🧠 Stroke Prediction Analysis

A machine learning project to predict the likelihood of stroke using health data. This project showcases end-to-end data science skills from EDA to deployment, with a focus on real-world healthcare applications.

Enter health information such as age, glucose level, BMI, etc., and get an instant stroke risk prediction using a machine learning model trained on real data.

📚 Table of Contents

• Dataset
• Exploratory Data Analysis
• Smoking Status Analysis
• Residence & Work Type Analysis
• Data Preprocessing
• Model Building
• Deployment
• Folder Structure
• Author

---

📁 Dataset

• Source: Kaggle - Stroke Prediction Dataset
• Records: 5,110
• Target Variable: stroke (1 = stroke occurred, 0 = no stroke)
• ⚠️ Class Imbalance: Only ~4.9% of patients had a stroke

---

📊 Exploratory Data Analysis (EDA)

Initial steps involved:

• Checked column types, missing values, and value counts
• Identified 201 missing values in the bmi column
• Detected significant class imbalance (~95% no stroke)
• Created visualizations:
  • Stroke Count Distribution
  • Stroke Distribution by Gender

---

🚬 Stroke Distribution by Smoking Status

Analyzed stroke prevalence across:
"formerly smoked", "never smoked", "smokes", and "unknown".

🔍 Observations:

• "Never smoked" is the largest group.
• "Formerly smoked" shows slightly higher stroke proportion.
• "Smokes" has fewer stroke cases than expected.
• "Unknown" still includes stroke cases.

💡 Insight: Stroke risk spans across all smoking categories — confirming that stroke is multifactorial.

---

🏘️ Residence & Work Type Analysis

🏡 Residence Type:

• Urban stroke rate: ~5.2%
• Rural stroke rate: ~4.5%
• 💡 Slightly higher in urban areas — possibly due to stress or better diagnosis access.

💼 Work Type:

• Self-employed: Highest stroke proportion (~7.9%)
• Private/Govt jobs: ~5%
• Children: Minimal strokes
• 💡 Self-employed individuals may experience more health risks due to irregular schedules or lack of access to care.

---

🧼 Data Preprocessing

Steps completed:

• ✅ Filled missing bmi values using median imputation
• ✅ Encoded categorical variables using pd.get_dummies
• ✅ Normalized numeric features with StandardScaler
• ✅ Split dataset into train/test (80/20)

---

🤖 Model Building

⚠️ Problem:

• Severe class imbalance (~95% no-stroke vs ~5% stroke)
• Baseline models had high accuracy but failed to detect stroke cases

---

🔄 SMOTE: Synthetic Minority Over-sampling Technique

Used SMOTE to synthetically create more stroke samples in the training set.

✅ Result:

• Balanced training set: 50% stroke / 50% no-stroke
• Enabled models to learn patterns in the minority class

---

📈 Model Comparison

🔹 Logistic Regression (with SMOTE)

Metric	Value
Accuracy	74.85%
Recall (Stroke=1)	0.80 ✅
Precision	0.14
F1-Score	0.24

🎯 Best for real-world screening where false negatives are costly
✅ Recommended for healthcare settings

---

🌲 Random Forest (with SMOTE)

Metric	Value
Accuracy	91.78% ✅
Recall (Stroke=1)	0.14 ❌
Precision	0.15
F1-Score	0.14

⚠️ Prioritizes overall accuracy, but misses most stroke cases

---

🧠 Takeaway:

• Random Forest has better accuracy
• But Logistic Regression is better for identifying stroke patients — critical in healthcare

---

💾 Model Saving

Final models were saved using joblib:

joblib.dump(rf_smote, 'models/random_forest_stroke_model.joblib')
joblib.dump(scaler, 'models/standard_scaler.joblib')

🚀 Deployment 

## 🚀 Live Demo

The stroke prediction web app is live and can be accessed here:  
🔗 [https://stroke-prediction-analysis.onrender.com](https://stroke-prediction-analysis.onrender.com)

### 🛠️ How to Use
1. Go to the [Live App](https://stroke-prediction-analysis.onrender.com/)
2. Enter health details like age, BMI, glucose levels, etc.
3. Click **"Predict Stroke Risk"**
4. See the result: 🚨 "Stroke Risk" or 😊 "No Stroke"

⚙️ **Custom Threshold**
Lowered the prediction threshold to `0.3` (from default `0.5`) to **prioritize recall** and catch more true stroke cases — important in healthcare where false negatives are riskier than false positives.

### 📈 ROC-AUC Evaluation

To further assess the model's ability to distinguish between stroke and non-stroke cases, we plotted the **Receiver Operating Characteristic (ROC) Curve** and calculated the **Area Under the Curve (AUC)**.

#### ✅ Cross-validated AUC scores (on SMOTE-resampled training data):
[0.9840, 0.9981, 0.9966, 0.9972, 0.9967]
Mean AUC Score: 0.9945

This high score indicates that the model is **very effective** at separating stroke from non-stroke classes in training, thanks to balanced resampling using **SMOTE**.

#### 🧪 ROC Curve on Test Set:
- **AUC = 0.76**
- Shows solid performance on unseen data.
- While lower than the training score (as expected), it still reflects **meaningful predictive power** in identifying stroke risks.

> ℹ️ **Why AUC drops on test data**: SMOTE balances only the training data. The test set retains the original class imbalance, making it harder to predict stroke cases. This drop is natural and reflects real-world conditions.

ROC Curve :

<img src="notebooks/roc_curve.png" width="500"/>


project/
│
├── data/                  # Raw dataset
├── notebooks/             # EDA and modeling notebooks
├── scripts/               # Preprocessing and training scripts
├── templates/             # HTML templates for Flask app
├── models/                # Saved .joblib models
├── app.py                 # Flask backend
├── README.md              # Project summary and documentation
└── requirements.txt       # Python dependencies

---

## 🧪 How to Use

1. Visit the [Live App](https://stroke-prediction-analysis.onrender.com/)
2. Enter patient info into the form
3. Submit to receive real-time stroke risk prediction (based on trained ML model)

---

## ⚙️ Tech Stack

- Python 3.11
- Flask
- Scikit-learn
- SMOTE (Imbalanced-learn)
- Pandas, NumPy, Matplotlib
- Hosted on Render

## 🎓 What I Learned

This project deepened my understanding of real-world data science workflows, particularly in healthcare prediction problems. Key takeaways:

- **End-to-End Pipeline:** I built a full ML pipeline from EDA → preprocessing → model training → deployment, reinforcing the importance of structure and iteration.

- **Class Imbalance & SMOTE:** I learned how imbalanced datasets can lead to misleading model performance and how SMOTE can help improve recall — a critical metric in healthcare applications.

- **Evaluation Trade-offs:** I understood how accuracy isn't always the best metric. In this case, **recall** was prioritized because missing a stroke case is more dangerous than a false positive.

- **Model Deployment:** I implemented my first Flask app and deployed it using **Render**, gaining confidence in converting notebooks into usable applications.

- **Git & Version Control:** Managed my project with Git and GitHub, learning to track experiments, sync changes, and maintain a clean structure.

- **Communication:** Writing markdown summaries helped me practice translating technical insights into business-relevant language.

This was more than just a machine learning task — it was a full-stack learning experience.

👤 Author

Vivek Raghunathan
🔗 LinkedIn
💻 GitHub: @vivekr25