🩺 Diabetes Patients Readmission Prediction

A Machine Learning Approach to Predict 30-Day Hospital Readmission Risk

📌 Introduction

Hospital readmission among diabetes patients is a major challenge for healthcare systems. Early detection of patients with a high probability of being readmitted enables hospitals to plan interventions, reduce costs, and improve patient outcomes.

This project builds a machine learning model to predict whether a diabetes patient will be readmitted within 30 days using historical patient records, medical attributes, and clinical indicators.

---

🎯 Project Objective

To develop a predictive model that classifies patients into:

• Readmitted < 30 days (high-risk)
• Readmitted > 30 days
• Not readmitted

Primary ML Target Variable:
readmitted_30_days → 1 (readmitted within 30 days), 0 (not readmitted)

---

🧪 Dataset Summary

The dataset contains 100,000+ hospital encounters and includes:

• Patient demographics
• Diagnoses & procedure codes
• Laboratory test results
• Outpatient / inpatient visit history
• Medication and insulin-related features
• Admission & discharge information

Data Cleaning Performed:

✔ Handling missing values
✔ Fixing inconsistent categories
✔ Reducing noise in diagnosis codes
✔ Creating new engineered features for better model accuracy

---

📊 Exploratory Data Analysis (Key Findings)

🧱 Correlation Heatmap
Most numerical features show low correlation → dataset is primarily categorical.

👵 Age Distribution
Majority of patients fall in the 50–80 age group, highlighting higher chronic diabetes prevalence.

🔁 Readmission Distribution

• “Not readmitted” → majority class
• “Readmitted < 30 days” → minority class requiring class-imbalance handling

⭐ Important Predictors:

• Number of prior inpatient visits
• Number of outpatient visits
• Number of diagnoses
• Change in diabetes medication
• Insulin dosage
• Time spent in the hospital

---

📉 Model Evaluation (Charts)

Generated evaluation charts include:

• Confusion matrix
• ROC curve
• F1-score comparison
• Model accuracy comparison
• Random Forest feature importance

These visualizations help interpret model performance and identify improvement areas.

---

🛠️ Feature Engineering

Key transformations performed:

✔ Label encoding & one-hot encoding
✔ SMOTE oversampling for minority class
✔ Scaling numerical variables
✔ Grouping diagnosis codes
✔ Creating composite features:

• Total number of visits
• Diabetes medication change flags
• Chronic illness indicators

Final processed dataset:
feature_engineered_diabetes.csv

---

🤖 Machine Learning Models

The following models were trained and compared:

Logistic Regression Random Forest Classifier XGBoost Classifier

---

⭐ Best Performing Models

• XGBoost → highest accuracy
• Random Forest → most stable & interpretable

Both models performed well in identifying high-risk patients.

---

📈 Results Summary

• Accuracy: ~86%
• ROC-AUC: ~0.83
• F1-Score: strong for majority class, significantly improved for minority class after SMOTE
• Interpretability: insights extracted from feature importance

The model delivers meaningful predictions that can support hospital decision-making and help reduce readmission rates.