Multivariate outlier detection applied to multiply imputed laboratory data.

Area of Science:

• Clinical laboratory science
• Biostatistics
• Data science

Background:

• Multivariate outlier detection identifies patients with unusual patterns in clinical lab data.
• Missing data in clinical datasets are often handled by single imputation, which can underestimate variability.
• Multiple imputation methods aim to address the limitations of single imputation.

Purpose of the Study:

• To evaluate multivariate outlier detection methods applied to multiply imputed clinical laboratory safety data.
• To compare the performance of different multiple imputation techniques in the presence of missing data.
• To assess the accuracy of outlier detection results based on various analysis methods.

Main Methods:

• Generating clinical laboratory data sets with varying proportions of missing data (4, 7, 12, 30 dimensions).
• Applying eight different multiple imputation methods to handle missing values.
• Utilizing Mahalanobis distance and generalized principal component analysis for outlier detection on imputed datasets.

Main Results:

• Performance of multivariate outlier detection techniques on multiply imputed data was analyzed.
• Comparison of eight multiple imputation methods was conducted across different missing data scenarios.
• Measures for assessing the accuracy of missing data results were introduced.

Conclusions:

• Multivariate outlier detection can be effectively applied to multiply imputed clinical laboratory data.
• The choice of multiple imputation method impacts the performance of outlier detection.
• Accurate assessment of missing data imputation is crucial for reliable clinical data analysis.