Digital twin survival calibration to enable cross-disease interpretation of early clinical trials.

Abstract

e14507 Background: Early-phase oncology programs increasingly explore multiple disease indications, yet overall survival outcomes cannot be directly compared across cancers due to fundamentally different baseline mortality patterns. This creates uncertainty when interpreting early survival signals and planning cross indication expansion. A digital twin survival model was developed that calibrates survival time across disease sites using non parametric hazard alignment, without modeling treatment-specific effects. Methods: Publicaly available overall survival data were obtained from the TCGA PanCancer survival resource via UCSC Xena. Three disease cohorts were analyzed: melanoma, non-small cell lung cancer, and renal cell carcinoma (n = 1,900 patients total). For each cohort, survival distributions were estimated non parametrically and transformed into cumulative hazard function. Digital twin survival time was generated by mapping source survival time to the target disease time scaleat matched cumulative hazard, yielding patient level counterfactual baseline survival trajectories. Translation performance was evaluated using time-wrap functions and survival curve overlays. Results: Baseline survival differed markedly across disease sites, with median overall survival of 1,175 days for melanoma, 523 days for Non small cell lung cancer, and 1,175 days for renal cell carcinoma. When translating melanoma survival to the lung cancer hazard scale, 365 days of melanoma survival corresponded to 324 days in lung cancer, and 730 days corresponded to 559 days. The median digital twin survival of melanoma patients expressed on the lung cancer time scale was 862 days. After transformation, digital twin survival closely matched the observed lung cancer survival distribution, demonstrating accurate cross-disease calibration of baseline survival distributions. Conclusions: Digital twin survival modeling provides a quantitative framework for calibrating baseline survival across disease sites without assuming equal treatment effects. This enables hazard-calibrated interpretation of early survival signal, supports indication prioritization, and informs cross- disease dose-expansion planning in early drug development. The approach is fully reproducible using public data and is applicable across therapeutic modalities.