Deep Learning-based Early Prediction of Heart Disease using Clinical Healthcare Data

Abstract

Cardiovascular Diseases (CVD) are the number one cause of death globally and are responsible for a significant proportion of all global deaths annually, and are particularly prevalent in LMICs where diagnostic resources are scarce. Accurate and early prediction of heart disease risk from routinely collected clinical information can enable early intervention, minimize diagnostic delay, and reduce long-term treatment. This article presents a hybrid deep learning model that processes one-dimensional Convolutional Neural Networks (CNNs) for local feature extraction, a Long Short-Term Memory (LSTM) network for temporal and sequential modeling of the longitudinal clinical measurements, and a fully connected classification head with an attention mechanism to attend to clinically salient features. The model is trained and tested using a combined dataset consisting of the Cleveland, Hungarian, Switzerland, and Long Beach VA subsets of the UCI Heart Disease repository, along with the Framingham and Z-Alizadeh Sani datasets to enhance the model's generalizability. The missing values are handled using multiple imputation; the Synthetic Minority Oversampling Technique (SMOTE) is used to combat class imbalance, and z-score normalization is applied to scale the features. Experimental results are then shown, which includes an accuracy of 94.6%, sensitivity of 93.8%, specificity of 95.1%, F1-score of 94.2% and an Area Under the Receiver Operating Characteristic Curve (AUC-ROC) of 0.978, outperforming the deep learning baselines (plain CNN, plain LSTM, deep neural network) and classical machine learning models (logistic regression, random forest, support vector machine, XGBoost) when tested under the same conditions. Using SHAP (SHapley Additive exPlanations) model interpretation, the most influential parameters are chest pain type, maximum heart rate achieved, number of major vessels colored by fluoroscopy, and thalassemia status, in accordance with known clinical knowledge. The outcomes indicate that the proposed framework can be a promising decision support tool in primary care and resource-constrained environments, and that it is interpretable and computationally efficient.