Machine Learning-Based Asthma Attack Prediction Models From Routinely Collected Electronic Health Records: Systematic...

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Respiratory Medicine

Background:

• Electronic health records (EHRs) offer valuable historical patient data for developing predictive tools.
• Machine learning (ML) coupled with EHRs presents an opportunity to create early warning systems for asthma attacks.
• Existing studies have explored ML-based tools for asthma attack prediction.

Purpose of the Study:

• To critically evaluate machine learning (ML) models developed using EHRs for predicting asthma attacks.
• To assess the methodologies, performance, and reporting of ML models in asthma attack prediction.

Main Methods:

• Systematic literature search of PubMed and Scopus (2012-2023) for studies using EHRs to predict asthma attacks.
• Inclusion criteria focused on EHR data, asthma attack outcomes, and comparative ML model performance.
• Excluded non-English, non-research, and papers lacking methodological detail; summarized studies on data preprocessing, algorithms, validation, explainability, and implementation.

Main Results:

• 17 studies met the inclusion criteria, revealing significant heterogeneity in asthma attack definitions.
• Extreme class imbalance was common (76%), with only 38% addressing it.
• Gradient boosting was the top-performing ML method (59%); 82% used model explanation, but no studies followed reporting guidelines or underwent prospective validation.

Conclusions:

• The field of ML-based asthma attack prediction using EHRs is underdeveloped.
• Key challenges include data heterogeneity, class imbalance, lack of external validation, and suboptimal reporting.
• Addressing these technical challenges is crucial for advancing clinical adoption of these predictive tools.