Stroke Prediction Model

A machine learning project that predicts stroke risk in patients using logistic regression. The model is designed to prioritize recall — minimizing missed stroke cases — making it suitable for clinical screening support.

---

Project Structure

project/
├── data/
│   └── processed/
│       ├── X_train.csv
│       ├── X_test.csv
│       ├── y_train.csv
│       └── y_test.csv
├── models/
│   ├── model.pkl
│   └── threshold.pkl
├── images/
│   ├── stroke_vs_gender.png
│   ├── stroke_vs_residence_type.png
│   ├── stroke_vs_job_type.png
│   ├── stroke_vs_marriage_status.png
│   ├── stroke_vs_smoking_status.png
│   ├── stroke_vs_heart_disease.png
│   ├── stroke_vs_bmi.png
│   ├── distribution_of_bmi.png
│   ├── distribution_of_ages.png
│   └── tpr_vs_fpr.png
├── notebooks/
│   ├── preprocessing_and_eda.ipynb
│   └── model.ipynb
└── README.md

---

Dataset

Source: Stroke Prediction Dataset on Kaggle (fedesoriano/stroke-prediction-dataset), downloaded via kagglehub.

• 5,110 patients, 12 columns (including target)
• Class imbalance: 4,861 non-stroke (95.1%) vs 249 stroke (4.9%)
• Missing values: bmi has 201 missing values; smoking_status has "Unknown" entries treated as NaN

---

Notebooks

1. preprocessing_and_eda.ipynb

Exploratory Data Analysis — generates bar plots of stroke rate across various patient segments:

• Stroke vs Gender
• Stroke vs Residence Type
• Stroke vs Job Type
• Stroke vs Marriage Status
• Stroke vs Smoking Status
• Stroke vs Heart Disease
• Stroke vs BMI Group (Underweight / Normal / Overweight / Obese)
• Distribution of BMI values
• Distribution of Ages

Preprocessing & Splitting:

• Drops id and the temporary bmi_group columns
• Replaces smoking_status = "Unknown" with NaN
• Stratified 80/20 train/test split (random_state=42)
• Saves processed splits to data/processed/

---

2. model.ipynb

Pipeline (via imblearn.Pipeline):

Step	Details
Categorical preprocessing	Mode imputation + One-Hot Encoding (drop="if_binary")
Numerical preprocessing	Median imputation + Standard Scaling
Binary preprocessing	Mode imputation (pass-through)
Class imbalance	SMOTE oversampling (random_state=42)
Model	LogisticRegression(max_iter=10000)

Hyperparameter Tuning via RandomizedSearchCV (21 iterations, 3-fold CV, scored on ROC-AUC):

{
    "model__C"            : [0.001, 0.01, 0.1, 1, 10, 100, 1000],
    "model__l1_ratio"     : [0.1, 0.5, 1],
    "model__solver"       : ["saga"],
    "model__class_weight" : ["balanced", None]
}

Threshold Optimization — the decision threshold is tuned to maximize precision while enforcing recall ≥ 0.94:

• Best threshold: ~0.164
• Best recall: 0.96
• Confusion matrix: [[432, 540], [2, 48]]

The model trades high false positives for near-complete capture of true stroke cases — appropriate for medical screening.

Serializes model.pkl and threshold.pkl to the models/ directory.

---

Installation

pip install -r requirements.txt

---

Usage

Step 1 — Run EDA & Preprocessing

Open and run notebooks/preprocessing_and_eda.ipynb. This downloads the dataset, generates EDA plots, and saves the processed train/test splits.

Step 2 — Train the Model

Open and run notebooks/model.ipynb. This trains the pipeline, tunes the threshold, and saves model.pkl and threshold.pkl.

---

Notes

• It is not a medical tool. It is just a student project a practice project.
• Feature names and values must match exactly (e.g., 'Residence_type' with capital R, 'formerly smoked' lowercase).
• Missing values in input are handled automatically by the pipeline's imputers.

Disclaimer

This project is for educational purposes only. It is not intended for clinical use or medical decision-making. Always consult a qualified medical professional.

Note: Built while learning - newer projects reflect improved practices and fixed mistakes