探索全球前沿学术脉络

Configurable Clinical Information Extraction with Agentic RAG: What Works, What Breaks, and Why

arXiv:2606.19602v1 Announce Type: new Abstract: Patient contexts span hundreds of heterogeneous documents and thousands of structured data points, yet the document-level metadata that AI systems need for retrieval and triage is absent or incomplete. Standard retrieval-augmented generation fails on this data, mishandling temporal reasoning, cross-document dependencies, and missing metadata. We deploy ACIE (Agentic Clinical Information Extraction) at University Medicine Essen: an on-premise agentic RAG pipeline that reasons over complete patient contexts and grounds every answer in source passages for clinician verification. We quantify the metadata gap, trace the architectural decisions it shaped, and evaluate extraction alongside an independent retrospective lymphoma registry study, in which nuclear-medicine physicians verify every extracted value against its cited sources. Across 7,326 judgments, clinicians accepted 96.5\% of extractions, with per-type acceptance ranging from 80\% to 99\%.

Development and External Validation of a Machine Learning Model for 10-Year Ischemic Stroke Risk Prediction in Diverse Populations

Importance: Machine-learning models for ischemic stroke risk prediction are rarely validated across ancestrally distinct cohorts, and the contributions of polygenic risk scores (PRS) and self-reported race in such models remain unclear. Objective: To develop and externally validate a 10-year ischemic stroke risk model and quantify the incremental contributions of laboratory trajectories, PRS, and self-reported race and ethnicity across populations. Design, Setting, and Participants: Retrospective cohort study with model development in the All of Us (AoU) Research Program (n = 34,987; 1,920 incident strokes) and external validation in the BioMe Biobank at Mount Sinai (n = 10,693; 107 incident strokes). Adults aged 45 years or older with at least 1 year of pre-baseline electronic health record data were anchored to a January 2010 baseline with 10-year follow-up. Exposures: Three XGBoost model tiers added laboratory feature trajectories (M2) and 20 PRS (M3) to clinical baseline features (M1); evaluated under race-blind and race-aware specifications. Main Outcomes and Measures: First inpatient ischemic stroke within 10 years; discrimination (area under the receiver operating characteristic curve [AUROC]) and calibration (observed-to-expected [O/E] ratio). Results: In the AoU test partition (n = 6,998; 384 cases), M3 achieved an AUROC of 0.813 (95% CI, 0.788-0.837), outperforming the Revised Framingham Stroke Risk Profile (AUROC difference, 0.164) and Pooled Cohort Equations (AUROC difference, 0.181; both P < 0.001). Discrimination transferred to BioMe (AUROC, 0.745), but predictions were systematically high (aggregate O/E ratio, 0.12 vs 1.00 in AoU), consistent with intercept-shift miscalibration; BioMe-fitted intercept recalibration restored calibration in African American and Hispanic participants but not European American participants. The PRS contribution was significant only among Hispanic participants in BioMe (AUROC difference, 0.042; P = 0.003), with no significant within-stratum gain in the other 5 cohort-by-race combinations. Adding self-reported race produced small gains when combined with PRS (BioMe AUROC difference, 0.022; P = 0.034; AoU AUROC difference, 0.006; P = 0.052) but not when added without PRS. Conclusions and Relevance: A machine-learning ensemble combining clinical, laboratory, and polygenic features outperformed traditional risk scores by 0.16 to 0.18 AUROC and retained discriminative validity in an ancestrally distinct external cohort but required site-specific recalibration of absolute risk. The marginal contribution of self-reported race overlapped with polygenic signal, supporting per-ancestry calibration over universal race-aware model deployment.