Learning Covariate Relations in Disease Progression Models Using Symbolic Neural Networks.

Area of Science:

• Biostatistics
• Machine Learning
• Computational Biology

Background:

• Covariate modeling in disease progression is crucial for individual outcome predictions.
• Current methods struggle with predefined functions, leading to poor covariate selection and biased models.
• Existing approaches have scalability issues with high-dimensional data due to combinatorial complexity.

Purpose of the Study:

• To develop a novel, automated method for identifying covariate models in disease progression.
• To overcome limitations of predefined parametric functions and combinatorial challenges in current methodologies.
• To improve the accuracy and efficiency of covariate selection and parameter optimization.

Main Methods:

• Utilized symbolic neural networks to simultaneously identify parametric covariate functions and optimize Markov chain model parameters.
• Employed stepwise pruning of dense symbolic networks to generate human-readable covariate functions.
• Applied the methodology to a type 2 diabetes patient dataset for disease progression modeling.

Main Results:

• The novel method successfully identified covariate relationships and optimized model parameters.
• The resulting model demonstrated predictive performance comparable to state-of-the-art methods.
• The automated approach achieved similar predictive accuracy while utilizing fewer covariates.

Conclusions:

• Symbolic neural networks offer an effective approach for automated covariate model identification in disease progression.
• The proposed methodology enhances model interpretability and predictive accuracy.
• This automated approach represents a significant advancement for high-dimensional covariate modeling in clinical research.