Performance comparison of multi-label learning algorithms on clinical data for chronic diseases.

Area of Science:

• Medical Informatics
• Machine Learning
• Data Science

Background:

• Classifying chronic diseases in patients is crucial for clinical decision-making.
• Multi-label learning effectively models overlapping medical conditions common in chronic illnesses.
• Existing multi-label algorithms require performance comparison for analyzing sequential clinical data.

Purpose of the Study:

• To compare the performance of state-of-the-art multi-label learning algorithms.
• To analyze multivariate sequential clinical data from chronically ill patients.
• To enhance the evaluation of new algorithms for disease classification.

Main Methods:

• Utilized a summary statistics approach for processing sequential clinical data.
• Extracted interpretable features linked to patient medical records.
• Evaluated algorithms including ML-kNN, AdaBoostMH, binary relevance, classifier chains, HOMER, and RAkEL on the MIMIC-II dataset.

Main Results:

• Binary relevance approaches demonstrated optimal performance in most scenarios, particularly for disease detection.
• Binary relevance methods exhibit scalability for large datasets and are easy to implement.
• The RAkEL algorithm performed well in ranking labels by dominant disease, despite scalability issues with large datasets.

Conclusions:

• Binary relevance algorithms are highly effective for chronic disease classification and detection.
• RAkEL offers value in prioritizing diagnoses based on disease dominance.
• This comparative study aids in selecting appropriate multi-label learning methods for clinical data analysis.