Hospital Readmission Prediction

Binary classification to identify diabetic patients at risk of readmission within 30 days of discharge, using clinical encounter data from 130 US hospitals.

---

Dataset

UCI — Diabetes 130-US Hospitals (1999–2008)

• 101,766 patient encounters, 50 features
• Target: readmitted within 30 days (<30) vs all other outcomes
• Class imbalance: 11.4% positive class

---

Preprocessing

• Dropped columns with >80% missing values: weight (97%), max_glu_serum (95%), A1Cresult (83%)
• Dropped zero-variance columns: examide, citoglipton
• Removed patients discharged to hospice or deceased (discharge codes 11, 13, 14, 19, 20, 21)
• ICD-9 diagnosis codes mapped to 8 clinical categories (circulatory, respiratory, diabetes, etc.)
• Age brackets encoded as ordinal integers; remaining categorical columns one-hot encoded
• Missing values in race, diag_1/2/3 imputed with mode or placeholder

Final dataset: 99,343 rows, 93 columns

---

Modelling

Imbalance handling: RandomOverSampler applied inside each cross-validation fold via imblearn.Pipeline to prevent data leakage

Model selection: 5-fold stratified cross-validation on training data (80/20 split)

Model	CV ROC-AUC	CV Recall
Logistic Regression	0.646 ± 0.010	0.526 ± 0.010
Random Forest	0.640 ± 0.008	0.014 ± 0.002

Logistic Regression selected: higher ROC-AUC, substantially better recall across all folds.

Threshold tuning: default 0.5 threshold adjusted to 0.509 to achieve recall ≥ 0.50 while maximising precision.

---

Results

Model: Logistic Regression + RandomOverSampler | Threshold: 0.509

Class	Precision	Recall	F1
Not readmitted	0.92	0.70	0.79
Readmitted <30d	0.18	0.51	0.26

ROC-AUC: 0.649 | Average Precision: 0.209

Of 2,263 high-risk patients in the test set, 1,144 were correctly flagged.

---

Limitations

ROC-AUC of ~0.65 is consistent with published results on this dataset (Strack et al., 2014). 30-day readmission is driven largely by factors absent from hospital records: social support, housing, post-discharge care quality, and medication adherence. This is a screening tool, not a diagnostic.

---

Structure

hospital_readmission/
├── data/
│   ├── diabetic_data.csv          # Raw data (download from UCI)
│   └── cleaned_data.csv
├── notebooks/
│   ├── 01_setup_and_first_look.ipynb
│   ├── 02_eda.ipynb
│   └── 03_modelling_v2.ipynb
└── outputs/

---

Setup

pip install pandas scikit-learn imbalanced-learn matplotlib seaborn jupyter

Download diabetic_data.csv from UCI (link above) and place in data/.

---

Reference

Strack et al. (2014). Impact of HbA1c Measurement on Hospital Readmission Rates. BioMed Research International.