Adapting Trauma Outcome Prediction Models to Individual Facilities using Transfer Learning

Abstract

With the increasing availability of big data and advanced computational techniques, machine learning (ML) models are becoming common in medicine and healthcare. Generalizable models, models that can be applied to any setting or patient cohort, are described as a goal of ML, yet sacrifices in performance are required to demonstrate such broad applicability. To date, there is a gap in the use of modeling techniques that can learn overarching patterns from larger data sets that, when incorporated with local data, can better model facility-specific trends. Transfer learning (TL) techniques incorporate two disparate data sets, a source and a target. TL techniques learn patterns from the initial data set (the source) and apply relevant knowledge to the modeling task of a second data set (the target). We evaluated the use of TL in trauma outcomes prediction modeling at the level of the individual hospital to assess the impact of this approach when using small record sets. We considered two feature set variations when developing our logistic regression predictive models: the shared feature set between the source and the target and the union of the two feature sets. Compared to baseline approaches, TL-based models did not result in consistent improvements in predictive model performance.