Integrating tumor growth models to improve late-stage lung cancer outcome prediction using early-stage clinical trial data.

Abstract

8532 Background: Immunotherapy has substantially improved clinical outcomes but poses distinct challenges for response assessment due to heterogeneous and non-monotonic tumor kinetics. RECIST 1.1 rely on unidimensional lesion measurements that may incompletely capture spatial heterogeneity which can result in limited characterization of tumor growth. Tumor growth modeling provides a complementary, quantitative framework for capturing longitudinal tumor dynamics. Different frameworks have been proposed to model total disease progression and response to treatment from total tumor volume. However, model selection may be challenging with each model capturing distinct growth mechanisms. In this work, we present a framework integrating complementary tumor growth models using a neural network classifier to predict late-stage outcomes from early-stage timepoints. Methods: We retrospectively aggregated deidentified data from 417 anonymized subjects with metastatic non-small cell lung cancer (NSCLC) treated with immunotherapy. Serial CT datasets were acquired at five timepoints over a 30-week period and lesions were annotated by expert radiologists according to RECIST 1.1 criteria. Tumor burden was derived from radiologist annotations using a previously validated AI method that reconstructs 3D lesion volumes from bidimensional data. For each subject, Modified Gompertz (MG) and Stein-Claret (SC) models were implemented to estimate intrinsic tumor growth parameters. Both models were fitted using early on-treatment data up to 21 weeks and used to predict response to treatment at week 30. To evaluate complementarity and the ability of early tumor dynamics to predict radiological response, model parameters were combined and used to train a two-layer neural network for radiological outcome classification. Performance of the combined MG+SC was compared with single-model classifiers on response (CR/PR) vs non-response (SD/PD) and progression (PD) vs non-progression (CR/PR/SD) using accuracy and sensitivity metrics. Results: The cross-validated accuracy in classifying responders and non-responders at week 30 was 81.1% from the combined MG+SC model (sensitivity=79.2%; specificity=83.0%), higher than both individual models (accuracy = 77.8% and 68.0% for MG and SC respectively). The combined modeling framework confirmed higher accuracy (70.7%) when distinguishing progressive disease from the non-progressor group (sensitivity=64.5%, specificity=76.5%) when compared to both MG (69.1%) and SC (66.8%). Conclusions: This work showed that integrating early-stage dynamics from complementary mechanistic growth models improves week-30 radiographic outcome prediction over individual models in metastatic NSCLC immunotherapy trials. This framework can enable earlier identification of patient-level response that may support treatment adaptation.