Instance-based classifiers applied to medical databases: diagnosis and knowledge extraction

Area of Science:

• Artificial Intelligence
• Machine Learning
• Medical Informatics

Background:

• Instance-based (IB) learning methods are explored for their potential in clinical decision support.
• Knowledge discovery in clinical databases requires effective diagnostic tools and knowledge extraction methods.

Purpose of the Study:

• To evaluate the feasibility and performance of IB learning classifiers as diagnostic tools.
• To assess the utility of IB classifiers for knowledge extraction in clinical databases.
• To introduce and evaluate a novel classifier, the prototype exemplar learning classifier (PEL-C).

Main Methods:

• Five IB classifiers, including two exemplar-based, one prototype-based, and two hybrid (one novel: PEL-C), were applied to three distinct clinical databases.
• Cross-validation techniques were employed to assess classifier performance using accuracy, sensitivity, specificity, and conciseness metrics.
• The number and type of instances representing diagnostic classes were analyzed for knowledge extraction capabilities.

Main Results:

• The optimized k-nearest neighbour classifier (k-NNC) and PEL-C demonstrated the best classification performance.
• k-NNC utilized 100% of the database for representation, while PEL-C achieved comparable performance using only ~3% of the data.
• PEL-C provided the most insightful knowledge extraction, yielding class representations combining prototypical and atypical clinical cases.

Conclusions:

• IB methods, specifically optimized k-NNC and PEL-C, are suitable for clinical decision support systems and nosological knowledge extraction.
• PEL-C offers advantages in terms of compact and meaningful class descriptions, beneficial for storage and knowledge extraction.
• Further validation in diverse clinical domains is recommended to confirm these findings.