Medicine and Health

Data leakage inflates prediction performance in connectome-based machine learning models

M. Rosenblatt, L. Tejavibulya, et al.

This research by Matthew Rosenblatt, Link Tejavibulya, Rongtao Jiang, Stephanie Noble, and Dustin Scheinost delves into the critical issue of data leakage in neuroimaging predictive modeling. By examining five types of leakage across four datasets, the study unveils how feature selection and repeated subject leakage can dramatically skew prediction outcomes, particularly in smaller datasets. Discover the nuances of leakage's impact and its significance for achieving valid results!