Survival prediction model for right-censored data based on improved composite quantile regression neural network

Area of Science:

• Biostatistics
• Machine Learning
• Medical Informatics

Background:

• Statistical inference for right-censored data is crucial in medical diagnosis.
• Existing survival analysis models may not fully capture complex data patterns.

Purpose of the Study:

• To propose an improved composite quantile regression neural network framework (rcICQRNN) for survival prediction with right-censored data.
• To enhance the model with inverse probability weighting and a whale optimization algorithm (WOA) for hyperparameter tuning.
• To introduce a binary whale optimization algorithm (BWOA) for variable selection in high-dimensional data.

Main Methods:

• Developed the rcICQRNN model integrating composite quantile regression and a multi-hidden layer feedforward neural network loss function.
• Employed inverse probability weighting for survival prediction.
• Utilized the WOA algorithm for hyperparameter optimization and integer/One-Hot encoding for classification features.
• Proposed the BWOA method for variable selection in high-dimensional datasets.

Main Results:

• The rcICQRNN-5 model showed suitability for simulated datasets.
• The WOA-rcICQRNN-30 model with One-Hot encoding performed well on NKI70 breast cancer data.
• Optimal results for the METABRIC dataset were achieved with feature selection at k=15.
• Cross-dataset validation indicated stable C-index results using One-Hot encoding.

Conclusions:

• The proposed rcICQRNN prediction model offers a flexible and stable approach for analyzing right-censored data.
• The model demonstrates practical utility in biomedicine, insurance, and financial economics.
• One-Hot encoding generally leads to more stable predictive performance.