Language-supervised electrocardiogram interpretation with vision transformer Authors: Xue Zhou, Tianhui Li, Keijiro Nakamura, Hiromasa Hayama, Shing-Hong Liu, Tetsuo Sasano, Wenxi Chen, Xin Zhu Published: 2026 ABSTRACT PurposeMost artificial intelligence models for electrocardiogram (ECG) interpretation rely on task-specific supervised learning, requiring dedicated labeled datasets for each diagnostic task and limiting scalability across diverse cardiac conditions. This study investigated whether vision-language learning from paired ECG recordings and summary reports could enable transferable ECG representations without task-specific supervised training.MethodsWe developed ECG-VLM, a contrastive vision-language model consisting of a Transformer-based image encoder and a text encoder, trained on 800,034 paired 10-s ECG recordings and summary reports from the MIMIC-IV-ECG database. Cardiac condition detection was performed using natural-language prompts without condition-specific retraining or fine-tuning. Performance was evaluated on an internal test set and externally validated using the LSAS-ECG dataset (45,151 recordings) across 14 cardiac conditions, together with retrieval and subgroup analysis.ResultsECG-VLM achieved diagnostic performance across 14 cardiac conditions, with AUROC ranging from 0.679 to 0.981 on the internal test set and from 0.691 to 0.991 on external validation. Performance was strongest for rhythm abnormalities (mean AUROC 0.948 internally and 0.921 externally) and lower for morphological conditions. Compared with the previously reported signal-based ECG-CLIP framework, ECG-VLM achieved comparable performance for rhythm abnormalities while improving AUPRC for several morphological conditions. Performance remained broadly consistent across external validation and demographic and clinical subgroups.ConclusionThese findings demonstrate the feasibility of vision-language supervision from paired ECG recordings and summary reports for learning transferable ECG representations without task-specific supervised training. This framework provides a basis for future extension of vision-language learning to ECG images encountered in routine clinical practice.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1007/s13755-026-00482-7. Introduction Cardiovascular disease remains a leading cause of mortality worldwide, and the 12-lead electrocardiogram (ECG) is indispensable for its diagnosis by providing information on cardiac rhythm, conduction, and morphology [1, 2]. Despite its clinical indispensability, ECG interpretation remains labor-intensive and requires substantial expertise, with considerable variability in diagnostic performance across clinicians and training levels [3, 4]. These challenges are further exacerbated by shortages and unequal distribution of cardiovascular specialists worldwide [5]. Artificial intelligence has demonstrated promising performance in automated ECG analysis. However, most existing AI models are developed using supervised learning, requiring task-specific labeled datasets and dedicated model training for individual diagnostic tasks [6]. Such dependence on annotated data may limit their scalability across diverse cardiac conditions and poses additional challenges for rare or underrepresented conditions, where labeled examples are often scarce. An additional practical challenge arises from the way ECGs are stored and exchanged across healthcare systems. In many clinical environments, ECGs are stored and transmitted not as digital waveform files but as images, including printed reports, scanned documents, and photographs. Signal-based deep learning methods presuppose access to structured digital waveforms, which may be unavailable in these settings. This constraint is particularly relevant in several clinical contexts: (1) resource-limited or legacy healthcare settings where paper-based ECG records may predominate and digital waveform archiving may be limited; (2) cross-institutional data exchange scenarios where proprietary or incompatible waveform formats can preclude direct signal access; and (3) retrospective research utilizing historical ECG archives that exist only in printed or scanned form. Even when waveform data are available, multiple coexisting ECG storage standards and vendor-specific implementations can complicate interoperability across healthcare systems [7]. Two broad strategies have emerged to address this gap. One seeks to recover digital waveforms from ECG images through digitization and signal reconstruction pipelines [8–10]. However, this process introduces additional sources of error, and standard paper ECG layouts, such as the 3 × 4 format, contain only short lead segments rather than the complete 10-s recording, requiring dedicated reconstruction algorithms to recover the missing portions of the signal. Another strategy performs diagnostic inference directly from ECG images without waveform reconstruction. Recent studies have demonstrated that image-based deep learning models, including vision transformers, can achieve competitive performance for cardiac disease diagnosis [11–13]. Among these, Vaid et al. demonstrated that standardized ECG images rendered from digital waveforms provide an effective representation for large-scale vision transformer pre-training [13]. Nevertheless, these image-based approaches remain predominantly task-specific, requiring dedicated labeled datasets and separate model development or fine-tuning for individual diagnostic tasks. Developing a task-agnostic vision-language framework for image-based ECG interpretation is a natural step toward more scalable diagnostic systems, even if the initial framework is built using standardized image representations derived from digital waveforms. Contrastive Language–Image Pretraining (CLIP) [14] offers a promising alternative to task-specific supervised learning. By learning aligned representations between images and natural language through contrastive objectives, CLIP enables transferable visual representations that can be adapted to diverse downstream tasks without requiring separate model retraining for each task. This paradigm has demonstrated considerable performance in medical imaging domains including chest radiography [15] and echocardiography [16], yet its application to ECG images remains unexplored. One factor that makes ECG interpretation amenable to this framework is the routine generation of paired ECG–report data during clinical practice. In many healthcare systems, ECG recordings are accompanied by structured clinical reports produced either automatically by the U.S. Food and Drug Administration (FDA)-cleared ECG acquisition systems or by interpreting clinicians. These naturally occurring dual-modal pairs provide abundant supervision without additional labeling effort. Furthermore, the standardized terminology used in ECG reports to describe rhythm, conduction, or morphological findings provides consistent semantic supervision for contrastive representation learning [17]. Additionally, most existing CLIP models are pre-trained on natural image–text pairs, whereas ECG images differ substantially from natural photographs in both appearance and information content. Rather than depicting objects or scenes, ECG images encode diagnostic information through precise