Survival Analysis
Time-to-event is its own problem: the question is not whether something happens but when, and some subjects leave the study before it does. The page goes from introductions, to a library, to deep learning methods, and ends with papers and their implementations.
The same notes are in Life Time Value (LTV).
The introductions come first. A good introduction is the archived set of lecture notes. For Survival Analysis the Intro by Sushma Dhamodharan explains why the method is needed: logistic regression studies whether risk factors are associated with presence or absence of disease, but when the interest is how a factor affects time to an event, or when subjects drop out and their outcome is unknown, logistic regression is not appropriate. The brief intro by Jia Qing points out that despite the name, survival analysis is not only for survival data and covers a wide area of applications. Intuition is Anurag Pandey's intuition and implementation in Python, framed around the business questions the technique can answer. What is? is Kunal Sawarkar's note that a method now behind models of millennials' social media behavior has been around for almost half a millennium.
With the idea in place, the library is Survival scikit kit, the introduction to survival analysis with scikit-survival.
Deep learning is the next step. Survival Analysis: Leveraging Deep Learning for Time-to-Event Forecasting comes in two parts, Part I & II; Part II is Lina Faik on survival models for predicting when an event occurs, with use cases such as predictive maintenance, customer churn in marketing analytics, and patient re-hospitalization. The figure below is from that series.
The papers then compare and implement those methods. (Paper) A comparison of machine learning methods for survival analysis of high-dimensional clinical data for dementia prediction is the Scientific Reports comparison. For DeepSurv, This repository is an unofficial PyTorch implementation of DeepSurv: Personalized Treatment Recommender System Using A Cox Proportional Hazards Deep Neural Network. We reimplement the experiments in the paper, which is followed by Github, and the detailed understanding is available on my Blog. The official repo describes DeepSurv as a deep learning approach to survival analysis, and the blog post explains that survival models relate patient features to treatment outcome, where earlier models such as the linear Cox proportional hazards model needed medical expertise for feature engineering. DeepSurv Keras is an implementation of DeepSurv using Keras. When several events compete, DeepHit: A Deep Learning Approach to Survival Analysis with Competing Risks is the code for that model, and Comparing DeepHit and DeepSurv models on the SUPPORT dataset puts the two side by side. Survival analysis with PyTorch is the pycox library that closes the list.
Deprecated links
These links and images no longer work. The original wording is kept here. A same-resource copy, when one was checked, is used above.
A good introduction for Survival Analysis. This address no longer opens: http://www.stat.columbia.edu/~madigan/W2025/notes/survival.pdf
Intuition. This address no longer opens: https://towardsdatascience.com/survival-analysis-intuition-implementation-in-python-504fde4fcf8e
Survival Analysis: Leveraging Deep Learning for Time-to-Event Forecasting Part I & II. This address no longer opens: https://towardsdatascience.com/survival-analysis-predict-time-to-event-with-machine-learning-part-i-ba52f9ab9a46
Deep Learning for Survival Analysis Authors: Laura Löschmann, Daria Smorodina. This address no longer opens: https://humboldt-wi.github.io/blog/research/information_systems_1920/group2_survivalanalysis/
Last updated
Was this helpful?