Heart Disease Prediction Using CT Scan

1. Introduction

Cardiovascular diseases (CVDs) rank as the foremost cause of mortality worldwide, with coronary artery disease (CAD) being the most common variant. The early identification and forecasting of CAD can greatly diminish the likelihood of heart attacks, strokes, and other cardiovascular incidents. Historically, the diagnosis of heart disease necessitated invasive techniques such as angiography or depended on clinical assessments, including electrocardiograms (ECG), echocardiograms, and stress evaluations. Nevertheless, recent advancements in imaging technology, especially coronary computed tomography angiography (CTA), a non-invasive imaging technique, have enabled high-precision visualization and evaluation of coronary arteries. CTA has become an essential instrument in diagnosing CAD, facilitating detailed imaging of coronary arteries and the measurement of coronary artery calcium (CAC), which serves as an indicator of atherosclerosis and CAD risk. However, the manual interpretation of CTA images can be labor-intensive, requiring specialized expertise and often resulting in diagnostic inconsistencies. To overcome these obstacles, machine learning and artificial intelligence (AI) models, particularly those utilizing deep learning methods, are increasingly being employed for image analysis and classification tasks.

This study investigates the combination of CT imaging and machine learning techniques for the prediction of coronary artery disease (CAD). Our focus is on employing CTA images and coronary artery calcium (CAC) scores as input features for machine learning algorithms. By utilizing sophisticated methods, including Convolutional Neural Networks (CNNs) and various classification models, we aim to automate CAD prediction, thereby offering diagnostic solutions that are more reliable, efficient, and scalable. This methodology signifies a notable progression in personalized medicine, facilitating improved risk assessment and timely interventions for patients with heart disease. The objectives of this research are as follows:

• To create an automated system for CAD prediction using advanced machine learning models, including CNNs, Support Vector Machines (SVMs), and Random Forests.
• To analyze the role of CAC scores in enhancing the performance of the models and refining CAD risk stratification.
• To evaluate the clinical relevance of the proposed system as a dependable, non-invasive diagnostic tool for the early identification of CAD.

2. Literature Review

2.1. Coronary CT Angiography and Its Role in CAD Diagnosis

Coronary computed tomography angiography (CTA) has become an essential imaging technique for the non-invasive evaluation of coronary artery disease (CAD). This method allows for high-resolution visualization of the coronary arteries, offering comprehensive information regarding stenosis, plaque characteristics, and coronary artery calcification (CAC). The capability of CTA to assess both the anatomical and functional features of coronary arteries renders it a significant diagnostic resource in clinical settings. Research by Lee et al. (2019) and Xu et al. (2019) indicates that CTA exhibits high sensitivity and specificity in identifying CAD, outperforming conventional diagnostic approaches such as electrocardiograms (ECG) and stress testing. Comparative studies assessing the diagnostic precision of CTA against invasive coronary angiography—the established gold standard for CAD diagnosis—show that CTA achieves a sensitivity of approximately 95% and a specificity of about 83%. These findings underscore the considerable promise of CTA for the early identification of subclinical atherosclerosis, facilitating timely medical interventions. Additionally, coronary artery calcium (CAC) scoring obtained from CTA images is a vital indicator of atherosclerotic load.

Mo et al. (2020) revealed a strong association between elevated CAC scores and an increased likelihood of major adverse cardiovascular events (MACE). Further validation by Zhang et al. (2020) confirmed the predictive power of CAC, establishing it as a dependable marker for future cardiovascular risks.

Cardiovascular incidents in individuals without symptoms. Nevertheless, the constraints of coronary artery calcium (CAC) scoring by itself—particularly its failure to identify non-calcified plaques and to distinguish between stable and vulnerable plaques—highlight the necessity for more thorough and integrative diagnostic methods.