temporal and amplitude relationships across twelve leads. Whether visual representations learned from natural image–text data can transfer effectively to this highly structured visual domain therefore remains unclear. In this work, we investigate whether a CLIP-based vision–language framework can learn transferable representations for ECG interpretation from paired ECG images and their corresponding summary reports. To establish a controlled foundation for this investigation, the ECG images used in this study are rendered from standardized digital waveforms under controlled preprocessing conditions, rather than derived from authentic clinical sources such as printed ECGs, scanned documents, or smartphone photographs. This study therefore aims to demonstrate the feasibility of vision–language pretraining for standardized image-based ECG representations and to provide a methodological foundation for future extension to authentic clinical ECG images with greater format and quality variability. Methods Datasets and study population The datasets and cohort construction were identical to those described in our previous ECG-CLIP work [18]. Briefly, model development and internal evaluation were conducted using the MIMIC-IV-ECG database [19, 20], which comprises 800,034 paired 12-lead ECG recordings and corresponding machine-generated ECG summary reports from 161,352 unique patients. Each ECG recording was recorded for 10-s at a sampling frequency of 500 Hz. As previously reported [18], these reports have highly standardized linguistic characteristics, with limited template and sentence diversity. The dataset was partitioned at the patient level into training, validation, and internal test sets for model optimization, model selection, and internal evaluation, respectively. Because MIMIC-IV-ECG does not provide structured ECG-level diagnostic labels, labels for 14 cardiac conditions were extracted from the corresponding ECG summary reports following the procedure described previously [18]. Briefly, diagnostic terms were identified using rule-based matching incorporating full condition names, abbreviations, and synonymous expressions, and subsequently binarized (presence = 1, absence = 0). The mappings between report expressions and cardiac conditions are provided in Supplementary Table 1. External validation was performed using the Large-Scale 12-lead ECG Database for Arrhythmia Study (LSAS-ECG) [21, 22], which contains 45,151 12-lead ECG recordings with physician-adjudicated diagnoses encoded using SNOMED-CT in the ECG header files. Ground-truth labels for the 14 cardiac conditions shared with MIMIC-IV-ECG were extracted for external evaluation, with condition-to-code mappings summarized in Supplementary Table 1. Demographic information for MIMIC-IV-ECG was obtained by linking records with the MIMIC-IV-ECG-Ext-ICD database [23, 24] using study identifiers, whereas demographic information for LSAS-ECG was extracted directly from ECG header files. Compared with MIMIC-IV-ECG, LSAS-ECG included a younger population, while sex distributions were broadly comparable. Detailed cohort characteristics, demographic distributions, and condition prevalence are summarized in Supplementary Table 2 and were previously reported in detail [18]. ECG image preprocessing and text preprocessing Raw 12-lead ECG signals were downsampled to 100 Hz and transformed into standardized 2D images. Each recording was first bandpass-filtered using a fourth-order Butterworth filter with a passband of 0.5–50 Hz to suppress baseline wander and high-frequency noise. Lead-wise min–max normalization was then applied to each lead independently to ensure all leads remain visually represented within the image regardless of amplitude differences, followed by a 0.5 offset to center waveforms around the mid-axis. In cases where individual leads were absent, the corresponding normalized values were set to zero. Processed waveform data were rendered as 224 × 224-pixel images, consistent with the input resolution of the ViT-B/32 encoder. Summary report texts were cleaned to remove special characters and standardized to lowercase, then tokenized using CLIP’s byte-pair encoding (vocabulary size 49,408) with a maximum context length of 77 tokens, following the same procedure as described in [18]. To assess the potential impact of context-length truncation, the token length distribution of all 800,034 summary reports in the training corpus was examined. The median token length was 19 tokens, with the 99th percentile at 68 tokens. Only 2,648 reports (0.33%) exceeded the 77-token context limit and were truncated. Model architecture and training We propose ECG-VLM, a vision-language model for ECG image interpretation. ECG-VLM is adapted from the CLIP architecture [14], which was pre-trained on a large-scale corpus of image-text pairs sourced from the internet, acquiring generalizable representations of visual concepts and their correspondence with natural language descriptions. This pre-training endows CLIP with a well-structured joint embedding space where semantically related image and text inputs are mapped to proximal representations. We leveraged this capability by adapting CLIP to paired ECG images and reports. ECG-VLM consists of a Vision Transformer (ViT-B/32) image encoder and a Transformer text encoder. The text encoder comprises 12 layers with a hidden dimension of 512 and 8 attention heads, processing tokenized reports with a maximum context length of 77 tokens. Both encoders project their outputs into a shared embedding space through learned linear projection layers. The pre-trained CLIP weights were used for initialization, and the entire network was fine-tuned on the training set. For each mini-batch containing N paired ECG images and reports, cosine similarity scores were computed for all N × N image and text embedding pairs. The model was optimized using a symmetric bidirectional contrastive objective, in which cross-entropy loss was computed for image-to-text retrieval and text-to-image retrieval, and the final loss was defined as the average of the two. A learnable temperature parameter τ scaled the cosine similarity scores prior to the loss computation, sharpening the contrast between matched and unmatched pairs and providing more effective gradient signals during training. τ was jointly optimized with all other model parameters throughout fine-tuning. Model optimization was performed using stochastic gradient descent with a momentum of 0.9, a learning rate of 1 × 10⁻4, and a batch size of 256 for 30 epochs. Model checkpoints were saved every 500 mini-batches, and the checkpoint with the lowest validation loss was selected for subsequent analysis. ECG-VLM for cardiac condition diagnosis Following training, ECG-VLM was deployed for cardiac condition diagnosis by exploiting the learned alignment between the image and text embedding spaces. For each cardiac condition, diagnosis was formulated as a binary classification task. A condition-specific positive prompt (Supplementary Table 1) was paired against a negative prompt (“other”), and both were encoded by the text encoder to produce label embeddings. The cosine similarity between the ECG image embedding and each label embedding was then scaled by the learned temperature parameter τ and normalized