Abstract
End-of-life care (EoLC), such as advance care planning, advance directives, hospice, and palliative care consults, can improve patient quality of life and reduce costs, yet such interventions are underused. Machine learning-based prediction models show promise in identifying patients who may be candidates for EoLC based on increased risk of short-term (less than 1 year) mortality. Clinical decision support systems using these models can identify candidate patients at a time during their care when care teams can increase the provision of EoLC.Evaluate changes in the provision of EoLC with implementation of a machine learning-based mortality prediction model in an academic health center.A clinical decision support system based on a random forest machine learning mortality prediction model is described. The system was implemented in an academic health system, first in the medical intensive care unit, then house-wide. An interrupted time series analysis was performed over the 16 weeks prior to and 43 weeks after the implementations. Primary outcomes were the rates of documentation of advance directives, palliative care consultations, and do not attempt resuscitation (DNAR) orders among encounters with an alert for PRISM score over 50% (PRISM positive) compared with those without an alert (PRISM negative).Following a steep preintervention decline, the rate of advance directive documentation improved immediately after implementation. However, the implementations were not associated with improvements in any of the other primary outcomes. The model discrimination was substantially worse than that observed in model development, and after 16 months, it was withdrawn from production.A clinical decision support system based on a machine learning mortality prediction model failed to provide clinically meaningful improvements in EoLC measures. Possible causes for the failure include system-level factors, clinical decision support system design, and poor model performance.