Predicting Heart Disease: Harnessing Machine Learning for Early Detection - Kumaraguru College of Liberal Arts & Science (KCLAS)

INTRODUCTION

Heart disease is one of the leading causes of mortality worldwide, claiming millions of lives each year. With the rising prevalence of heart conditions, there is an urgent need for more efficient and predictive healthcare solutions. Traditional diagnostic methods, while effective, can be timeconsuming and resource-intensive, often delaying crucial interventions. Our project, Predictive Analytics Framework for Heart Disease Detection via Machine Learning, seeks to address this issue by developing a machine learning model capable of predicting the presence of he art disease based on a patient’s medical profile. By leveraging machine learning, we offer a powerful tool for early detection, enabling healthcare providers to take timely action and potentially save lives through earlier, more personalized interventions.

1.The Heart of the Problem: Why Predict Heart Disease?

Heart disease is a silent killer. Many patients remain undiagnosed until significant damage has already been done, and in some cases, the disease progresses without any outward symptoms. Medical professionals rely on various diagnostic methods, but they can be time-consuming, expensive, and limited by access to resources. Machine learning offers an opportunity to streamline the diagnostic process by analyzing multiple patient features at once, predicting the likelihood of heart disease more quickly and effectively.

Our team aimed to create a machine learning model that could analyze a combination of 13 medical features, such as age, chest pain type, cholesterol levels, and resting blood pressure, to determine the likelihood of heart disease. The dataset for this project was sourced from Kaggle and included real patient data, with the target variable indicating either the presence (1) or absence (0) of heart disease.

The project posed an exciting challenge: could we build a model that not only achieved high accuracy but was also interpretable enough to be useful in real-world healthcare applications? Machine learning models are often viewed as black boxes, but by carefully selecting the right model and preprocessing the data, we aimed to create something that doctors could potentially use to assist in diagnosis.

2. The Solution: How We Built Our Model

To tackle this problem, we decided to use a Decision Tree Classifier, a well-known algorithm for classification tasks. Decision trees are particularly useful because they provide a visual representation of decision-making, making the model more interpretable for medical professionals who may not be familiar with machine learning.

Before diving into training the model, we conducted data preprocessing to ensure the dataset was clean and ready for analysis. This included encoding categorical features like chest pain type into numerical formats using one-hot encoding. We also performed exploratory data analysis (EDA), which revealed strong correlations between certain features and the likelihood of heart disease. For instance, the chest pain type showed a significant positive correlation with heart disease, while maximum heart rate exhibited a negative correlation, indicating that lower maximum heart rates are associated with higher risk.

Once the data was prepared, we split it into training and testing sets. The challenge here was ensuring that our model could generalize well to unseen data. Overfitting was a concern, especially when using small or imbalanced datasets. In this case, the dataset had fewer instances of patients without heart disease, which could skew the results and make the model more likely to predict heart disease even in low-risk cases.

To overcome this, we used SMOTE (Synthetic Minority Over-sampling Technique) to balance the dataset. SMOTE generates synthetic samples from the minority class (patients without heart disease), helping the model learn to identify both classes equally well. This technique significantly improved our model’s performance, as we’ll discuss in the results.

3. Model Performance: How Well Did It Work?

Our evaluation process involved several metrics to assess how well the model performed. These included accuracy, precision, recall, and F1 score. Initially, we trained the model without SMOTE and then compared it to the SMOTE-augmented model.

Regular Training Results:

Accuracy: 72.13%
Precision for heart disease presence (class 1): 79.41% F1
score for absence of heart disease (class 0): 0.6667 SMOTE-

Augmented Results:

Accuracy: 89.05%
Precision for heart disease presence (class 1): 88.76%
F1 score for absence of heart disease (class 0): 0.9009

These results highlight the importance of balancing datasets, particularly in medical prediction models where false positives or negatives can have serious consequences. With SMOTE, the model was not only more accurate but also better at correctly identifying patients who did not have heart disease. This is crucial in healthcare applications, as a false positive diagnosis could lead to unnecessary tests and stress for patients, while a false negative could delay potentially life-saving treatments.

One of the key challenges we encountered was ensuring that the model didn’t become too complex, leading to overfitting. Initially, the decision tree grew too deep, overfitting to the training data and performing poorly on new data. By adjusting the hyperparameters, such as the maximum depth of the tree, we were able to strike a balance between accuracy and generalizability.

4. Challenges and Insights Gained

Throughout the development of our predictive model, we encountered various challenges, the most significant of which was dealing with the imbalanced dataset. Without balancing the dataset, the model leaned towards predicting heart disease even when it wasn’t present, leading to inflated accuracy scores that weren’t truly reflective of the model’s ability to generalize. Applying SMOTE was a pivotal moment in overcoming this issue, improving both precision and recall across both classes.

Another key insight was the importance of feature selection. While some features like chest pain type were obviously crucial, others such as maximum heart rate emerged as strong predictors of heart disease after performing EDA. These findings underscored the value of data exploration in machine learning, where seemingly less important features can provide valuable insights into patient health.

CONCLUSION

Our project demonstrates the incredible potential of machine learning in transforming healthcare, particularly in predicting heart disease. By utilizing models like decision trees and incorporating techniques such as SMOTE, we can significantly enhance early detection efforts and make more accurate predictions based on patient data. The implications of this work are vast: from enabling earlier diagnosis to crafting more personalized treatment plans, machine learning could become an indispensable tool in modern medicine.

In the future, we aim to test our model on larger and more diverse datasets to ensure its robustness. Additionally, we plan to integrate real-time data from wearable health devices, which could further enhance the model’s accuracy and usability in real-world clinical settings. As technology continues to advance, so too does the potential for machine learning to reshape how we approach healthcare, offering more proactive, data-driven care for patients around the world.

– By

Anamika, Dharshinisrii, Jeyanth, Anjana, and Akshay Roopan