CALM-ECG: Toward Accurate and Explainable ECG Analysis Through Deep Learning and Vision-Language Model Integration

Deep learning has advanced the analysis of electrocardiograms (ECG), but clinical deployment remains limited by two factors: the lack of reliable confidence estimates and the opacity of predictions. We propose CALM-ECG, a unified framework that integrates deep learning, conformal prediction (CP), and a fine-tuned vision-language model (VLM) to achieve both uncertainty-aware classification and clinically meaningful interpretation. CP generates calibrated prediction sets, enabling the system to separate confident from ambiguous cases. When the conformal predictor produces a singleton set, the fine-tuned VLM generates an explanation for the predicted class, while in the prediction set that contains multiple candidate labels, the sample is routed to a fine-tuned VLM that runs inference to generate predictions and aligned explanations. To enable domain specialization, we construct a dataset and fine-tune a multimodal VLM with parameter-efficient LoRA. Experiments on two benchmark ECG datasets show that CALM-ECG outperforms 17 competitive baselines, improving balanced accuracy by up to +22.2% and macro-F1 by +4.5 points. The fine-tuned VLM produces clinically coherent explanations with strong human-verified factual accuracy. These results demonstrate that combining calibrated uncertainty with multimodal reasoning yields a more accurate, reliable, and interpretable ECG analysis pipeline.