Key result
ARO-FTAMNet outperforms deep baselines for heart disease prediction, achieving an AUROC of ~0.86.
Why the study?
Most machine-learning studies claiming to predict heart disease are evaluated only on the same small training dataset, which provides little information about how a model will perform on new patients.
Does the ARO-FTAMNet machine-learning framework improve heart disease prediction performance on an unseen external patient dataset compared to standard deep learning baselines?
Cross-Sectional (n=3,308)
Yes
Does the ARO-FTAMNet machine-learning framework improve heart disease prediction performance on an unseen external patient dataset compared to standard deep learning baselines?
A compact, feature-gated multi-scale neural network optimized by Artificial Rabbits Optimization generalizes substantially better to an unseen external heart-disease cohort (AUROC 0.8582, MCC 0.5460) than standard deep learning architectures.
ARO-FTAMNet shows promising external generalization for heart disease prediction; hypothesis-generating and should not yet change practice.
Heart disease remains the leading cause of death worldwide, and most machine-learning studies claiming to predict it are evaluated only on the same small training dataset. Such results say little about how a model will perform on new patients. This paper presents ARO (Artificial Rabbits Optimization)– FTAMNet (Future Tuned Adaptive Meta Network), a prediction framework trained on a mixed (hybrid) heart-disease corpus and tested on a separate patient set the model never encounters during training. The framework has two parts. The first is ARO, which simultaneously searches for a small, useful set of input attributes and good network settings, using a single fitness score that rewards ranking quality and balanced agreement while penalizing large feature sets. The second is FTAMNet, a compact feed-forward network with four working blocks: a dense projection, an adaptive feature gate that learns how much of each internal signal to retain, a residual learning block, and a three-branch multi-scale stage whose outputs are merged before a sigmoid decision layer. The decision threshold is not fixed at 0.5; it is tuned on validation data only and then frozen. On the external test set of 1,025 records, the framework achieves an accuracy of 0.7727, a precision of 0.7901, a recall of 0.7586, an F1-score of 0.7740, an AUROC (Area Under the Receiver Operating Characteristic Curve) of 0.8582, an AUPRC (Area Under The Precision-Recall Curve) of 0.8617, and a Matthews correlation coefficient of 0.5460. Three deep baselines built on the same data [Deep Multi-Layer Perceptron (DeepMLP), Residual Multi-Layer Perceptron (ResMLP), and Attention-based Multi-Layer Perceptron (AttMLP)] fall far behind, with MCC (Matthews Correlation Coefficient) values ranging from −0.005 to 0.078, indicating that plain deep networks do not transfer across cohorts. Removing ARO from the pipeline raises precision but cuts recall by about 15 percentage points, so the optimizer mainly buys sensitivity, which is what a screening tool needs. The study shows that a small, well-tuned network with an honest external test is more useful than a large model scored on its own data. Bangladesh Journal of Medical Science Vol. 25 No. 04 October’26 Page: 1099-1114
No takes yet. Share an insight, caveat, or question.
Patel et al. (2026) conducted a cross-sectional in Heart disease (n=3,308). ARO-FTAMNet (Artificial Rabbits Optimization–Future Tuned Adaptive Meta Network) vs. Deep baseline neural networks (DeepMLP, ResMLP, AttMLP) and unoptimized FTAMNet was evaluated on Heart disease prediction performance on the unseen external test cohort (AUROC and Matthews correlation coefficient). ARO-FTAMNet achieved an AUROC of 0.8582, accuracy of 77.27%, and MCC of 0.5460 on 1,025 external heart disease records, outperforming deep baselines (MCC -0.005 to 0.078).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: