Propensity score prediction for electronic healthcare databases using Super Learner and High-dimensional Propensity...

Area of Science:

• Pharmacoepidemiology
• Health Informatics
• Statistical Learning

Background:

• The optimal prediction model varies with data distribution.
• Super Learner (SL) is an ensemble method using cross-validation to select models.
• SL's performance in large electronic healthcare databases needs thorough evaluation.

Purpose of the Study:

• Evaluate Super Learner (SL) for propensity score (PS) prediction in electronic healthcare databases.
• Compare SL performance against individual prediction models.
• Introduce and assess a novel strategy combining SL with high-dimensional propensity score (hdPS).

Main Methods:

• Applied SL to three electronic healthcare databases for PS prediction.
• Utilized a library of parametric and nonparametric candidate models.
• Combined SL with the hdPS algorithm for a novel prediction strategy.
• Assessed predictive performance using negative log-likelihood, AUC, and time complexity.

Main Results:

• Optimal individual algorithms varied across datasets.
• SL adapted to datasets, improving predictive performance over single models.
• The SL-hdPS combination demonstrated consistent prediction performance.
• AUC and negative log-likelihood metrics indicated varying performance across models.

Conclusions:

• Super Learner (SL) effectively predicts propensity scores in large healthcare databases.
• The combination of SL with hdPS is a promising and consistent method for PS estimation.
• This approach holds potential for prediction modeling in pharmacoepidemiology and comparative effectiveness research.