via softmax over the two candidates, yielding a condition detection probability. This procedure requires no modification to the trained model. Cardiac conditions expressible in natural language can be evaluated by substituting the corresponding prompt. For multi-class general ECG classification, a set of class-specific descriptors replaced the single positive prompt. Performance evaluation Retrieval evaluation Retrieval performance was evaluated following the multi-sampling evaluation framework established in our previous work on ECG-CLIP [18], comprising 100 independent runs with different random seeds. As in [18], reports were stratified according to their occurrence frequency in the corpus into rare (frequency = 1, n = 22,722 ECGs), common (frequency = 2–10, n = 15,851 ECGs), and frequent (frequency > 10, n = 42,111 ECGs) categories. Performance was assessed using Recall@k (k ∈ {1, 5, 10}) and mean reciprocal rank (MRR) [25, 26]. Under the instance-level criterion used in our previous work [18], a retrieval was considered correct only when the retrieved report corresponded to the original paired report of the query ECG. Although each ECG was paired with a unique report instance, for the common and frequent categories multiple such ECG-report pairs shared identical report text content. In these cases, retrieval performance was additionally evaluated using a content-based matching criterion, under which a retrieval was regarded as correct at rank k if any of the top-k retrieved reports had report text identical to that of the query ECG’s reference report, regardless of whether it was paired with the same ECG. Correspondingly, MRR was calculated as the reciprocal rank of the first retrieved report with identical report content. Cardiac condition detection performance Cardiac condition detection performance was evaluated on both the internal test set and the external validation set using area under the receiver operating characteristic curve (AUROC), area under the precision-recall curve (AUPRC), accuracy, sensitivity, specificity, precision, and F1 score. For each cardiac condition, the decision threshold was determined on the validation set using the sensitivity-constrained optimization strategy described in our previous work [18]. Candidate thresholds achieving a sensitivity of at least 0.8 were first identified, and the threshold yielding the highest F1 score among these candidates was selected. If no threshold satisfied the sensitivity constraint, the threshold with the highest F1 score was used. It should be noted that this threshold selection procedure requires labeled validation examples for each cardiac condition and therefore constitutes condition-specific calibration, distinct from the zero-shot ranking evaluation reflected by AUROC and AUPRC. Confidence intervals (95% CI) for all performance metrics were estimated using stratified bootstrap resampling with 1,000 iterations while preserving the original class distribution. A fixed random seed was used throughout the evaluation to ensure reproducibility. Comparison with previously reported models Performance of the proposed model was compared with that of the previously published ECG-CLIP model [18] across the 14 cardiac conditions evaluated in both studies. In addition, comparisons with the two supervised models (Supervised model 1: Ribeiro et al. [27]; Supervised model 2: Lu et al. [28]) reported in [18] were performed for the six cardiac conditions evaluated by all models (atrial fibrillation/flutter (AF), sinus tachycardia (ST), sinus bradycardia (SB), right bundle branch block (RBBB), left bundle branch block (LBBB), and 1 st degree atrioventricular block (1dAVb). Performance metrics for the previously published models were obtained from the corresponding results reported in [18]. Subgroup analysis Cardiac condition detection performance was further evaluated across demographic and clinical subgroups to characterize whether performance was consistent across patient populations. Age was categorized into 12–44, 45–74, and ≥75 years for the internal test set, with an additional pediatric subgroup (0–11 years) included for the external validation set. Sex-specific analysis was performed for male and female patients. Performance was also evaluated across different clinical care settings by stratifying the internal test cohort into three care intensity groups: (1) ICU inpatients (ICU), (2) non-ICU inpatients (non-ICU), and (3) patients without inpatient encounters (outpatient). This study was reported in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis + Artificial Intelligence (TRIPOD+AI) statement [29]. Results Retrieval metrics Retrieval performance under the instance-level criterion decreased monotonically with report frequency, from recall@1 of 0.413 for rare reports to 0.073 for frequent reports (Table 1). Under content-based matching, applied to the common and frequent groups, retrieval performance improved substantially across all metrics relative to the instance-level criterion, particularly for frequent reports, where recall@1 increased from 0.073 to 0.282 (Table 1). Table 1Retrieval performance across report frequency groups under instance-level and content-based matching criteriaReport groupCriterionRecall@1Recall@5Recall@10MRRRareInstance0.413 (0.392, 0.431)0.729 (0.711, 0.744)0.828 (0.812, 0.840)0.547 (0.529, 0.562)CommonInstance0.288 (0.273, 0.304)0.658 (0.641, 0.677)0.790 (0.774, 0.808)0.443 (0.430, 0.456)Content0.376 (0.360, 0.392)0.675 (0.658, 0.693)0.796 (0.780, 0.815)0.501 (0.486, 0.515)FrequentInstance0.073 (0.067, 0.081)0.228 (0.216, 0.242)0.324 (0.310, 0.337)0.140 (0.132, 0.148)Content0.282 (0.267, 0.300)0.439 (0.422, 0.459)0.535 (0.518, 0.557)0.349 (0.334, 0.364)Values are point estimates (95% CI) derived from 100 repeated evaluation runs with independently sampled subsets. Instance-level (exact) criterion required retrieval of the exact paired report; content-based matching, applied only to the common and frequent groups, considered a retrieval correct if the retrieved report was equivalent in content to the paired report. MRR mean reciprocal rank Table 1Retrieval performance across report frequency groups under instance-level and content-based matching criteriaReport groupCriterionRecall@1Recall@5Recall@10MRRRareInstance0.413 (0.392, 0.431)0.729 (0.711, 0.744)0.828 (0.812, 0.840)0.547 (0.529, 0.562)CommonInstance0.288 (0.273, 0.304)0.658 (0.641, 0.677)0.790 (0.774, 0.808)0.443 (0.430, 0.456)Content0.376 (0.360, 0.392)0.675 (0.658, 0.693)0.796 (0.780, 0.815)0.501 (0.486, 0.515)FrequentInstance0.073 (0.067, 0.081)0.228 (0.216, 0.242)0.324 (0.310, 0.337)0.140 (0.132, 0.148)Content0.282 (0.267, 0.300)0.439 (0.422, 0.459)0.535 (0.518, 0.557)0.349 (0.334, 0.364)Values are point estimates (95% CI) derived from 100 repeated evaluation runs with independently sampled subsets. Instance-level (exact) criterion required retrieval of the exact paired report; content-based matching, applied only to the common and frequent groups, considered a retrieval correct if the retrieved report was equivalent in content to the paired report. MRR mean reciprocal rank Retrieval performance across report frequency groups under instance-level and content-based matching criteria Values are point estimates (95% CI) derived from 100 repeated