Development of Type 2 Diabetes Mellitus Phenotyping Framework Using Expert Knowledge and Machine Learning Approach

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Research Informatics

Background:

• Automated phenotyping from electronic health records (EHRs) is crucial for improving medical care and type 2 diabetes mellitus (T2DM) research.
• Existing T2DM phenotyping algorithms often lack the accuracy needed for effective screening or identifying clinical research participants.
• There is a growing demand for precise T2DM phenotyping methods to advance research and patient management.

Purpose of the Study:

• To develop a practical phenotyping framework integrating expert knowledge and machine learning for T2DM patient identification.
• To create two distinct phenotyping algorithms: one optimized for screening and another for identifying research subjects.
• To enhance the accuracy and utility of phenotyping algorithms for T2DM research and clinical applications.

Main Methods:

• A hybrid approach combining rule-based expert knowledge for excluding clear controls and machine learning for complex cases.
• Development of binary classification algorithms to accurately determine T2DM status in patients.
• Introduction of novel evaluation metrics, including area under the precision-sensitivity curve (AUPS) tailored for high sensitivity and high positive predictive value.

Main Results:

• Phenotyping algorithms developed using the proposed framework demonstrated superior performance compared to existing baseline algorithms.
• The framework successfully generated two types of algorithms, adaptable for either screening or research subject identification based on tuning.
• The developed algorithms proved effective in extracting T2DM patients for retrospective studies.

Conclusions:

• A novel, easily implementable phenotyping framework was developed, supported by appropriate evaluation metrics aligned with user objectives.
• The phenotyping algorithms derived from this framework are valuable tools for identifying T2DM patients in retrospective research settings.
• This approach offers a significant advancement in the practical application of phenotyping for T2DM research and clinical data analysis.