CardioSSL: When Does Self-Supervision Help Heart Sound Screening? A Controlled Benchmark

Phonocardiogram (PCG) recordings are cheap and easy to collect, which makes them attractive for cardiovascular screening where echocardiography is unavailable. What is scarce is the expert label, which is the standard motivation for self-supervised learning (SSL). Despite its benefit and growing interests, no controlled benchmark exists for comparing SSL objectives under scarce data labels and distribution shift. We benchmark five structurally distinct SSL objectives (SimCLR, BYOL, MAE, DINO, JEPA) against two controls, a randomly initialized encoder and a supervised convolutional network, under one shared backbone, one fine-tuning procedure, and one evaluation protocol. We evaluate at 10, 25, 50 and 100 percent of the training labels, with paired significance testing and multiple-comparison correction at every fraction, and we test transfer to an external pediatric cohort. Three findings emerge. First, the benefit of SSL is real and confined to the low-label regime: at 10 percent of labels, JEPA and MAE gain 9.44 and 8.89 balanced-accuracy points over the control (Holm-corrected $p<0.001$, Cohen's $d>3$), and the advantage decays monotonically until no method is significantly better than the control at full supervision. Second, the gain is attributable to self-supervision rather than to the backbone: a supervised convolutional control matches the transformer control at every label fraction. Third, strong in-distribution performance does not survive a cohort shift: all seven conditions reach AUROC 0.938 to 0.954 on the in-distribution test set, yet fall to 0.49 to 0.52 when transferred to CirCor, a completely different benchmarking dataset, even though a model trained within CirCor itself reaches AUROC 0.636. We argue that PCG-SSL should be evaluated as a label-efficiency method under external validation, not as an accuracy improvement under full supervision.