evaluation runs with independently sampled subsets. Instance-level (exact) criterion required retrieval of the exact paired report; content-based matching, applied only to the common and frequent groups, considered a retrieval correct if the retrieved report was equivalent in content to the paired report. MRR mean reciprocal rank Cardiac condition detection performance Detection performance varied across the 14 cardiac conditions on the internal test set (Table 2). AUROC ranged from 0.679 to 0.981, whereas AUPRC ranged from 0.130 to 0.877. Accuracy ranged from 0.508 to 0.990, sensitivity from 0.791 to 0.936, specificity from 0.460 to 0.990, precision from 0.105 to 0.786, and F1 score from 0.186 to 0.838. Sensitivity exceeded 0.79 for all cardiac conditions, whereas precision was below 0.20 for atrioventricular block (AVB, 0.105), supraventricular tachycardia (SVT, 0.108), and prolonged QT interval (PQTI, 0.118). External validation across the same 14 cardiac conditions showed a similar overall pattern. AUROC (range: 0.691–0.991) and sensitivity (0.633–0.945) were comparable to those of the internal test set, whereas AUPRC (0.058–0.982), precision (0.016–0.952), and F1 score (0.032–0.940) showed lower values and greater variability (Fig. 1). Table 2Performance of ECG-VLM on internal testing and external validationCardiac conditionEvaluation metricsInternal testingExternal validationAtrial fibrillation/flutterAUROC0.925 (0.923–0.928)0.915 (0.911–0.919)AUPRC0.712 (0.703–0.722)0.831 (0.824–0.837)Accuracy0.878 (0.876–0.880)0.833 (0.830–0.836)Sensitivity0.814 (0.806–0.822)0.860 (0.853–0.867)Specificity0.887 (0.885–0.889)0.826 (0.822–0.829)Precision0.505 (0.497–0.512)0.579 (0.571–0.586)F1 score0.623 (0.616–0.630)0.692 (0.685–0.698)Supraventricular tachycardiaAUROC0.977 (0.949–0.994)0.959 (0.952–0.965)AUPRC0.436 (0.385–0.487)0.469 (0.444–0.495)Accuracy0.990 (0.989–0.991)0.965 (0.963–0.967)Sensitivity0.849 (0.782–0.912)0.701 (0.668–0.735)Specificity0.990 (0.990–0.991)0.969 (0.968–0.971)Precision0.108 (0.088–0.128)0.272 (0.251–0.293)F1 score0.192 (0.159–0.223)0.392 (0.368–0.416)Sinus tachycardiaAUROC0.979 (0.978–0.980)0.963 (0.961–0.964)AUPRC0.804 (0.795–0.813)0.821 (0.812–0.830)Accuracy0.953 (0.952–0.955)0.905 (0.903–0.908)Sensitivity0.848 (0.840–0.856)0.885 (0.877–0.892)Specificity0.963 (0.962–0.965)0.909 (0.907–0.912)Precision0.685 (0.675–0.694)0.651 (0.642–0.661)F1 score0.758 (0.750–0.765)0.750 (0.742–0.757)Sinus bradycardiaAUROC0.981 (0.980–0.982)0.991 (0.990–0.992)AUPRC0.877 (0.871–0.884)0.982 (0.981–0.984)Accuracy0.956 (0.954–0.957)0.956 (0.954–0.958)Sensitivity0.897 (0.892–0.903)0.928 (0.924–0.932)Specificity0.964 (0.963–0.965)0.973 (0.971–0.975)Precision0.786 (0.778–0.793)0.952 (0.949–0.955)F1 score0.838 (0.832–0.842)0.940 (0.937–0.942)Premature ventricular complexAUROC0.926 (0.922–0.930)0.922 (0.908–0.935)AUPRC0.568 (0.555–0.580)0.202 (0.169–0.236)Accuracy0.898 (0.896–0.900)0.887 (0.884–0.890)Sensitivity0.792 (0.781–0.803)0.826 (0.782–0.869)Specificity0.905 (0.903–0.907)0.888 (0.885–0.891)Precision0.351 (0.342–0.360)0.046 (0.040–0.052)F1 score0.487 (0.477–0.496)0.087 (0.077–0.098)Premature atrial complexAUROC0.898 (0.893–0.902)0.779 (0.766–0.791)AUPRC0.338 (0.322–0.352)0.142 (0.127–0.158)Accuracy0.842 (0.840–0.845)0.761 (0.757–0.765)Sensitivity0.797 (0.784–0.810)0.659 (0.635–0.682)Specificity0.844 (0.842–0.847)0.764 (0.760–0.768)Precision0.190 (0.184–0.196)0.077 (0.072–0.082)F1 score0.307 (0.299–0.316)0.138 (0.130–0.146)Atrioventricular blockAUROC0.697 (0.690–0.704)0.691 (0.679–0.704)AUPRC0.130 (0.125–0.135)0.079 (0.071–0.087)Accuracy0.508 (0.504–0.511)0.291 (0.287–0.295)Sensitivity0.807 (0.796–0.817)0.914 (0.901–0.927)Specificity0.486 (0.482–0.489)0.269 (0.265–0.273)Precision0.105 (0.102–0.108)0.042 (0.040–0.045)F1 score0.186 (0.181–0.190)0.081 (0.077–0.085) 1 st degree atrioventricular blockAUROC0.836 (0.832–0.841)0.835 (0.824–0.846)AUPRC0.248 (0.238–0.259)0.168 (0.149–0.189)Accuracy0.726 (0.723–0.730)0.542 (0.538–0.547)Sensitivity0.791 (0.780–0.802)0.899 (0.883–0.916)Specificity0.722 (0.719–0.725)0.533 (0.529–0.538)Precision0.165 (0.160–0.170)0.047 (0.045–0.050)F1 score0.273 (0.267–0.280)0.090 (0.085–0.095)Right bundle branch blockAUROC0.963 (0.961–0.964)0.963 (0.958–0.969)AUPRC0.767 (0.758–0.777)0.390 (0.361–0.417)Accuracy0.937 (0.936–0.939)0.943 (0.941–0.945)Sensitivity0.801 (0.791–0.810)0.866 (0.839–0.892)Specificity0.949 (0.948–0.951)0.944 (0.942–0.946)Precision0.576 (0.565–0.586)0.184 (0.171–0.198)F1 score0.670 (0.661–0.678)0.304 (0.285–0.324)Left bundle branch blockAUROC0.935 (0.932–0.937)0.882 (0.856–0.906)AUPRC0.588 (0.581–0.595)0.347 (0.312–0.379)Accuracy0.894 (0.892–0.896)0.949 (0.947–0.952)Sensitivity0.936 (0.926–0.945)0.633 (0.568–0.692)Specificity0.893 (0.890–0.895)0.951 (0.949–0.953)Precision0.235 (0.227–0.243)0.065 (0.055–0.074)F1 score0.376 (0.366–0.386)0.118 (0.101–0.134)Left ventricular hypertrophyAUROC0.679 (0.675–0.684)0.735 (0.727–0.743)AUPRC0.504 (0.499–0.509)0.478 (0.467–0.488)Accuracy0.549 (0.546–0.553)0.529 (0.525–0.534)Sensitivity0.823 (0.813–0.831)0.927 (0.906–0.946)Specificity0.518 (0.514–0.521)0.523 (0.519–0.528)Precision0.166 (0.162–0.169)0.027 (0.025–0.030)F1 score0.276 (0.271–0.281)0.053 (0.049–0.057)ST-T changeAUROC0.721 (0.717–0.726)0.761 (0.748–0.773)AUPRC0.372 (0.364–0.380)0.077 (0.069–0.086)Accuracy0.526 (0.522–0.529)0.357 (0.353–0.361)Sensitivity0.824 (0.818–0.830)0.945 (0.931–0.957)Specificity0.460 (0.456–0.464)0.342 (0.337–0.346)Precision0.251 (0.247–0.255)0.036 (0.034–0.038)F1 score0.385 (0.380–0.390)0.070 (0.066–0.074)Prolonged QT intervalAUROC0.829 (0.822–0.836)0.757 (0.732–0.781)AUPRC0.295 (0.282–0.311)0.105 (0.084–0.128)Accuracy0.709 (0.706–0.712)0.612 (0.607–0.616)Sensitivity0.798 (0.786–0.811)0.742 (0.700–0.783)Specificity0.704 (0.701–0.708)0.610 (0.606–0.615)Precision0.118 (0.114–0.122)0.016 (0.015–0.019)F1 score0.205 (0.199–0.211)0.032 (0.029–0.036)Low QRS voltageAUROC0.734 (0.728–0.739)0.748 (0.732–0.763)AUPRC0.202 (0.195–0.210)0.058 (0.052–0.063)Accuracy0.550 (0.547–0.554)0.473 (0.468–0.478)Sensitivity0.807 (0.799–0.816)0.858 (0.835–0.879)Specificity0.526 (0.523–0.530)0.464 (0.459–0.469)Precision0.138 (0.135–0.141)0.036 (0.034–0.039)F1 score0.236 (0.231–0.241)0.070 (0.066–0.074)Values represent mean values with 95% confidence intervals shown in parentheses. Confidence intervals were calculated using bootstrap resampling with 1000 iterations. AUROC Area Under the Receiver Operating Characteristic curve, AUPRC Area Under the Precision-Recall CurveFig. 