Predictive Alarm Prevention by Forecasting Threshold Alarms at the Intensive Care Unit.

Patient monitors at intensive care units produce too many alarms – most of them being unnecessary. Medical staff becomes desensitised and ignores alarms. This phenomenon is called alarm fatigue and it negatively influences for both patients and staff. Some alarms are due to an acute and unforeseeable events but others are the result of a continued trend and hence foreseeable. We present a system that forecasts alarms – at least the foreseeable share – and transforms them into scheduled tasks. To achieve this, we use time-series models to forecast the patient’s vital parameters and check whether the forecast violates the corresponding alarm threshold. The vital parameter measurements and alarm data stem from MIMIC-III but go through extensive preprocessing before the actual forecasting can take place. The result is a proof of concept but unfit for productive use. Lack of alarm data and low sampling frequencies for vital parameters impair alarm forecasting. Our work shows that gated recurrent unit models generally perform best for this task. A next step towards productive use is evaluating the approach on vital parameter data with higher time-resolution.