Benchmarking the KEYNOTE-811 standard of care using AI-reconstructed external control arms in advanced gastric cancer.

Abstract

4021 Background: Randomized controlled trials in rare cancers and biomarker-enriched populations face significant recruitment burdens. External control arms (ECA) derived from historical Individual Patient Data (IPD) can mitigate these challenges, yet cross-trial heterogeneity often confounds comparisons. This study evaluates the feasibility of an AI-powered pipeline to reconstruct IPD from historical HER2+ advanced gastroesophageal adenocarcinoma trials and validates the cohort against a modern standard-of-care benchmark (KEYNOTE-811). Methods: An LLM-computer vision pipeline digitized KM curves from 10 historical trials (2011–2022) and the KEYNOTE-811 control arm (2023). IPD was reconstructed using a modified Guyot algorithm with risk-table guidance and least-squares optimization. Accuracy was validated against published HR and median survival. A propensity score matched (PSM) analysis (1:1 nearest-neighbor) compared the KEYNOTE-811 control arm (n=348) to historical controls (n=395). Sensitivity analyses addressed geographic imbalance (34% vs. 95% Asian) and early-period confounding via 3- and 6-month landmarking. Results: Technical validation demonstrated high fidelity: median OS concordance was 100%, with a reconstructed-to-published OS HR delta of only 0.01 (0.85 vs 0.84). The initial PSM suggested significantly worse outcomes for the KEYNOTE-811 control (OS HR 1.69, p<0.0001). However, this discrepancy was primarily driven by regional heterogeneity. After region-specific matching, OS HR stabilized at 1.13 (p=0.40). A 6-month landmark analysis further resolved early-period bias, achieving statistical equivalence for OS (HR 1.26, p=0.057) and PFS (HR 0.97, p=0.84). Residual imbalances in ECOG (SMD 0.25) and primary site (SMD 0.32) persisted but did not negate the trend toward equivalence. Conclusions: AI-powered IPD reconstruction achieves high fidelity to original trial data. While naive pooled ECAs may display era-specific survival artifacts, rigorous causal inference frameworks can resolve cross-trial heterogeneity. This study confirms that AI-reconstructed ECAs are viable tools for benchmarking novel therapies when paired with robust methodological guardrails for regional and baseline covariate alignment. PSM sensitivity analyses with covariate balance. Analysis OS HR (95% CI) PFS HR (95% CI) Pairs Key Imbalances (SMD >0.2) Main PSM 1.69 (1.37–2.08)*** 1.30 (1.05–1.61)* 217 Age (0.86), ECOG (0.42), Primary site (0.49), Region (0.65) + Geographic matching 1.13 (0.85–1.50) 0.87 (0.66–1.16) 119 ECOG (0.25), Primary site (0.32) + 3-month landmark 1.16 (0.93–1.45) 1.17 (0.93–1.48) 194/177 ECOG (0.25), Primary site (0.32) + 6-month landmark 1.26 (0.99–1.60) 0.97 (0.73–1.30) 174/115 ECOG (0.25), Primary site (0.32) ***p<0.001, *p<0.05; SMD = standardized mean difference; All other covariates SMD <0.2.