1Diagnostic performance of ECG-VLM on the internal test set and external validation set across 14 cardiac conditions. Radar plots show a AUROC, b AUPRC, c accuracy, d sensitivity, e specificity, and f F1 score. Each axis represents one cardiac condition: sinus bradycardia (SB), sinus tachycardia (ST), supraventricular tachycardia (SVT), atrial fibrillation (AF), low QRS voltage (LowQRS), prolonged QT interval (PQTI), ST-T change (STTC), left ventricular hypertrophy (LVH), left bundle branch block (LBBB), right bundle branch block (RBBB), 1 st degree atrioventricular block (1dAVb), atrioventricular block (AVB), premature atrial complex (PAC), and premature ventricular complex (PVC), with radial gridlines indicating performance values from 0 to 1.0. Blue lines denote internal testing and red lines denote external validation Table 2Performance of ECG-VLM on internal testing and external validationCardiac conditionEvaluation metricsInternal testingExternal validationAtrial fibrillation/flutterAUROC0.925 (0.923–0.928)0.915 (0.911–0.919)AUPRC0.712 (0.703–0.722)0.831 (0.824–0.837)Accuracy0.878 (0.876–0.880)0.833 (0.830–0.836)Sensitivity0.814 (0.806–0.822)0.860 (0.853–0.867)Specificity0.887 (0.885–0.889)0.826 (0.822–0.829)Precision0.505 (0.497–0.512)0.579 (0.571–0.586)F1 score0.623 (0.616–0.630)0.692 (0.685–0.698)Supraventricular tachycardiaAUROC0.977 (0.949–0.994)0.959 (0.952–0.965)AUPRC0.436 (0.385–0.487)0.469 (0.444–0.495)Accuracy0.990 (0.989–0.991)0.965 (0.963–0.967)Sensitivity0.849 (0.782–0.912)0.701 (0.668–0.735)Specificity0.990 (0.990–0.991)0.969 (0.968–0.971)Precision0.108 (0.088–0.128)0.272 (0.251–0.293)F1 score0.192 (0.159–0.223)0.392 (0.368–0.416)Sinus tachycardiaAUROC0.979 (0.978–0.980)0.963 (0.961–0.964)AUPRC0.804 (0.795–0.813)0.821 (0.812–0.830)Accuracy0.953 (0.952–0.955)0.905 (0.903–0.908)Sensitivity0.848 (0.840–0.856)0.885 (0.877–0.892)Specificity0.963 (0.962–0.965)0.909 (0.907–0.912)Precision0.685 (0.675–0.694)0.651 (0.642–0.661)F1 score0.758 (0.750–0.765)0.750 (0.742–0.757)Sinus bradycardiaAUROC0.981 (0.980–0.982)0.991 (0.990–0.992)AUPRC0.877 (0.871–0.884)0.982 (0.981–0.984)Accuracy0.956 (0.954–0.957)0.956 (0.954–0.958)Sensitivity0.897 (0.892–0.903)0.928 (0.924–0.932)Specificity0.964 (0.963–0.965)0.973 (0.971–0.975)Precision0.786 (0.778–0.793)0.952 (0.949–0.955)F1 score0.838 (0.832–0.842)0.940 (0.937–0.942)Premature ventricular complexAUROC0.926 (0.922–0.930)0.922 (0.908–0.935)AUPRC0.568 (0.555–0.580)0.202 (0.169–0.236)Accuracy0.898 (0.896–0.900)0.887 (0.884–0.890)Sensitivity0.792 (0.781–0.803)0.826 (0.782–0.869)Specificity0.905 (0.903–0.907)0.888 (0.885–0.891)Precision0.351 (0.342–0.360)0.046 (0.040–0.052)F1 score0.487 (0.477–0.496)0.087 (0.077–0.098)Premature atrial complexAUROC0.898 (0.893–0.902)0.779 (0.766–0.791)AUPRC0.338 (0.322–0.352)0.142 (0.127–0.158)Accuracy0.842 (0.840–0.845)0.761 (0.757–0.765)Sensitivity0.797 (0.784–0.810)0.659 (0.635–0.682)Specificity0.844 (0.842–0.847)0.764 (0.760–0.768)Precision0.190 (0.184–0.196)0.077 (0.072–0.082)F1 score0.307 (0.299–0.316)0.138 (0.130–0.146)Atrioventricular blockAUROC0.697 (0.690–0.704)0.691 (0.679–0.704)AUPRC0.130 (0.125–0.135)0.079 (0.071–0.087)Accuracy0.508 (0.504–0.511)0.291 (0.287–0.295)Sensitivity0.807 (0.796–0.817)0.914 (0.901–0.927)Specificity0.486 (0.482–0.489)0.269 (0.265–0.273)Precision0.105 (0.102–0.108)0.042 (0.040–0.045)F1 score0.186 (0.181–0.190)0.081 (0.077–0.085) 1 st degree atrioventricular blockAUROC0.836 (0.832–0.841)0.835 (0.824–0.846)AUPRC0.248 (0.238–0.259)0.168 (0.149–0.189)Accuracy0.726 (0.723–0.730)0.542 (0.538–0.547)Sensitivity0.791 (0.780–0.802)0.899 (0.883–0.916)Specificity0.722 (0.719–0.725)0.533 (0.529–0.538)Precision0.165 (0.160–0.170)0.047 (0.045–0.050)F1 score0.273 (0.267–0.280)0.090 (0.085–0.095)Right bundle branch blockAUROC0.963 (0.961–0.964)0.963 (0.958–0.969)AUPRC0.767 (0.758–0.777)0.390 (0.361–0.417)Accuracy0.937 (0.936–0.939)0.943 (0.941–0.945)Sensitivity0.801 (0.791–0.810)0.866 (0.839–0.892)Specificity0.949 (0.948–0.951)0.944 (0.942–0.946)Precision0.576 (0.565–0.586)0.184 (0.171–0.198)F1 score0.670 (0.661–0.678)0.304 (0.285–0.324)Left bundle branch blockAUROC0.935 (0.932–0.937)0.882 (0.856–0.906)AUPRC0.588 (0.581–0.595)0.347 (0.312–0.379)Accuracy0.894 (0.892–0.896)0.949 (0.947–0.952)Sensitivity0.936 (0.926–0.945)0.633 (0.568–0.692)Specificity0.893 (0.890–0.895)0.951 (0.949–0.953)Precision0.235 (0.227–0.243)0.065 (0.055–0.074)F1 score0.376 (0.366–0.386)0.118 (0.101–0.134)Left ventricular hypertrophyAUROC0.679 (0.675–0.684)0.735 (0.727–0.743)AUPRC0.504 (0.499–0.509)0.478 (0.467–0.488)Accuracy0.549 (0.546–0.553)0.529 (0.525–0.534)Sensitivity0.823 (0.813–0.831)0.927 (0.906–0.946)Specificity0.518 (0.514–0.521)0.523 (0.519–0.528)Precision0.166 (0.162–0.169)0.027 (0.025–0.030)F1 score0.276 (0.271–0.281)0.053 (0.049–0.057)ST-T changeAUROC0.721 (0.717–0.726)0.761 (0.748–0.773)AUPRC0.372 (0.364–0.380)0.077 (0.069–0.086)Accuracy0.526 (0.522–0.529)0.357 (0.353–0.361)Sensitivity0.824 (0.818–0.830)0.945 (0.931–0.957)Specificity0.460 (0.456–0.464)0.342 (0.337–0.346)Precision0.251 (0.247–0.255)0.036 (0.034–0.038)F1 score0.385 (0.380–0.390)0.070 (0.066–0.074)Prolonged QT intervalAUROC0.829 (0.822–0.836)0.757 (0.732–0.781)AUPRC0.295 (0.282–0.311)0.105 (0.084–0.128)Accuracy0.709 (0.706–0.712)0.612 (0.607–0.616)Sensitivity0.798 (0.786–0.811)0.742 (0.700–0.783)Specificity0.704 (0.701–0.708)0.610 (0.606–0.615)Precision0.118 (0.114–0.122)0.016 (0.015–0.019)F1 score0.205 (0.199–0.211)0.032 (0.029–0.036)Low QRS voltageAUROC0.734 (0.728–0.739)0.748 (0.732–0.763)AUPRC0.202 (0.195–0.210)0.058 (0.052–0.063)Accuracy0.550 (0.547–0.554)0.473 (0.468–0.478)Sensitivity0.807 (0.799–0.816)0.858 (0.835–0.879)Specificity0.526 (0.523–0.530)0.464 (0.459–0.469)Precision0.138 (0.135–0.141)0.036 (0.034–0.039)F1 score0.236 (0.231–0.241)0.070 (0.066–0.074)Values represent mean values with 95% confidence intervals shown in parentheses. Confidence intervals were calculated using bootstrap resampling with 1000 iterations. AUROC Area Under the Receiver Operating Characteristic curve, AUPRC Area Under the Precision-Recall Curve Performance of ECG-VLM on internal testing and external validation Values represent mean values with 95% confidence intervals shown in parentheses. Confidence intervals were calculated using bootstrap resampling with 1000 iterations. AUROC Area Under the Receiver Operating Characteristic curve, AUPRC Area Under the Precision-Recall Curve Diagnostic performance of ECG-VLM on the internal test set and external validation set across 14 cardiac conditions. Radar plots show a AUROC, b AUPRC, c accuracy, d sensitivity, e specificity, and f F1 score. Each axis represents one cardiac condition: sinus bradycardia (SB), sinus tachycardia (ST), supraventricular tachycardia (SVT), atrial fibrillation (AF), low QRS voltage (LowQRS), prolonged QT interval (PQTI), ST-T change (STTC), left ventricular hypertrophy (LVH), left bundle branch block (LBBB), right bundle branch block (RBBB), 1 st degree atrioventricular block (1dAVb), atrioventricular block (AVB), premature atrial complex (PAC), and premature ventricular complex (PVC), with radial gridlines indicating performance values from 0 to 1.0. Blue lines denote internal testing and red lines denote external validation Grouping the 14 cardiac conditions into rhythm abnormalities (AF, SVT, ST, SB, premature ventricular complex (PVC), premature atrial complex (PAC)), conduction disorders (AVB, 1dAVb, RBBB, LBBB), and morphological conditions (left ventricular hypertrophy (LVH), ST-T change (STTC), prolonged QT interval (PQTI), low QRS voltage (LowQRS)) revealed consistent differences in performance across all metrics (Table 2, Fig. 1). Rhythm abnormalities showed the highest overall detection performance, with a mean AUROC of 0.948 on the internal test set and 0.921 on the external validation set, and five of the six conditions achieved AUROC values above 0.90 on both datasets. Conduction disorders showed intermediate performance (mean AUROC 0.858 internal and 0.843 external), whereas morphological conditions consistently exhibited the lowest AUROC (0.741 internal and 0.750 external) among the three