A0914
Title: Machine learning to the mean and its correction: An application to imaging-based brain age prediction
Authors: Shuo Chen - University of Maryland (United States) [presenting]
Hwiyoung Lee - University of Maryalnd (United States)
Abstract: A commonly observed issue in machine learning models predicting continuous outcomes is reported, referred to as "machine learning to the mean''. When applying a machine learning model built on a training dataset to a testing dataset, the difference between the predicted continuous outcome and the true value is often negatively correlated with the true value. For observations with continuous outcomes much smaller or greater than the mean, the predicted values tend to be automatically warped toward the mean. It is shown that this issue can be caused by the commonly objective function of minimizing the square loss. A new constrained strategy is proposed to correct the bias and develop computationally efficient algorithms for implementation. The new approach is applied to predicting brain age by brain imaging data while addressing the well-known issue of chronological age-related bias in brain age prediction.
