Autopopulus: A Novel Framework for Autoencoder Imputation on Large Clinical Datasets.

Area of Science:

• Health Informatics
• Machine Learning
• Clinical Data Science

Background:

• Electronic health records (EHRs) increase patient data accessibility for clinical decision support.
• Missing data in EHRs pose significant challenges, potentially invalidating predictive models.
• Machine learning (ML) imputation methods offer a promising solution for estimating missing values.

Purpose of the Study:

• Introduce Autopopulus, a novel framework for designing and evaluating autoencoder architectures for efficient data imputation.
• Develop and assess ML-based imputation techniques for large-scale clinical datasets.
• Identify imputation methods that enhance the performance of downstream predictive models.

Main Methods:

• Developed Autopopulus, a framework implementing existing and novel autoencoder imputation methods.
• Included a new technique outputting a range of estimated values instead of point estimates.
• Demonstrated a workflow for selecting appropriate imputation methods based on user needs.

Main Results:

• Autopopulus enables efficient imputation on large clinical datasets.
• Identified imputation methods that accurately impute missing data.
• Determined imputation methods that optimize predictive model performance for chronic kidney disease (CKD) progression.

Conclusions:

• Autopopulus provides a robust framework for evaluating imputation strategies in clinical data.
• The choice of imputation method significantly impacts the accuracy of predictive models.
• This work facilitates informed decisions for handling missing data in EHRs to improve clinical predictions.