Automatically Detecting Acute Myocardial Infarction Events from EHR Text: A Preliminary Study.

Area of Science:

• Cardiology
• Medical Informatics
• Machine Learning

Background:

• The Worcester Heart Attack Study (WHAS) monitors acute myocardial infarction (AMI) trends.
• Manual data assessment in WHAS is time-consuming.
• Automating AMI data extraction from electronic health records (EHR) is needed.

Purpose of the Study:

• To develop supervised machine learning (ML) models for automated AMI detection in EHR.
• To address data sparseness challenges in ML models for cardiovascular surveillance.

Main Methods:

• Annotated 105 EHR discharge summaries for AMI information with high inter-annotator agreement (Cohen's κ > 0.74).
• Applied Conditional Random Fields (CRFs), a state-of-the-art supervised ML model, for AMI detection.
• Evaluated various feature engineering approaches, including cluster-based word features.

Main Results:

• Achieved high agreement in manual annotation of EHR data.
• Conditional Random Fields (CRFs) demonstrated effectiveness for AMI detection.
• Cluster-based word features yielded the highest performance in AMI detection models.

Conclusions:

• Supervised ML, particularly CRFs with cluster-based features, can automate AMI detection in EHR.
• This approach can enhance the efficiency and accuracy of population-based cardiovascular surveillance.
• Automated methods support ongoing analysis of acute myocardial infarction trends.