A model-agnostic framework for dataset-specific selection of missing value imputation methods in pain-related...

Area of Science:

• Biomedical Data Science
• Statistical Modeling
• Pain Research Informatics

Background:

• Missing value imputation is crucial in biomedical data analysis but lacks dataset-specific optimization.
• Existing imputation techniques are often not tailored to the unique characteristics of cross-sectional numerical data.

Purpose of the Study:

• To propose a systematic framework for selecting optimal imputation methods for cross-sectional numerical biomedical data.
• To develop a method for assessing imputation accuracy using artificially generated missing values and novel diagnostic references.

Main Methods:

• Generated artificial missing values to assess reconstruction accuracy of various imputation algorithms.
• Introduced 'poisoned' (biased) and 'calibrating' (noisy) imputation methods for objective evaluation.
• Tested the framework on synthetic and four biomedical datasets (pain-focused) using 29 imputation methods.

Main Results:

• The framework successfully identified the most suitable imputation technique for each dataset.
• Multivariate imputation methods generally demonstrated superior performance compared to univariate approaches.
• Quantifiable thresholds for acceptable imputation errors were established using poisoned and calibrated references.

Conclusions:

• The proposed framework provides practical, reproducible guidelines for imputing missing values in biomedical research, particularly pain studies.
• It enhances data integrity and analytical robustness by enabling informed imputation method selection.
• The model-agnostic framework is available as the open-source R package 'opImputation'.