Predicting Heart Disease with Machine Learning: How I Built a 9… by Ayotunde OniPredicting Heart Disease with Machine Learning: How I Built a 9… by Ayotunde Oni

Predicting Heart Disease with Machine Learning: How I Built a 9…

Ayotunde Oni

Completed work

Data Analyst

Data Visualizer

ML Engineer

Matplotlib

Python

scikit-learn

Artificial Intelligence

Predicting Heart Disease with Machine Learning: How I Built a 97% Accurate Model Using Random Forest

Ayotunde Oni

5 min read

3 hours ago

Heart disease claims 17.9 million lives annually, making early detection not just a technical challenge, but also a global imperative. In this project, I harnessed machine learning to predict heart disease with remarkable accuracy, merging data science with real-world impact. My goal? Build a classifier that excels at identifying heart disease cases (high recall) while maintaining top-tier accuracy. Using a dataset generously provided by David Lapp on Kaggle, I crafted a Random Forest model that hit the mark. Let’s dive into the journey.

Step 1: Preprocessing the Data

Great models start with great data. I loaded the dataset using pandas:

import pandas as pddf = pd.read_csv('heart.csv')print(df.head())print(df.info())

This dataset was a gem. There were no missing values, just 13 numerical features like age, chol (cholesterol), and thalach (max heart rate), paired with a binary target: 1 for heart disease, 0 for none. I split it into features (X) and target (y):

X = df.drop('target', axis=1)y = df['target']

With only ~300 samples, I chose a 40% test set to ensure robust evaluation:

from sklearn.model_selection import train_test_splitX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.4, random_state=14)

This larger test set reduces overfitting risk and tests generalization — a calculated trade-off for reliability.

Step 2: Casting a Wide Net — Model Selection

Next, I tested six classifiers, grouped by scaling needs, to find the best performer:

Scale-Insensitive Models

Random Forest Classifier: Captures non-linear relationships, perfect for complex medical data.

Gaussian Naive Bayes: A simple probabilistic model suited to smaller datasets.

Gradient Boosting Classifier: Iteratively boosts weak learners for high accuracy.

from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifierfrom sklearn.naive_bayes import GaussianNBforest = RandomForestClassifier(random_state=14)forest.fit(X_train, y_train)gnb = GaussianNB()gnb.fit(X_train, y_train)gbc = GradientBoostingClassifier(random_state=14)gbc.fit(X_train, y_train)

Scale-Sensitive Models

K-Nearest Neighbors (KNN): Uses local patterns, intuitive for medical features.

Logistic Regression: A baseline for binary classification.

Support Vector Classifier (SVC): Shines in high-dimensional data with clear separation.

These required scaling with StandardScaler:

from sklearn.preprocessing import StandardScalerfrom sklearn.neighbors import KNeighborsClassifierfrom sklearn.linear_model import LogisticRegressionfrom sklearn.svm import SVCscaler = StandardScaler()X_train_scaled = scaler.fit_transform(X_train)X_test_scaled = scaler.transform(X_test)knn = KNeighborsClassifier()knn.fit(X_train_scaled, y_train)log_reg = LogisticRegression(random_state=14)log_reg.fit(X_train_scaled, y_train)svc = SVC(probability=True, random_state=14)svc.fit(X_train_scaled, y_train)

Why split them? Scaling ensures fairness — KNN and SVC struggle without it, while Random Forest thrives unscaled.

Step 3: The Scoreboard — Model Evaluation

I evaluated models on accuracy (overall correctness) and recall (catching true positives), critical for medical applications. Here’s the lineup:

