Machine learning can accurately predict which patients will not live beyond 30 days after discharge from the ER, giving these patients time to discuss end-of-life care with family members and hospice professionals.
Meanwhile the advance warning can inform healthcare organizations’ decision-making around appropriate allocation of resources.
The findings come from a study conducted by researchers in Sweden and the U.S. and published online Aug. 10 in BMJ Open.
The team trained six supervised machine-learning models on EHR and administrative data from more than 65,000 emergency care episodes.
To validate the models’ performance, they used a dataset of more than 55,000 ER care records from a separate facility to which the models were not exposed.
Noting that mortality occurred in 0.21% of cases in the training set and in 0.15% in the validation set, the authors report that four of their machine-learning models predicted all-cause 30-day mortality with excellent discrimination on the validation set.
Further, the models outperformed manual indexes conventionally used for predicting short-term mortality.
The machine-learning model with the best performance had sensitivity of 87% and specificity of 86%.
In their discussion section, the authors point out that their algorithms were trained on real-world data flowing from routine delivery of care, demonstrating the practicality of putting the models into everyday clinical use.
“Buying into the hypothesis that patients who are given an opportunity to communicate their end-of-life preferences are more likely to receive end-of-life care that is in line with their preferences, we aimed to train supervised machine learning models to identify patients at the end of life,” they write.
“Our ambition is that the final models can subsequently be used to systematically identify patients who may benefit from a discussion about end-of-life care without significantly adding to the workload of healthcare practitioners.”