A scalable discrete-time survival model for neural networks

Area of Science:

• Medical prediction
• Machine learning in healthcare
• Survival analysis

Background:

• Neural networks are increasingly used for medical prediction.
• Survival data, including follow-up time and event indicators, is crucial for accurate modeling.
• Existing models may not fully leverage survival data, leading to information loss.

Purpose of the Study:

• Introduce Nnet-survival, a discrete-time neural network model for survival data.
• Enable efficient training on large datasets and generation of predicted survival curves.
• Provide a flexible model where hazard rates vary with follow-up time.

Main Methods:

• Developed a discrete-time survival model integrated with neural networks (Nnet-survival).
• Employed maximum likelihood estimation with mini-batch stochastic gradient descent (SGD) for training.
• Implemented the model in the Keras deep learning framework.

Main Results:

• Nnet-survival demonstrated strong performance on both simulated and real-world medical data.
• The model showed rapid convergence due to SGD, suitable for large datasets.
• Compared favorably against established models such as Cox-nnet and Deepsurv.

Conclusions:

• Nnet-survival offers a flexible and efficient approach for medical prediction using survival data.
• The model's ability to handle time-varying hazards enhances predictive accuracy.
• Availability of source code facilitates adoption and further research in deep learning for survival analysis.