Effects of Non-IID Distributions in Lung Cancer Data on Survival Prediction with Federated Ensemble Learning

A common challenge in Federated Learning (FL) is that distribution shifts between clients, or Non-IIDness, decrease global model performance. Non-IIDness means that data is not independently and identically distributed between participating sites. Stronger distributional shifts lead to greater reductions of model performance. We have implemented an FL algorithm to compare Random Forest (RF) and AdaBoost in various non-IID scenarios using sequencing data from lung cancer patients. Therefore, we systematically shifted the class label distributions among clients over several iterations. Both RF and AdaBoost performed worse in highly non-IID scenarios than on balanced data, and RF significantly outperformed AdaBoost. Our FL algorithm offers potential for model personalization and fine tuning, and could be applied to other datasets including clinical data.