groups, although this category showed similar AUROC values between the internal and external datasets. Performance comparison Compared with the previously published ECG-CLIP model, ECG-VLM showed largely comparable detection performance for rhythm abnormalities on both the internal test set and external validation set (Supplementary Fig. 1). Performance differences were more apparent for conduction disorders, where ECG-CLIP generally achieved higher AUROC and AUPRC. In contrast, ECG-VLM showed consistently higher AUPRC for several morphological conditions, particularly LVH and STTC, on both datasets. Overall, performance differences between ECG-VLM and ECG-CLIP were smallest for rhythm abnormalities and greater for conduction disorders, whereas ECG-VLM generally showed higher performance for morphological conditions. Compared with the two supervised models, ECG-VLM achieved broadly comparable AUROC across the six cardiac conditions evaluated by all models (AF, ST, SB, RBBB, LBBB, and 1dAVb), with only modest differences for most conditions (Supplementary Fig. 2). AUPRC was consistently lower than that of the supervised models on the internal test set but became more comparable on the external validation set, where ECG-VLM matched or exceeded the supervised models for several conduction disorders. Overall, ECG-VLM achieved AUROC comparable to the supervised models, whereas differences in AUPRC were more evident on the internal test set and became smaller on external validation. Subgroup analysis Overall, cardiac condition detection performance remained broadly consistent across demographic and clinical subgroups. Across the internal test set (Fig. 2, Supplementary Tables 3–5), AUROC decreased modestly with increasing age (mean 0.891, 0.870, and 0.831 for patients aged 12–44, 45–74, and ≥75 years, respectively). Performance was nearly identical between male and female patients. Similarly, only minor variation was observed across care settings, with slightly lower AUROC in ICU patients than in non-ICU inpatients and outpatients. AUPRC did not mirror the age-related AUROC trend, instead peaking in the 45–74 years group. External validation showed similar sex-specific performance and broadly comparable age-related patterns among the shared age groups (12–44, 45–74, and ≥75 years) (Supplementary Fig. 3, Supplementary Tables 6–7). The pediatric subgroup (0–11 years), evaluated only in the external cohort, represented the main exception, achieving the highest AUROC but the lowest AUPRC among all age groups.Fig. 2Diagnostic performance of ECG-VLM on the internal test set stratified by age, sex, and care intensity across 14 cardiac conditions. Radar plots show a AUROC stratified by age group, b AUROC stratified by sex, c AUROC stratified by care intensity, d AUPRC stratified by age group, e AUPRC stratified by sex, and (f) AUPRC stratified by care intensity. Each axis represents one cardiac condition: sinus bradycardia (SB), sinus tachycardia (ST), supraventricular tachycardia (SVT), atrial fibrillation (AF), low QRS voltage (LowQRS), prolonged QT interval (PQTI), ST-T change (STTC), left ventricular hypertrophy (LVH), left bundle branch block (LBBB), right bundle branch block (RBBB), 1 st degree atrioventricular block (1dAVb), atrioventricular block (AVB), premature atrial complex (PAC), and premature ventricular complex (PVC), with radial gridlines indicating performance values from 0 to 1.0 in increments of 0.2. In a and d, blue, red, and purple lines denote patients aged 12–44, 45–74, and ≥75 years, respectively; in (b) and (e), blue lines denote male patients and red lines denote female patients; in c and f, blue, red, and purple lines denote ICU, non-ICU, and outpatient care settings, respectively Diagnostic performance of ECG-VLM on the internal test set stratified by age, sex, and care intensity across 14 cardiac conditions. Radar plots show a AUROC stratified by age group, b AUROC stratified by sex, c AUROC stratified by care intensity, d AUPRC stratified by age group, e AUPRC stratified by sex, and (f) AUPRC stratified by care intensity. Each axis represents one cardiac condition: sinus bradycardia (SB), sinus tachycardia (ST), supraventricular tachycardia (SVT), atrial fibrillation (AF), low QRS voltage (LowQRS), prolonged QT interval (PQTI), ST-T change (STTC), left ventricular hypertrophy (LVH), left bundle branch block (LBBB), right bundle branch block (RBBB), 1 st degree atrioventricular block (1dAVb), atrioventricular block (AVB), premature atrial complex (PAC), and premature ventricular complex (PVC), with radial gridlines indicating performance values from 0 to 1.0 in increments of 0.2. In a and d, blue, red, and purple lines denote patients aged 12–44, 45–74, and ≥75 years, respectively; in (b) and (e), blue lines denote male patients and red lines denote female patients; in c and f, blue, red, and purple lines denote ICU, non-ICU, and outpatient care settings, respectively Discussion Our findings suggest that vision-language supervision derived from ECG-report pairs has the potential to support multi-condition ECG interpretation without task-specific supervised training. ECG-VLM achieved diagnostic performance that varied meaningfully across cardiac condition categories: performance was strongest for rhythm abnormalities, intermediate for conduction disorders, and more limited for morphological conditions, with precision remaining low for several conditions as a consequence of the sensitivity-prioritizing threshold selection strategy. Performance was nonetheless broadly consistent across independent external validation and demographic and clinical subgroups. These findings suggest that summary reports may provide sufficient semantic supervision to learn transferable ECG representations applicable to multiple diagnostic tasks from a single trained model. Unlike conventional supervised approaches, which require dedicated labeled datasets and separate model development for each diagnostic task, the proposed framework enables condition evaluation through natural-language prompts without retraining, suggesting that vision-language learning may offer a scalable alternative for expanding ECG diagnostic coverage. A finding of this study was the consistently better performance for rhythm abnormalities than for morphological conditions across both the internal and external datasets. Importantly, a similar category-dependent performance gradient was also observed in our previously reported signal-based ECG-CLIP framework [18], despite the different input modality, suggesting that this pattern is unlikely to be specific to image-based ECG representations. One possible explanation is that these condition categories rely on fundamentally different diagnostic criteria. Rhythm abnormalities are generally identified from characteristic rhythm patterns and described using standardized diagnostic terminology, such as atrial fibrillation and sinus bradycardia, according to established ECG interpretation guidelines [30]. In contrast, many morphological conditions depend on relatively subtle quantitative waveform characteristics, including ST-segment deviation, T-wave morphology, interval measurements, and voltage-based criteria defined by clinical thresholds [31, 32]. While the diagnostic conclusions are routinely