Learning Personalized Treatment Rules from Electronic Health Records Using Topic Modeling Feature Extraction.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Personalized Medicine

Background:

• Individualized treatment rules (ITRs) aim to tailor therapies for chronic disorders, but rules learned from randomized controlled trials (RCTs) often lack real-world generalizability.
• Electronic health records (EHRs) offer valuable data but present challenges like confounding and selection bias for learning valid ITRs.
• Precision medicine requires robust methods to derive ITRs from diverse patient populations.

Purpose of the Study:

• To develop and validate a novel machine learning method for estimating optimal ITRs from EHR data.
• To address confounding and selection bias inherent in observational EHR studies.
• To improve the generalizability and clinical utility of ITRs for personalized treatment strategies.

Main Methods:

• A matching-based machine learning approach was employed to estimate ITRs from EHRs.
• Latent Dirichlet Allocation (LDA) was used to extract interpretable features (topics and weights) from medication and diagnosis codes.
• The method incorporated matching for confounding reduction and LDA features for enhanced treatment optimization.

Main Results:

• The proposed method successfully estimated ITRs from EHR data, outperforming uniform treatment strategies in cross-validation.
• The inclusion of LDA-based features led to a greater reduction in post-treatment complications for type 2 diabetes (T2D) patients.
• The approach demonstrated improved treatment optimization by augmenting the feature space with clinically relevant topics.

Conclusions:

• Machine learning methods, particularly those incorporating topic modeling on EHR data, can effectively generate generalizable ITRs.
• This approach offers a promising avenue for advancing precision medicine in chronic disease management.
• The developed method provides a robust framework for learning optimal treatment strategies from real-world data while mitigating biases.