Handling missing values in healthcare data: A systematic review of deep learning-based imputation techniques

Area of Science:

• Healthcare Data Science
• Machine Learning in Medicine
• Biostatistics

Background:

• Effective handling of missing data is crucial for reliable clinical research and decision-making.
• Deep learning (DL) based imputation techniques are increasingly developed to address complex and diverse data.
• This review systematically evaluates DL imputation methods to guide healthcare researchers.

Approach:

• A systematic literature search was conducted across five major databases for DL imputation models.
• Articles were analyzed based on data types, model architectures, imputation strategies, and comparisons with non-DL methods.
• An evidence map was created to visualize DL model adoption across different data types.

Key Points:

• Tabular static (29%) and temporal (40%) data were most frequently addressed by DL imputation.
• Specific DL architectures like autoencoders and recurrent neural networks dominate certain data types.
• Integrated imputation strategies were prevalent for tabular temporal and multi-modal data.
• DL methods generally demonstrated superior imputation accuracy compared to non-DL approaches.

Conclusions:

• DL imputation models offer diverse network structures tailored to specific healthcare data types.
• While not universally superior, DL models can achieve satisfactory results for particular datasets.
• Ongoing challenges include the portability, interpretability, and fairness of current DL imputation techniques.