Disease severity assessment and ordering of patients in ICU by using a Bayesian network

Abstract

This paper aims to develop a realistic triage system to better quantify a patient’s disease severity for the evaluation of admission or discharging. A good triage can reduce loads of doctors and draw attention of staffs to critical conditions. However, existing systems score on readings of vital signs and the superficial scores usually are apart from doctors’ judgement. Instead of summing up rating score, we take a Bayesian network approach to estimate the source diseases that lead to the observed vital signs, such as temperature, lactate, HCT, and CRP, etc. Because the purpose of this assessment is not making a correct diagnosis, the source diseases are only stratified to four disease categories. Based on the reading of vital signs, Bayes belief network inferences the probability distributions of the severity for each one of the four disease categories. Finally, the four distributions are then sufficient to rank a patient’s final severity by a probabilistic decision framework. Diffing from traditional paper based evaluation, our method is required to use computer to perform the computation. Our triage results closely match doctors’ judgement. Sensitivity and specificity are improved significantly, comparing to traditional APACH II systems. Absolute and relative assessment gains were calculated and proved to be practical.