The intensive care unit accommodates patients in a very critical health condition. Therefore, their vital parameters are constantly monitored. The purpose of monitoring is to detect deterioration at an early stage and to intervene timely. Medical staff must be able to rely on monitors that indicate deviations from predefined threshold ranges by means of acoustic signals (alarms). However, these normal ranges are usually not adapted to the individual patient. This leads to a high number of false alarms, which not only burden clinical staff but also pose a major risk to patient safety. Decreased response times can have serious consequences for patient health. The ECRI Institute regularly lists alarm hazards as one of the top ten technology issues in healthcare, highlighting its importance.
This thesis aims to help ensure that vital parameter thresholds are better tailored to individual patients in the future. To this end, I examine the heart rate and blood pressure thresholds contained in MIMIC- IV – a database containing ICU patient records. The analysis focuses on different influencing factors, such as demographics or medication. I include external information sources like SNOMED CT to enrich the MIMIC-IV data with additional medical knowledge. I then present an approach to automate vital parameter thresholds, considering specific patient characteristics. I focus mainly on the gradient boosting algorithm CatBoost and compare the results to other tree-based models. In order to create explainable predictions, I evaluate the feature importance with SHAP values.
I found that MIMIC-IV predominantly contains default thresholds and thus confirms the issue of alarm management described above. However, some vital parameter trends are reflected in the thresh- olds. The heart rate thresholds show more variance than the blood pressure thresholds. Patient char- acteristics tend to only reflect the vital parameter trend in one of the two threshold values. This indi- cates that thresholds were only adapted after an alarm occurred. Diagnosis and medication display the most alignment of vital parameter and threshold trends. Demographic characteristics only sporadically reflect trends from vital parameters.
CatBoost achieves considerably better results than the Decision Tree and Random Forest models. The lowest mean absolute error is obtained for the low heart rate threshold, the highest for the high systolic blood pressure threshold. The evaluation of feature importance shows a high influence of structural, hospital-specific factors such as the first care unit. However, this can also be seen as a grouping of specific diagnoses. The fact that the diagnosis features in models without structural fea- tures show the greatest influence supports this theory. Only a few features show vital parameter or threshold-specific influences. In most cases, the importance of a feature can be transferred to the other models. This improves the possibility of transferring the approach to local hospital data. An exception to the transferability of feature importance is the medication. Catecholamines show a higher influence on the blood pressure models, while general anesthetics affect the heart rate threshold pre- dictions more strongly.
This thesis provides a guidance that enables future projects to automate patient thresholds. Future efforts should focus on the acquisition of data containing highly patient-centered thresholds. A model trained on these data could reduce the burden of false alarms on medical staff and patients across different hospitals.