Preparing a Clinical Support Model for Silent Mode in General Internal Medicine
Proceedings of the 5th Machine Learning for Healthcare Conference, PMLR 126:950-972, 2020.
The general internal medicine (GIM) ward oversees the recovery of ill patients, excluding those who require intensive attention. Clinicians provide full recoveries, or when appropriate, end-of-life care. We hope to eliminate unexpected deaths in the GIM ward, promptly transfer patients who require escalated care to the intensive care unit, and proactively address deteriorating health to minimise ICU transfers. We describe a clinical decision support system which accesses labs, vitals, administered medications, clinical orders, and specialty consults. Using an ensemble of linear, gated recurrent unit (GRU) and GRU-decay (GRU-D) models, we are able to achieve a positive predictive value of 0.71 while successfully identifying 40% of patients who will experience a future adverse event. We believe that this tool will be useful in shift scheduling and discharging patients, in addition to warning clinicians of risk of decompensation. We note the lessons we learned in transitioning from a high performing model to deployment in silent mode, and all results reported in this paper report on data immediately preceding silent mode.