Random feature baselines provide distributional performance and feature selection benchmarks for clinical and 'omic machine learning

원문 초록

Identifying predictive features from high-dimensional datasets is a major task in biomedical research. However, it is difficult to determine the robustness of selected features. Here, we investigate the performance of randomly chosen features, what we term "random feature baselines" (RFBs), in the context of disease risk prediction from blood plasma proteomics data in the UK Biobank. We examine two published case studies predicting diagnosis of (1) dementia and (2) hip fracture. RFBs perform similarly to published proteins of interest (using the same number, randomly chosen). We then measure the performance of RFBs for all 607 disease outcomes in the UK Biobank, with various numbers of randomly chosen features, as well as all proteins in the dataset. 114/607 outcomes showed a higher mean AUROC when choosing 5 random features than using all proteins, and the absolute difference in mean AUC was 0.075. 163 outcomes showed a higher mean AUROC when choosing 1000 random features than using all proteins, and the absolute difference in mean AUC was 0.03. Incorporating RFBs should become part of ML practice when feature selection or target discovery is a goal.