Learning statistical models of phenotypes using noisy labeled training data.

Area of Science:

• Computational Health Informatics
• Machine Learning in Healthcare
• Electronic Health Records Research

Background:

• Traditional patient phenotyping relies on labor-intensive rule-based definitions.
• Machine learning for electronic phenotyping is hindered by limited labeled training data.

Purpose of the Study:

• To demonstrate the feasibility of using semi-automatically labeled training sets for machine learning-based phenotype model creation.
• To develop an efficient alternative to manual data labeling for phenotyping.

Main Methods:

• Utilized a keyword list to generate noisy labeled training data for phenotypes.
• Trained L1 penalized logistic regression models on comprehensive patient medical record data.
• Evaluated model performance against a gold standard for chronic and acute diseases.

Main Results:

• Achieved high precision and accuracy for Type 2 diabetes mellitus (0.90, 0.89) and myocardial infarction (0.86, 0.89) models.
• Demonstrated the feasibility of learning phenotype models from imperfectly labeled data.
• Model performance was comparable to validated rule-based definitions.

Conclusions:

• The proposed method offers a viable alternative to manual labeling for creating statistical phenotype models.
• This approach can expedite research utilizing large observational healthcare datasets.
• The method facilitates the creation of local phenotype models efficiently.