Cost-sensitive case-based reasoning using a genetic algorithm: application to medical diagnosis.

Area of Science:

• Machine Learning
• Medical Informatics
• Artificial Intelligence

Background:

• Conventional case-based reasoning (CBR) is widely used in medicine but cannot account for asymmetric misclassification costs.
• Unequal costs of misdiagnosing illness versus non-illness limit CBR's applicability in real-world medical scenarios.
• There is a need for a CBR model that can dynamically incorporate varying misclassification costs.

Purpose of the Study:

• To develop and evaluate a novel cost-sensitive case-based reasoning (CSCBR) technique.
• To integrate unequal misclassification costs into the CBR framework.
• To enhance the applicability of CBR in medical decision-making by addressing asymmetric error costs.

Main Methods:

• Proposed a cost-sensitive case-based reasoning (CSCBR) method incorporating unequal misclassification costs.
• Utilized a genetic algorithm (GA) to dynamically adjust optimal classification and neighbor selection cut-off points.
• Applied and compared CSCBR with GA against C5.0 and CART on five medical datasets.

Main Results:

• CSCBR demonstrated lower total misclassification costs compared to other cost-sensitive methods in multiple cases.
• Paired t-tests confirmed statistically significant reductions in total misclassification costs for CSCBR on several datasets.
• Despite GA limitations (instability, overfitting), CSCBR with GA generally outperformed alternative methods.

Conclusions:

• The proposed CSCBR method effectively incorporates unequal misclassification costs and optimizes neighbor selection via GA.
• CSCBR proves cost-sensitive, with total misclassification costs remaining low even with increasing false absence costs.
• CSCBR achieved the lowest total misclassification costs on four out of five datasets, showing significant improvements.