MVIRA: A model based on Missing Value Imputation and Reliability Assessment for mortality risk prediction

Area of Science:

• Healthcare Informatics
• Machine Learning in Medicine
• Predictive Analytics

Background:

• Accurate mortality risk prediction from electronic health records (EHR) is crucial in healthcare.
• Missing values in medical datasets significantly hinder predictive model performance.
• Existing imputation methods often neglect the fidelity and confidence of interpolated values, impacting model reliability.

Purpose of the Study:

• To develop a novel model, MVIRA, for enhanced mortality risk prediction from EHR data.
• To address the challenge of missing data in EHR by incorporating imputation and reliability assessment.
• To improve the accuracy and trustworthiness of patient mortality risk predictions in clinical settings.

Main Methods:

• Proposed the Missing Value Imputation and Reliability Assessment (MVIRA) model.
• Utilized a combination of variational autoencoder and recurrent neural networks for missing value imputation.
• Incorporated Monte Carlo Dropout for uncertainty quantification and reliability assessment of predictions.

Main Results:

• Validated the MVIRA model on the MIMIC-III and MIMIC-IV public datasets.
• Demonstrated superior performance of the MVIRA model compared to existing methods.
• Showcased improved accuracy in mortality risk prediction across various metrics.

Conclusions:

• The MVIRA model effectively enhances the accuracy of mortality risk prediction.
• The model provides a reliable assessment of prediction uncertainty.
• MVIRA can assist medical institutions in better assessing patient conditions and risks.