Classifying highly imbalanced ICU data.

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Data mining for clinical outcomes

Background:

• Intensive Care Unit (ICU) data often presents highly imbalanced datasets, where the outcome of interest (e.g., patient mortality) is rare.
• Accurate prediction of patient discharge status (alive or deceased) is critical for effective ICU management and resource allocation.

Purpose of the Study:

• To evaluate and compare the performance of common data mining methods in predicting patient mortality from imbalanced ICU datasets.
• To assess the impact of varying misclassification cost ratios (MCR) on the predictive accuracy of these methods.

Main Methods:

• Comparison of logistic regression, discriminant analysis, Classification and Regression Tree (CART), C5, and Support Vector Machines (SVM).
• Evaluation using metrics such as specificity, recall, precision, F-measure, and confusion entropy (CEN).
• Application of misclassification cost ratios (MCR) to address data imbalance and Hand's measure for overall method comparison.

Main Results:

• C5 and SVM demonstrated superior performance over other methods in classifying imbalanced ICU data.
• At a MCR of 100, C5 achieved the highest recall, while SVM exhibited the highest specificity and lowest confusion entropy (CEN).
• Logistic regression performed best according to Hand's measure, indicating its robustness across different evaluation criteria.

Conclusions:

• Utilizing misclassification cost ratios (MCR) significantly enhances the classification performance of data mining methods for imbalanced medical data.
• While MCR improves overall classification, the F-measure and precision did not show improvement with increased MCR values.
• The choice of data mining method and evaluation metric is crucial for accurately predicting rare events in clinical settings.