documented, the accompanying reports in our dataset generally do not explicitly include the quantitative measurements underlying these findings, and morphological diagnoses are often further qualified by descriptors specifying the affected leads or anatomical regions (e.g., inferior ST-T changes, lateral ST-T changes, and low QRS voltage in limb leads). These characteristics may increase semantic variability in the language supervision available for contrastive learning. Taken together, these observations raise the possibility that the observed performance gradient reflects differences in the clinical and linguistic characteristics of the diagnostic categories, rather than limitations specific to either input modality. Against this common category-dependent performance gradient, comparison with the signal-based ECG-CLIP model [18] revealed that the relative strengths of the two representations differed across diagnostic categories. Performance for rhythm abnormalities was largely comparable, whereas ECG-CLIP retained an advantage for conduction disorders and ECG-VLM achieved higher AUPRC for several morphological conditions, most consistently for LVH and STTC, across both datasets. Although the reasons for these differences remain uncertain, one possible explanation is that image-based representations preserve the spatial organization and inter-lead relationships of the standard clinical ECG layout, potentially facilitating recognition of morphological patterns that are interpreted by considering waveform appearance across multiple leads. In contrast, signal-based representations may better preserve the precise temporal information required for interval-based measurements, which could contribute to their advantage for conduction disorders. These observations suggest that the two representations capture complementary aspects of ECG information and motivate future work exploring hybrid signal-image vision-language models that integrate the strengths of both modalities. Comparison with the two supervised models retrained on the same dataset [18, 27, 28], restricted to the six shared diagnostic tasks, revealed a more nuanced pattern than suggested by AUROC alone. While ECG-VLM consistently yielded lower AUPRC than both supervised models on the internal test set, particularly for right and left bundle branch block, this performance gap became substantially smaller in the external cohort. One explanation is the marked difference in disease prevalence between the two datasets. The prevalence of RBBB and LBBB decreased from 7.9% to 1.4% and from 3.4% to 0.5%, respectively (Supplementary Table 2), and because AUPRC is intrinsically dependent on class prevalence, such shifts reduce the attainable AUPRC for all models evaluated on the external cohort [33]. Beyond prevalence differences, the two datasets also differ in label provenance: internal labels were derived from machine-generated ECG summary reports via rule-based text matching, whereas external labels in LSAS-ECG are physician-adjudicated SNOMED-CT-encoded diagnoses. This difference in labeling methodology represents an additional source of interpretive uncertainty when comparing internal and external performance, as observed metric differences may partly reflect systematic variation in label quality and diagnostic granularity rather than true differences in model generalizability. However, the relatively smaller decline observed for ECG-VLM also raises the possibility that vision-language supervision may encourage representations that are less tightly coupled to the label distribution of the training cohort than conventional task-specific supervised learning. Although the present study cannot disentangle these effects, the findings suggest that robustness to dataset shift may represent an important dimension for future comparisons between foundation models and conventional supervised approaches, ideally using external cohorts with controlled prevalence differences or multiple independent validation datasets. Beyond the performance comparison with the signal-based ECG-CLIP framework, the image-based approach considered in the present study has a practical motivation related to the availability and format of ECG data. Signal-based frameworks such as ECG-CLIP require access to digital waveform data, which may not be available when ECGs are retained only as visual records, such as printed reports that are subsequently scanned or photographed, paper-based historical archives, or images exchanged across institutions without the underlying waveform data. In such settings, image-based analysis provides a means of working directly with the ECG representation that is available, without requiring recovery of the original waveform. The present study provides an initial assessment of the feasibility of vision-language pretraining for standardized image-based ECG representations, laying the groundwork for future extension of this framework to authentic clinical image formats, including scanned ECGs and smartphone photographs. Diagnostic performance also remained broadly consistent across the demographic and clinical subgroups examined, despite the model being trained without explicit demographic information. This finding suggests that the ECG representations learned through vision-language supervision may generalize across the age, sex, and care-setting distributions represented in the current cohorts. Nevertheless, age and sex are well-established determinants of ECG interpretation [30, 34, 35]. Rather than indicating that demographic information is unnecessary, these results suggest that ECG-VLM has the potential to learn clinically meaningful representations without explicit demographic inputs, while also motivating future work to investigate whether incorporating demographic metadata as an additional modality further improves performance, particularly for conditions with known age- or sex-dependent ECG manifestations. The retrieval analysis also provides insight into future directions for ECG-language pretraining. Applying a content-based matching criterion improved retrieval performance, particularly for frequently occurring report templates, indicating that many retrievals considered incorrect under instance-level evaluation were nevertheless clinically equivalent. This observation reflects an intrinsic characteristic of routine ECG reports, which are highly standardized and frequently share identical or near-identical semantic content across different patients. Under CLIP’s symmetric InfoNCE objective, however, every unpaired ECG-report combination within a mini-batch is treated as a negative example regardless of semantic similarity, such that semantically equivalent ECG-report pairs are explicitly separated in the embedding space. This false-negative problem has recently attracted increasing attention in medical vision-language pretraining, motivating content-aware and soft-label contrastive objectives that reduce penalties between semantically matched pairs [36, 37]. More broadly, our findings raise the possibility that ECG reports may be more naturally modeled as semantic contexts rather than isolated ECG-specific instances. Extending ECG representation learning with language supervision beyond strict instance-level correspondence toward context-level semantic