Evaluation Protocols and On-Chip Feasibility in Embedded Electrocardiogram Classification: A Systematic Review

Embedded electrocardiogram (ECG) classifiers are compared through classification quality and hardware cost, while the evaluation protocol that produced the quality value is treated as a matter of statistical fairness without hardware consequence. The opposite position is developed here. The data partition (intra-subject, patient-specific, inter-patient, inter-session) fixes the variability a classifier must absorb, and therefore the capacity it requires, the memory it occupies and the class of device able to run it. Web of Science, Scopus, IEEE Xplore, PubMed and ScienceDirect were searched (last on 21 September 2026) and 54 studies were included: 30 embedded implementations on microcontroller, FPGA and ASIC targets and 24 algorithmic studies spanning five model families. Risk of bias was assessed with a five-domain instrument and the results were synthesized narratively without pooling. The primary evidence is paired. Five studies evaluated one model under more than one partition with database and preprocessing held constant, and the loss on unseen subjects ranged from 6.2 percentage points of accuracy to a macro-averaged F1 score falling from 0.81 to 0.35 in a model whose accuracy remained near 0.98. Across studies, group mean accuracy was 97.9% for intra-subject (n=12) and 92.5% for inter-patient (n=2) evaluation, a descriptive difference of 5.4 points between groups that differ in database, class set and preprocessing. Deployed footprints spanned 1.3 to 186.2 kB, and five of the twelve designs with a determinable footprint exceeded the 16 kB on-chip budget of an entry-class microcontroller, beyond which every inference requires external accesses one to two orders of magnitude more costly. The capacity cost of subject-separated evaluation is formulated as a mechanism-based hypothesis (the generalization tax), shown to be consistent with every paired observation, confronted with the on-chip memory of named devices, and tied to the measurement that would quantify it, a capacity sweep under a fixed inter-patient partition. The conclusions apply to the beat-level classification of individuals not seen in training.