Outlier Removal in Model-Based Missing Value Imputation for Medical Datasets.

Area of Science:

• Data Science
• Machine Learning
• Medical Informatics

Background:

• Real-world medical datasets frequently exhibit missing attribute values, necessitating imputation techniques.
• Standard missing value imputation can be compromised by noisy data or outliers, leading to unreliable estimations.
• Instance selection is explored as a preprocessing step to potentially enhance imputation accuracy.

Purpose of the Study:

• To evaluate if combining instance selection with missing value imputation outperforms imputation alone in medical datasets.
• To identify optimal combinations of instance selection and imputation algorithms for improved data quality.

Main Methods:

• Utilized three instance selection algorithms: DROP3, Genetic Algorithm (GA), and Instance-Based 3 (IB3).
• Employed three imputation algorithms: K-Nearest Neighbors Imputation (KNNI), Multilayer Perceptron (MLP), and Support Vector Machine (SVM).
• Experimentally compared the performance of combined methods against imputation alone across different medical data types.

Main Results:

• Instance selection positively impacted missing value imputation for numerical medical datasets.
• Specific combinations of instance selection and imputation methods improved results for mixed data types.
• Instance selection did not consistently benefit imputation for categorical medical datasets.

Conclusions:

• Combining instance selection with missing value imputation is a viable strategy to enhance data quality in medical datasets.
• The effectiveness of instance selection depends on the data type (numerical, mixed, categorical) and the chosen algorithms.
• Further research may explore tailored approaches for categorical data to leverage instance selection benefits.