding the model to handle many more sub-populations or more complex biomarker relationships will increase the computational load.
  • Future Directions: The authors plan to implement automated estimation of the number of sub-populations and explore more complex notations to handle even higher-dimensional longitudinal data.

Summary updated

Abstract

Modelling the progression of Degenerative Diseases (DD) is essential for detection, prevention, and treatment, yet it remains challenging due to the heterogeneity in disease trajectories among individuals. Factors such as demographics, genetic conditions, and lifestyle contribute to diverse phenotypical manifestations, necessitating patient stratification based on these variations. Recent methods like Subtype and Stage Inference (SuStaIn) have advanced unsupervised stratification of disease trajectories, but they face potential limitations in robustness, interpretability, and temporal granularity. To address these challenges, we introduce Disease Progression Modelling and Stratification (DP-MoSt), a novel probabilistic method that optimises clusters of continuous trajectories over a long-term disease time-axis while estimating the confidence of trajectory sub-types for each biomarker. We validate DP-MoSt using both synthetic and real-world data from the Parkinson's Progression Markers Initiative (PPMI). Our results demonstrate that DP-MoSt effectively identifies both sub-trajectories and subpopulations, and is a promising alternative to current state-of-the-art models.
BibTeX
@inproceedings{viani2024disease,
  author = {Alessandro Viani and Boris A. Gutman and Emile d’Angremont and Marco Lorenzi},
  title = {{Disease Progression Modelling and Stratification for detecting sub-trajectories in the natural history of pathologies: application to Parkinson's Disease trajectory modelling}},
  booktitle = {LDTM/MMMI/ML4MHD/ML-CDS @ MICCAI},
  year = {2024},
  doi = {10.1007/978-3-031-84525-3\_1},
}
Figure