Spatiotemporal Modeling of Heterogeneous Disease Progression

267 concepts215 papers11 proposals

Traditional clinical medicine often operates under a monolithic assumption: a disease is viewed as a single entity that follows a predictable, linear path from onset to late-stage severity. However, in complex chronic and neurodegenerative conditions, this assumption breaks down. Patients sharing the same diagnosis frequently exhibit vastly different symptoms, biomarker profiles, and rates of decline.

To resolve this, a powerful computational paradigm has emerged: data-driven disease progression modeling. Instead of treating individual variability as statistical noise, these machine learning frameworks explicitly disentangle a disease into distinct spatial subtypes and temporal stages. By learning directly from multi-modal patient data, these models transform cross-sectional snapshots into clear, individualized trajectories.

The Confounding Problem of Subtypes and Stages

To understand why modeling disease heterogeneity is difficult, imagine observing a crowded highway from a satellite. You see cars moving at different positions, but you do not know their origins, destinations, or speeds.

In clinical datasets, researchers face a similar challenge. If a patient presents with moderate cortical thinning, are they in the late stages of a mild, slow-progressing variant of a disease, or are they in the early stages of an aggressive, fast-progressing subtype?

Standard clustering techniques (like $k$-means) fail here because they group patients based on a static snapshot of their current state. They group an early-stage patient of Subtype A with a late-stage patient of Subtype B simply because their current symptom severity matches.

Static Clustering (Flawed):
[Early Subtype A]  <-- Grouped Together -->  [Late Subtype B]

Trajectory Modeling (Correct):
Subtype A Path:  [Early] ---------> [Moderate] ---------> [Late]
Subtype B Path:  [Early] -> [Late] (Fast Progression)

As highlighted by Nicholas C. Firth et al. in their study on Longitudinal neuroanatomical and cognitive progression of posterior cortical atrophy, diseases like Alzheimer's display profound phenotypic heterogeneity. A variant like posterior cortical atrophy targets visual processing networks rather than typical memory centers, progressing along an entirely different spatial and temporal timeline. To map these pathways accurately, models must simultaneously solve two unknowns: What path is the patient on (subtype), and how far along that path are they (stage)?

The Mechanism: Subtype and Stage Inference

A primary breakthrough in solving this joint optimization problem is the Subtype and Stage Inference (SuStaIn) algorithm. Developed to map complex spatiotemporal trajectories, SuStaIn conceptualizes disease progression as a sequence of discrete cumulative events—such as a specific brain region shifting from a normal volume to a threshold of severe atrophy.

The algorithm operates via a generative probabilistic framework:

  1. Event-Based Modeling: For a given disease pathway, the progression is defined by the order in which biomarkers become abnormal.
  2. Probabilistic Formulation: The model calculates the likelihood of a patient's data given a specific sequence of events. It assumes that the true state of a patient is a latent variable corresponding to a specific stage (the number of events that have occurred so far).
  3. Simultaneous Optimization: Using optimization techniques like Markov Chain Monte Carlo (MCMC), the algorithm iteratively updates both the number of distinct trajectories (subtypes) and the expected ordering of events within each subtype.

This approach allows the model to map out distinct evolutionary branches of a disease. For instance, Mar Estarellas et al. applied a missing-data-enabled variant of SuStaIn in Multimodal subtypes identified in Alzheimer’s Disease Neuroimaging Initiative participants by missing-data-enabled subtype and stage inference, successfully identifying five distinct Alzheimer's subtypes characterized by unique progression patterns. A "Cognitive" subtype demonstrated rapid transition from mild impairment to dementia, whereas a "Subcortical" subtype progressed much more slowly.

The power of this mechanism lies in its versatility across different pathologies. The same underlying mathematical framework was leveraged by William J. Scotton et al. in Uncovering spatiotemporal patterns of atrophy in progressive supranuclear palsy using unsupervised machine learning to identify two distinct patterns of brain atrophy in progressive supranuclear palsy, and by Alexandra L. Young et al. in Characterizing the Clinical Features and Atrophy Patterns of MAPT -Related Frontotemporal Dementia With Disease Progression Modeling to map mutation-specific trajectories in frontotemporal dementia.

Continuous Trajectories and Individualized Deviations

While event-based models excel at ordering discrete pathological milestones, other frameworks treat progression as a continuous, multi-resolution process.

Instead of forcing an individual to fit perfectly into a rigid population average, hierarchical latent variable models allow for personalized adjustments. In A Framework for Individualizing Predictions of Disease Trajectories by Exploiting Multi-Resolution Structure, Peter Schulam et al. proposed a model that structures disease progression into three layers:

$$\text{Individual Trajectory} = \text{Population Trend} + \text{Subtype-Specific Path} + \text{Individual Deviation}$$

The model utilizes Gaussian Processes to capture continuous-time changes. The population trend provides a coarse baseline, the latent subtype layer captures shared phenotypic variations, and an individual-specific random effect accounts for a patient's unique baseline and rate of decline. Crucially, this framework can filter out short-term clinical "noise"—such as a temporary drop in lung function due to an acute infection—to uncover the true underlying trajectory of a chronic condition.

Why This Paradigm Works

Data-driven disease progression modeling works because it shifts the analytical focus from absolute state to rate and direction of change.

By treating clinical data as a spatiotemporal coordinate system, these models achieve several key capabilities:

  • They unmix cross-sectional data: They allow large, single-visit datasets to be stitched together into a continuous timeline, bypassing the need to wait decades for comprehensive longitudinal cohorts.
  • They recognize localized drivers: As shown by Daniel C. Alexander et al. in Identifying and evaluating clinical subtypes of Alzheimer's disease in care electronic health records using unsupervised machine learning, unsupervised clustering of electronic health records can isolate highly specific clinical profiles—such as an "Anxiety, Depression, and Early-Onset" subtype—that exhibit significantly accelerated decline.
  • They enable precision clinical trials: Instead of testing a drug on a heterogeneous pool of patients where conflicting trajectories wash out the therapeutic effect, researchers can use these models to enroll individuals who are on the exact spatiotemporal path the treatment is designed to target.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_disease_heterogeneity_d399f16d,
  title = {disease heterogeneity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/disease-heterogeneity/rdr_d399f16df30541dca6f9697e7b2909eb},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_disease_heterogeneity_d399f16d_v1,
  title = {disease heterogeneity},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/disease-heterogeneity/rdr_d399f16df30541dca6f9697e7b2909eb/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake