Non-linear missing data imputation for healthcare data via index-aware autoencoders.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Healthcare data analytics offers opportunities for discovering hidden patterns to enhance clinical decision-making.
• Predictive models are vital for extracting insights from medical data, but missing values hinder their performance.
• Autoencoder models are commonly used for missing data imputation but have limitations.

Purpose of the Study:

• To evaluate the limitations of standard autoencoder models in medical data imputation.
• To propose and assess modified autoencoder models for improved imputation performance.
• To enhance the accuracy of disease classification using imputed medical data.

Main Methods:

• Developed and implemented modified autoencoder models for non-linear data imputation.
• Compared the proposed models against five established imputation techniques.
• Evaluated performance using six diverse medical datasets and five classification methods.
• Assessed imputation effectiveness across varying degrees of missing data ratios.

Main Results:

• The proposed non-linear imputation model demonstrated superior performance compared to existing methods.
• The model achieved the highest disease classification accuracy across all tested datasets.
• Performance improvements were consistent across all tested missing data ratios.

Conclusions:

• Modified autoencoder models offer a significant advancement for handling missing data in healthcare.
• Accurate data imputation using advanced autoencoders directly translates to improved clinical predictive modeling.
• This approach holds promise for enhancing diagnostic accuracy and clinical decision support systems.