Speaker
Description
Personalized forecasting in oncology requires models that are both mechanistically interpretable and effective with limited data. I will present recent unpublished results on predicting treatment response in mouse tumor models using a library of mechanistic growth and treatment models, including 1D formulations (exponential, logistic, weak Allee, strong Allee, and piecewise exponential growth) and 2D models with sensitive/resistant subpopulations under related growth structures, plus a frequency-dependent evolutionary game model. These models are applied to two longitudinal datasets: a chemotherapy dataset from NSG mice and a radiotherapy dataset from WT and SIRPα-deficient mice. Models are evaluated by both goodness of fit and ability to predict outcome (relapse versus durable control). Prediction performance is high across both datasets. For chemotherapy, strong Allee model parameters are highly informative: using only one measurement before and one after treatment, prediction achieves balanced accuracy 0.946 and AUC 1.0. For radiotherapy, the same sparse design yields balanced accuracy 0.874 and AUC 0.964. These results suggest that low-dimensional mechanistic models, combined with machine learning, can provide early, data-efficient forecasts of treatment response and identify interpretable biomarkers for adaptive oncology.
