Scalable Electronic Phenotyping For Studying Patient Comorbidities.

Area of Science:

• Health Informatics
• Computational Biology
• Medical Data Science

Background:

• Millions of Americans live with multiple chronic conditions, complicating medical decision-making.
• Current methods for identifying patient cohorts with comorbidities lack scalability and generalizability.
• Retrospective cohort studies are the primary basis for treatment decisions in patients with multiple conditions.

Purpose of the Study:

• To develop a supervised machine learning algorithm for learning comorbidity phenotypes.
• To enable scalable and generalizable patient cohort generation without manual training sets.
• To evaluate the impact of training data characteristics on model performance.

Main Methods:

• A supervised machine learning algorithm was developed to learn comorbidity phenotypes.
• Myocardial infarction (MI) and type-2 diabetes (T2DM) patient cohorts were generated using ICD9 codes.
• LASSO logistic regression models were trained and assessed based on training sample size, physician input, and clinical text features.

Main Results:

• The algorithm achieved performance comparable to keyword-based labeling using only ICD9 codes.
• Increased training sample size and inclusion of physician input compensated for the absence of clinical text features.
• The best-performing model incorporated clinical text features alongside a large training sample size.

Conclusions:

• Machine learning offers a scalable and generalizable solution for identifying patient cohorts with comorbidities.
• The proposed algorithm effectively learns comorbidity phenotypes, reducing reliance on manual annotation.
• Optimizing training data, including clinical text and sufficient sample size, enhances model performance for complex conditions.