Skip to content

Latest commit

 

History

History
102 lines (80 loc) · 4.71 KB

README.md

File metadata and controls

102 lines (80 loc) · 4.71 KB
Raw

Survival Analysis

  • Author: Lina Faik
  • Creation date: February 2023
  • Last update: April 2023

Objective

This repository contains the code and notebooks used to train survival models to tackle real-world predictive problems. It was developed as an experimentation project to support the explanation blog posts around survival models. For more information, you can find the articles here:

  1. Part I - Survival Analysis: Predict Time-To-Event With Machine Learning

    Practical Application to Customer Churn Prediction

  2. Part II - Survival Analysis: Leveraging Deep Learning for Time-to-Event Forecasting

    Practical Application to Rehospitalization

You can find all my technical blog posts here.

Project Description

Data

The project consists of two use cases. Each one is described in a different article.

The data used in part 1 is from Kaggle. They are related to a subscription-based digital product offering for financial advice that includes newsletters, webinars, and investment recommendations. More specifically, the data consist of the following information:

  • Customer sign-up and cancellation dates at the product level
  • Call center activity
  • Customer demographics
  • Product pricing info

The data used in part 2 is from Kaggle and described in this research paper. It was collected from patients admitted over a period of two years at Hero DMC Heart Institute in India.
The data consists of information about the patient including:

  • Demographics: age, gender, locality (rural or urban)
  • Patient history: smoking, alcohol, diabetes mellitus, hypertension, etc.
  • Lab results: hemoglobin, total lymphocyte count, platelets, glucose, urea, creatinine, etc.

Code structure

datasets # folder containing the initial datasets
├── customer_subscription # used for the use case described in part 1
│   ├── customer_cases.csv
│   ├── customer_info.csv
│   ├── customer_product.csv
│   ├── customer_info.csv
├── hospitalisation # used for the use case described in part 2
│   ├── HDHI Admission data.csv
│   ├── HDHI Mortality data.csv
│   ├── HDHI Pollution data.csv
│   ├── table_headings.csv
notebooks
├── 01_data_preprocessing_customer_subscription.ipynb # clean and prepare data in part 1
├── 02_data_exploration_customer_subscription.ipynb # explore the data in part 1
├── 03_modeling_survival_ml_customer_subscription.ipynb # train multiple models in part 1
├── 04_evaluation_customer_subscription.ipynb # evaluate models in part 1
├── 11_data_preprocessing_customer_hospitalisation.ipynb # clean and prepare data in part 2
├── 12_data_exploration_customer_hospitalisation.ipynb # explore the data in part 1
├── 13_modeling_survival_ml_hospitalisation.ipynb # train multiple models in part 1
├── 14_evaluation_customer_hospitalisation.ipynb # evaluate models in part 1
outputs
├── data
│   ├── customer_subscription_clean.csv # pre-processed data in part 1
│   ├── hdhi_clean.csv # pre-processed data in part 2
│   ├── scaler.pkl # fitted scaler
│   ├── imputation_values.pkl # values used for importation
│   ├── train_x.pkl # features used to train models
│   ├── train_y.pkl # target from the train set
│   ├── val_x.pkl # features used to evaluate models
│   ├── val_y.pkl # target from the validation set
├── models # folder containing the trained models
├── model_scores.csv # model performance in part 1
├── model_scores_dl.csv # model performance in part 2
src
├── train.py # general functions to train models           
├── train_survival_ml.py # functions to train survival models
├── train_survival_deep.py # functions to train deep learning survival models
├── evaluate.py # functions to evaluate models

How to Use This Repository?

Requirement

The code relies on the following libraries:

scikit-survival==0.19.0 
plotly==4.14.3
torch==1.13.1
torchtuples==0.2.2
pycox==0.2.3

Experiments

To run experiments, you need to run the notebooks in the order suggested by their names. The associated code is in the src directory.