Random Forest led the pack, but I confirmed its dominance with ROC curves and AUC scores:

from sklearn.metrics import roc_curve, roc_auc_scoreimport matplotlib.pyplot as pltmodels = [forest, gnb, gbc, knn, log_reg, svc]labels = ['Random Forest', 'Gaussian NB', 'Gradient Boosting', 'KNN', 'Logistic Regression', 'SVC']plt.figure(figsize=(10, 6))for model, label in zip(models, labels):    X_eval = X_test if model in [forest, gnb, gbc] else X_test_scaled    y_pred_prob = model.predict_proba(X_eval)[:, 1]    fpr, tpr, _ = roc_curve(y_test, y_pred_prob)    auc = roc_auc_score(y_test, y_pred_prob)    plt.plot(fpr, tpr, label=f'{label} (AUC = {auc:.3f})')plt.plot([0, 1], [0, 1], 'k--')plt.xlabel('False Positive Rate')plt.ylabel('True Positive Rate')plt.title('ROC Curves: Heart Disease Prediction')plt.legend()plt.grid(True, linestyle='--', alpha=0.7)plt.show()

Visual Spotlight: ROC Curves

Random Forest: Near-vertical rise to (0, 1), then flat along the top — AUC 0.999. Nearly flawless.

Gradient Boosting: Steep but less perfect, slightly lower AUC.

SVC: Strong start, more gradual later — solid but outpaced.

Gaussian NB, KNN, Logistic Regression: Moderate curves, trailing the leaders.

Random Forest’s AUC of 0.999 highlights its ability to nail true positives with minimal false positives.

Step 4: Tuning the Winner — Grid Search

With Random Forest in the lead, I optimized its hyperparameters using GridSearchCV. I tested max_depth, min_samples_split, max_features, and n_estimators. Most parameters had minimal impact, but n_estimators — the number of trees — stood out. After exploring 200, 250, and 500, I found 500 trees delivered peak performance: 0.978 accuracy.

from sklearn.model_selection import GridSearchCVfrom sklearn.ensemble import RandomForestClassifierparam_grid = {    'n_estimators': [200, 250, 500],    'max_depth': [None, 10, 20],    'min_samples_split': [2, 5],    'max_features': ['auto', 'sqrt'],    'random_state': [14]}forest = RandomForestClassifier()grid_search = GridSearchCV(estimator=forest, param_grid=param_grid, cv=3, n_jobs=-1, verbose=2)grid_search.fit(X_train, y_train)best_forest = grid_search.best_estimator_print(f'Best Parameters: {grid_search.best_params_}')print(f'Test Score: {best_forest.score(X_test, y_test):.3f}')

The final model settled on n_estimators=500 as the key driver of success.

Step 5: Peeking Inside — Feature Importance

What fuels this model? Here’s a visualization that shows what.

import numpy as npimport matplotlib.pyplot as pltfeature_importance = best_forest.feature_importances_features = X.columnsdf_importance = pd.DataFrame({'Feature': features, 'Importance': feature_importance})df_importance = df_importance.sort_values('Importance', ascending=True)plt.figure(figsize=(10, 6))colors = plt.cm.YlGn(np.linspace(0, 1, len(features)))plt.barh(df_importance['Feature'], df_importance['Importance'], color=colors)plt.xlabel('Importance')plt.title('Feature Importance in Heart Disease Prediction')plt.grid(True, linestyle='--', alpha=0.7)plt.tight_layout()plt.show()

Key Insights:

Top Features:

cp (chest pain): ~0.14 — diagnostic royalty.

thal (thalassemia): ~0.12 — a heart health staple.

thalach (max heart rate): ~0.11 — vital sign power.

Mid-Tier: ca (~0.10), oldpeak (~0.09), age (~0.08).

Lower Impact: chol (~0.07) to fbs (~0.01).

Chest pain and thalassemia lead, aligning with medical intuition about heart disease risk.

Lessons Learned

Model Diversity: Testing six classifiers revealed Random Forest’s edge — explore widely!

Scaling Nuances: KNN and SVC need scaling; Random Forest doesn’t — know your tools.

Tuning Pays Off: 500 trees outperformed 200 — optimization matters.

Explainability: Feature importance connects tech to medicine — don’t just predict, interpret.

Future Frontiers

Deeper Tuning: Experiment with max_depth or Random Search for efficiency.

Feature Engineering: Test interactions like age * thalach.

Stacking: Combine Random Forest and Gradient Boosting for a hybrid boost.

More Data: Larger datasets could enhance generalization.

