Missing data techniques in classification for cardiovascular dysautonomias diagnosis.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Data Mining

Background:

• Missing data (MD) is prevalent in e-health datasets, necessitating pre-processing for data mining (DM) decision systems.
• Effective MD handling is crucial for reliable diagnostic tools in healthcare.

Purpose of the Study:

• To evaluate the impact of different missing data techniques on classification system accuracy for cardiovascular dysautonomias diagnosis.
• To compare deletion versus k-nearest neighbors (KNN) imputation for handling missing values.

Main Methods:

• Compared four classifiers: Random Forest (RF), Support Vector Machines (SVM), C4.5 decision tree, and Naive Bayes (NB).
• Utilized two MD techniques: deletion and KNN imputation.
• Conducted 216 experiments across three missingness mechanisms (MCAR, MAR, NMAR) and varying MD percentages (10-90%) on an autonomic nervous system (ANS) dataset.

Main Results:

• K-nearest neighbors (KNN) imputation significantly enhanced the accuracy rates of all four classifiers compared to data deletion.
• Increasing percentages of missing data negatively impacted classification performance across all tested mechanisms and techniques.
• Accuracy rates consistently decreased as the percentage of missing data increased.

Conclusions:

• KNN imputation is a superior strategy to deletion for improving classification accuracy in cardiovascular dysautonomia diagnosis.
• Minimizing missing data is critical for optimizing the performance of machine learning models in clinical decision support systems.