A distribution-free smoothed combination method to improve discrimination accuracy in multi-category classification

Area of Science:

• Biostatistics
• Machine Learning
• Medical Diagnostics

Background:

• Combining multiple diagnostic tests enhances accuracy.
• Maximizing area under the receiver operating characteristic curve (AUC) is common for binary classification but computationally intensive with many biomarkers.
• Existing methods struggle with multi-category classification and large biomarker sets.

Purpose of the Study:

• To develop an efficient linear combination method for multi-category diagnostic classification.
• To address the computational challenges of existing AUC maximization methods.
• To improve diagnostic accuracy using a large number of biomarkers.

Main Methods:

• Developed a novel method maximizing a smooth approximation of the empirical Hyper-volume Under Manifolds (HUM).
• Approximated HUM using sigmoid and normal cumulative distribution functions for smooth optimization.
• Employed efficient gradient-based algorithms for parameter estimation.

Main Results:

• The proposed method yields consistent coefficient estimates under regularity conditions.
• Asymptotic normality of coefficient estimates was derived.
• Simulation studies demonstrated the method's effectiveness compared to existing approaches.
• The method was successfully applied to two real medical datasets.

Conclusions:

• The developed method offers an efficient and effective approach for multi-category diagnostic classification.
• It overcomes the computational limitations of traditional AUC maximization methods.
• This technique holds promise for improving diagnostic accuracy in complex medical scenarios.