SMAGS-LASSO: A Novel Feature Selection Method for Sensitivity Maximization in Early Cancer Detection.

Area of Science:

• Machine Learning
• Biomarker Discovery
• Cancer Diagnostics

Background:

• Traditional machine learning prioritizes overall accuracy, not clinical needs for early cancer detection.
• High sensitivity and specificity are crucial for effective cancer screening tools.

Purpose of the Study:

• Develop a feature selection algorithm to maximize sensitivity at a given specificity.
• Optimize machine learning for early cancer detection by prioritizing clinical metrics.

Main Methods:

• Introduced SMAGS-LASSO, combining a Sensitivity Maximization at a Given Specificity (SMAGS) framework with L1 regularization.
• Utilized a custom loss function and parallel optimization techniques for simultaneous feature selection and sensitivity optimization.
• Validated on synthetic datasets and real-world colorectal cancer biomarker data, comparing against LASSO and random forest.

Main Results:

• SMAGS-LASSO achieved perfect sensitivity (1.00) at 99.9% specificity on synthetic data, significantly outperforming LASSO (0.19).
• On colorectal cancer data, SMAGS-LASSO showed a 21.8% improvement over LASSO and 38.5% over random forest at 98.5% specificity.
• The algorithm effectively selected minimal biomarker panels while maintaining high performance.

Conclusions:

• SMAGS-LASSO facilitates the creation of minimal biomarker panels for early cancer detection.
• The method achieves high sensitivity at predefined specificity thresholds, improving diagnostic performance.