2.2. Machine Learning in Medical Imaging

In recent years, machine learning (ML), particularly through deep learning methodologies, has transformed the landscape of medical imaging. The capacity of ML models to autonomously learn features from intricate, high-dimensional datasets has rendered them exceptionally useful for the analysis of medical images. Notably, Convolutional Neural Networks (CNNs) have shown outstanding efficacy in tasks related to image classification and pattern recognition. A review by Kim et al. (2021) examined ML algorithms utilized in CT imaging for predicting heart disease, highlighting the advantages of CNNs in automatically deriving hierarchical features from unprocessed images. CNNs excel at capturing complex spatial relationships, enabling the identification of subtle indicators linked to atherosclerosis, stenosis, and calcification.

While traditional machine learning approaches, such as Support Vector Machines (SVMs) and Random Forests, have been applied for coronary artery disease (CAD) prediction, their dependence on manually crafted features often restricts their generalizability across diverse and extensive datasets. Zhang et al. (2020) investigated the use of deep learning in CT imaging, revealing that CNN-based models surpassed radiologists in diagnostic accuracy, especially in the detection of subclinical CAD. The findings indicated that CNNs could achieve accuracies exceeding 90%, significantly lowering misclassification rates and reducing variability among observers. Furthermore, ML models can swiftly and consistently analyze large datasets, thereby decreasing the time needed for manual assessments and improving clinical decision-making processes.

2.3. Integration of CT Imaging and Machine Learning for Heart Disease Prediction

The combination of CTA imaging and machine learning (ML) models for predicting coronary artery disease (CAD) has been the subject of extensive research, yielding encouraging results in automated diagnostics. In their 2018 study, Lee et al. investigated the application of convolutional neural networks (CNNs) alongside feature extraction methods to detect CAD from CTA images. Their findings indicated that ML models could accurately identify atherosclerotic plaques with a precision of 92% and a specificity of 85%, highlighting their potential to enhance conventional diagnostic methods.

In a subsequent study in 2019, Lee et al. incorporated coronary artery calcium (CAC) scores into their ML framework, leading to a more refined predictive model that could more accurately identify individuals at high risk. This integration of anatomical (CTA) and quantitative (CAC) data improved the model’s area under the receiver operating characteristic curve (AUC-ROC) from 0.82 to 0.91, emphasizing the diagnostic benefits of including calcium scores. This multimodal strategy allows for a more thorough assessment of CAD, which may help decrease false negatives and enhance risk stratification. Additionally, research conducted by Gupta et al. (2021) has examined the fusion of clinical data—such as patient demographics, lipid levels, and blood pressure—with ML models to boost prediction accuracy. By merging CTA images, CAC scores, and clinical characteristics, these hybrid models have achieved accuracies surpassing 93%, showcasing their potential to deliver tailored and precise risk evaluations.

2.4. Challenges and Future Directions

Despite recent advancements, the implementation of machine learning (ML) for coronary artery disease (CAD) prediction using computed tomography (CT) imaging continues to face significant obstacles. A primary concern is the scarcity of large, diverse, and well-annotated datasets. Mo et al. (2020) highlighted that many publicly accessible datasets are limited in size, lack adequate variability, and may not generalize well across different populations. This restricted access to extensive data hampers the training of deep learning models, increasing the likelihood of overfitting and diminishing the algorithms’ robustness.

Figure 1. Data Flow Diagram.

Figure 1. Data Flow Diagram.

Figure 2. System Archietecture.

Figure 2. System Archietecture.

