Authors:
Denise Schmidt
;
Jonas Chromik
and
Bert Arnrich
Affiliation:
Hasso Plattner Institute, University of Potsdam, Germany
Keyword(s):
Intensive Care Unit, Patient Monitor Alarm, Alarm Fatigue, Threshold Alarm, Threshold Forecasting, CatBoost, SHAP, MIMIC-IV.
Abstract:
Intensive care unit staff relies on patient monitors to identify critical conditions. The monitors trigger alarms as soon as the patient’s vital parameters deviate from predefined threshold ranges. However, these ranges are usually not adapted to the individual patient. High numbers of false alarms burden clinical staff and pose a major risk to patient safety. We propose a recommender system for threshold values to enable a patient-centered monitoring system. This can reduce false alarms caused by default monitoring settings. We employ CatBoost – a gradient boosting algorithm – to predict blood pressure and heart rate thresholds. We use SHAP values to evaluate the importance of different patient characteristics, diagnoses, or medications. Several patient characteristics show an impact on the model output: Diagnoses, first care unit, vital parameter measurements, and the amount of general anaesthetics are the most important features in all threshold models. The recommendations of our
system deviate from the actual thresholds by approximately 3.5 bpm for the heart rate and 4.9 mmHg for the blood pressure thresholds. Blood pressure thresholds have a higher variance which leads to larger errors. However, the underlying data is not very patient-centered and we require better alarm data to further improve threshold recommendation.
(More)