A Relative Gap Analysis of Machine-Learning Frameworks for Heart-Disease Prediction and the Case for an Optimization-Driven Lightweight Transformer

Ten years of machine-learning work on heart-disease prediction has chased a single number classification accuracy on one small cohort, usually the 303-patient Cleveland set. The best of these studies now report accuracies near 98%, and the newest add explainability and synthetic oversampling on top. We argue that this progress measures the wrong thing. Three properties decide whether a risk model ever reaches a patient in a low-resource clinic-does it hold across populations it was not tuned on, can a clinician see why it decided, and will it run on the phone in a health worker's pocket-and almost none of the accuracy first literature measures any of them. This paper reads the field as a search that optimized one axis and left three unpriced. We survey the accuracy-first tradition, the parallel track of metaheuristic feature selection, and the small but growing green-AI literature, and we name five concrete gaps that follow. We then test a recent design, SFO-LiteTransMLP, against all five. Because it selects features with Synergistic Fibroblast Optimization, feeds them to a two-layer transformer of 0.42 million parameters, carries SHAP inside the pipeline rather than beside it, and is validated across four cohorts with leave-one-dataset-out transfer, it is the first model in this line that can be judged on every axis at once-and it clears the bar on all of them at roughly one-fiftieth the parameter cost of its nearest deep competitor.