Approximating the risk score for disease diagnosis using MARS.

Area of Science:

• Biostatistics
• Medical Diagnostics
• Machine Learning in Healthcare

Background:

• Accurate disease diagnosis often relies on combining multiple diagnostic markers.
• Previous methods for risk score approximation, like McIntosh and Pepe's, have limitations in ROC curve range and require prior information.
• Diagnostic tests are imperfect, and single markers rarely outperform others uniformly.

Purpose of the Study:

• To evaluate Multivariate Adaptive Regression Splines (MARS) as a tool for approximating risk scores in disease diagnosis.
• To address limitations of existing methods, particularly concerning the range of Receiver Operating Characteristic (ROC) curves.
• To provide an easily implementable and interpretable method for combining multiple diagnostic markers.

Main Methods:

• Utilized Multivariate Adaptive Regression Splines (MARS) to approximate the risk score, which is the probability of disease conditional on multiple markers.
• Employed simulation studies to assess the performance of MARS.
• Compared MARS with existing two-step procedures for risk score approximation.

Main Results:

• MARS successfully approximates risk scores when combining multiple markers, especially when ROC curves from different tests cross.
• The MARS-derived ROC curve is defined across the entire range of false-positive rates (0,1), overcoming limitations of prior methods.
• The MARS approach is shown to be easy to implement and offers intuitive interpretation.

Conclusions:

• MARS is a valuable tool for disease screening and diagnosis by effectively combining multiple markers.
• The MARS method enhances diagnostic accuracy and provides a complete ROC curve without needing extra prior information.
• This approach offers practical advantages in implementation and interpretability for clinical applications.