Predicting Cardiovascular Disease with CNNs: A Comparative Study of Raw Tabular and Heatmap Transformed Data

A significant global health concern, cardiovascular diseases (CVDs) necessitate precise predictive models for early detection. Using structured tabular data, this study investigates the prediction of heart disease using Convolutional Neural Networks (CNNs). Two methods were compared: a 2D CNN trained on 3x4 grayscale heatmap images created from the same data, and a 1D CNN applied to reshaped feature vectors with input shape (12,1). Accuracy, ROC AUC, and F1 Score were the metrics used to assess both models. With an accuracy of 69.89%, F1-score of 64.28%, and ROC AUC of 69.91%, the 1D CNN performed better than the 2D CNN, indicating superior feature relationship learning. The 2D CNN performed worse on all metrics, suggesting that the accuracy of the model might not be improved by transforming structured data into image-like formats. These findings highlight the significance of data representation and demonstrate that 1D CNNs perform better when processing structured clinical data for CVD prediction tasks.