[ML] Patient Re-Admission Prediction
Develop a binary classification model to predict 30-day readmission risk for diabetes patients, enabling early intervention and improved patient outcomes.
Objective
Develop a binary classification model to predict if a discharged diabetes patient will be readmitted within 30 days.
Enable proactive monitoring and preventive measures for severe cases to improve survival rates.
Dataset Overview
Source: [UCI Machine Learning Repository]
Description: Readmission status (binary: 1 or 0) for diabetes patients.
Total Records: 101,766
Number of Features: 24
Problem Definition
Challenge: Increasing readmissions among discharged diabetes patients.
Expected Impact:
Early identification of severe cases.
Improved patient outcomes through proactive interventions.
Approach
Session 1
Data preparation using encoding techniques (e.g., OneHotEncoder).
Create a baseline model.
Session 2
Analyze the impact of class weights on errors.
Hyperparameter tuning using LightGBMClassifier and Optuna.
Session 3
Perform detailed error analysis, including cohort-based error grouping.
Investigate feature importance.
Key Highlights
Feature creation with OneHotEncoder.
LightGBMClassifier for modeling and Optuna for hyperparameter optimization.
Comprehensive error analysis, including general and cohort-based methods.