Endostrat: a multicenter study of AI-assisted two-level risk stratification for gastric precancerous and early neoplastic lesion screening

Background In large-scale endoscopic screening, neoplastic lesions constitute only a small fraction of the cases examined, making detection susceptible to reader fatigue. Most existing AI systems rely on image-level annotations and single-stage classification, limiting patient-level assessment and safety-aware review prioritization. We developed and validated EndoStrat, a two-stage AI-assisted risk stratification system that triages low-risk patients and prioritizes high-risk patients for focused endoscopic review. Methods This retrospective, multicenter study included 8,829 patients (approximately 220,000 white-light endoscopic images) from three centers with distinct pathological compositions. Patients were classified into four categories: normal/superficial gastritis, atrophic gastritis/intestinal metaplasia (Atrophy/IM), low-grade intraepithelial neoplasia (LGIN), and high-grade intraepithelial neoplasia/early gastric cancer (HGIN/EGC). Center 1 was split into a training cohort ( n = 3,555) and an internal validation cohort ( n = 890); Center 2 ( n = 2,896) and Center 3 ( n = 1,488) served as independent external validation cohorts. EndoStrat employs a Level-1 screening module that triages patients into low-risk and review-required groups using patient-level weak supervision and a Level-2 stratification module that outputs a continuous risk score integrating image features with age and sex for priority-based review. Two comparison methods—an image-level classifier and a four-class multiple instance learning classifier—were evaluated under identical conditions. Results Across the three validation cohorts, Level 1 achieved AUCs of 0.770, 0.720, and 0.674, which were numerically higher than those of both comparators. At the exact fixed operating threshold of 0.22, the HGIN/EGC flag rate was 93.0% or higher in all cohorts, and the neoplasia miss rate ranged from 5.8% to 8.1%. Among patients with normal/superficial pathology, 21.1%–26.1% were classified as low risk; across all patients, the corresponding overall low-risk assignment rates ranged from 11.2% to 19.6%. In a module-level analysis including all patients with pathology other than Normal/Superficial irrespective of Level-1 classification, the Level-2 continuous risk score showed a clear monotonic gradient across Atrophy/IM, LGIN, and HGIN/EGC (all p < 0.001), with concordance indices of 0.907–0.967 after clinical features were incorporated. Conclusions EndoStrat showed consistent risk-stratification performance across three validation cohorts with different pathological compositions. These retrospective findings suggest that EndoStrat may have potential to support risk-based review prioritization while retaining a high proportion of neoplastic cases in the review-required group. Prospective multicenter studies are needed to determine its effects on clinical workflow, diagnostic performance, and patient outcomes.