Cardiac-Net: A Wearable ECG Prototype and Hybrid CNN–Transformer Model for Atrial Fibrillation Detection in Long-Term ECG Recordings

Abstract

Postoperative atrial fibrillation (POAF) is a leading cause of stroke and mortality after cardiac surgery, motivating continuous long-term ECG monitoring.This paper presents Cardiac-Net, a wearable ECG system (ESP32-S3/AD8232 BLE device with a Flutter mobile gateway; 128 Hz nominal/85 Hz effective BLE sampling rate, 12.6 ms mean latency, zero over-the-air loss of transmitted packets, 3.11 mJ per transmitted sample) combined with a hybrid 1D CNN–Transformer classifier and a proposed cloud-assisted monitoring architecture; the cloud inference and monitoring-dashboard layer is a design-level proposal only and has not been implemented. Bench signal fidelity, validated against a BIOPAC MP36 reference using a Fluke ProSim 8 simulator across five noise conditions, yields a Pearson Correlation Coefficient of 0.9222 and BPM errors below 0.2% under clean conditions, within the 5% heart-rate-error threshold specified by IEC 60601-2-47. The classifier was trained and evaluated on the Long-Term Atrial Fibrillation Database (84 recordings) using a record-wise (patient-wise) data split with verified zero overlap between training, validation, and test recordings; 5-fold record-wise cross-validation spanning all 84 recordings yields AUC =0.953±0.016, F1 =0.894±0.030, precision =0.880±0.051, and recall =0.915±0.073; re-evaluating at the device’s actual 85 Hz effective throughput confirms comparable discrimination (AUC =0.946±0.010) alongside a real increase in false-alarm burden. A clinical evaluation protocol beyond segment-level accuracy shows that most AF episodes are detected (episode-level sensitivity 0.839±0.222 by fold; per-patient median 1.000, IQR [0.925, 1.000]), but this is not uniform across patients: 7.2% (6/83) fall below chance-level detection. Combined with a false-alarm burden of 86±38 alerts per patient-day and a prevalence-adjusted positive predictive value of only 0.28–0.44 at 5–10% assumed AF prevalence, alarm calibration and patient-level reliability both remain barriers to unsupervised clinical deployment. Zero-shot evaluation on the independent MIT-BIH Atrial Fibrillation Database, with no fine-tuning, yields an ensemble AUC of 0.834, indicating generalization beyond the training distribution alongside a measurable domain gap. Post-hoc Grad-CAM and Integrated Gradients attributions are consistent with physiologically grounded decision-making, concentrating on RR-interval irregularity and QRS morphology. Cardiac-Net thus contributes a validated wearable acquisition layer and a rigorously, record-wise cross-validated detection model, together with an honest characterization of the false-alarm and calibration challenges that remain before postoperative clinical deployment.