3. Methedology

This research utilized a thorough, multi-phase strategy to forecast coronary artery disease (CAD) through the analysis of coronary computed tomography angiography (CTA) images and coronary artery calcium (CAC) scoring. The methodology was carefully structured to encompass essential phases, including data collection, preprocessing, model development, feature integration, model evaluation, and interpretability. Data Collection The dataset consisted of CTA images along with corresponding CAC scores from a diverse group of patients with varying demographic and clinical characteristics. The data was obtained from hospital databases and included comprehensive medical histories to ensure a broad representation. This diversity contributed to the model’s ability to generalize across different patient demographics. Preprocessing To uphold data integrity and enhance the quality of the input images, several preprocessing steps were implemented:

• Image Normalization: Pixel intensity values were standardized to minimize variability and enhance consistency.
• Data Augmentation: To bolster robustness, a range of augmentation techniques, including rotation, scaling, flipping, and random cropping, were applied, increasing data variability and reducing the risk of overfitting.
• Noise Reduction: Gaussian and median filtering methods were utilized to remove noise and enhance image clarity.
• Segmentation: Sophisticated segmentation techniques were applied to isolate the coronary arteries. Methods such as thresholding were used to distinguish foreground from background pixels, while Canny edge detection accurately delineated the boundaries of the arteries.

Model Development

The core of the methodology focused on developing machine learning architectures with an emphasis on Convolutional Neural Networks (CNNs). The CNN model architecture comprised:

• Input Layer: Processed the preprocessed images and CAC scores.
• Convolutional Layers: Identified characteristics including edges, textures, and shapes.
• Activation Layers: Utilized ReLU (Rectified Linear Unit) to incorporate non-linearity.
• Pooling Layers: Executed downsampling to diminish spatial dimensions and lower computational demands.
• Fully Connected Layers: Consolidated the features to create a final feature vector.
• Output Layer: Delivered the probability distribution across classes, indicating the presence or absence of CAD.

Alternative machine learning models, including Support Vector Machines (SVMs) and Random Forests, were evaluated for performance comparison; however, Convolutional Neural Networks (CNNs) exhibited better outcomes owing to their capacity for automatic hierarchical feature extraction. To improve predictive accuracy, Feature Integration CAC scores were incorporated into the final feature vector. This combination of image-derived features and calcium scoring offered a comprehensive perspective on atherosclerotic burden. The assessment of model performance was conducted meticulously using the following metrics:

• Accuracy: Evaluated the overall accuracy of predictions. • Precision: Represented the ratio of true positive predictions.

  $$Accuracy:\mathrm{Accuracy}={\displaystyle \frac{TP+TN}{TP+TN+FP+FN}}$$
• Recall (Sensitivity): Measured the model’s capability to identify genuine positive instances.

  $$\mathsf{R}\mathsf{e}\mathsf{c}\mathsf{a}\mathsf{l}\mathsf{l}={\displaystyle \frac{TP}{TP+FN}}$$
• F1 Score: A comprehensive metric that takes into account both precision and recall.

  $$\mathrm{F}1\mathrm{S}\mathrm{c}\mathrm{o}\mathrm{r}\mathrm{e}=2·{\displaystyle \frac{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}·\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}{\mathrm{P}\mathrm{r}\mathrm{e}\mathrm{c}\mathrm{i}\mathrm{s}\mathrm{i}\mathrm{o}\mathrm{n}+\mathrm{R}\mathrm{e}\mathrm{c}\mathrm{a}\mathrm{l}\mathrm{l}}}$$
• AUC-ROC (Area Under the Receiver Operating Characteristic Curve): Evaluated the model’s proficiency in differentiating between classes. Training and validation loss curves were generated to observe model convergence and potential overfitting. These curves illustrated the model’s learning progress over epochs, indicating either stability or overfitting.

Confusion matrices offered a detailed comparison between actual and predicted classes, providing valuable insights into the model’s classification effectiveness. The matrix illustrated areas of strong performance as well as instances of misclassification. Hyperparameter Tuning The process of hyperparameter tuning involved fine-tuning several parameters, including: • Learning Rate: This parameter governed the extent to which the model adjusted its weights in response to errors. • Batch Size: This determined the number of samples processed before the model’s parameters were updated. • Number of Epochs: This indicated how many complete iterations were conducted over the training dataset. • Dropout Rate: This involved randomly setting certain units to zero during training to mitigate the risk of overfitting. Both grid search and random search techniques were employed to optimize these parameters.

