Comparison of classification methods that combine clinical data and high-dimensional mass spectrometry data

Area of Science:

• Biomarker discovery in clinical cancer research.
• Proteomic profiling using mass spectrometry.
• Development of predictive models for binary outcomes.

Background:

• Identifying novel diagnostic and prognostic biomarkers is crucial in cancer research.
• Mass spectrometry generates high-dimensional proteomic profiles, often exceeding the number of individuals studied.
• Proteomic data can complement or extend traditional clinical variables.

Purpose of the Study:

• To evaluate and compare the predictive performance of new and existing models.
• To assess models that integrate mass spectrometry data with classical clinical variables.
• To conduct the study within the framework of binary prediction tasks.

Main Methods:

• Utilized simulated and real-world datasets (proteomic markers of steatosis).
• Employed penalization methods (Ridge, Lasso) and dimension reduction techniques (PLS) for high-dimensional data.
• Applied sparse PLS for combined data strategies in binary classification.

Main Results:

• Evaluated methods based on mean classification rate and predictive marker selection accuracy.
• Compared performance across clinical-only, mass spectrometry-only, and combined data models.
• Demonstrated the utility of proposed methods in handling high-dimensional proteomic data.

Conclusions:

• Models integrating both clinical and mass spectrometry data can outperform single-data-source models.
• The benefit of combined data models is most pronounced when dataset sample size is sufficiently large.
• This approach enhances the selection of true predictive markers for binary outcomes.