An Expedited Chart Review Process for Large Database Studies Using Natural Language Processing and Multiwave Adaptive...

Area of Science:

• Health Informatics
• Biostatistics
• Data Science in Healthcare

Background:

• Validating code-based algorithms in large claims databases is crucial for enhancing analysis.
• Manual chart review of electronic health records is time-consuming and resource-intensive.
• Outcome misclassification can bias results in inferential studies.

Purpose of the Study:

• To describe an expedited process for validating code-based algorithms.
• To introduce efficiency through natural language processing (NLP) and adaptive sampling.
• To illustrate the process with a case study validating an algorithm for intentional self-harm in patients with obesity.

Main Methods:

• Utilized natural language processing (NLP) to reduce human reviewer time per chart.
• Implemented a multi-wave adaptive sampling approach with pre-defined stopping criteria.
• Validated a claims-based outcome algorithm for intentional self-harm in an obesity cohort.

Main Results:

• The NLP-assisted annotation reduced review time per chart by 40%.
• Adaptive sampling with a stopping rule would have avoided reviewing 77% of patient charts.
• Sufficient precision of performance characteristics was maintained with limited compromise.

Conclusions:

• The described approach facilitates routine validation of code-based algorithms.
• This enhances the understanding of reliability in findings from database studies.
• Improves the robustness of health research using large datasets.