Machine Learning Algorithms to Predict Mortality and Allocate Palliative Care for Older Patients With Hip Fracture.

The data were administrative claims and rehabilitation assessment records, so they lacked laboratory results, medications, and social or behavioral information. The sample was fee-for-service Medicare patients, so the models may not apply to people in Medicare Advantage plans. The data had a lag time, and the paper says the models would require validation in electronic medical records. The participants were mostly female and white, although the paper says that reflects the national hip fracture population. The decision threshold was chosen by a statistical index and did not take clinical judgment into account. The study did not examine how predictors change over time during follow-up. The models were evaluated by splitting the same data into folds, and the paper says they would require validation in the electronic medical record. The paper says implementation and translation should be pilot-tested to demonstrate benefits.

The supplied metadata lists NIH extramural research support. No author conflicts of interest are reported in the supplied text.

Do not read these models as ready-made tools for deciding who should get palliative care. They were built from 2014 to 2015 Medicare claims data, require validation in the electronic medical record, and the paper says implementation should be pilot-tested in inpatient rehabilitation facilities.