Enhancing Medical Data Imputation Using a Denoising Autoencoder with Missing-Neighborhood Perturbation
IEEE Transactions on Nanobioscience
|July 16, 2026
Summary
This study introduces Multiple Imputation based on Neighborhood Perturbation Denoising Autoencoder (MI_NPDAE) for accurate imputation of missing clinical data. MI_NPDAE effectively handles complex missingness patterns, outperforming existing methods and improving downstream predictive tasks.
Area of Science:
- Medical Informatics
- Machine Learning
- Data Science
Background:
- Missing data in clinical datasets presents challenges for accurate analysis.
- Non-random missingness, where data is not missing by chance, reduces the efficacy of standard imputation models.
Purpose of the Study:
- To develop a robust and generalized imputation method for tabular medical data.
- To address the limitations of existing models in handling diverse and non-random missing data patterns.
Main Methods:
- Proposed Multiple Imputation based on Neighborhood Perturbation Denoising Autoencoder (MI_NPDAE).
- Leveraged neighborhood information to identify optimal donor records.
- Employed denoising autoencoders with perturbed inputs and additive noise for robust feature learning and adaptability to various missingness patterns.
Main Results:
- MI_NPDAE demonstrated superior performance compared to baseline methods across different missing data mechanisms and ratios.
- Achieved consistently lower imputation errors on both public and clinical datasets.
- Imputed data significantly enhanced the performance of downstream predictive tasks.
Conclusions:
- MI_NPDAE offers a robust and adaptable solution for imputing missing values in complex clinical datasets.
- The method shows practical value in improving the accuracy of clinical data analysis and predictive modeling.