Read Every Single Note (RESiN): Coverage-First Clinical Data Abstraction Beyond RAG

Abstraction of complex data from clinical notes is a process more akin to the reconstruction of a patient's journey than the retrieval of atomic facts. Standard methods of retrieval-augmented generation (RAG) first select a subset of information from the entire patient record to then inform a response. This design is efficient for simple and static data elements but is poorly matched to the abstraction of longitudinal repeated measurements, or of complex clinical events with inter-temporal dependencies, where evidence may be scattered across years of incomplete and inconsistent notes. Motivated by the relentless improvement in price-performance of LLM inference, we present Read Every Single Note (RESiN), a coverage-first system that reverses the typical order of operations. It processes every valid clinical note, generates an intentionally overcomplete set of candidate events, then shifts the bulk of work to the back end to harmonize and reconcile raw abstractions into a synoptic record of real-world evidence. A modular design accommodates various cancer and tumor types and hundreds of different data elements; a companion interface speeds human verification of results. RESiN was validated on four oncology cohorts totaling 71 patients with 47,000 notes, abstracting: 138 of 139 reference treatments for metastatic renal-cell carcinoma; 98.4\% event-matching accuracy across 1,903 breast cancer clinical events; 95\% field-level accuracy across 100 data fields for chimeric antigen receptor T-cell therapy in leukemia; and 97\% accuracy across 70 biomarker annotations in lymphoma. We demonstrate this system, currently in pilot at a national cancer center, and forecast near-future optimizations in cost and design to realize continuous abstraction for all patients.