Towards application of one-class classification methods to medical data

Area of Science:

• Biomedical data analysis
• Machine learning
• Pattern recognition

Background:

• One-class classification (OCC) is crucial for distinguishing a target class from all other objects in biomedical applications like diagnosis and medical data analysis.
• Existing state-of-the-art OCC techniques face challenges with diverse data types, particularly nominal variables.

Purpose of the Study:

• To experimentally compare a typicality test approach for one-class classification (OCC) against established methods.
• To evaluate the performance of OCC techniques on various biomedical datasets, including those with non-continuous data.

Main Methods:

• A typicality test-based one-class classification (OCC) method was developed and evaluated.
• The typicality approach was compared against Gaussian, mixture of Gaussians, naive Parzen, Parzen, and support vector data description (SVDD) methods.
• Twelve experimental datasets with multiple classes were utilized, treating each class as the target class in turn.

Main Results:

• The typicality test approach demonstrated good performance in one-class classification (OCC) tasks.
• This method proved effective for high-dimensional data.
• The typicality approach is versatile, applicable to continuous, discrete, and nominal data types.

Conclusions:

• The typicality test offers a robust and flexible solution for one-class classification (OCC) in biomedical contexts.
• Its ability to handle diverse data types, including nominal variables, surpasses limitations of current state-of-the-art methods.