Patient and Database Shift in Lightweight ECG Arrhythmia Classification: A Controlled Evaluation of Noise,...

Efficient microcontroller execution does not ensure reliable electrocardiogram classification across patients and databases. We conducted a controlled benchmark of three convolutional architectures under a corrected subject-disjoint MIT-BIH split, external INCART testing, Noise Stress Test Database artifacts, 16-bit floating-point (FP16) and full-integer 8-bit (INT8) conversion, and classifier-only invocation on an ESP32 without external RAM. Five fixed seeds were used. On the MIT-BIH test cohort (DS2), INT8 four-class macro-F1 was 0.474-0.498 and supraventricular (S) recall was 0.150-0.180. Mean S-class area under the receiver operating characteristic curve (AUROC) was below 0.5 in all six architecture-regime groups, although all corresponding record-cluster intervals included 0.5. Within-record ranking could remain high despite poor pooled ranking. On INCART, S-class ranking was stronger, with all six patient-cluster AUROC intervals above 0.5, but the representative mixed-noise depthwise-separable network (DS-CNN) achieved 0.797 recall with 0.074 precision. Including record 201, which shares a subject with test record 202, increased DS2 S-recall point estimates by 2.49-14.68 percentage points. In the primary architecture panel, FP16 and INT8 met the engineering tolerances in all 120 comparisons. DS-CNN required 9.49 ms per classifier invocation on the ESP32 using preprocessed, annotation-centered inputs. Acquisition, QRS detection, preprocessing, current, and energy were not measured. Patient-dependent S-score behavior and external false-positive predictions limited this panel more than numerical precision or microcontroller resources.