Classification of patients by severity grades during triage in the emergency department using data mining methods

Area of Science:

• Medical Informatics
• Machine Learning
• Emergency Medicine

Background:

• Accurate patient severity classification in emergency departments is crucial for resource allocation and treatment prioritization.
• Traditional methods often rely on subjective physician assessments, leading to potential variability.

Purpose of the Study:

• To evaluate the feasibility of using data mining techniques for classifying emergency department patient severity.
• To compare the performance of machine learning algorithms against physician assessments and baseline classifiers.

Main Methods:

• Applied Naïve Bayes and C4.5 algorithms to classify 402 emergency department patient records into five severity grades.
• Compared classifier accuracy against physician assessments, random classification, and maximal-prevalence class selection.
• Utilized metrics including positive predictive value, sensitivity, specificity, and entropy change.

Main Results:

• Data mining classifiers achieved a mean accuracy of 52.94%, significantly better than random classification (34.60%).
• Accuracy increased to 85.42% when allowing for a one-grade deviation, matching physician consensus.
• Naïve Bayes showed better performance on unbalanced datasets, and learning from physician consensus improved results.

Conclusions:

• Computerized classification models for emergency department patient severity are feasible using data mining.
• Consensus-driven physician data enhances model performance compared to individual physician data.
• Naïve Bayes or C4.5 can be employed, with Naïve Bayes preferred for unbalanced datasets. Ambiguity in intermediate grades affects both physician and classifier accuracy.