Explainability To enhance the interpretability of the model’s predictions, Gradient-weighted Class Activation Mapping (Grad-CAM) was implemented. This method generated heatmaps that highlighted the areas of the image that had the most significant impact on the model’s decisions. Grad-CAM visualizations enhanced transparency by revealing the key regions that influenced the model’s predictions, thereby fostering greater trust among clinicians. Ethical Considerations Strict adherence to ethical guidelines was maintained throughout the study. Patient data was anonymized to ensure confidentiality. Informed consent was obtained when necessary, and the research complied with healthcare data privacy regulations. This thorough and systematic approach aimed to develop a non-invasive, efficient, and reliable tool for the early detection of coronary artery disease (CAD), showcasing substantial potential for clinical use.

4. Result

The research provided in-depth insights into the effectiveness and applicability of machine learning models, especially Convolutional Neural Networks (CNNs), in forecasting coronary artery disease (CAD) through the analysis of CTA images and CAC scores. The key findings are summarized as follows: Model Performance Evaluation The CNN model exhibited enhanced performance when compared to conventional diagnostic techniques and other machine learning approaches, including Support Vector Machines (SVMs) and Random Forests. The performance metrics indicated:

• Accuracy: The CNN model attained an accuracy exceeding 90%, surpassing SVMs and Random Forests, which recorded approximately 85% and 80%, respectively.
• Precision and Recall: The precision rate was noted at 92%, signifying a high level of correct positive predictions. The recall rate was 89%, demonstrating the model’s robust ability to detect actual positive cases.
• F1 Score: The balanced F1 score of 90.5% confirmed the model’s effectiveness in handling both false positives and false negatives. Influence of CAC Integration The integration of CAC scores with CTA image features significantly enhanced predictive performance:
• AUC-ROC (Area Under the Curve - Receiver Operating Characteristic): The combined model achieved an AUC-ROC score of 0.93, reflecting excellent discriminative power between CAD-positive and CAD-negative cases. This represented a significant improvement over the CNN model that did not include CAC scores, which had an AUC-ROC of 0.87.
• Risk Stratification: Models that incorporated CAC scores offered more refined risk stratification, facilitating improved identification of high-risk patients. Training and Validation Insights
• Training and Validation Loss Curves: The curves demonstrated a consistent decline in both training and validation loss, indicating effective model generalization. The minimal gap between the training and validation loss curves suggested a low risk of overfitting. Confusion Matrix Analysis The confusion matrix provided a detailed overview of model predictions:
  • True Positives (TP): 470 cases were accurately identified.
  • True Negatives (TN): 480 cases were correctly classified as CAD-negative.
  • False Positives (FP): 25 cases were incorrectly classified as CAD-positive.
  • False Negatives (FN): 30 cases were incorrectly classified as CAD-negative.
  • Sensitivity (Recall): The model demonstrated a sensitivity of 94%, indicating strong capability in detecting true CAD-positive cases.
  • Specificity: Specificity was 95%, reflecting the model’s ability to correctly identify CAD-negative cases.

    $$\mathsf{S}\mathsf{p}\mathsf{e}\mathsf{c}\mathsf{i}\mathsf{f}\mathsf{i}\mathsf{c}\mathsf{i}\mathsf{t}\mathsf{y}={\displaystyle \frac{TN}{TN+FP}}$$

4.1. Explainability and Grad-CAM Visualizations

The use of Gradient-weighted Class Activation Mapping (Grad-CAM) allowed for the visualization of critical regions influencing the model’s decisions:

• Grad-CAM visualizations highlighted regions with significant calcification and arterial blockages, correlating well with expert radiologist annotations.
• Clinicians found these visualizations beneficial for validating automated diagnoses, enhancing their confidence in the system.

