Modeling how diseases progress

Chronic diseases such as Alzheimer’s unfold over decades, but the data we can collect is usually cross-sectional: one measurement per person, at whatever stage they happen to be in. Event-based models turn that limitation into a question of inference — if each patient is a snapshot from a shared underlying trajectory, we can reconstruct the trajectory from the snapshots.

This line of work is led by Hongtao Hao, who built it through his PhD and continues it now. Every model below ships as an installable package, because a method nobody can run is not a method.

As far as we are aware, these models are the state of the art for event-based disease progression modeling. Evaluated across 9,000 synthetic datasets and real ADNI data, our stage-aware model significantly outperforms prior event-based methods — Gaussian mixture EBM, kernel density EBM, and discriminative EBM among them — at both recovering the order of disease events and assigning patients to stages. One finding worth stating plainly, since it cuts against the usual instinct: the simpler Gaussian-based models consistently beat the more complicated KDE-based ones.

The models

Stage-aware modeling (SA-EBM). Standard event-based models treat every biomarker as either “affected” or “not affected.” We formulate the intuition that a disease increasingly impacts more cognitive and biological factors as it progresses, and show that modeling stage directly improves recovery of the progression sequence. Paper · GitHub · pip install pysaebm

Subtypes (Bayesian EBM). Diseases do not progress the same way in every patient, but neither do they vary arbitrarily — there are typically a few recurring subtypes. We infer subtype and stage jointly. Paper · GitHub · pip install bebms

Mixed pathology (JPM). Most event-based models assume one disease per person. In reality, several pathologies often progress at once, so we model them jointly rather than forcing a single explanation. Paper · GitHub · pip install pyjpm

Learned inference (TEMPO). A transformer trained on data simulated from the probabilistic model performs inference faster and more accurately than our original method — while the probabilistic model remains essential, both as the source of the training data and as the thing that makes the result interpretable. Paper (CHIL 2026) · GitHub

Collaboration

The work is joint with collaborators in neurology and neuroimaging, and is developed in the open — code and data live on the JPCCA GitHub organization. Recent results were presented at ML4H and the NeurIPS Time Series for Health workshop, and TEMPO appeared at CHIL 2026.