Agentic echocardiography report structuring for cardiovascular prediction and disease association discovery

Abstract

Noisy echocardiography reports limit data-driven cardiovascular research because measurements that matter for downstream modelling remain in scanned images, multilingual narratives or semi-structured tables. This study presents an end-to-end agentic pipeline that takes such reports through to calibrated multilabel cardiovascular prediction, with separate extraction-side external evaluation and preliminary prediction-side probing. C-SARA-TM, an agentic report-structuring system with trace memory, clinically weighted optimal-transport routing and targeted repair, was applied to 500 OCR-digitized Chinese reports from the Second Xiangya Hospital as the extraction-side external cohort. Strict extraction accuracy reached 0.86, compared with 0.72 for the single-pass baseline and 0.635–0.758 for same-cohort rule-based, clinical-language-model and open-weight language-model comparators, with mean reviewer-level manual acceptability of 97.9% and a steady-state cost of 1.156 model calls per report. Downstream modelling used a MIMIC-derived cohort with 12,183 admissions, 22 structured predictors and 28 cardiovascular labels. We developed Dominance Hybrid, a calibrated label-aware multilabel prediction method that combines a sparse phi-graph hybrid learner with calibrated tree-ensemble baselines under label-specific convex weights. Dominance Hybrid achieved the highest prespecified composite score on internal testing (all-28: 0.630; echo-direct: 0.786), with additional same-split MIMIC sensitivity baselines reported in the main tables and Supplementary Information. A 50-case Xiangya cohort with report-derived labels served as a preliminary prediction-side cross-institutional probe limited to three echo-derived conditions (mean AUROC 0.756; mean AUPRC 0.608) without retraining or recalibration. The multilabel structure also produced a data-driven disease-association graph capturing 36 stable cardiovascular association patterns, including coupled valvular, ventricular and atrial pathologies that align with established cardiac pathophysiology. The pipeline is a reusable research resource for noisy retrospective echocardiography archives, covering report structuring, calibrated multilabel prediction and data-driven disease-association discovery.