A calibrated temporal reference map of disease progression
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Understanding the evolution of human illness requires capturing the temporal directionality of disease progression, yet existing biomedical reference maps largely describe cross-sectional states or static comorbidity. We introduce a directed, probability-ranked map (i.e., a knowledge-base) of clinical progression derived from population-scale longitudinal electronic health records.
Methods
The knowledge-base was constructed from de-identified EHRs of 295,678 individuals across the Mass General Brigham system, yielding 435,240 phenotype-pair-duration associations via temporal Spearman correlation. To distinguish biological progression from administrative artefact at scale, we distilled a locally deployed MedGemma labeling function into two complementary classifiers: a RF capturing local episodic signal and a GNN aggregating global network topology via message passing. Their outputs were combined as an unweighted late-fusion average. Classifier confidence was systematically evaluated against pairwise genome-wide genetic correlation estimates from the UK Biobank as an independent biological reference standard.
Results
Both classifiers achieved comparable distillation fidelity on the 200-row development set (RF AUROC 0.772; GNN AUROC 0.769). Genetic support was concentrated in the highest confidence deciles, with both models achieving highly significant top-decile enrichment for validated genetic pleiotropy (RF: 1.36-fold, p < 0.001; GNN: 1.32-fold, p < 0.001), demonstrating that classifier confidence aligned with independent genomic support. The framework additionally identified two complementary classes of progression: acquired mechanical cascades with high classifier confidence but null genomic overlap (exemplified by musculoskeletal pain progressing to cardiac dysrhythmias beyond 90 days, ), and topological bridges structurally enforced by network architecture despite sparse local co-occurrence (exemplified by acute myocardial infarction to epilepsy within 0–14 days, P GNN = 0.930 versus P RF = 0.332).
Conclusions
By transitioning from static comorbidity networks to a confidence-ranked landscape of temporal trajectories, the map provides a biologically calibrated coordinate system for prioritising mechanistic, translational, and clinical investigation of disease progression.
