Research Square · Oct 27, 2026
Development and Temporal Validation of an Interpretable Machine‑Learning Model for Risk‑Stratified Endoscopy Triage in Population‑Based Esophageal Squamous Cell Carcinoma Screening
By Zhao T, Gao F, He X et al.
Background Esophageal squamous cell carcinoma (ESCC) imposes a substantial disease burden in high‑incidence regions including China. Population‑based endoscopic screening can detect early‑stage ESCC, yet universal endoscopy brings heavy workload pressure on screening programmes. Existing questionnaire‑based ESCC risk‑prediction models often rely on random‑split internal validation rather than temporal evaluation, and few translate model outputs into actionable triage metrics for screening management. We aimed to develop and temporally validate an interpretable machine‑learning model for risk‑stratified triage to identify prevalent screen‑detected ESCC and estimate the potential for reductions in endoscopy workload. Methods We retrospectively linked pathology and questionnaire records from a population‑based screening programme for adults aged 40‑75 years. Index endoscopies in 2024‑2025 formed the development cohort, and January‑June 2026 formed an independent temporal‑validation cohort. We considered 45 candidate predictors, handled missing data via multiple imputation, performed feature screening using L1 penalization, and compared ten machine‑learning algorithms. Class‑weighted learning was the primary strategy to address outcome imbalance. The final LightGBM ensemble was Platt‑calibrated. Model performance was assessed using AUROC, AUPRC, Brier score, calibration metrics, decision‑curve analysis, and threshold‑specific clinical utility. Results The cohort included 66,431 participants: 53,874 in development (844 ESCC events) and 12,557 in temporal validation (171 events). Temporal AUROC was 0.8605, AUPRC 0.0901, and Brier score 0.0129. At the development‑derived cutoff of 0.00729, sensitivity was 95.9%, specificity 54.8%, NPV 99.90%, and the referral rate 45.9%. Per 1000 screened participants, 541 endoscopies may be deferred and 0.56 ESCC cases may be missed under this threshold. Age was the dominant explanatory feature, with smaller contributions from sex, smoking, prior gastroscopy, education, diet, oral health, and family history. Conclusions A class‑weighted, Platt‑calibrated LightGBM model demonstrated promising performance for high‑sensitivity risk‑stratified triage in temporal validation. Independent geographic external validation and prospective impact evaluation are required prior to any real‑world clinical deployment.