A doctor holds a stack of test results taken at different times. Each sheet on its own is just a snapshot of a single day, but what you want is a whole movie of how the disease unfolded. A neural network called Tempo learns to glue these scattered frames into a smooth reel.
Standard event-based models (EBMs) are like an archaeologist who can only say that a pot is older than a coin, but not by how many centuries. Tempo works differently. It is trained on tens of thousands of synthetic patient stories where the order and speed at which deviations appear in biomarkers (measurable signs of disease) are known in advance. Like a child learning to walk, falling at first, the network develops a feel for typical disease trajectories.
On synthetic tests, Tempo made half as many errors in event ordering as previous leaders, and in cases with dozens of biomarkers, almost three times fewer. When fed real data from patients with Alzheimer's disease (the ADNI project), it built the following chain: first the temporal lobes shrink, then amyloid, a sticky protein that clogs the space between cells, accumulates, and only then does memory noticeably weaken. At later stages, another protein, tau, appears, and brain destruction accelerates sharply. This sequence matches what doctors already knew, but now it was obtained without a single manual hint.
Previously, for each new hypothesis about the course of a disease, scientists had to derive complex formulas. Tempo, by contrast, is like a universal camera that doesn't care what it's shooting: faces, landscapes, or microbes. It quickly runs through different assumptions and shows which one fits the data best. The stack of test results in the doctor's hands finally turns into a real movie.
as written by the authors arXiv:2604.23368
Event-Based Models (EBMs) infer biomarker progression from cross-sectional data but typically only as ordinal sequences and rely on rigid model assumptions. We propose \textsc{Tempo}, a Transformer architecture that learns both ordinal and continuous event sequences through simulation-based supervised learning. \textsc{Tempo} uses two Transformer modules: one treats biomarkers as tokens to infer event sequencing; the other treats patients as tokens, representing each by their per-biomarker abnormality profile, to infer patients' disease stages. On synthetic benchmarks, \textsc{Tempo} reduces normalized Kendall's Tau distance by 52.89\% and staging MAE by 25.33\% compared to state-of-the-art SA-EBM, with larger reductions in high-dimensional settings (58.88\% and 61.10\%). Applied to ADNI, \textsc{Tempo} recovers a biologically plausible Alzheimer's progression: early medial temporal atrophy, followed by amyloid accumulation and cognitive decline, and late-stage tau pathology with terminal acceleration of global neurodegeneration -- broadly consistent with established disease models. \textsc{Tempo} also eliminates the need to derive custom inference algorithms and enables rapid empirical comparison of generative hypotheses.
The word "biomarker" sounds fancy, but it's really just a measurable trace of a disease. Blood sugar level is as much a biomarker as the amount of a harmful protein in cerebrospinal fluid.
The idea of predicting a patient's future from scattered data is reminiscent of Philip K. Dick's "Minority Report," but instead of three precogs, a neural network trained on millions of simulations does the job.