DeepSurv
Also known as: Neural network survival, DL survival model
DeepSurv is a deep neural network approach to survival analysis that learns personalized survival distributions directly from data. Introduced by Katzman et al. in 2018, it extends the Cox proportional hazards model using deep learning to capture complex, nonlinear relationships between covariates and survival outcomes. It solves the problem of modeling heterogeneous treatment effects and time-to-event predictions in high-dimensional settings.
Key highlights
- Captures complex, nonlinear relationships between covariates and survival outcomes automatically
- Produces interpretable personalized hazard estimates and survival curves for individual subjects
- Handles high-dimensional covariate spaces naturally through deep learning representation learning
- Enables discovery of treatment-by-covariate interactions without manual feature engineering
- Provides probabilistic predictions suitable for clinical decision support
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use DeepSurv when you have high-dimensional covariates, expect nonlinear or interaction effects between variables, or need personalized survival predictions. It is particularly valuable in precision medicine applications where treatment effects vary by individual characteristics. Assume proportional hazards approximately hold, and that you have sufficient data for the network to learn reliably. Prefer DeepSurv over Cox regression when preliminary exploration suggests the linearity assumption is violated.
Strengths & limitations
- Captures complex, nonlinear relationships between covariates and survival outcomes automatically
- Produces interpretable personalized hazard estimates and survival curves for individual subjects
- Handles high-dimensional covariate spaces naturally through deep learning representation learning
- Enables discovery of treatment-by-covariate interactions without manual feature engineering
- Provides probabilistic predictions suitable for clinical decision support
- Requires substantially more data than traditional Cox models to estimate accurately
- Network architecture and hyperparameter selection add complexity to the modeling process
- Theoretical guarantees on proportional hazards assumption are weaker than Cox regression
- Computational cost is higher due to neural network training
- Feature importance interpretation is less straightforward than traditional methods
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
How does DeepSurv differ from Cox regression?
Cox regression assumes a linear relationship between log-hazards and covariates. DeepSurv replaces this linear assumption with a neural network, allowing it to learn arbitrary nonlinear transformations. Both use the same partial likelihood objective, but DeepSurv can capture interactions and complex patterns automatically.
What is the baseline hazard in DeepSurv, and how is it estimated?
The baseline hazard (hazard when all covariates are zero) is not explicitly estimated by the network. Instead, the Breslow estimator computes it from the training data after the network is fitted, using the residuals. This decoupling allows DeepSurv to focus on learning covariate effects.
How many subjects do I need to train a reliable DeepSurv model?
There is no fixed rule, but deep learning typically requires more samples than traditional methods. A practical guideline is to have at least 50–100 events per covariate (or per hidden unit in small networks). With fewer subjects, simpler models like Cox regression are preferable.
Can DeepSurv handle time-varying covariates?
Standard DeepSurv (and Cox regression) assumes covariates are fixed. Extensions using recurrent neural networks (RNNs) or transformers can model time-varying covariates, but these are more advanced variants.
How should I choose the network architecture?
Start with one or two hidden layers (32–128 units each) and tune via cross-validation using the concordance index. Deeper networks can overfit; regularization (dropout, L1/L2) is essential. Early stopping on a validation set prevents overfitting.
What software packages implement DeepSurv?
The original DeepSurv code is available in TensorFlow. Modern implementations include lifelines (Python), pycox, and DeepLearningLifeExpectancy. Each has different features and computational backends.
Sources
- 1.Faraggi, D., & Simon, R. (1995). A neural network model for survival data. Statistics in Medicine, 14(1), 73–82.
- 2.Katzman, J. L., et al. (2018). DeepSurv: Personalized treatment recommender system using a Cox proportional hazards deep neural network. Journal of Machine Learning Research, 40, 40–51.
- 3.Lee, C., Zame, W., Yoon, J., & van der Schaar, M. (2018). Deephit: A deep learning approach for dynamic survival analysis. AAAI Conference on Artificial Intelligence, 32(1).
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). DeepSurv. ScholarGate. https://scholargate.app/survival/deepsurv