Adaptive Multi-Wave Sampling for Efficient Chart Validation

Area of Science:

• Health Informatics
• Biostatistics
• Clinical Research Methods

Background:

• Healthcare data, including electronic health records and claims, are crucial for patient characterization and outcome identification.
• Computable phenotypes require validation through chart review studies to assess their measurement characteristics, such as positive predictive value.

Purpose of the Study:

• To develop and evaluate an adaptive method for validating computable phenotypes using healthcare data.
• To minimize the number of charts required for review while ensuring accurate measurement characteristics.

Main Methods:

• An adaptive sampling approach (Neyman's sampling) was compared against random and stratified random sampling.
• Frequentist confidence bands and Bayesian credible intervals were proposed for sequential evaluation of the quantity of interest.
• A tool was developed to predict the stopping time, i.e., the number of charts needed for review completion.

Main Results:

• Bayesian credible intervals demonstrated tighter intervals compared to frequentist confidence bands.
• Simple random sampling performed comparably to Neyman's sampling in efficiency.
• The adaptive approach effectively minimizes the number of charts reviewed.

Conclusions:

• Random sampling combined with Bayesian credible intervals offers an efficient strategy for validating outcomes in both binary and continuous data.
• This approach enhances the reliability of computable phenotypes derived from healthcare data.