Forest-EMCBE: an evolutionary ensemble learning algorithm for multiclass diagnosis of bacterial pneumonia using the...

Area of Science:

• Medical Informatics
• Machine Learning
• Computational Biology

Background:

• Rapid bacterial pneumonia diagnosis is critical but hindered by traditional methods' time constraints.
• Machine learning offers potential for medical diagnosis but struggles with complex, imbalanced datasets.
• Class imbalance in complete blood count (CBC) data presents a significant challenge for accurate bacterial pneumonia classification.

Purpose of the Study:

• To address multiclass imbalanced CBC datasets for bacterial pneumonia diagnosis.
• To propose a novel ensemble learning algorithm, Forest of Evolutionary Multi-Classifiers Based on Bagging with Error-Correcting Output Coding (Forest-EMCBE).
• To enhance classifier generalization through an integrated three-layer structure.

Main Methods:

• Developed Forest-EMCBE, integrating Multi-Objective Genetic Algorithm, Error-Correcting Output Codes (ECOC), and balanced sampling.
• Trained the diagnostic model on a CBC dataset with 1,457 samples across 4 bacterial pneumonia classes.
• Compared Forest-EMCBE against 11 state-of-the-art algorithms.

Main Results:

• Forest-EMCBE demonstrated superior performance on the CBC dataset.
• The proposed algorithm outperformed all 11 compared state-of-the-art methods.
• Shapley value analysis identified key features like age, gender, and neutrophil percentage impacting predictions.

Conclusions:

• Forest-EMCBE is effective for diagnosing bacterial pneumonia from imbalanced CBC data.
• The algorithm offers improved generalization and accuracy compared to existing methods.
• Feature importance analysis provides insights into predicting infections by different bacterial species.