A preclustering-based ensemble learning technique for acute appendicitis diagnoses.

Area of Science:

• Medical Informatics
• Machine Learning
• Diagnostic Accuracy

Background:

• Acute appendicitis diagnosis is challenging due to difficulties in timely and accurate detection.
• Imbalanced class distribution in training data poses a significant hurdle for diagnostic algorithms.
• Missed diagnoses of acute appendicitis lead to patient risk and increased healthcare resource utilization.

Purpose of the Study:

• To introduce and evaluate a novel preclustering-based ensemble learning (PEL) technique for acute appendicitis diagnosis.
• To address the challenge of imbalanced sample learning in the classification of acute appendicitis.
• To enhance the accuracy and reduce bias in diagnosing acute appendicitis.

Main Methods:

• The proposed PEL technique utilizes undersampling to balance majority-class instances.
• Preclustering groups similar majority-class instances, enabling selection of representative samples.
• Ensemble learning is integrated to further improve diagnostic performance.
• The technique was evaluated using 574 clinical cases against benchmark methods and scoring systems.

Main Results:

• PEL demonstrated superior effectiveness and reduced bias compared to benchmark techniques.
• The PEL technique exhibited higher sensitivity for identifying positive acute appendicitis than the Alvarado scoring system.
• PEL showed improved specificity in identifying negative acute appendicitis and higher overall sensitivity and specificity than resampling approaches.
• The area under the curve (AUC) for PEL reached 0.619, indicating the best overall performance.

Conclusions:

• The PEL technique effectively addresses imbalanced sample learning challenges in acute appendicitis diagnosis.
• PEL proved to be less biased towards positive or negative classes compared to benchmark methods.
• The proposed technique offers improved overall diagnostic effectiveness compared to existing scoring systems and resampling-based classification methods.