Prospects for AI-Enhanced ECG as a Unified Screening Tool for Cardiac and Non-Cardiac Conditions -- An Explorative Study in Emergency Care

TL;DR

This exploratory study assesses whether a single 12-lead ECG can serve as a unified screening signal for predicting a broad spectrum of cardiac and non-cardiac discharge diagnoses. Using the public MIMIC-IV-ECG-ICD-ED data, the authors train multi-label classifiers (comparing XResNet1d50 and S4) to predict 1076 ICD-10 codes, achieving AUROCs above 0.8 for hundreds of codes, including 81 cardiac and 172 non-cardiac categories, with strong results in the ED setting. External validation on CODE-test corroborates performance across diverse cardiac conditions, and the S4 architecture consistently outperforms the convolutional baseline across training/evaluation scenarios; the best ED-focused model achieves a macro AUROC of about $0.774$ on ED-to-ED predictions. The work argues for the potential of AI-enhanced ECG as a scalable screening tool in emergency care, while acknowledging limitations inherent to label proxies, confounding factors, and distribution shifts, and outlining directions toward explainability and richer multi-modal inputs. Overall, the study provides evidence that ECG data, when paired with large-scale, well-curated public datasets and rigorous evaluation, can support broad, fine-grained screening across diverse clinical conditions, with meaningful implications for ED triage and patient profiling.

Abstract

Current deep learning algorithms designed for automatic ECG analysis have exhibited notable accuracy. However, akin to traditional electrocardiography, they tend to be narrowly focused and typically address a singular diagnostic condition. In this exploratory study, we specifically investigate the capability of a single model to predict a diverse range of both cardiac and non-cardiac discharge diagnoses based on a sole ECG collected in the emergency department. We find that 253, 81 cardiac, and 172 non-cardiac, ICD codes can be reliably predicted in the sense of exceeding an AUROC score of 0.8 in a statistically significant manner. This underscores the model's proficiency in handling a wide array of cardiac and non-cardiac diagnostic scenarios which demonstrates potential as a screening tool for diverse medical encounters.

Figures (6)

• Figure 1: Schematic illustration depicting the proposed workflow, featuring two patient use cases. First, consider Patient 10002430, who does not undergo hospital admission, we try to infer the ED discharge diagnosis from the initial 10s of ECG. This snippet is fed into our deep learning model, which outputs probabilities for each of the most reliably predictable ICD-10 codes (e.g., 439 codes with AUROC exceeding 0.8). Second, consider the arrival of Patient 10014652 to the Emergency Department (ED), where a variety of ECG recordings are obtained during both the ED stay and subsequent hospital admission. In this particular scenario, our objective is to predict the hospital discharge diagnosis, again based on the first 10 seconds of the recorded ECG. This approach allows us to leverage the most accurate clinical ground truth and to connect it to the first recorded ECG of the patient, which can provide valuable information for decisions at the ED.
• Figure 2: Schematic summary of the dataset composition and distribution of ICD codes across the dataset. (A) From the main MIMIC-IV-ECG-ICD-ED database of 184,700 samples across 83,738 patients, we utilize records of 75,339 patients for training, records of 4,195 patients for model selection in the validation stage, and records of 4,204 patients for testing. The median ECG records per patient is 1, however, the distribution is long-tailed with a maximum of ECG records per patient at 171. (B) represents the distribution of ICD codes according to chapters (all percentages as relative fractions compared to the dataset size), where chapter IX (Circulatory system diseases) is the most strongly represented chapter with 17.7%, closely followed by chapter XXI (Health system and status) with 17%, we present in supplementary material the distribution of cardiac conditions within chapter IX (Circulatory system diseases categories).
• Figure A.1: Schematic representation of the two model architectures considered in this work: (A) XResNet1d50 (B) S4-model.
• Figure A.2: Representation of the distribution of cardiac conditions in the ED dataset within chapter IX (Circulatory system diseases categories) at the 3rd digit level including all of its descendants, where category I25 (Chronic ischemic heart disease) is the most represented category with 19.6%, closely followed by I50 (heart failure) with 15.3%, I48 (Atrial fibrillation and flutter) with 13.5%, and I10 (Essential hypertension) with 10.7%.
• Figure A.3: Representation of the distribution of the full dataset instead of just the ED subset: Descriptive statistics on the MIMIC-IV-ECG-ICD dataset obtained by joining the MIMIC-IV-ECG with diagnoses from the clinical MIMIC-IV dataset: (A) represents the distribution of chapters, whereas (B) represents the distribution with the cardiovascular chapter IX.