Trajectories matter: Discovery and validation of ordered EHR sequences that inform clinical risk predictions

Objective To test whether temporally ordered clinical sequences mined from EHRs improve the understanding of downstream adverse outcomes versus unordered bag-of-codes representations. Materials and Methods Within the NIH All of Us Controlled Tier, we mined frequent ordered event pairs (A->B) and tested for risk elevation of later adverse events (C) to create three-event (A->B->C) sequences. The algorithm used observation-period-aware indexing, a maximum 5-year A->B gap, and 90-day latency before follow-up. Comparators were (B without prior A)->C (primary), B->A->C, (A without B)->C, and a calendar-time baseline. Covariates were balanced with inverse probability of treatment weighting (IPTW) Weighted Aalen-Johansen estimated cumulative incidence at 1, 2, and 5 years. Discovery and confirmation were analyzed separately with global Benjamini-Hochberg false discover rate (BH-FDR). Results Of 633,545 persons, 432,617 met eligibility (≥365 observed days) and were split 70/30 into discovery and confirmation. We mined 3,066,183 trajectories from the discovery set by combining 340,687 sufficiently supported A->B pairs with each of 9 curated adverse third events. Due to compute constraints, we tested 20,565 of these trajectories (0.67%) for risk elevation. After discovery FDR, 234 trajectories advanced, and 39 validated for at least one time horizon in confirmation. At 5 years, the median risk ratio (RR) was 3.96 versus baseline and 2.18 versus B->C with no prior A. Reverse-order checks were feasible for 89.7% of hypotheses; the median (A->B->C) vs (B->A->C) RR was 1.21. Discussion Ordered trajectories captured clinically coherent pathways where temporal order added information beyond diagnosis presence alone. Conclusion Trajectory mining and confirmation reveal actionable, risk pathways that complement conventional risk models.