JMIR Publications

ABSTRACT

Background:

The initial timely decision for the patient is important for acute illness. However, there are only few studies that determined the prognosis of patients with insufficient laboratory data at the initial stage of treatment.

Objective:

To develop and validate time adaptive prediction models for patients’ severity of illness in the emergency department (ED) using highly sparse laboratory test data (order status and test result) based on the machine learning approach.

Methods:

This is a retrospective study and used emergency department data from a tertiary academic hospital in Seoul, Korea. Two different models were developed with laboratory test data: the order status only (OSO) model and order status and result (OSR) model. The binary composite adverse outcome was used, indicating whether patients died or were hospitalized in the intensive care unit. The models were evaluated by various performance criteria, including the area under the receiver operating characteristic curve (AUC) and balanced accuracy (BA). Clinical usefulness was examined by determining the positive and negative likelihood ratios (LR+ and LR-).

Results:

Of 5,622 eligible patients in the ED (mean age [standard deviation], 55.1 [17.6] years; 2,913 [51.8%] women), the model development cohort and validation cohort included 3,936 and 1,686 patients, respectively. The OSR (AUC=0.85, BA=0.79) model generally had better performance than the OSO model (AUC=0.75, BA=0.72), but the difference in performance was not significant in specificity (0; P=.36 for difference). The OSR model was more informative than the OSO model in patient’s predicted outcome with low risk and high risk (P< .001 for difference in both LR- and LR+).

Conclusions:

The adverse outcome of febrile patients at the early stage could be predicted using highly sparse order status and results of laboratory test data through the machine learning based models. This prediction tool has its value for sharing information, leading dynamic communication, and preventing medical errors in consequence while treating the same patient simultaneously.