Answers are grounded in this paper and cite their sources.
Background: Mental health care often requires clinicians to reason over patient journeys: longitudinal patterns of symptoms, diagnoses, treatments, behaviors, social context, and responses to care. Yet these journeys are rarely documented as readily available, explicit timelines. Clinically important temporal information is often embedded in unstructured discharge narratives, where events may be repeated, narrated out of order, or expressed through vague and relative time references. This makes temporal recovery difficult for clinicians and computational systems, and can limit the reliability of downstream summarization. Methods: We present CliniCIRCA, a multi-stage large language model (LLM) framework for Calendar-anchored, Imprecision-aware Reconstruction of Clinical Annals. CliniCIRCA decomposes patient-journey reconstruction into three stages: extracting atomic clinical events from raw discharge summaries, assigning calendar-grounded temporal labels with explicit uncertainty, and rendering the verified timeline as a date-grouped clinical summary. The framework is designed to make temporal reasoning inspectable rather than hidden inside a single summarization step. Results: From 14,882 MIMIC-III mental-health-coded admissions, we construct a benchmark of 52 discharge summaries on which CliniCIRCA produces 15,891 temporally tagged events. Following clinician-defined rubrics, human annotation with clinician adjudication corrects 629 errors, producing verified gold-standard timelines. These corrected timelines then drive a temporally grounded summarization stage that compresses each source 1.52 times into a date-grouped chronological record while preserving temporal labels and uncertainty. We further scale the framework to generate 1,000 silver-standard timelines and evaluate them as supervision for smaller open-weight models. Compared with zero- and few-shot prompting, instruction tuning generally improves five open-weight models across event extraction, temporal tagging, and summarization under both silver and clinician-verified evaluations. Conclusions: Auditable temporal reconstruction can serve as both a clinician-readable representation of complex mental health narratives and a reusable supervision source for smaller models. These findings suggest that intermediate, temporally grounded representations may help bridge clinical summarization, timeline generation, and scalable model adaptation, while preserving the need for expert adjudication in high-stakes clinical natural language processing (NLP).
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: