A Comparative Study on Imputation Techniques: Introducing a Transformer Model for Robust and Efficient Handling of...

Area of Science:

• Clinical Data Science
• Machine Learning for Healthcare
• Biomedical Signal Processing

Background:

• Missing data in clinical datasets is a common issue, arising from non-response, data errors, or processing mistakes.
• Inaccurate imputation of missing values can lead to biased analyses, reduced statistical power, and unreliable findings.
• Traditional methods like Zero, Mean, and k-Nearest Neighbors (KNN) imputation often fail to capture complex data patterns.

Purpose of the Study:

• To develop and evaluate a novel imputation model using transformer-based architectures for clinical datasets.
• To address the limitations of traditional imputation methods in handling complex data, specifically EEG signal amplitude data.
• To improve the accuracy and reliability of clinical data imputation.

Main Methods:

• A novel imputation model employing transformer-based architectures (TabTransformer) was developed.
• The model was trained exclusively on complete electroencephalography (EEG) signal amplitude data from PhysioNet and CHB-MIT datasets.
• The model's performance was evaluated against traditional methods (Zero, Mean, KNN) using error metrics and R2 scores.

Main Results:

• The proposed transformer-based model achieved high accuracy in predicting missing EEG amplitude data.
• Achieved R2 scores of 0.993 for the PhysioNet dataset and 0.97 for the CHB-MIT dataset.
• Demonstrated significant performance improvements over Zero, Mean, and KNN imputation methods.

Conclusions:

• Transformer-based imputation models offer a powerful approach to handling missing data in complex clinical datasets.
• The proposed model effectively captures intricate patterns in EEG amplitude data, enhancing dataset integrity.
• This advancement holds transformative potential for improving the utility and reliability of clinical data analysis.