To understand how a chronic condition like Alzheimer's disease, schizophrenia, or liver fibrosis develops, medical researchers ideally want to track individuals over decades, measuring biological markers at regular intervals. However, true longitudinal studies are logistically challenging, incredibly expensive, and prone to high participant dropout rates.
Instead, clinical researchers usually have access to large cross-sectional datasets—single-timepoint "snapshots" of many different patients captured at various stages of health and decline. The core insight of event-based disease progression modeling is that a large population of static snapshots can be aligned to reconstruct a continuous, global timeline of a disease. By treating the variation across a population as representative of different stages of a single underlying process, these models recover temporal dynamics without requiring multi-decade follow-up data.
Discretizing the Timeline: What is an Event?
How do we turn static, continuous measurements (such as brain volume, cognitive scores, or blood biomarkers) into a clear chronological sequence? The foundational step is to redefine progression as a series of discrete biological events. An "event" occurs when a specific biomarker transitions from a healthy state to an abnormal state.
As shown by Yuya Saito et al. in Temporal Progression Patterns of Brain Atrophy in Corticobasal Syndrome and Progressive Supranuclear Palsy Revealed by Subtype and Stage Inference (SuStaIn), these thresholds can be defined mathematically using z-scores, which measure standard deviations away from a healthy control population. In this framework, an event might be defined as "the caudate nucleus reaching a z-score of 2."
Alternatively, Neil P. Oxtoby et al. in Targeted Screening for Alzheimer's Disease Clinical Trials Using Data-Driven Disease Progression Models demonstrate that kernel density estimation can be used to model the data distributions of healthy and diseased populations directly. This data-driven approach avoids the need for rigid, expert-defined clinical thresholds by learning the progression order directly from the distribution of observed values.
Inferring the Sequence via Probabilistic Ordering
Once biomarkers are framed as discrete events, the challenge becomes finding their most likely chronological order. The core assumption of the Event-Based Model (EBM) is monotonicity: once a biomarker becomes abnormal, it remains abnormal throughout the disease course. This principle was leveraged by Peter A. Wijeratne et al. in A Multi-Study Model-Based Evaluation of the Sequence of Imaging and Clinical Biomarker Changes in Huntington's Disease to map out how structural brain changes consistently precede clinical symptoms.
To build intuition, imagine two biomarkers, $A$ and $B$. If we look across our cross-sectional cohort and find many patients where $A$ is abnormal but $B$ is normal, and almost no patients where $B$ is abnormal but $A$ is normal, we can intuitively infer that event $A$ happens before event $B$.
In practice, human biology is noisy, and individuals deviate from the norm. To handle this, Jonathan Huang et al. formulated a probabilistic framework in Probabilistic Event Cascades for Alzheimer's disease that searches for a central, optimal sequence of events by maximizing the likelihood of the observed cross-sectional data over a distribution of possible permutations.
Because the space of all possible permutations grows factorially with the number of biomarkers, researchers use Markov Chain Monte Carlo (MCMC) sampling to explore the sequence space. This allows the model to quantify the exact uncertainty of the inferred timeline, providing a robust measure of confidence for the inferred disease pathways, an approach detailed by Alexandra L. Young et al. in Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference.
Disentangling "What" from "When": Subtyping and Staging
A major complication in modeling disease progression is patient heterogeneity. For instance, two patients might present different symptoms not because one is further along the timeline, but because they belong to entirely different disease variants.
To solve this, Alexandra L. Young et al. developed the Subtype and Stage Inference (SuStaIn) algorithm. SuStaIn simultaneously solves two problems: it clusters patients into distinct spatial subtypes ("what" path they are on) and determines their relative progression stage ("when" or how far along that path they are).
This framework has revealed hidden trajectories across numerous conditions:
- Schizophrenia: Yuchao Jiang et al. used SuStaIn in Neuroimaging biomarkers define neurophysiological subtypes with distinct trajectories in schizophrenia to uncover two distinct brain atrophy pathways—one starting in the cortex and another in the hippocampus—which predicted patient responses to different therapies.
- Progressive Supranuclear Palsy: William J. Scotton et al. applied it in Uncovering spatiotemporal patterns of atrophy in progressive supranuclear palsy using unsupervised machine learning to identify distinct image-based subtypes that correlate with clinical presentations.
- Liver Fibrosis: A Suzuki et al. demonstrated in Modeling Liver Fibrosis Progression in Patients With Viral Hepatitis Using the Machine Learning Tool Subtype and Stage Inference (SuStaIn) that the model could map a hierarchical decline across entirely different physiological systems, showing that serum biomarkers fail first, followed by functional contrast uptake decline, and finally physical tissue atrophy.
Alternative Geometries for Trajectory Reconstruction
While discrete event-based models are highly intuitive, other approaches map cross-sectional data into continuous spaces. For instance, Sergey E. Golovenkin et al. developed the ClinTrajan package in Trajectories, bifurcations and pseudotime in large clinical datasets: applications to myocardial infarction and diabetes data, which applies Elastic Principal Graphs to treat static patient snapshots as a continuous landscape, mapping branching trajectories and calculating a continuous "pseudotime" metric.
In another approach, Iain G. Johnston et al. developed the HyperTraPS algorithm in Precision identification of high-risk phenotypes and progression pathways in severe malaria without requiring longitudinal data, utilizing mutual information to recover probabilistic transition pathways between symptom combinations in acute conditions like severe malaria.
Whether continuous or discrete, these models fundamentally change clinical trial design. By mapping a patient's single snapshot to an exact position on an inferred timeline, clinicians can "screen in" individuals at the precise stage where a therapeutic intervention is most likely to succeed.
Key Foundational Work
An Event-Based Disease Progression Model and Its Application to Familial Alzheimer's Disease (2011) — contributed 4 concept elements to this direction.
Probabilistic Event Cascades for Alzheimer's disease (2012) — contributed 1 concept elements to this direction.