Comparative evaluation of deep learning optimizers for cardiovascular risk prediction using clinical features from an Indian hospital dataset

Background: Cardiovascular disease (CVD) remains a leading cause of morbidity and mortality globally, with particularly high prevalence and earlier onset in South Asian populations. Routine clinical features such as age, blood pressure, serum cholesterol, chest pain type, and exercise induced ST changes offer predictive value for automated risk stratification. Deep learning (DL) models provide powerful tools for this purpose; however, optimizer selection, a critical determinant of model convergence and performance, has been insufficiently studied in clinical contexts.

Methods: We analyzed a publicly available dataset comprising 1000 patient records (918 complete cases) from a multispecialty hospital in India, with 12 demographic, clinical, and diagnostic features and a binary outcome for cardiovascular disease presence. Preprocessing included outlier assessment, standardization, and categorical encoding. Descriptive statistics and non-parametric tests were applied to evaluate feature distributions and associations. A feedforward neural network was trained using six optimizers- SGD, RMSprop, Adam, Adagrad, Adamax, and Nadam under 5-fold stratified cross-validation. Performance was assessed using accuracy, precision, recall, F1-score, and ROC–AUC.

Results: RMSprop, Adam, and Nadam consistently outperformed others, achieving mean AUC ≈ 0.99 and F1 ≈ 0.97, while Adagrad underperformed (AUC = 0.97, F1 = 0.94). Statistical analyses confirmed significant associations of resting blood pressure, serumcholesterol, maximum heart rate, oldpeak, and chest pain type with cardiovascular disease status. Multicollinearity diagnostics indicated independence across predictors, supporting model stability.

Conclusions: Optimizer choice critically influences DL model performance for cardiovascular disease prediction. Adaptive optimizers (RMSprop, Adam, Nadam) demonstrated superior discrimination and reliability, underscoring the importance of explicit optimizer benchmarking in clinical ML workflows.