(T-039) Predicting Tumor Growth Dynamics and Overall Survival with Deep Learning Model

Eric Song – Genmab; Kashvi Srivastava – University of Michigan; Kinjal Sanghavi – Genmab; Chris Le Gallo – Genmab; Chaitali Passey – Genmab; Summer Feng – Genmab

Objectives: Traditional tumor growth inhibition (TGI) models have been widely used to derive tumor dynamics metrics that inform overall survival (OS) predictions. However, they typically require complete longitudinal tumor size data and rely on predefined parametric assumptions. This work applied and extended a deep learning (DL) model for tumor growth dynamics (TGD) capable of learning trajectories from truncated tumor size data and generating kinetic parameters predictive of OS. We also evaluated its OS predictive performance in comparison to traditional TGI-OS based approaches.

Methods: We utilized 7 clinical studies across several tumor types with ~700 patient-level data, including demographics, baseline lab assessments, pharmacokinetic (PK) parameters, and pre-treatment clinical history, etc. A deep learning architecture TDNODE[1] (with RNN encoder and Neural-ODE decoder) was adapted to model tumor dynamics from longitudinal sum-of-longest-diameter tumor size data and extract individualized kinetic parameters. These latent parameters from decoder were used as covariates in Cox proportional hazards models for OS analysis. To simulate early clinical data settings, we trained models using truncated observations (e.g., limited to weeks post-baseline), while traditional TGI models were fit using complete longitudinal data. Additionally, baseline covariates associated with best overall response were identified using machine learning (ML) based feature selection methods: Boruta[2], Recursive Feature Elimination (RFE), and LASSO. The intersection of covariates selected by these methods was incorporated into the TGD-DL model to enhance tumor dynamics predictions.

Results: The TGD-DL model demonstrated progressively improved accuracy in tumor trajectory prediction as the observation window lengthened. When trained on the entire patient data, it achieved an RMSE of 3.33 and an R² of 0.995, compared to the TGI model’s RMSE of 2.70 and R² of 0.996. Incorporating selected baseline covariates into the TGD-DL model further reduced the RMSE to 3.27. In OS analysis, the TGI based Cox model achieved a C-index of 0.70 ± 0.03, Brier score of 0.18 ± 0.03, and mean time-dependent AUC of 0.77 ± 0.04. In comparison, the TGD-DL based Cox model outperformed these benchmarks, yielding 0.77 ± 0.02, 0.14 ± 0.03, and 0.85 ± 0.04 respectively.

Conclusions: The TGD-DL model effectively learns tumor growth dynamics from truncated data and generates latent parameters predictive of OS. While TGI models fit tumor size well with full data, TGD-DL offers greater flexibility and improved OS prediction, especially with ML selected covariates. Future work will integrate longitudinal PK profiles to further enhance predictive accuracy.

Citations: [1] Laurie, M. & Lu, J. Explainable deep learning for tumor dynamic modeling and overall survival prediction using Neural-ODE. npj Syst Biol Appl 9, 58 (2023).
[2] Kursa et al, Feature Selection with the Boruta Package. Journal of Statistical Software, 36(11)

Keywords: Tumor Growth Dynamics, Overall Survival Analysis, Deep Learning