Sleep apnea test prediction based on Electronic Health Records.

Area of Science:

• Health Informatics
• Machine Learning in Healthcare
• Predictive Analytics

Background:

• Obstructive Sleep Apnea (OSA) diagnosis typically occurs later in life via Polysomnography.
• Early identification of individuals at risk for OSA is desirable for timely intervention.
• Electronic Health Records (EHR) offer a rich data source for developing predictive models.

Purpose of the Study:

• To develop predictive models using EHR data to identify individuals likely to undergo sleep apnea testing after age 50.
• To investigate the impact of EHR data variability and control matching on prediction accuracy.
• To evaluate the effectiveness of temporal variable selection methods and subgroup modeling.

Main Methods:

• Utilized Electronic Health Records (EHR) data to predict future sleep apnea testing.
• Developed and compared four classifiers: 1-CNN, LSTM, Random Forest, and Logistic Regression.
• Introduced the RankLi method for temporal variable selection based on t-test divergence scores.
• Investigated control matching strategies based on the number of EHR records and subgroup modeling.

Main Results:

• Control matching using the number of EHR records was found to be crucial for model performance.
• Modeling separate subgroups based on EHR record count improved prediction effectiveness.
• Deep learning models, specifically 1-CNN, demonstrated superior balanced accuracy and AUC.
• In males, the 1-CNN model achieved 90% balanced accuracy and 93% AUC at age 50 with 100 temporal variables.

Conclusions:

• Predictive modeling using EHR data can facilitate early identification of individuals at risk for Obstructive Sleep Apnea.
• The number of EHR records significantly impacts control matching and predictive model performance.
• Deep learning approaches, particularly 1-CNN, show strong potential for accurate OSA risk prediction in diverse populations.