Cox-nnet v2.0: improved neural-network-based survival prediction extended to large-scale EMR data

Area of Science:

• Computational biology
• Machine learning in healthcare

Background:

• Cox-nnet, a neural network for prognosis prediction, was initially developed for genomics data.
• Electronic Medical Records (EMR) offer vast potential for population health studies but require efficient analytical tools.

Purpose of the Study:

• To introduce Cox-nnet version 2.0 (Cox-nnet v2.0), an optimized method for prognosis prediction.
• To enhance the efficiency and interpretability of Cox-nnet for large-scale population and EMR datasets.
• To incorporate feature importance and coefficient direction for better model understanding.

Main Methods:

• Cox-nnet v2.0 utilizes a neural network architecture for survival analysis.
• The method incorporates permutation-based feature importance and coefficient direction.
• The updated model was evaluated on kidney transplantation and SUPPORT datasets.

Main Results:

• Cox-nnet v2.0 demonstrated up to a 32-fold reduction in training time on a dataset of 10,000 individuals.
• The method achieved superior prediction accuracy compared to Cox-PH (P<0.05) on kidney transplantation data.
• Comparable superior performance was observed on the 8,000-individual SUPPORT dataset.

Conclusions:

• Cox-nnet v2.0 significantly improves efficiency and interpretability for survival prediction.
• The method is well-suited for analyzing large-scale population data, including EMR.
• Cox-nnet v2.0 represents a valuable tool for survival prediction in EMR-based research.