Integrating Stochastic Cancer Dynamics and Artificial Intelligence for Oncolytic Virotherapy Analysis

This is an evaluation and benchmarking study of an artificial neural network (ANN) surrogate for an existing stochastic model of cancer dynamics under oncolytic virotherapy; it does not propose a new stochastic cancer model or a clinical decision tool. The model considers the interactions between uninfected tumor cells, infected tumor cells, free virus particles and immune cells, with the random fluctuations in the system modeled via stochastic perturbation. Trajectories simulated with the Euler–Maruyama method are used to train and test a feed-forward ANN that predicts a tumor-burden proxy. In the original single-trajectory experiment, the ANN achieves an RMSE of 0.000547675 and an R2 of 1.000000 on the training subset, while validation yields an RMSE of 0.000827163 and an R2 of 0.995484. The test set, which is the chronologically later segment of the same trajectory and not an independent realization, however, has an R2 of −2.089698, indicating poor ability to generalize to the held-out segment of the simulated trajectory. We show that this failure is caused by extrapolation in time on a monotonically decaying trajectory and that, in this one-step design, a simple persistence predictor is more accurate than the ANN. In a revised benchmark with 70 independent stochastic realizations split by trajectory, a prediction horizon of 0.1 and increment targets, the ANN achieves a test RMSE of 0.0442±0.0008 (mean ± 95% CI over ten training runs, i.e., weight-initialization uncertainty) and outperforms persistence, autoregressive and Gaussian-process predictors and, modestly, linear regression (paired Wilcoxon signed-rank test over 20 test trajectories, p=0.0017; p<0.05 in all 10 individual training runs). This is a predictive skill on model-generated data, not a clinical predictive skill. Monte Carlo ensembles quantify the response of the tumor burden to the noise intensity, and control scenarios show that, within the assumed parameterized system, virotherapy accelerates the decline of the burden (area under the curve 32% smaller than in a no-virus control and 28% smaller than in an infection-disabled control with the same initial state) but does not change its equilibrium level. The importance of independent testing and of simple baselines is also highlighted. Clinical effectiveness cannot be inferred from the analysis as it is based on simulated data. Future research should involve experimental or clinical observations for the model calibration and external validation.