A deep learning-based, unsupervised method to impute missing values in electronic health records for improved patient...

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Machine Learning

Background:

• Missing values in electronic health records (EHRs) pose a significant challenge for accurate patient management and clinical decision-making.
• Deep learning offers advanced techniques for addressing data imputation issues in complex healthcare datasets.

Purpose of the Study:

• To develop and evaluate a novel deep learning-based unsupervised method for imputing missing values in EHRs.
• To assess the imputation effectiveness and predictive efficacy of the proposed method for peritonitis patient management.
• To compare the performance of the new method against prevalent benchmark techniques.

Main Methods:

• A deep autoencoder framework was utilized, incorporating missing data patterns, patient data relationships, and temporal dynamics.
• A novel loss function was developed for error calculation and regularization.
• A comparative evaluation was conducted using a dataset of 27,327 patient records against benchmark imputation methods.

Main Results:

• The proposed deep learning method demonstrated superior imputation performance, achieving 5.3-15.5% lower imputation errors compared to benchmark techniques.
• Imputed data from the proposed method significantly improved the prediction of patient readmission, length of stay, and mortality (2.7-11.5% improvement).
• The method showed viability, effectiveness, and clinical decision support utility.

Conclusions:

• The developed deep learning method effectively imputes missing values in EHRs, reducing imputation biases.
• The approach enhances predictive accuracy for critical patient outcomes, supporting clinical decision-making.
• This method offers a valuable tool for analyzing and utilizing EHRs in various clinical scenarios.