ObserveTrace: Explicit Evidence States for Partially Observed EHR Decisions
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
At the moment a renal-risk medication is ordered, no local creatinine can mean never measured, advertised by an external exchange but unavailable, or too old for dosing; a note-derived value can be present yet uncertain. Standard code–value timelines collapse these conditions even though they imply different evidence needs and safety actions. ObserveTrace makes the observation process a first-class model interface: activated evidence requirements become canonical, round-trip-parseable events with value, units, time, provenance, and one of five evidence states. The same event object drives a 7B decoder, constrained alerts, frozen safety rules, and reviewer cards. Observability-aware denoising (OAD) reconstructs workflow-conditioned evidence losses, while incremental clinical consistency (ICC) regularizes adjacent-prefix decisions only when a frozen clinical protocol detects no rule or threshold change. Across 847K high-risk orders from 312K patients at seven health systems, ObserveTrace-7B records 3.2 (2.8–3.7) queue-reaching misses per 1,000 eligible orders, versus 3.8 (3.3–4.4) for an information-, backbone-, and training-budget-matched structured long-context model (paired difference -0.6 (-1.0 to -0.2), ); its false-alert point estimate is also lower, 42.3 versus 43.7. In a fully crossed study, replacing a plain timeline with the complete interface transformation lowers misses from 9.1 to 5.1 before either objective. On 128,640 held-out gate-stable prefix pairs, ICC lowers argmax changes from 0.218 (0.206–0.231) to 0.129 (0.121–0.138), while the action-update rate on 34,920 gate-changing pairs is 0.924 (0.912–0.935). Endpoint concordance is 95.6% with blinded re-adjudication agreement ; a later 184.4K-order silent study preserves 2.9–3.4 logged queue-reaching misses per 1,000 across Epic, Cerner, and MEDITECH workflows. These results establish explicit evidence state as an effective common boundary for partially observed longitudinal decisions.
- Keywords:
- biomedical and clinical ai, ObserveTrace