Adjusting Covariate Misclassification in Electronic Health Records-Based Machine Learning Prediction Models

Area of Science:

• Biomedical Informatics
• Health Services Research
• Machine Learning in Healthcare

Background:

• Electronic health records (EHRs) contain valuable data but are prone to misclassification errors.
• These errors can introduce bias into predictive models, affecting clinical decision-making.
• Accurate prediction of patient outcomes, such as adherence to lung cancer screening, is crucial.

Purpose of the Study:

• To develop and evaluate methods for adjusting misclassification errors in EHR-derived covariates.
• To reduce bias in predictive modeling using group-wise and individualized weights.
• To enhance the accuracy of predictive models for lung cancer screening adherence.

Main Methods:

• Developed methods using group-wise and individualized weights based on sensitivity and specificity to adjust EHR covariates.
• Applied logistic regression, XGBoost, and neural networks to predict lung cancer screening adherence.
• Utilized natural language processing (NLP) for Lung-RADS category extraction and adjusted it using kernel and multinomial regression weights.
• Compared adjusted models against unadjusted (naïve) and true value (oracle) models using Area Under the Receiver Operating Characteristic (AUROC) curves.

Main Results:

• Adjusted models significantly outperformed naïve models across various validation set sizes (10%, 20%, 30%).
• AUROC improvements ranged from 0.3% to 10.4% compared to naïve models.
• Adjusted models reduced the performance gap compared to oracle models to 2.0%-7.5%.
• Individualized weights demonstrated more precise error correction than group-wise weights.

Conclusions:

• The developed framework effectively mitigates misclassification bias in EHR-derived covariates.
• The methods enhance predictive accuracy for lung cancer screening adherence without requiring extensive manual data review.
• This approach offers a scalable solution for improving the reliability of predictive models using real-world clinical data.