Self-Training With Quantile Errors for Multivariate Missing Data Imputation for Regression Problems in Electronic...

Area of Science:

• Machine Learning
• Data Science
• Medical Informatics

Background:

• Missing data is a prevalent challenge in real-world machine learning applications.
• Existing imputation methods include statistical approaches (mean, expectation-maximization, MICE) and machine learning techniques (MLP, k-NN, decision trees).

Purpose of the Study:

• To impute numeric medical data, including physical and laboratory values.
• To develop an effective data imputation strategy using self-training for scarce medical data environments.

Main Methods:

• Proposed a progressive self-training method to gradually increase available data for model training.
• Employed pseudolabeling: models trained on complete data predict missing values, and valid predictions are incorporated back into the complete dataset.
• Iteratively repeated the prediction and incorporation process until a stopping condition was met, evaluating pseudolabel accuracy by its impact on model performance.

Main Results:

• Self-training with Random Forest (RF) demonstrated up to 12% lower mean squared error and 0.1% higher Pearson correlation coefficient compared to pure RF.
• Statistical tests (Friedman, Wilcoxon signed-rank) confirmed the significant improvement of self-training over Multiple Imputations by Chained Equations (MICE) and mean imputation (p < .05 and p = 3.05e-5, respectively).

Conclusions:

• Self-training shows statistically significant improvements in imputing missing values, particularly for medical datasets.
• Further validation in real-world machine learning systems and refinement of pseudolabel evaluation methods are warranted for future research.