The Finish Line: Impact Unlocked

This journey produced a Random Forest Classifier with 0.978 accuracy, 0.972 recall, and a 0.999 AUC — a near-perfect tool for heart disease prediction. With 97.2% recall, it could help clinicians spot high-risk patients early, potentially saving lives. From clean data to actionable insights, every step built a winner. Thanks to David Lapp’s Kaggle dataset, this project thrived. Next, I’ll push for even sharper models — because in healthcare, every edge counts.

What’s your data challenge? Share below — I’m all ears!

Personal Note

As a machine learning engineer passionate about healthcare, this project blended challenge and purpose perfectly. It’s polished, data-rich, and ready to inspire!

Like this project

Completed work

Posted Mar 18, 2025

Heart disease claims 17.9 million lives annually, making early detection not just a technical challenge, but also a global imperative. In this project, I harne…

Likes

Views

Predicting Heart Disease with Machine Learning: How I Built a 9…

Ayotunde Oni

Completed work

Data Analyst

Data Visualizer

ML Engineer

Matplotlib

Python

scikit-learn

Artificial Intelligence

Predicting Heart Disease with Machine Learning: How I Built a 97% Accurate Model Using Random Forest

Ayotunde Oni

5 min read

3 hours ago

Step 1: Preprocessing the Data

Great models start with great data. I loaded the dataset using pandas:

import pandas as pddf = pd.read_csv('heart.csv')print(df.head())print(df.info())

X = df.drop('target', axis=1)y = df['target']

With only ~300 samples, I chose a 40% test set to ensure robust evaluation:

from sklearn.model_selection import train_test_splitX_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.4, random_state=14)

This larger test set reduces overfitting risk and tests generalization — a calculated trade-off for reliability.

Step 2: Casting a Wide Net — Model Selection

Next, I tested six classifiers, grouped by scaling needs, to find the best performer:

Scale-Insensitive Models

Random Forest Classifier: Captures non-linear relationships, perfect for complex medical data.

Gaussian Naive Bayes: A simple probabilistic model suited to smaller datasets.

Gradient Boosting Classifier: Iteratively boosts weak learners for high accuracy.

from sklearn.ensemble import RandomForestClassifier, GradientBoostingClassifierfrom sklearn.naive_bayes import GaussianNBforest = RandomForestClassifier(random_state=14)forest.fit(X_train, y_train)gnb = GaussianNB()gnb.fit(X_train, y_train)gbc = GradientBoostingClassifier(random_state=14)gbc.fit(X_train, y_train)

Scale-Sensitive Models

K-Nearest Neighbors (KNN): Uses local patterns, intuitive for medical features.

Logistic Regression: A baseline for binary classification.

Support Vector Classifier (SVC): Shines in high-dimensional data with clear separation.

These required scaling with StandardScaler:

from sklearn.preprocessing import StandardScalerfrom sklearn.neighbors import KNeighborsClassifierfrom sklearn.linear_model import LogisticRegressionfrom sklearn.svm import SVCscaler = StandardScaler()X_train_scaled = scaler.fit_transform(X_train)X_test_scaled = scaler.transform(X_test)knn = KNeighborsClassifier()knn.fit(X_train_scaled, y_train)log_reg = LogisticRegression(random_state=14)log_reg.fit(X_train_scaled, y_train)svc = SVC(probability=True, random_state=14)svc.fit(X_train_scaled, y_train)

Why split them? Scaling ensures fairness — KNN and SVC struggle without it, while Random Forest thrives unscaled.

Step 3: The Scoreboard — Model Evaluation

I evaluated models on accuracy (overall correctness) and recall (catching true positives), critical for medical applications. Here’s the lineup:

