Stroke Risk Prediction

A machine learning project that analyzes patient health data to predict stroke risk using feature engineering, clustering, and classification models.

Project Overview

This project explores a dataset of ~15,000 patient records containing symptoms and health indicators to:

• Engineer meaningful risk indices from raw symptom data
• Segment patients into risk profiles using clustering
• Build and compare classification models for stroke risk prediction

Dataset

The dataset (strokeX.csv) contains patient-level health indicators:

Feature	Description
age	Patient age (years)
gender	Male/Female
15 symptom indicators	Binary (0/1): chest_pain, high_bp, irregular_heartbeat, short_breath, fatigue, dizziness, swelling, neckjaw_pain, excess_sweating, persistent_cough, nausea_vomiting, chest_discomfort, cold_extremities, sleep_apnea, anxiety
stroke_risk_pct	Estimated stroke risk (0-100%)
at_risk	Binary target variable (1 = at risk, 0 = not at risk)

Feature Engineering

Three custom risk indices were developed:

Index	Formula	Purpose
ANRI	stroke_risk_pct / age	Identifies patients with unusually high risk for their age
SBI	Sum of all symptoms	Quantifies total symptom burden
CHRI	0.4×high_bp + 0.4×irregular_heartbeat + 0.2×ANRI	Composite cardiovascular risk score

Key Finding: SBI shows strong correlation (r=0.68) with stroke risk. Patients with 7+ symptoms have 99.3% average risk.

Clustering Analysis

K-Means clustering (k=3) segmented patients into distinct risk profiles:

• Low Risk: Younger patients, fewer symptoms
• Medium Risk: Middle-aged, moderate symptom burden
• High Risk: Older patients, high symptom count

GMM clustering validated these groupings with ARI > 0.8.

Model Performance

Seven classification models were evaluated:

Model	AUC	F1 Score	Accuracy
CatBoost	1.000	0.993	0.991
HistGradientBoosting	0.999	0.992	0.990
XGBoost	0.999	0.989	0.986
Gradient Boosting	0.998	0.982	0.978
Logistic Regression	0.998	0.980	0.975
Random Forest	0.997	0.977	0.972
KNN	0.950	0.891	0.870

Best Model: CatBoost achieved near-perfect classification.

Top Predictive Features

1. SBI (Symptom Burden Index)
2. Age
3. Total symptoms
4. Sleep apnea, chest pain, high BP

Requirements

pip install -r requirements.txt

Key dependencies: pandas, numpy, scikit-learn, catboost, xgboost, matplotlib, seaborn

Usage

Open and run stroke_risk.ipynb in Jupyter Notebook or JupyterLab.

Conclusions

1. Symptom count is the strongest predictor - Patients with 7+ symptoms have near-certain stroke risk
2. Age amplifies risk - Older patients with multiple symptoms face significantly higher risk
3. Gradient boosting models excel - CatBoost outperformed traditional ML approaches
4. Clustering validates clinical intuition - Risk stratification aligns with low/medium/high groupings