Key Takeaways & Clinical Implications
- Cancer Diagnostic Accuracy: Achieved 100% sensitivity and an AUC of 0.986 for detecting malignant lesions (carcinoma, high-grade dysplasia, high-grade lymphoma) with 100% Negative Predictive Value (NPV) and 97.2% specificity.
- Significant Reduction in Reporting Delay: Identified 6 unflagged cancer cases originally triaged as routine that would have otherwise faced reporting delays of over 60 days.
- Pre-Malignant Lesion Detection: Demonstrated 100% sensitivity (AUC of 0.990) for low-grade dysplasia/adenoma and 94.9% sensitivity (AUC of 0.967) for intestinal metaplasia.
- Rapid Triage: Generated automated, prioritized triage lists within minutes of slide availability across three health board networks.
Executive Summary
This prospective multicentre study across three Welsh health boards evaluated the deployment of an AI-based pathology triage system across 982 consecutive gastric biopsies. By generating rapid triage lists within minutes of slide scanning, the AI achieved 100% sensitivity for cancer lesions and pre-malignant low-grade dysplasia. The platform successfully prioritized high-risk cases classified as routine based on the clinical recommendations from endoscopy, which would otherwise have experienced diagnostic delays of 10 to 60+ days. Thus, the AI demonstrated its potential capacity to significantly improve diagnostic prioritization and reduce waiting times in routine NHS pathology workflows.
Background & Objective
Pathology services in Wales process large numbers of gastric biopsies, many triaged as routine and often waiting over 60 days for reporting. Only cases clinically marked for urgent indications are prioritised. This study evaluated whether a clinical pathology AI tool could improve triage and enable earlier prioritisation of high-risk cases.
Methods
Between March and June 2025, an AI-based triage system was deployed across three hospital networks and prospectively analysed 982 consecutive gastric biopsies. Cases were assigned to routine or urgent pathways according to standard practice. Diagnostic performance was assessed for cancer-related pathology (carcinoma, high-grade dysplasia, high-grade lymphoma) and for pre-malignant findings including low-grade dysplasia/adenoma and intestinal metaplasia. The AI system was configured to maximise sensitivity and negative predictive value (NPV), generating triage lists within minutes of slide availability.
Results
- Cancer: The AI demonstrated 100% sensitivity, 97.2% specificity, 63.5% PPV, 100% NPV, and an AUC of 0.986. All confirmed cancer cases were correctly triaged with 0% false negatives and 2.7% false positives. The system identified 6 cancer cases initially marked as routine that would otherwise have waited more than 60 days and flagged 34 urgent cases, including one originally diagnosed as benign, that could have experienced delays of over 10 days.
- Low-Grade Dysplasia / Adenoma: The AI performance was 100% sensitivity, 98.0% specificity (2.5% to 4.5% FP), 41.4% PPV, 100% NPV, and an AUC of 0.990, successfully triaging all 12 positive cases.
- Intestinal Metaplasia: The AI demonstrated 94.9% sensitivity, 98.5% specificity, 89.4% PPV, 99.3% NPV, and an AUC of 0.967, correctly triaging 54 of 56 positive routine cases (96.4%) and 39 of 42 urgent cases (92.9%).
Conclusions
AI-supported triage demonstrated high sensitivity for malignant and premalignant gastric lesions, with minimal false positives. By rapidly identifying high-risk cases that would otherwise face delays of 10–60 days, the system may help reduce waiting times and support more efficient diagnostic prioritisation.
