Comparative analysis, applications, and interpretation of electronic health record-based stroke phenotyping methods

Area of Science:

• Biomedical informatics
• Clinical research methodology
• Artificial intelligence in healthcare

Background:

• Accurate identification of acute ischemic stroke (AIS) patient cohorts is critical for clinical investigations.
• Current phenotyping methods are laborious and lack generalizability.
• Electronic health records (EHRs) offer a new approach for automated cohort identification.

Purpose of the Study:

• To systematically compare machine learning algorithms and case-control combinations for phenotyping AIS patients using EHR data.
• To evaluate the generalizability and accuracy of automated phenotyping methods.
• To assess the utility of diagnosis codes in training classifier models.

Main Methods:

• Developed and evaluated machine learning models using structured EHR data from a tertiary-care hospital system.
• Tested 75 different case-control and classifier combinations for AIS patient identification.
• Externally validated models using UK Biobank data to detect AIS patients without diagnosis codes.

Main Results:

• Machine learning models achieved a mean AUROC of 0.963 and average precision of 0.790 with minimal feature processing.
• Classifiers trained with AIS diagnosis codes and controls without cerebrovascular disease codes yielded the best F1 score (0.832).
• External validation showed significant enrichment (60-150 fold) of AIS patients without diagnosis codes in top-predicted cohorts.

Conclusions:

• Machine learning algorithms provide a generalizable and accurate method for identifying AIS patients without manual feature curation.
• Diagnosis codes can be effectively used to train classifier models when a specific AIS patient set is unavailable.