Survival analysis for PyTorch, censoring included

What it does

pycox is a PyTorch-based library for survival analysis—predicting when an event such as failure, death, or churn will happen rather than simply whether it will. It packages nine methods, ranging from continuous-time Cox variants like DeepSurv and Cox-Time to discrete-time approaches like DeepHit and LogisticHazard, along with IPCW evaluation metrics and preprocessing utilities. The library also ships with several real-world event-time datasets accessible through a pycox.datasets module.

The interesting bit

Most deep-learning libraries treat classification or regression as the default; pycox treats right-censored data as first-class, offering both continuous and discrete hazard models under one roof. It is built on top of the author’s torchtuples training utility, which means the training loop abstractions are intentionally separated from the survival logic.

Key highlights

• Nine implemented methods including CoxPH (DeepSurv), Cox-Time, DeepHit, MTLR, and PCHazard, covering both continuous and discrete time.
• Evaluation suite with time-dependent concordance, IPCW Brier score, and administrative censoring metrics via EvalSurv.
• Built-in datasets such as METABRIC, FLCHAIN, and KKBox churn data available through pycox.datasets.
• Preprocessing tools and example notebooks that mix autoencoders with survival networks and CNNs on MNIST survival tasks.

Verdict

Worth a look if you are doing medical informatics, churn prediction, or reliability engineering with PyTorch and need more than a baseline Cox model. Skip it if you are already happy with scikit-survival or need a framework-agnostic solution.