312P Prospective multi-centre evaluation of a guideline-grounded conversational AI to streamline multidisciplinary tumour boards

Abstract

Background: Multidisciplinary tumour boards (MTBs) are time-intensive, capacityconstrained, and absent or inconsistent in many hospitals.We built a retrievalaugmented, guideline-locked conversational assistant that ingests a guideline PDF and structured patient data to generate page-cited, auditable recommendations.Methods: Prospective, paired, observer-blinded deployment at six hospitals.Consecutive adult oncology cases were processed and discussed at the MTB.The assistant was guideline-locked to the NCCN Guidelines and Ukrainian recommendations.Patient data was ingested from the e-health system, de-identified, and normalized prior to inference.Co-primary endpoints: (i) acceptable concordance with MTB advice (non-inferiority margin -5 percentage points for AI-MTB risk difference) and (ii) time-to-recommendation (TTR).Secondary endpoints: citation density, System Usability Scale (SUS), auto-flagging of missing/conflicting data, and safety events.Deterministic citation checks, full audit trail, and GDPR-aligned processing were enforced.Results: 412 cases in eight tumour groups: breast 28.2% (116), colorectal 22.1% (91), lung 20.4% (84), prostate 10.0% (41), gynaecological 8.0% (33), head & neck 4.6% (19), urothelial 3.9% ( 16), other 2.9% (12).Overall agreement with MTB was 95.9% (395/412): full 82.8%(341/412), partial 13.1% (54/412); non-agreement 4.1% (17/ 412).Adjudicated errors: MTB human-factor 8/17 (missed biomarker, incomplete chart, misapplied guidance); AI limitations 7/17 (retrieval gaps, local-nuance restrictions, ambiguous e-health fields).Only 2/412 wrong for both; joint AI+MTB would be correct in 410/412 (99.5%).Median TTR: AI 3.8 min (IQR 2.9-5.4)vs MTB 22.1 min (15.6-31.8);ratio 0.17; p<0.001.Citations: median 6/case.Auto-flags 26.7% (110/412).SUS 81.2; 88% found useful/very useful.No safety incidents. Conclusions:A guidline-grounded assistant achieved non-inferior agreement while cutting TTR 5.8.Complementary AI/MTB errors imply near-complete accuracy in tandem, enabling triage of routine cases and pre-structured review -especially where MTBs are absent.Scale with version-lock, audit, GDPR; evaluate outcomes and cost.

Ähnliche Arbeiten