In clinical practice, patients are often sorted into binary buckets: healthy or sick, control or case. While this simplifies diagnosis, it contradicts the fluid reality of biology. Most chronic conditions—from Alzheimer’s disease to lung dysfunction—do not appear overnight; they evolve along a continuous pathological spectrum. Pirhaji et al. argue that these binary classifications fail to capture the nuanced transcriptomic changes that occur as a condition worsens, potentially missing critical therapeutic targets that only appear in transitional states.
Recent research has shifted toward continuous disease progression modeling, a framework that treats illness as a trajectory rather than a destination. By moving beyond discrete labels, researchers are developing tools that can identify subclinical decline, predict individual rates of deterioration, and visualize the subtle "morphing" of anatomy as pathology takes hold.
The Limitations of Discrete Staging
Standard diagnostic frameworks often rely on rigid thresholds, such as the GOLD stages for COPD or clinical cut-offs for amyloid deposition in the brain. However, these "hard" boundaries create significant issues for both clinicians and machine learning models.
First, they ignore within-class variation. As Kang et al. and Kang et al. note, two patients labeled with "Alzheimer’s Disease" may be at opposite ends of the cognitive spectrum, confusing models that treat them as identical. Second, thresholds can hide early-stage risks. Kim et al. demonstrated that patients who test "negative" for amyloid plaques by traditional standards still face significant cognitive decline if they show subthreshold accumulation. Similarly, Gregg et al. found that key predictors of lung disease are active long before a patient meets the binary criteria for a COPD diagnosis.
Quantifying the Progression Score
To replace discrete stages, researchers have introduced the concept of a Disease Progression Score (DPS). Instead of a category, the DPS is a continuous value that represents where a patient sits on the timeline of a disease.
Schmidt-Richberg et al. pioneered this by using quantile regression to create "growth charts" for biomarkers, allowing clinicians to see not just how far a patient has progressed, but how fast they are moving. Scelsi et al. expanded this by fusing multimodal data—such as PET scans and hippocampal volume—into a unified score, which helped uncover new genetic loci like LCORL that influence the speed of decline.
Visualizing the Latent Trajectory
One of the most intuitive ways to understand the disease spectrum is through generative modeling. If a disease is a continuous path, we should be able to "walk" along it.
Yang et al. developed Surreal-GAN, which uses generative adversarial networks to decompose complex brain patterns into continuous, interpretable dimensions. This allows researchers to visualize how different regions of the brain atrophy at different rates. Similarly, Dravid et al. used latent traversal in GANs to morph healthy lung X-rays into diseased ones, providing a high-resolution map of how visual features evolve. This approach moves interpretability away from static "heatmaps" and toward dynamic simulations of disease evolution, as also explored by Adeli et al. through 3D warping "simulator" networks.
Learning from Ordinal Relationships
Even when continuous data is scarce, the natural order of disease stages (e.g., Healthy $\rightarrow$ Mild Impairment $\rightarrow$ Dementia) provides a powerful signal. Ordinal learning leverages this sequence to structure the model's internal logic.
The HOPE framework, proposed by Wang et al., uses "ordinal prototypes" to map the pathway from mild cognitive impairment to Alzheimer's. By teaching the model that these states are ordered, it learns a more robust metric space than simple classification. Baek et al. applied a similar logic to data imputation, using Ordinal Contrastive Learning (OCL) to fill in missing brain scan data by respecting the biological trajectory.
A critical challenge in these models is ensuring they reflect biological reality—for example, neurodegeneration is generally irreversible. Wang et al. addressed this by introducing a monotonicity constraint (RMLP), which prevents models from erroneously predicting that a patient is "getting better" due to noisy data, ensuring the predicted risk score only increases over time.
Toward Personalized Flight Paths
The ultimate goal of modeling the pathological continuum is precision medicine. By treating disease as a spectrum, models can predict individualized "flight paths" for patients. Jung et al. demonstrated this by fusing imaging and clinical scores to forecast future cognitive decline years in advance.
Whether it is identifying subclinical tuberculosis in rural populations (Fehr et al.) or distinguishing between subtle grades of tumor malignancy (Ganz et al.), the shift toward continuous modeling allows AI to respect the inherent fluidity of human health. A reader should finish understanding that in the eyes of modern research, a diagnosis is not a box, but a coordinate on a long and evolving map.
Key Foundational Work
Analysis and detection of speech under sleep deprivation (2006) — contributed 1 concept elements to this direction.
Regression of Liver Fibrosis in Cases of Chronic Liver Disease Type C: Quantitative Evaluation by Using Computed Image Analysis (2004) — contributed 1 concept elements to this direction.