supervision therefore represents a promising direction for future foundation models. Several limitations of this study should be acknowledged. Firstly, the ECG images used for model development and evaluation were synthetically rendered from standardized digital 12-lead ECG waveforms under controlled preprocessing conditions, rather than derived from authentic clinical image sources such as printed ECGs, scanned documents, or smartphone photographs. It is important to distinguish the ECG representation from the underlying diagnostic task. The present study addressed conventional 12-lead ECG interpretation, differing from signal-based approaches by representing the ECG as standardized images rather than digital waveforms. The results therefore demonstrate the feasibility of vision-language learning for this image-based representation of 12-lead ECGs. However, real-world clinical ECG images exhibit substantial additional variability, including diverse paper layouts (e.g., 3×4, 2×6, or rhythm-strip formats), grid backgrounds, annotations, folds, scanning artifacts, and variations in image quality, lighting, and camera perspective, none of which were represented in the present dataset. Consequently, the present study should be viewed as establishing a vision-language learning framework for standardized image representations of conventional 12-lead ECG recordings, rather than as a demonstration of performance on authentic clinical ECG images. This provides a methodological foundation for future studies that systematically investigate extension to real-world formats, including robustness evaluation under degraded image quality, diverse paper layouts, and acquisition-related artifacts. Secondly, the fixed-resolution image representation used in this study may not fully preserve the fine-grained temporal and amplitude resolution present in the original digital waveforms. Standard 12-lead ECG images spanning a full 10-s recording across twelve leads inevitably compress substantial temporal detail into a limited pixel grid, which may blur subtle features such as precise interval boundaries, high-frequency notching, or small-amplitude deflections that are more directly accessible in signal-based representations. This resolution constraint may partly explain the relative advantage of signal-based ECG-CLIP for conduction disorders, which depend on precise interval measurements, and reinforces the motivation for hybrid signal-image approaches discussed above, wherein higher temporal fidelity from signal inputs may complement the spatial and morphological representation captured by image-based encoding. Future work could explore higher-resolution or multi-scale image encoding strategies, such as patch-based or hierarchical vision transformers that preserve finer spatial detail within each lead, to evaluate whether increased image resolution narrows this gap while retaining the practical advantages of image-based ECG representation. Thirdly, the diagnostic labels used for model development and internal evaluation were derived from machine-generated reports in the MIMIC-IV-ECG dataset using rule-based text matching rather than expert re-adjudication of each recording. Although this approach enables large-scale label extraction, it may introduce label noise arising from ambiguous phrasing, negation, or synonymous expressions not fully captured by the matching rules, potentially affecting the precision of the reported detection metrics. In contrast, the external LSAS-ECG dataset provides physician-adjudicated diagnostic labels, reducing this source of labeling uncertainty during external validation. This difference in label provenance between the two databases means that direct comparisons of performance metrics between the internal and external datasets should be interpreted with caution, as observed differences may reflect not only population-level factors such as demographic composition and disease prevalence, but also systematic differences in label quality between the two datasets. A further consideration concerns the choice of text encoder. ECG-VLM adopts the standard CLIP text encoder, which uses a general-domain English byte-pair encoding vocabulary rather than a tokenizer trained on biomedical text. Domain-adapted language models such as those underlying BiomedCLIP are pre-trained on biomedical corpora and may better represent clinical terminology. The use of a general-domain text encoder was motivated by consistency with the ECG-CLIP framework and by the highly standardized and lexically constrained nature of machine-generated MIMIC-IV-ECG reports, which exhibit limited vocabulary diversity compared with free-text clinical notes. Nevertheless, whether a biomedical text encoder would further improve learned ECG-text alignments remains an open question and represents a direction for future investigation. Regarding the 77-token context-length limit, analysis of the full training corpus confirmed that the median report length was 19 tokens and that only 2,648 reports (0.33%) exceeded the context limit, suggesting that truncation was unlikely to have had a substantial impact on model training. Furthermore, external validation was conducted using a single external cohort. While this dataset differs from MIMIC-IV-ECG in demographic composition and disease prevalence, generalizability to healthcare systems, acquisition protocols, and geographically or ethnically diverse populations remains to be established through validation on additional independent cohorts. Precision was also low for several conditions, in part reflecting the threshold selection strategy adopted in this study, which prioritized sensitivity over positive predictive value. This suggests the model in its current configuration is better suited to prioritizing recordings for clinician review than to autonomous diagnosis. In addition, the threshold selection procedure required labeled validation examples for each cardiac condition and therefore involved condition-specific calibration. Evaluating a fixed condition-independent threshold or assessing threshold transfer across datasets would provide further insight into the practical applicability of the framework and should be investigated in future studies. Conclusion This study demonstrates the feasibility of a vision-language framework trained on paired ECG images and summary reports to learn transferable ECG representations without task-specific supervised training. ECG-VLM achieved promising diagnostic performance across diverse cardiac conditions, with broadly consistent performance in external validation and across demographic and clinical subgroups, particularly for rhythm abnormalities. By using ECG images rendered from standardized digital waveforms, this work provides a basis for extending vision-language learning to real-world clinical ECG image formats, including scanned ECGs, printed ECGs, and smartphone photographs. Future work incorporating real-world image variability, context-aware contrastive learning objectives, and prospective multicenter clinical evaluation will be important for translating this framework toward practical clinical deployment. Supplementary Information Below is the link to the electronic supplementary material.Supplementary file1 (PDF 994 KB) Supplementary file1 (PDF 994 KB)