AI Reads Heart Rhythms Well on Paper, but Clinical Proof Still Lags

Artificial intelligence can now detect heart disease from electrocardiograms with reported accuracy often between 90 and 99 percent, but a new systematic review warns that almost none of these models have been tested in real hospitals with real patients. The review, published in Discover Artificial Intelligence, analyzed 75 peer-reviewed studies published between 2021 and 2025 that applied machine learning, deep learning, and explainable AI to ECG-based cardiovascular disease prediction. Its central finding is a striking gap between laboratory performance and clinical readiness: models excel on curated benchmark datasets, yet relatively few have undergone external or prospective clinical validation, which the authors identify as the major barrier to bringing AI-assisted ECG diagnosis into routine practice.

The review followed PRISMA 2020 guidelines and began with 3,432 records drawn from six databases, including ScienceDirect, IEEE Xplore, Wiley Online Library, Nature Portfolio, MDPI, and Google Scholar. After removing 686 duplicates, screening titles and abstracts, and assessing full texts against strict inclusion criteria, the team led by Sabit Ahamed Preanto and colleagues narrowed the field to 75 studies. Only English-language articles in Q1 or Q2 journals qualified, and conference papers were excluded because they typically report preliminary findings with abbreviated methods. The authors acknowledge that this quality-focused strategy may have introduced publication bias, potentially omitting emerging techniques first presented at major AI conferences.

The performance figures across the included studies were consistently high. Reported diagnostic accuracy commonly ranged from approximately 90 to 99 percent, sensitivity from roughly 81 to 99 percent, and specificity from about 84 to 99 percent. Area under the receiver operating characteristic curve values reached approximately 0.98 to 0.99 in several studies, and F1-scores exceeded 0.90 on multiple benchmark datasets. However, the review emphasizes that these numbers varied substantially depending on dataset characteristics, disease category, preprocessing strategy, model architecture, validation protocol, and evaluation metrics, making direct comparison between studies unreliable.

Dataset choice emerged as a decisive factor. The most frequently used resources were the MIT-BIH Arrhythmia Database, PTB-XL, PhysioNet, CPSC 2018, and the UK Biobank, supplemented by hospital-based cohorts such as Chang Gung Memorial Hospital data and specialized collections like the INCART 12-lead database and CPSC-2021. PTB-XL contains 21,837 recordings from 18,885 patients, while the UK Biobank holds ECG-related data from roughly 56,000 individuals, making it valuable for population-level risk stratification. The review found that curated benchmark datasets produce cleaner signals and more consistent annotations, which inflates accuracy, but models trained only on such data may fail when confronted with the noise, device variation, and patient heterogeneity of real clinical environments.

On the modeling side, no single architecture dominated the field. Convolutional neural networks were widely used for morphological and image-like ECG representations, long short-term memory networks for temporal signal modeling, and transformer-based models for capturing long-range dependencies across multiple leads. Traditional machine learning methods, including support vector machines, random forests, and logistic regression, remained relevant when structured features were available, offering interpretability and lower computational cost. Hybrid CNN-LSTM architectures combined spatial and temporal feature extraction, with one study reporting 99.56 percent accuracy, though at the price of increased computational complexity. The review concludes that model suitability is context-dependent rather than universally superior, shaped by input format, dataset size, and clinical task.

Recent algorithmic innovations have pushed performance further. Attention mechanisms allow transformer models to focus on clinically important ECG segments, with one attention-based model reaching 98.6 percent accuracy for arrhythmia detection on large datasets. Transfer learning from pretrained architectures such as ResNet and VGG16 has shown promise when labeled data are scarce. Wavelet and Fourier transforms improve time-frequency feature extraction, autoencoders support denoising, and generative adversarial networks can synthesize minority-class samples to counter class imbalance. Yet the authors caution that synthetic ECG signals may introduce non-physiological artifacts, and models trained on them risk learning dataset-specific quirks rather than genuine cardiac electrophysiology, so GAN-based augmentation requires careful validation against real patient data.

Explainable AI featured prominently in the reviewed literature, with SHAP the most frequently used technique, followed by Grad-CAM and LIME. These post-hoc methods let clinicians see which parts of an ECG signal drive a prediction, such as QRS complex regions highlighted for ventricular abnormalities or ST-segment attributions linked to ischemia. But the review raises a fundamental concern: post-hoc explanations approximate model behavior rather than revealing its internal logic, and different methods can produce inconsistent explanations for the same input. Highlighted signal regions do not always correspond to physiologically meaningful components, and formal validation of XAI outputs by independent cardiologists remains rare, leaving a gap between algorithmic interpretability and clinically validated interpretability.

Several persistent challenges limit clinical translation. Class imbalance skews models toward majority classes and reduces sensitivity to rare but dangerous conditions such as ventricular fibrillation. Demographic underrepresentation in datasets like MIT-BIH, which offers limited minority and pediatric coverage, raises fairness concerns, since models trained on predominantly male cohorts may perform worse in female patients. Computational demands of deep architectures, including the quadratic cost of transformer self-attention, hinder deployment on wearable devices and in resource-constrained settings. The review also flags a subtle methodological trap: many studies use record-level rather than patient-level data splitting, allowing recordings from the same patient to appear in both training and test sets, which can leak information and inflate performance estimates.

The authors argue that future progress depends less on chasing accuracy and more on building trustworthy systems. They call for large-scale multi-institutional ECG repositories, federated learning to train models across hospitals without sharing sensitive patient data, domain adaptation techniques to bridge distribution shifts between devices and populations, and lightweight architectures using pruning, quantization, and knowledge distillation for real-time and edge deployment. They also advocate moving beyond post-hoc explanations toward intrinsically interpretable models, incorporating uncertainty estimation through Bayesian or ensemble methods, and conducting prospective, clinician-in-the-loop trials that test whether AI explanations align with expert cardiology judgment.

The review’s bottom line is measured but clear. Machine learning and deep learning have genuinely transformed what is possible in automated ECG analysis, supporting tasks from arrhythmia and myocardial infarction detection to heart failure prediction, pediatric and fetal ECG analysis, and wearable monitoring. Progress so far, however, has been driven primarily by methodological improvements rather than clinically validated deployment. Until models demonstrate robust performance across diverse populations, acquisition devices, and healthcare settings, and until ethical, legal, and regulatory questions of bias, privacy, and accountability are resolved, AI-assisted ECG diagnosis should be viewed as a promising decision-support tool awaiting its proving ground, not a replacement for expert cardiologists.

Subject of Research: Machine learning and explainable AI for ECG-based cardiovascular disease prediction

Article Title: A systematic review of machine learning and explainable artificial intelligence for electrocardiogram based cardiovascular disease prediction

Article References: Preanto, S. A., Bijoy, M. H. I., Paul, T., Khan, F., Hasan, M. Z., Arefin, M. S., Dhar, P. K., Ohidujjaman, & Samad, M. A. (2026). A systematic review of machine learning and explainable artificial intelligence for electrocardiogram based cardiovascular disease prediction. Discover Artificial Intelligence, 6(1), Article 977. https://doi.org/10.1007/s44163-026-02070-w

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02070-w

Keywords: electrocardiogram, cardiovascular disease, machine learning, deep learning, explainable AI, systematic review, arrhythmia detection, clinical validation, MIT-BIH, PTB-XL, SHAP, wearable ECG