Improving Mortality Models in the ICU with High-Frequency Data

Background: Assessment of the performance of Intensive Care Units (ICU) is of vital importance for an effective
healthcare system. Such assessment ensures that the limited resources of the healthcare system are allocated
where they are most needed. Severity scoring systems are employed for this purpose and improving these systems
is a continuing area of research which has focused on the use of more complex techniques and new
variables.
Objectives: This paper investigates whether scoring systems could be improved through use of metrics which
better summarise the high frequency data collected by automated systems for patients in the ICU.
Methods and Data: 3128 admissions to the Gold Coast University Hospital ICU are used to construct three logistic
regressions based on the most widely used scoring system (APACHE III) to compare performance with and
without predictors leveraging available high frequency information. Performance is assessed based on model
accuracy, calibration, and discrimination. High frequency information was considered for existing pulse and
mean arterial pressure physiology fields and resulting models compared against a baseline logistic regression
using only APACHE III physiology variables.
Results: Model discrimination and accuracy were better for models which included high frequency predictors,
with calibration remaining good in all cases. The most influential high frequency summaries were the number of
turning points in a patient’s mean arterial pressure or pulse in the first 24 h of ICU admission.
Conclusions: The findings indicate that scoring systems can be improved by better accounting for high frequency
data.