A new AI framework called CliniCIRCA can rebuild the timeline of a psychiatric patient's care from messy hospital notes that never state when anything actually happened.
Researchers built a multi-stage LLM pipeline, formally named Calendar-anchored, Imprecision-aware Reconstruction of Clinical Annals, that reads unstructured discharge summaries and tags each clinical event with an estimated date even without explicit timestamps. Working from 14,882 mental health admissions in the MIMIC-III database, the team built a 52-summary benchmark and had the system extract 15,891 temporally tagged events. A clinician review caught and corrected 629 errors, producing a verified gold-standard dataset. The corrected timelines then fed a summarization step that compresses each record by about 1.52 times into a chronological, date-grouped account.
Mental health records are notoriously scattered, with biological, psychological, and social events described in prose that carries no clean timestamps, forcing clinicians to piece together history by hand. The team also scaled the pipeline to generate 1,000 machine-labeled silver-standard timelines and used them to instruction-tune five open-weight models, which then beat simple zero- and few-shot prompting on event extraction, temporal tagging, and summarization.
It is a narrow, unglamorous fix for untangling dates in hospital paperwork, but that is exactly the kind of grunt work that decides whether AI is useful in a clinic instead of just a demo.