Semisupervised Deep Learning Techniques for Predicting Acute Respiratory Distress Syndrome From Time-Series Clinical...

Area of Science:

• Computational biology
• Medical informatics
• Machine learning in healthcare

Background:

• Many hospitalized COVID-19 patients develop acute respiratory distress syndrome (ARDS).
• Clinical decision support is needed for pandemic management, especially with limited early data.

Purpose of the Study:

• Develop machine learning algorithms using semisupervised learning (SSL) to predict ARDS.
• Enable early ARDS detection in general and COVID-19 patient populations with scarce labeled data.

Main Methods:

• Applied SSL techniques to 29,127 patient encounters from 7 US hospitals (May 2019-May 2021).
• Utilized recurrent neural networks with time-series electronic health record data.
• Focused on peripheral oxygen saturation <97% to predict subsequent ARDS development.

Main Results:

• The median time from initial low oxygen to respiratory failure was 21 hours.
• SSL-augmented models improved ARDS prediction (AUC 0.78) compared to labeled data alone (AUC 0.73).
• The best model, trained on all data, achieved an AUC of 0.84.

Conclusions:

• Unlabeled data can significantly enhance machine learning model performance for ARDS prediction.
• SSL is valuable when labeled data for predicting ARDS is scarce or costly.
• This approach supports clinical decision-making during early pandemic stages.