Understanding neural networks using regression trees: an application to multiple myeloma survival data.

Area of Science:

• Computational statistics
• Machine learning in healthcare

Background:

• Neural networks are powerful data analysis tools but offer limited interpretability.
• Understanding the relationship between neural network outputs and input variables (covariates) is challenging.
• Interpreting complex models is crucial, especially in clinical settings for patient stratification.

Purpose of the Study:

• To present a straightforward method for interpreting neural network outputs.
• To utilize regression trees as a tool to enhance the explainability of neural networks.
• To apply this interpretation technique to censored survival data in multiple myeloma research.

Main Methods:

• Employing readily available software to integrate regression trees with neural networks.
• Developing a practical workflow for visualizing and understanding neural network predictions.
• Focusing the application on censored survival data analysis for clinical relevance.

Main Results:

• Demonstrated an accessible technique for decoding neural network outputs using regression trees.
• Successfully applied the method to identify distinct prognostic groups in multiple myeloma patients.
• Validated the utility of regression trees for improving neural network interpretability.

Conclusions:

• Regression trees provide an effective and easy-to-use method for understanding neural network outputs.
• This approach facilitates the identification of patient subgroups with varying prognoses, particularly in survival analysis.
• The technique is broadly applicable to both censored and uncensored data analysis scenarios.