Random Forest led the pack, but I confirmed its dominance with ROC curves and AUC scores:

from sklearn.metrics import roc_curve, roc_auc_scoreimport matplotlib.pyplot as pltmodels = [forest, gnb, gbc, knn, log_reg, svc]labels = ['Random Forest', 'Gaussian NB', 'Gradient Boosting', 'KNN', 'Logistic Regression', 'SVC']plt.figure(figsize=(10, 6))for model, label in zip(models, labels):    X_eval = X_test if model in [forest, gnb, gbc] else X_test_scaled    y_pred_prob = model.predict_proba(X_eval)[:, 1]    fpr, tpr, _ = roc_curve(y_test, y_pred_prob)    auc = roc_auc_score(y_test, y_pred_prob)    plt.plot(fpr, tpr, label=f'{label} (AUC = {auc:.3f})')plt.plot([0, 1], [0, 1], 'k--')plt.xlabel('False Positive Rate')plt.ylabel('True Positive Rate')plt.title('ROC Curves: Heart Disease Prediction')plt.legend()plt.grid(True, linestyle='--', alpha=0.7)plt.show()

Visual Spotlight: ROC Curves

Random Forest: Near-vertical rise to (0, 1), then flat along the top — AUC 0.999. Nearly flawless.

Gradient Boosting: Steep but less perfect, slightly lower AUC.

SVC: Strong start, more gradual later — solid but outpaced.

Gaussian NB, KNN, Logistic Regression: Moderate curves, trailing the leaders.

Random Forest’s AUC of 0.999 highlights its ability to nail true positives with minimal false positives.

Step 4: Tuning the Winner — Grid Search

from sklearn.model_selection import GridSearchCVfrom sklearn.ensemble import RandomForestClassifierparam_grid = {    'n_estimators': [200, 250, 500],    'max_depth': [None, 10, 20],    'min_samples_split': [2, 5],    'max_features': ['auto', 'sqrt'],    'random_state': [14]}forest = RandomForestClassifier()grid_search = GridSearchCV(estimator=forest, param_grid=param_grid, cv=3, n_jobs=-1, verbose=2)grid_search.fit(X_train, y_train)best_forest = grid_search.best_estimator_print(f'Best Parameters: {grid_search.best_params_}')print(f'Test Score: {best_forest.score(X_test, y_test):.3f}')

The final model settled on n_estimators=500 as the key driver of success.

Step 5: Peeking Inside — Feature Importance

What fuels this model? Here’s a visualization that shows what.

import numpy as npimport matplotlib.pyplot as pltfeature_importance = best_forest.feature_importances_features = X.columnsdf_importance = pd.DataFrame({'Feature': features, 'Importance': feature_importance})df_importance = df_importance.sort_values('Importance', ascending=True)plt.figure(figsize=(10, 6))colors = plt.cm.YlGn(np.linspace(0, 1, len(features)))plt.barh(df_importance['Feature'], df_importance['Importance'], color=colors)plt.xlabel('Importance')plt.title('Feature Importance in Heart Disease Prediction')plt.grid(True, linestyle='--', alpha=0.7)plt.tight_layout()plt.show()

Key Insights:

Top Features:

cp (chest pain): ~0.14 — diagnostic royalty.

thal (thalassemia): ~0.12 — a heart health staple.

thalach (max heart rate): ~0.11 — vital sign power.

Mid-Tier: ca (~0.10), oldpeak (~0.09), age (~0.08).

Lower Impact: chol (~0.07) to fbs (~0.01).

Chest pain and thalassemia lead, aligning with medical intuition about heart disease risk.

Lessons Learned

Model Diversity: Testing six classifiers revealed Random Forest’s edge — explore widely!

Scaling Nuances: KNN and SVC need scaling; Random Forest doesn’t — know your tools.

Tuning Pays Off: 500 trees outperformed 200 — optimization matters.

Explainability: Feature importance connects tech to medicine — don’t just predict, interpret.

Future Frontiers

Deeper Tuning: Experiment with max_depth or Random Search for efficiency.

Feature Engineering: Test interactions like age * thalach.

Stacking: Combine Random Forest and Gradient Boosting for a hybrid boost.

More Data: Larger datasets could enhance generalization.

The Finish Line: Impact Unlocked

What’s your data challenge? Share below — I’m all ears!

Personal Note

As a machine learning engineer passionate about healthcare, this project blended challenge and purpose perfectly. It’s polished, data-rich, and ready to inspire!

Like this project

Completed work

Posted Mar 18, 2025

Heart disease claims 17.9 million lives annually, making early detection not just a technical challenge, but also a global imperative. In this project, I harne…

Likes

Views