Causal-Temporal Representation Learning for Comorbidity Progression Risk in Longitudinal Primary-Care Data

Longitudinal primary-care records offer broad and scalable coverage for studying how chronic conditions unfold over time, but progression graphs derived from temporal co-occurrences are often distorted by time-varying confounding, surveillance intensity, and care-process artifacts. As a result, recorded downstream diagnoses may reflect health-care patterns rather than clinically meaningful progression risk. We introduce CHRONOS (Causal Heterogeneity Representations for Onset-centric Networks in Observational Sequences), a temporal Siamese representation learning network for confounding-adjusted disease progression discovery from observational primary-care sequences. For each comorbidity candidate (disease pair), CHRONOS emulates a target trial, aligns patients at the incident onset of the first disease in the pair, and estimates the effect of that onset on the risk of developing the second disease within a fixed follow-up horizon. Methodologically, the framework combines temporal sequence encoding, mutual-information-based balancing, a Siamese contrastive loss over patient pairs constructed from predicted Individual Treatment Effect (ITE) similarity, and cross-fitted doubly robust estimation to obtain effect-annotated progression edges. We evaluate CHRONOS on a real-world longitudinal primary-care cohort from the Campania region in Italy, collected from 2000 to 2018. As a result, the inferred network successfully emphasizes stable within-category continuity-of-care effects while filtering many cross-category temporal associations likely driven by surveillance and administrative processes. The results show that CHRONOS offers a framework for moving from associative temporal co-occurrence patterns to confounding-adjusted progression-risk relations in longitudinal observational health-care data.