Feature robustness in non-stationary health records: caveats to deployable model performance in common clinical machine learning tasks

Machine Learning for Healthcare Conference
Bret Nestor
Bret Nestor

I am studying the generalisability of machine learning models applied to healthcare. The dynamic and adaptive nature of healthcare is reflected in the data that is collected in electronic health records. Sometimes we can anticipate these changes, and other times we need the model to be robust to these changes so that their decisions are reliable. In order to integrate machine learning into clinical models, we must understand when it fails, where it fails, and whom it fails to serve.