Diabetes Readmission Analysis & Prediction

Which patients return within 30 days? Exploratory analysis, feature engineering and an XGBoost pipeline on encounters from 130 US hospitals.

Jun 2025Course project

~100kinpatient encounters

130US hospitals

~8%readmitted within 30 days

5-fold cross-validation

Exploratory analysis

A heavily imbalanced target

No readmission · ~75%After 30 days · 17%Within 30 days · 8%

FeaturesTreatment intensity, prior visits, diagnosis groups

Impute & scaleMissing values, standardisation

One-hot encodeClinical categories

TruncatedSVDSparse-matrix reduction

XGBoost30-day readmission

EvaluateROC AUC, class metrics, 5-fold CV

History predicts readmissionPast-year inpatient visits were the leading correlate.

Minority-class recall stayed lowClass-specific metrics therefore mattered more than overall scores.

Next: resampling & class weightsPrioritised to catch more 30-day returns.