[Paper Review] A Probabilistic Disease Progression Model for Predicting Future Clinical Outcome

Yingying Zhu, Mert R. Sabuncu|arXiv (Cornell University)|Mar 13, 2018

Health Systems, Economic Evaluations, Quality of LifeEconomics, Econometrics and Finance27 references4 citations

TL;DR

This paper proposes a probabilistic, Bayesian disease progression model that predicts future clinical outcomes in longitudinal, multi-modal data with irregular visit timing and missing entries. By modeling latent disease progression via a sigmoidal curve and integrating subject-specific attributes like genotype, the method achieves superior prediction accuracy over benchmarks on a 3,000+ subject Alzheimer’s dataset, while enabling interpretable insights into disease dynamics.

ABSTRACT

In this work, we consider the problem of predicting the course of a progressive disease, such as cancer or Alzheimer's. Progressive diseases often start with mild symptoms that might precede a diagnosis, and each patient follows their own trajectory. Patient trajectories exhibit wild variability, which can be associated with many factors such as genotype, age, or sex. An additional layer of complexity is that, in real life, the amount and type of data available for each patient can differ significantly. For example, for one patient we might have no prior history, whereas for another patient we might have detailed clinical assessments obtained at multiple prior time-points. This paper presents a probabilistic model that can handle multiple modalities (including images and clinical assessments) and variable patient histories with irregular timings and missing entries, to predict clinical scores at future time-points. We use a sigmoidal function to model latent disease progression, which gives rise to clinical observations in our generative model. We implemented an approximate Bayesian inference strategy on the proposed model to estimate the parameters on data from a large population of subjects. Furthermore, the Bayesian framework enables the model to automatically fine-tune its predictions based on historical observations that might be available on the test subject. We applied our method to a longitudinal Alzheimer's disease dataset with more than 3000 subjects [23] and present a detailed empirical analysis of prediction performance under different scenarios, with comparisons against several benchmarks. We also demonstrate how the proposed model can be interrogated to glean insights about temporal dynamics in Alzheimer's disease.

Motivation & Objective

• To develop a flexible, probabilistic model that predicts future clinical trajectories in progressive diseases like Alzheimer’s despite irregular visit schedules and missing data.
• To integrate multiple clinical modalities—imaging, cognitive tests, biomarkers—into a unified generative framework.
• To enable personalized predictions by adapting to individual patient history, even with sparse or absent prior data.
• To uncover interpretable insights into disease dynamics, such as the impact of genetic risk factors like APOE on progression.
• To outperform static classification benchmarks by modeling full longitudinal trajectories rather than binary outcomes.

Proposed method

• The model uses a latent sigmoidal progression curve to represent unobserved disease pathology, parameterized by inflection point and slope.
• Clinical observations (e.g., MMSE, ADAS-Cog, CDR-SB) are modeled as conditionally independent noisy measurements of the latent sigmoid curve.
• Subject-level attributes (e.g., age, sex, APOE status) are used to predict the parameters of the sigmoid curve, enabling personalized progression modeling.
• An approximate Bayesian inference strategy is employed to estimate model parameters while marginalizing over latent disease states.
• The framework handles missing data and irregular time intervals naturally, without requiring imputation or fixed time grids.
• The model is trained on a large Alzheimer’s disease dataset (ADNI) with over 3,000 subjects, including healthy controls, MCI, and AD patients.

Experimental results

Research questions

• RQ1 Can a probabilistic model accurately predict future clinical scores in Alzheimer’s disease when data are sparse, irregular, and multi-modal?
• RQ2 How does the inclusion of subject-specific attributes (e.g., APOE genotype) improve prediction accuracy and personalization?
• RQ3 To what extent can the model reveal interpretable disease dynamics, such as the influence of risk factors on progression timing and rate?
• RQ4 How does the model’s performance compare to static classification benchmarks that predict only binary outcomes (e.g., MCI to AD within 2 years)?
• RQ5 Can the generative structure of the model support mechanistic queries about disease progression beyond prediction?

Key findings

• The proposed model achieved significantly better prediction accuracy than benchmark methods across all target biomarkers (MMSE, ADAS-Cog, CDR-SB) under varying historical data availability and prediction horizons.
• The model outperformed static classification baselines by modeling full longitudinal trajectories rather than fixed-time binary outcomes.
• Personalized predictions improved when historical data were available, demonstrating the model’s ability to adapt to individual patient history.
• The model enabled mechanistic interrogation, revealing quantifiable impacts of APOE genotype on disease progression rate and timing.
• The Bayesian framework allowed for uncertainty-aware forecasts and transparent interpretation of risk factor effects, unlike black-box deep learning models.
• The model’s performance remained robust even with missing data and irregular visit schedules, highlighting its suitability for real-world clinical data.