Matched Learning for Optimizing Individualized Treatment Strategies Using Electronic Health Records

Area of Science:

• Biostatistics
• Machine Learning
• Health Informatics

Background:

• Current treatment guidelines rely on randomized controlled trials (RCTs) focusing on average effects, which may not suffice for personalized medicine.
• Electronic health records (EHR) offer vast potential for individualized treatment decisions but present analytical challenges.
• Personalized medicine aims to tailor treatments based on patient-specific characteristics, moving beyond one-size-fits-all approaches.

Purpose of the Study:

• To propose and evaluate a novel machine learning approach, M-learning, for estimating individualized treatment rules (ITRs) from electronic health records (EHR).
• To address limitations of existing methods in handling confounding and accurately assessing individual treatment responses using real-world data.
• To develop a flexible framework accommodating various outcome types for personalized treatment effect estimation.

Main Methods:

• Developed M-learning, a machine learning method based on matching for estimating optimal ITRs from EHR data.
• Introduced matching-based value functions to compare matched individuals, accommodating continuous, ordinal, and discrete outcomes.
• Established theoretical properties of M-learning, including Fisher consistency and convergence rates.

Main Results:

• M-learning demonstrated superior performance compared to existing methods in simulation studies, particularly when propensity scores were misspecified or unmeasured confounders were present.
• The method effectively handles confounding by employing a matching strategy instead of inverse probability weighting.
• M-learning successfully estimated personalized second-line treatment rules for type 2 diabetes patients using real-world EHR data.

Conclusions:

• M-learning provides a robust and flexible framework for deriving individualized treatment rules from electronic health records.
• The proposed method enhances the potential of personalized medicine by improving the accuracy of treatment effect estimation in real-world settings.
• Application to type 2 diabetes suggests M-learning's utility in optimizing patient care and outcomes.