Deep Neural Network With a Smooth Monotonic Output Layer for Dynamic Risk Prediction

Area of Science:

• Biostatistics
• Machine Learning
• Cardiovascular Disease Research

Background:

• Risk prediction is crucial in survival analysis, with dynamic prediction incorporating longitudinal data.
• Existing methods may introduce bias due to parametric assumptions or discrete survival function approximations.

Purpose of the Study:

• To develop a novel deep neural network for nonparametric, dynamic risk prediction.
• To introduce the Smooth Monotonic Output Layer (SMOL) to avoid discretization and parametric model assumptions.

Main Methods:

• A deep neural network incorporating the novel Smooth Monotonic Output Layer (SMOL).
• SMOL utilizes B-splines to construct monotonic, differentiable functions for direct estimation of survival and cumulative distribution functions.
• Utilized data from the Cardiovascular Disease Lifetime Risk Pooling Project (LRPP).

Main Results:

• The proposed deep learning approach achieved state-of-the-art accuracy.
• Demonstrated superior performance in predicting individual-level risk for atherosclerotic cardiovascular disease.

Conclusions:

• The novel deep learning model with SMOL offers an accurate, nonparametric approach to dynamic risk prediction.
• This method effectively addresses limitations of existing survival analysis techniques for cardiovascular disease risk assessment.