Empirical Bayes logistic regression.

Area of Science:

• Biomedical data analysis
• Computational biology
• Proteomics and metabolomics

Background:

• Accurate patient disease status prediction is crucial for effective treatment.
• Mass spectrometry of serum samples offers a rich source of biomarkers.
• Developing robust diagnostic models from complex biological data presents challenges.

Purpose of the Study:

• To construct a diagnostic predictor for patient disease status using mass spectrometry data.
• To evaluate the performance of logistic regression models with L1 and ridge penalization.
• To optimize hyperparameter selection using cross-validation for improved predictive accuracy.

Main Methods:

• Utilized logistic regression with Bernoulli log-likelihood.
• Applied quadratic ridge and absolute L1 penalties for model regularization.
• Employed singular value decomposition to reduce variables for ridge penalization.
• Used 10-fold cross-validation for hyperparameter selection.
• Assessed predictive ability on a separate data subset.

Main Results:

• The developed diagnostic predictor demonstrated capability in classifying patient disease status.
• Penalized logistic regression models effectively handled high-dimensional mass spectrometry data.
• Cross-validation successfully identified optimal penalization hyperparameters.
• The model's predictive performance was validated on independent data.

Conclusions:

• Logistic regression with L1 or ridge penalization is a viable approach for disease status prediction from serum mass spectrometry.
• The methodology provides a framework for building accurate diagnostic tools from complex biological datasets.
• Further validation on larger, diverse cohorts is warranted to confirm clinical utility.