CKD Risk Prediction
Chronic kidney disease detection pipeline with gradient-boosting ensembles, survival analysis and SHAP explanations, plus a FastAPI web app for per-patient risk.
Overview
A machine learning pipeline for early CKD detection and risk stratification, with a web application that explains each individual prediction.
My role
Researcher
Problem
Early detection and risk stratification of chronic kidney disease from clinical data.
Approach
Pipeline: automated preprocessing and feature selection, SMOTE for class imbalance, stratified k-fold validation, ensemble classifiers and survival models. App: a FastAPI backend taking 12 clinical inputs (age, creatinine, eGFR, HbA1c, blood pressure and others), applying median imputation and scaling, averaging scikit-learn and Keras model outputs, and returning SHAP explanations rendered with Chart.js.
Results
- Accuracy 94.2%, AUC-ROC 0.96, sensitivity 92.8%, specificity 95.1% with 10-fold cross-validation on 10,000 patients.