Predicting Hypoxia Using Machine Learning: Systematic Review.

Area of Science:

• Medical Informatics
• Clinical Prediction Models
• Artificial Intelligence in Healthcare

Background:

• Hypoxia is a critical risk factor and indicator of declining inpatient health.
• Predicting hypoxic events is crucial for timely interventions and preventing patient deterioration.

Purpose of the Study:

• To systematically review and compare machine learning models for predicting hypoxic events in hospitalized patients.
• To analyze the methodology, predictive performance, and patient populations of existing studies.

Main Methods:

• Systematic literature search across major databases (Web of Science, Embase, MEDLINE, Google Scholar).
• Inclusion of studies using machine learning for hypoxia/hypoxemia prediction in hospitalized patients.
• Risk of bias assessment using the Prediction Model Risk of Bias Assessment Tool.

Main Results:

• 12 papers and 32 models were analyzed, revealing diverse methodologies and populations.
• Most studies (83%) had unclear or high risk of bias, limiting comparability.
• Overall predictive performance was moderate to high; deep learning models performed comparably or better than conventional ML.
• Models using only peripheral oxygen saturation, often with long short-term memory (LSTM), showed superior performance.

Conclusions:

• Machine learning models can accurately predict hypoxic events using retrospective data.
• Study heterogeneity and bias necessitate further validation studies for generalizability and reliable predictive performance assessment.