JMIR Publications

ABSTRACT

Background:

A high number of patients hospitalized with COVID-19 also develop Acute Respiratory Distress Syndrome (ARDS).

Objective:

In response to the need for clinical decision support tools to help manage the pandemic, we developed machine learning algorithms to predict ARDS in general and COVID-19 populations.

Methods:

Semi-supervised machine learning (SSL) techniques were also applied to 29127 encounters from patients admitted to seven United States hospitals from 5/1/2019-5/1/2021. A recurrent neural network (RNN) using a time series of electronic health record (EHR) data was applied at the time peripheral oxygen saturation (SpO2) falls below the normal range (< 97%) to predict subsequent development of ARDS in the remaining duration of the hospital stay. Model performance was assessed with regard to area under the receiver operating characteristic (AUROC) and area under the precision recall curve (AUPRC) on an external hold out test set.

Results:

In the whole dataset the median time between the first SpO2 measurement < 97% and subsequent respiratory failure was 21 hours. The AUC for predicting subsequent ARDS was 0.73 when training on a labeled dataset of 6930 patients, 0.78 when training on the labeled dataset that had been augmented with the unlabeled dataset of 16,173 patients using SSL techniques and 0.84 when training on the entire training set of 23,103 labeled patients.

Conclusions:

In the setting of time series inpatient data, with careful model training design, unlabeled data can be used to improve the performance of machine learning when labeled data is scarce or expensive. Clinical Trial: N/A