Relational machine learning for electronic health record-driven phenotyping.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Computational Biology

Background:

• Electronic health records (EHRs) contain vast, complex patient data crucial for research.
• Developing phenotyping algorithms from EHRs is time-consuming and requires clinical expertise.
• Existing methods often struggle with the relational nature of EHR data.

Purpose of the Study:

• To evaluate the efficacy of relational machine learning (ML), specifically inductive logic programming (ILP), for EHR-based phenotyping.
• To determine if ILP can automate and improve the identification of patient phenotypes from complex EHR data.
• To compare ILP's performance against traditional non-relational ML algorithms.

Main Methods:

• Developed phenotypic models for nine conditions using two ILP approaches and three WEKA-based non-relational ML algorithms (PART, J48, JRIP).
• Utilized International Classification of Diseases, Ninth Revision (ICD-9) codes for automated training data labeling.
• Assessed model performance using accuracy, precision, recall, F-Measure, and Area Under the Receiver Operating Characteristic (AUROC) curve on independent test cohorts.

Main Results:

• Successfully developed an automated approach for labeling training examples using ICD-9 codes.
• ILP-based models demonstrated superior overall performance in AUROC compared to non-relational methods (PART, J48, JRIP) with statistically significant p-values.
• Nine phenotypic models were evaluated for each machine learning approach.

Conclusions:

• Relational learning with ILP presents a viable and effective method for EHR-driven phenotyping.
• ILP can enhance phenotyping by generating clinically interpretable rules and assisting experts in defining intuitive phenotypes.
• This approach holds potential for improving patient risk stratification and medical research.