Predicting Missing Values in Medical Data via XGBoost Regression.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Biostatistics

Background:

• Laboratory test results are crucial for clinical investigation and medical research.
• Missing values in laboratory data limit the effectiveness of predictive models.
• Existing imputation methods have shown limited performance due to inadequate use of contextual information.

Purpose of the Study:

• To develop an advanced approach for imputing missing laboratory test values.
• To effectively leverage both individual and inter-variable contextual information for imputation.
• To enhance the utility of patient laboratory data for research and clinical applications.

Main Methods:

• A hybrid approach combining unsupervised prefilling with supervised extreme gradient boosting (XGBoost) was developed.
• The method leverages both longitudinal (temporal) and cross-sectional (inter-variable) contextual information.
• Experiments were conducted on approximately 8,200 patient records from the MIMIC-III dataset.

Main Results:

• The proposed model significantly outperformed baseline and state-of-the-art imputation methods.
• Improvements were observed across 13 commonly collected laboratory test variables.
• An average imputation improvement of over 20% in normalized root mean square deviation (nRMSD) was achieved.

Conclusions:

• Imputing missing temporal laboratory variables can be substantially enhanced using a prefilling strategy.
• Supervised training techniques effectively leverage both longitudinal and cross-sectional data context.
• The developed method offers a promising solution for addressing missing data challenges in clinical research.