Electronic health records are authored to support patient care, not to serve as a retrieval corpus. In longitudinal oncology notes the primary diagnosis is restated in every report, while pivotal events such as recurrence or therapy change may appear only once. A similarity‑based retriever therefore keeps surfacing the repeated diagnosis and misses the temporal transitions, creating temporal redundancy. We argue that longitudinal clinical retrieval should be framed as allocating a limited evidence budget over time rather than merely ranking by relevance. Accordingly we decompose the problem into three objectives: temporal evidence allocation, candidate reconstruction, and clinical validation.
In Study A we evaluated 13 clinical queries across 20 patient records using a two‑stage reconstruction pipeline. The Temporal‑15 configuration achieved a milestone‑slot recall of 0.538, substantially higher than the Global‑50 baseline’s 0.388. However, Temporal‑15 retrieved 496 evidence units per query versus 50 for the baseline, indicating that the gain cannot be attributed to allocation alone. When the evidence count is held constant, removing the prompted precedence rules raises schema‑order reversal rates from 0.010 to 0.080. A subsequent pilot with five patients and a single query, using fixed budgets, showed that broader temporal coverage does not consistently produce more dated milestones. These findings suggest that the amount of text a record contains for a given period is a poor proxy for that period’s evidential usefulness. Because expert reference timelines are unavailable, our metrics describe retrieval behavior rather than clinical accuracy. This report details the retrieval analysis underlying the Onco‑CONSTRUE framework and outlines a research agenda for budget‑constrained longitudinal clinical retrieval.
Review