Early Prediction of Overall Survival in Oncology Trials Using Tumor Dynamic Neural-ODE

Accurate early prediction of overall survival (OS) is crucial in oncology drug development, where late-stage trial failures remain common. Traditional tumor growth inhibition (TGI) models provide predictive biomarkers of OS but rely on restrictive assumptions that limit generalizability. We evaluate Tumor Dynamic Neural Ordinary Differential Equations (TDNODE), a pharmacology-informed neural network that learns patient-specific tumor kinetics from longitudinal tumor size data. Using 8,121 patients across 10 phase II/III atezolizumab trials in five tumor types, we show that TDNODE-derived kinetic metrics consistently outperform TGI metrics in OS prediction, achieving higher Concordance Indices and lower Integrated Brier Scores. Notably, accurate prediction is possible with as little as 16 weeks of tumor data, often surpassing TGI models on full trajectories. These results demonstrate that TDNODE generalizes across trials and tumor types while enabling early survival prediction, offering a promising framework to support oncology development and accelerate decision-making.