A tri-modal, offline, and explainable framework for 12-lead electrocardiogram interpretation: fusing deterministic rules, a local residual network, and an agentic vision-language model.
Abstract
OBJECTIVE Automated electrocardiogram (ECG) interpretation is increasingly delivered as cloud deep-learning services, trading data sovereignty, auditability and robustness for accuracy. This study tests whether an offline, explainable framework can remove that trade-off. Approach. Three independent interpreters feed a decision-fusion layer: a deterministic engine of 26 executable clinical criteria, each finding linked to a highlighted signal segment and a literature source; a compact one-dimensional residual network trained on PTB-XL and run on-device via the Open Neural Network Exchange runtime; and an optional agentic vision-language model reading a rendered ECG image behind a security screen with human-in-the-loop escalation. Format-agnostic ingestion, lead-aware capability tiering and a three-tier signal-quality gate precede interpretation. Evaluation covers diagnostic accuracy, on-device latency, noise resilience, modality ablation and zero-shot external validation. Main results. On the PTB-XL test fold (n = 2198) the network reaches a macro-averaged area under the receiver operating characteristic curve (AUROC) of 0.932 (95% confidence interval (CI) 0.918-0.943), matching published GPU baselines, while the complete pipeline runs in 108 ± 25 ms on a CPU with no network access. Performance degrades gracefully to a signal-to-noise ratio of 5 dB (AUROC 0.951) as the quality gate rejects uninterpretable signals, and rules and network agree only weakly (Cohen's kappa 0.26), indicating complementary errors. Applied without retraining to an independent cohort (Chapman-Shaoxing, n = 1599) the model retains a macro-AUROC of 0.898. In a small pilot (25 records; 30 rule/network disagreements) the general-purpose agentic model was the weakest adjudicator (accuracy 0.267, 95% CI 0.142-0.444; exact McNemar test versus rules, p = 0.09). Significance. Edge inference, intrinsic explainability and governed multi-modal fusion give competitive accuracy with auditable findings. The agentic layer is positioned for governance rather than diagnosis, motivating ECG-specialized models. The open-source system is a research prototype, not a medical device.