Machine learning for enhanced survival prediction from tumour growth inhibition data

Background: The two-stage modelling approach is often used to characterise the link between overall survival (OS) and tumour growth inhibition metrics (TGI) derived from longitudinal tumour size data. This study evaluated the predictive performance of a two-stage TGI-OS model using survival machine learning (ML) methods in comparison to standard Cox proportional hazards (Cox PH) regression.

Methods: Longitudinal tumour size data, measured as the sum of the longest diameters of target lesions, was available from 8162 patients across 14 phase II/II studies evaluating the immune checkpoint inhibitor atezolizumab in five solid tumour types. Individual tumour growth trajectories were modelled using a biexponential TGI model including four parameters: baseline tumour size (TS0), growth rate (KG), shrinkage rate (KS) and the proportion of treatment sensitive cells. The TGI model was fit for each cancer type using nonlinear mixed effects modelling, and additionally by treatment line in NSCLC. The empirical Bayes estimates of the TGI parameters were utilised as predictors of OS. Discrimination of random forest, gradient boosted trees and Cox PH methods were evaluated within each cancer type group using the concordance-index and repeated 10-fold cross-validation. Goodness of fit was assessed using visual diagnostic and predictive checks.

Results: Random forest outperformed Cox PH in four of the six tumour type groups, and in particular for first line non-small cell lung cancer (NSCLC) and renal cell carcinoma (RCC). Cox PH and random forest were equivalent in the remaining two groups. Visualisation of the Shapley Additive Explanation (SHAP) values revealed a nonlinear effect of the log tumour growth rate parameter (KG) on OS risk present in all tumour types. Use of restricted cubic splines with Cox PH enabled improvements in the prediction of the Cox PH model.

Conclusion: Machine learning techniques demonstrate the potential to enhance the predictive capacity of TGI-OS models by capturing nonlinear associations between TGI metrics and survival outcomes. This suggests the importance of considering nonlinearities in TGI-OS models across various solid cancer types. Further work is needed to identify potential sources of the nonlinearity.

Acknowledgement: This abstract is based on research using data from data contributors Roche that has been made available through Vivli, Inc. Vivli has not contributed to or approved, and is not in any way responsible for, the contents of this publication.