Semi-varying coefficient multinomial logistic regression for disease progression risk prediction

Area of Science:

• Statistics
• Biostatistics
• Medical Informatics

Background:

• Accurate risk prediction is crucial for managing chronic diseases and guiding treatment strategies.
• Existing models may not fully capture complex, non-linear interactions between predictive factors.
• Improved methods are needed for patient stratification based on future disease progression.

Purpose of the Study:

• To propose a novel risk prediction model utilizing semi-varying coefficient multinomial logistic regression.
• To develop a penalized local likelihood method for model selection and coefficient estimation.
• To enhance predictive modeling capabilities, particularly when non-linear predictor interactions are present.

Main Methods:

• Semi-varying coefficient multinomial logistic regression framework.
• Penalized local likelihood for model selection and estimation of functional and constant coefficients.
• Leave-one-out cross-validation for prediction accuracy assessment and a recalibration framework for risk reliability evaluation.

Main Results:

• Simulation studies demonstrated effective model selection with minimal errors (wrong or missing selections).
• The proposed model successfully classified early rheumatoid arthritis patients into distinct future disease progression risk groups.
• Leave-one-out cross-validation confirmed a high correct prediction rate, and the recalibration framework assessed risk reliability.

Conclusions:

• The semi-varying coefficient model offers an effective approach for risk prediction, especially with non-linear predictor interactions.
• The penalized local likelihood method provides robust model selection and estimation.
• This methodology enhances patient stratification for diseases like rheumatoid arthritis, improving clinical decision-making.