06 // MACHINE LEARNING

Machine Learning Model for Heart Attack Risk Prediction

A machine learning decision-support prototype developed for iHealth Edu that estimates heart attack risk from patient data and delivers the result through a Flask API.

Role: Machine Learning Engineer
Period: June–August 2025
Status: Completed Prototype
Product Integration: iHealth Edu

View GitHub Repository

[01]

Project Gallery

Terminal inference response showing low-risk prediction, estimated probability, supporting factors, and top 5 feature weights. — [01]Structured patient input parameters comprising 21 demographic, clinical, and lifestyle variables prior to model inference.

Terminal view displaying 21 clinical, demographic, and lifestyle feature inputs prior to model inference. — [01]Structured patient input parameters comprising 21 demographic, clinical, and lifestyle variables prior to model inference.

[02]

Project Overview

I developed this project as the machine learning decision-support component integrated into the live iHealth Edu website for healthcare workers. Model development used A. Panday's 2025 Heart Attack Prediction in Indonesia dataset from Kaggle, containing 158,355 observations and 21 predictors.

The workflow covers data preparation, class balancing, model comparison, evaluation, and inference deployment. Random Forest was selected for its leading accuracy and ROC-AUC, then serialized and served through a Flask REST API deployed with Docker on Ubuntu.

Dataset Source: A. Panday, Heart Attack Prediction in Indonesia, Kaggle, 2025

[03]

My Contribution

Prepared the dataset by encoding five categorical features, scaling the predictors, creating a stratified 80:20 split, and applying SMOTE only to the training data.
Trained and compared Random Forest, Linear SVM, K-Nearest Neighbors, and Logistic Regression, with hyperparameter tuning focused on F1-score.
Selected Random Forest based on its accuracy and ROC-AUC, serialized the model and preprocessing artifacts with Joblib, and built a Flask REST API for inference.
Deployed the inference service with Docker on Ubuntu and integrated its prediction output into iHealth Edu for healthcare workers.

This project gave me experience in building a machine learning workflow from data preparation to real product integration.

Tech Stack

Python
Scikit-learn
Pandas
SMOTE
Flask
Docker

[04]

System Scope

■01 // Data Preparation

158,355 observations
22 columns
21 predictors
5 categorical features
No missing values
Stratified 80:20 split
SMOTE on training data only

■02 // Model Evaluation

Model	Accuracy	Precision	Recall	F1-Score	ROC-AUC	Highlights / Notes
Random Forest[Selected]	71.93%	64.12%	68.15%	0.6607	0.8015	Selected model: highest accuracy (71.93%) and ROC-AUC (0.8015)
Logistic Regression	—	—	—	0.6618	—	Highest compared F1: 0.6618
K-Nearest Neighbors	—	—	70.40%	—	—	Highest compared Recall: 70.40%
Linear SVM	—	—	—	—	—	Compared model with no metric added

Random Forest[Selected]

Accuracy71.93%

Precision64.12%

Recall68.15%

F1-Score0.6607

ROC-AUC0.8015

Selected model: highest accuracy (71.93%) and ROC-AUC (0.8015)

Logistic Regression

Accuracy—

Precision—

Recall—

F1-Score0.6618

ROC-AUC—

Highest compared F1: 0.6618

K-Nearest Neighbors

Accuracy—

Precision—

Recall70.40%

F1-Score—

ROC-AUC—

Highest compared Recall: 70.40%

Linear SVM

Accuracy—

Precision—

Recall—

F1-Score—

ROC-AUC—

Compared model with no metric added

Random Forest was selected because it produced the highest accuracy and ROC-AUC in the documented comparison.

[CLAIM BOUNDARY // MEDICAL DECISION SUPPORT]

This prototype supports risk assessment and does not provide a clinical diagnosis or replace medical judgment.

■03 // Inference Integration

Joblib model and preprocessing artifacts
Flask REST API
Risk class and probability
Supporting factors
Global feature importance
Docker deployment on Ubuntu
iHealth Edu integration for healthcare workers

Model

Accuracy

Precision

Recall

F1-Score

ROC-AUC

Highlights / Notes

Random Forest[Selected]

71.93%

64.12%

68.15%

0.6607

0.8015

Selected model: highest accuracy (71.93%) and ROC-AUC (0.8015)

Logistic Regression

—

0.6618

—

Highest compared F1: 0.6618

K-Nearest Neighbors

—

70.40%

—

Highest compared Recall: 70.40%

Linear SVM

—

Compared model with no metric added