Machine learning approaches for electronic health records phenotyping: a methodical review

Area of Science:

• Biomedical Informatics
• Computational Biology
• Health Services Research

Background:

• Accurate phenotyping is crucial for utilizing electronic health records (EHRs) in biomedical research.
• Machine learning (ML) offers a scalable alternative to traditional rule-based phenotyping algorithms.
• This study systematically reviews ML-based phenotyping approaches in recent literature.

Purpose of the Study:

• To evaluate data sources, phenotypes, methods, and reporting/evaluation practices in ML-based phenotyping.
• To assess the performance of ML algorithms compared to rule-based methods.
• To identify gaps and future directions in ML-based phenotyping research.

Main Methods:

• Systematic literature search of PubMed and Web of Science (2018-2022).
• Screening of 850 articles, extracting 37 variables from 100 studies.
• Categorization of studies based on data sources, phenotypes, ML methods, and evaluation strategies.

Main Results:

• Most studies used single-institution data, including clinical notes.
• ML facilitated characterization of chronic conditions and nuanced phenotypes like social determinants of health.
• Supervised deep learning was prevalent; ML did not uniformly outperform rule-based methods, but deep learning showed marginal improvement.

Conclusions:

• Progress in ML-based phenotyping is evident, yet challenges remain.
• Focus on nuanced phenotypes, external validity, and multi-institution data is needed.
• Developing standards for reporting, evaluation, and handling algorithm errors is essential for advancing the field.