Speaker
Description
Limitations of human data in cancer research poses significant challenges for accurately predicting tumor growth and treatment outcomes. To address this, virtual patient cohorts are created to facilitate in silico exploration of new treatments. We propose a novel framework that integrates mathematical modeling, statistical data augmentation, noise reduction, and neural network-based trajectory tailoring to create a virtual patient cohort. Starting from 10 initial simulated patients, we apply a bootstrap technique with added noise to generate a cohort of 200 candidate patients. We then apply denoising techniques and process the candidate patients through a neural network, to sculpt the tumor growth trajectories to match the statistics of our simulated data. To benchmark our method, we compare it against cohort generation via Bayesian inference and sampling of posteriors, demonstrating that our framework is not only more computationally efficient but also more robust in handling noisy data. Finally we apply the pipeline to a dataset for oncolytic virotherapies and recapture the Kaplan Meier survival curves accurately. Our framework effectively handles noisy data, produces suitable virtual patient cohorts, and offers a scalable, computationally efficient solution for virtual patient cohort generation.
