Receiver operating characteristic curves and confidence bands for support vector machines

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Statistical Modeling

Background:

• Binary classification is crucial for biomedical decision-making, including disease diagnosis and treatment response prediction.
• Support Vector Machines (SVMs) are effective for high-dimensional data but do not directly yield probabilities for traditional Receiver Operating Characteristic (ROC) curve analysis.
• Existing methods for ROC curve construction with SVMs have underdeveloped theoretical properties.

Purpose of the Study:

• To develop a method for constructing confidence bands for SVM-based ROC curves.
• To provide theoretical justification for the SVM ROC curve construction.
• To demonstrate the utility of the proposed method in a breast cancer treatment response prediction model.

Main Methods:

• Utilizing a sequence of weighted Support Vector Machines (SVMs) to construct the ROC curve.
• Developing and justifying a method for confidence band construction for the SVM ROC curve.
• Conducting simulation studies to validate the confidence band method.
• Applying the method to a real-world case of predicting breast cancer treatment response.

Main Results:

• A novel method for constructing confidence bands for SVM ROC curves is proposed.
• Theoretical justification is provided, demonstrating uniform consistency of the estimated decision rule's risk function.
• Simulation studies confirm the effectiveness of the confidence band method.
• An illustrative predictive model for breast cancer treatment response was developed.

Conclusions:

• The proposed method enhances ROC curve analysis for SVMs, enabling analyses not possible with traditional probability thresholding.
• Confidence bands provide crucial statistical rigor for SVM ROC curve interpretation.
• This approach has significant potential for improving biomedical decision-making tools.