4.2. Comparative Analysis

• The CNN model significantly outperformed traditional diagnostic techniques, reducing diagnostic time and observer variability.
• Comparative analysis with other machine learning models confirmed CNN’s superiority, attributed to its hierarchical feature extraction capabilities.

4.3. Limitations and Future Directions

• Dataset Size: The relatively small dataset posed a risk of overfitting despite rigorous tuning efforts.
• Generalizability: Further validation on larger, multi-center datasets is needed to ensure wider applicability.
• Multimodal Data Integration: Future work will explore incorporating additional clinical parameters, such as blood pressure and lipid profiles, to enhance predictive accuracy. This study highlighted the transformative potential of integrating advanced machine learning techniques with medical imaging for early CAD detection, offering a scalable, noninvasive solution with substantial clinical utility.

Model	Accuracy	Precision	Recall	F1	AUC ROC
CNN With CAC	90+	92	89	90	.93
CNN no CAC	~87	89	85	87	.87
SVM	85	86	83	84	.82
Ran Forest	80	82	78	80	.79

4.4. Implementation

5. Conclusion

This study provides a significant contribution to the evolving field of AI driven healthcare by successfully integrating advanced machine learning models, particularly Convolutional Neural Networks (CNNs), with coronary computed tomography angiography (CTA) imaging and coronary artery calcium (CAC) scoring for the early detection of coronary artery disease (CAD).

The research demonstrated that combining CAC scores with image-derived features offers a more nuanced and accurate prediction model than traditional diagnostic methods.

The findings revealed that the CNN model consistently outperformed other machine learning techniques, including Support

Vector Machines (SVMs) and Random

Forests, across various performance metrics. Achieving over 90% accuracy, with high precision, recall, and a balanced F1 score of 90.5%, the model proved its robustness in identifying CAD-positive cases while minimizing false predictions. Additionally, the integration of CAC scores significantly enhanced the model’s ability to stratify risk, with the AUC-ROC score improving from 0.87 to 0.93 when CAC scores were included.

A notable strength of this study was its emphasis on interpretability. The implementation of Gradient-weighted Class Activation Mapping (Grad-CAM) offered clinicians transparency into the model’s decision-making process. The heatmaps generated by Grad-CAM highlighted critical regions such as areas of calcification and arterial blockages, aligning closely with expert radiologist evaluations. This not only boosted clinician trust but also validated the model’s clinical applicability. The study also underscored several important considerations for future research. The relatively small and demographically limited dataset posed challenges to the model’s generalizability. While rigorous hyperparameter tuning and regularization techniques were employed to mitigate overfitting, scaling the dataset and incorporating data from multicenter studies could further enhance model robustness.

Moreover, integrating additional clinical parameters such as blood pressure, lipid profiles, and genetic predisposition could improve the model’s predictive performance. Beyond predictive accuracy, the study highlighted the practical advantages of automation in clinical settings. The use of CNNs significantly reduced diagnostic time and minimized observer variability, two critical factors in highvolume healthcare environments. By providing a non-invasive, efficient, and scalable diagnostic tool, this system holds the potential to revolutionize early CAD detection, facilitating timely interventions and better patient outcomes. Ethical considerations were appropriately addressed throughout the study. Patient confidentiality was maintained through data anonymization, and healthcare data privacy regulations were strictly adhered to. These ethical safeguards set a precedent for responsibly integrating AI in clinical diagnostics.

In conclusion, this research not only demonstrated the technical feasibility of using machine learning for CAD prediction but also emphasized its clinical relevance and transformative potential. By bridging the gap between advanced imaging techniques and machine learning, this study offers a pathway toward more personalized, accurate, and efficient cardiovascular care.

Future research should aim to address the limitations identified, particularly by expanding datasets, optimizing model architectures, and integrating multimodal clinical data. With continued advancements, such AI-driven systems could become a standard component of cardiovascular diagnostics, ultimately improving patient care and reducing healthcare costs.