In traditional medicine, diagnosis is often a binary gatekeeper: you either have a disease or you don't. We use rigid thresholds—a specific blood sugar level for diabetes, a certain amyloid count for Alzheimer’s, or a visual "plus disease" marker for retinopathy. However, biological reality rarely follows a step function. Instead, health is a continuous, stochastic degradation of "health capital."
Recent research is shifting away from these discrete "buckets" toward continuous disease progression modeling. By treating disease as a fluid trajectory rather than a static state, researchers are uncovering subclinical stages, improving genetic discovery, and providing more precise windows for medical intervention.
The Binary Trap and Information Loss
The primary motivation for this shift is the realization that binary labels are a form of lossy compression. When we categorize a patient as "healthy control" or "diseased," we discard the nuance of the transitional phase.
Li Shen and colleagues demonstrated the cost of this compression in genetic research. In Genetic analysis of quantitative phenotypes in AD and MCI, they showed that using continuous "quantitative phenotypes"—like hippocampal volume—provides 4 to 10 times more statistical power than binary diagnoses. This increased sensitivity allowed them to discover novel risk genes that traditional case-control studies completely missed.
Similarly, Hyung‐Ji Kim found that patients who are clinically "amyloid-negative" but show subtle, subthreshold accumulation are still at significant risk for cognitive decline. By treating amyloid deposition as a continuous pathological process, clinicians can identify at-risk individuals years before they hit an arbitrary diagnostic cutoff.
Disentangling Heterogeneity with Generative Models
If disease is a spectrum, how do we map its dimensions? A major thrust of this research uses Generative Adversarial Networks (GANs) and Variational Autoencoders (VAEs) to "disentangle" the complex ways a body changes.
Zhijian Yang and Christos Davatzikos developed Surreal-GAN, which decomposes brain disease patterns into continuous, interpretable dimensions. Rather than sorting patients into discrete subtypes, Surreal-GAN captures the spatial and temporal continuum of atrophy. This allows for a "dimensional representation" where a patient might be 30% along one atrophy pathway and 10% along another, rather than being forced into a single category.
This approach is echoed in Smile-GAN, where Ilya M. Nasrallah and Jimit Doshi map how healthy brain anatomy fluidly transforms into diseased states. By modeling the transition as a manifold, these tools can predict future clinical decline more accurately than traditional biomarkers.
Health as a Stochastic Process
Beyond static mapping, some researchers model disease as a dynamic system. Antoine Chambaz and Mark J. van der Laan have adapted threshold regression models to epidemiology. In this framework, health is viewed as a stochastic process that drifts toward a "failure threshold." Environmental toxins, like asbestos, are modeled not just as causes of disease, but as "accelerants" that speed up this drift, providing a concrete metric for years of life lost.
To capture the temporal nature of this drift, newer frameworks like BrainODE use Neural Ordinary Differential Equations. By conditioning growth trajectories on age and cognitive status, BrainODE can forecast future brain atrophy from a single clinical observation, treating the brain's shape change as a continuous physical evolution.
From Diagnosis to Liability Scores
The practical output of these models is often a "liability score" or "severity gradient" rather than a label.
- Justin Cosentino and colleagues developed a model for COPD that generates a continuous liability score from raw exhalation curves, outperforming standard clinical metrics in predicting mortality.
- Ziyang Song and Yue Li introduced MixEHR-SAGE, which transforms messy electronic health records into precise, continuous risk scores, uncovering genetic drivers that binary billing codes ignore.
- Nooshin Yousefzadeh developed LAVA, which uses retinal fundus images to map a seven-stage disease continuum for Alzheimer’s, providing a diagnostic score that reflects the nuanced reality of clinical progression.
Why It Works: Capturing the Gray Zone
The intuition behind continuous modeling is that the "gray zone"—the space between healthy and diseased—is where the most actionable information lives.
Traditional models struggle with "borderline" cases because they assume a clear decision boundary. Unsupervised methods, such as those proposed by Roozbeh Mohit, use spectral methods to infer a disease spectrum directly from the data's structure. By acknowledging that patients exist on a continuum, these models can identify subclinical conditions—like "covert" cerebrospinal fluid dysfunction—that contribute to neurodegeneration long before symptoms appear.
Ultimately, this research direction suggests that the future of medicine isn't about deciding if a patient is sick, but determining where they are on their unique biological trajectory.