JMIR Publications

ABSTRACT

Background:

Sepsis is a life-threatening disease that early detection and prompt treatment are essential. Several studies developed models for predicting sepsis with deep learning, which showed higher performance for predicting sepsis. However, the true label of sepsis is obscure, whether it could be used as true infection.

Objective:

This study aims to evaluate the performance of continuous monitoring for early detection of bacteremia among in-hospital patients with a deep learning model

Methods:

As a retrospective cohort study, our dataset consecutively includes 36023 patients, who underwent general surgeries from October 2007 to December 2017 at tertiary referral hospital in South Korea. The primary outcome is the area under the receiver operating curve (AUROC) and area under the precision-recall curve (AUPRC) of detecting bacteremia by deep learning model, and the secondary outcome is the feature explainability of the model by occlusion analysis.

Results:

A total of 36023 patients were included in dataset in which 720 case of bacteremia were included. Our model showed an AUROC of 0.97 (95% CI : 0.974 - 0.981), and the AUPRC of 0.17 (95% CI: 0.147 - 0.203) for detecting bacteremia in surgical in-patients. For predicting bacteremia within the previous 24 hour period, the AUROC and AUPRC values were 0.93 and 0.15, respectively. Occlusion analysis showed that vital signs and laboratory measurements(e.g. kidney function test and white blood cell group) were the most important variables for detecting bacteremia.

Conclusions:

Deep learning model based on time-series electronic health records data had a high detective ability for bacteremia in surgical in-patients. The model may be able to assist clinicians in evaluating infection among in-patients, ordering blood cultures, and prescribing antibiotics with a real-time monitoring.