Logo: to the web site of Uppsala University

Rapid identification and localization of an acute coronary occlusion are vital to prevent myocardial damage, yet reliance on ST-segment ECG criteria misses many acute occlusion myocardial infarctions (OMI) and triggers unnecessary acute angiographies. Here, we present a trained and validated deep learning model using 540,372 emergency ECGs paired with definitive catheterization outcomes. The model has a C-statistic of ≥0.95 for OMI and ≥0.87 for non-OMI infarctions and can localize culprit lesions in the three main coronary branches, which can guide the angiographer. Performance is similar across age, sex, and ECG hardware subgroups. Obviating dependence on ST-elevations and troponins, this model for the identification and localization of OMI has the potential to shorten the time to reperfusion of an acute coronary occlusion and save resources. Because human oversight of OMI detection on the ECG is limited, randomized clinical trials with patient-relevant outcomes are warranted.

A major challenge in translating medical AI systems into clinical practice is their limited general-ization. In the field of physiological time series analysis, we propose a fine-tuning framework thatleverages multiple small annotated datasets from diverse domains to improve out-of-distributiongeneralization performance (OOD-GP). Through an ablation study, we demonstrate the perform-ance of our framework by evaluating the role of incorporating a greater number of independentdatasets for tine-tuning to improve OOD-GP. Our experiments involve thirteen publicly avail-able electrocardiogram and electroencephalogram datasets across four distinct tasks. In addition,we develop a method to measure the alignment of the latent space of target domains. We use thismethod to interpret our results, suggesting that multi-source domain training facilitates the learn-ing of robust cross-domain features while minimizing learning of shortcut features. To supportfurther research, we provide reproducible source code, establishing a framework and benchmarkfor studies on OOD-GP (https://github.com/EranZvuloni/MSD-generalization).

Cardiovascular diseases remain a major contributor to the global burden of healthcare, highlighting the importance of accurate and scalable methods for cardiac monitoring. Cardiac biosignals, most notably electrocardiograms (ECG) and photoplethysmograms, are essential for diagnosing, preventing and managing cardiovascular conditions across clinical and home settings. However, their acquisition varies substantially across scenarios and devices, whereas existing analytical models often rely on homogeneous datasets and static bespoke models, limiting their robustness and generalizability in diverse real-world contexts. Here we present a cardiac sensing foundation model (CSFM) that leverages transformer architectures and a generative masked pretraining strategy to learn unified representations from heterogeneous health records. CSFM is pretrained on a multimodal integration of data from various large-scale datasets, comprising cardiac signals from approximately 1.7 million individuals and their corresponding clinical or machine-generated text reports. The embeddings derived from CSFM act as effective, transferable features across diverse cardiac sensing scenarios, supporting a seamless adaptation to the varied input configurations and sensor modalities. Extensive evaluations across diagnostic tasks, demographic recognition, vital sign measurement, clinical outcome prediction and ECG question answering demonstrate that CSFM consistently outperforms traditional one-modal-one-task approaches. Notably, CSFM maintains favourable performance across both 12-lead and single-lead ECGs, as well as in scenarios involving ECG only, photoplethysmogram only or a combination of both. This highlights its potential as a versatile and scalable foundation for comprehensive cardiac monitoring.

Data-driven methods for electrocardiogram (ECG) interpretation are rapidly progressing. Large datasets have enabled advances in artificial intelligence (AI) based ECG analysis, yet limitations in annotation quality, size, and scope remain major challenges. Here we present CODE-II, a large-scale real-world dataset of 2,735,269 12-lead ECGs from 2,093,807 adult patients collected by the Telehealth Network of Minas Gerais (TNMG), Brazil. Each exam was annotated using standardized diagnostic criteria and reviewed by cardiologists. A defining feature of CODE-II is a set of 66 clinically meaningful diagnostic classes, developed with cardiologist input and routinely used in telehealth practice. We additionally provide an openly available subset: CODE-II-open, a public subset of 15,000 patients, and the CODE-II-test, a non-overlapping set of 8,475 exams reviewed by multiple cardiologists for blinded evaluation. A neural network pre-trained on CODE-II achieved superior transfer performance on external benchmarks (PTB-XL and CPSC 2018) and outperformed alternatives trained on larger datasets.

Contrastive learning is a widely adopted self-supervised pretraining strategy, yet its dependence on cohort composition remains underexplored. We present Contrasting by Augmented Patient Electrocardiograms (CAPE) foundation model and pretrain on four cohorts (n = 5,203,269), from diverse populations across three continents (North America, South America, Asia). We systematically assess how cohort demographics, health status, and population diversity influence the downstream performance for prediction tasks also including two additional cohorts from another continent (Europe). We find that downstream performance depends on the distributional properties of the pretraining cohort, including demographics and health status. Moreover, while pretraining with a multi-centre, demographically diverse cohort improves in-distribution accuracy, it reduces out-of-distribution (OOD) generalisation of our contrastive approach by encoding cohort-specific artifacts. To address this, we propose the In-Distribution Batch (IDB) strategy, which preserves intra-cohort consistency during pretraining, discourages learning of spurious cohort-specific features, and instead promotes clinically meaningful variability within cohorts. This leads to improved out-of-distribution robustness, with gains of 9-40% in downstream label prediction performance. This work provides insights into pretraining strategies for more clinically deployable and generalisable foundation models.

Transformer models pre-trained on sensory stimuli such as vision and audition have been shown to capture neural representations and to effectively predict brain responses to corresponding stimuli. Olfactory perception, however, remains a relatively underexplored sensory modality in this context. Understanding the alignment between chemical transformer models and brain activity during olfactory processing is particularly challenging due to the distributed nature of olfactory neural processing and the inherently subjective character of olfactory experience, which is strongly shaped by individual memory and prior exposure. In this work, we evaluate the representational alignment between pre-trained transformer models of chemical structure and functional magnetic resonance imaging recordings collected from human participants during olfactory perception. Using voxel-wise encoding models, we show that representations extracted from state-of-the-art molecular transformers exhibit strong alignment with olfactory brain responses, often exceeding handcrafted molecular fingerprints and physicochemical descriptors. We further analyze how this alignment varies across brain regions, model layers, and temporal response components, revealing distinct region-specific and temporal patterns. These findings represent a step toward understanding how that pre-trained chemical transformers capture meaningful structure relevant to human olfactory perception and provide insight into how molecular features relate to neural processing in the olfactory system.

BACKGROUND: Chagas disease (ChD), a neglected cardiovascular condition, affects 7.5 to 10.5 million people worldwide. Opportunistic screening during routine electrocardiography may allow earlier detection of unrecognized infections, enabling timely antiparasitic therapy or optimized cardiac care. This study prospectively evaluated the feasibility and diagnostic accuracy of an artificial intelligence (AI)-enabled ECG model enriched with 3 epidemiological questions (AI-ECG-EPI model) as an opportunistic screening for ChD in endemic regions of Brazil. METHODS: We conducted a prospective study embedded in the Telehealth Network of Minas Gerais, Brazil, across 107 municipalities in hyperendemic and endemic regions. From October 2023 to September 2024, adults undergoing routine tele-ECGs were eligible. The primary exposure was the output of the AI-ECG-EPI model: all positive individuals and a 3:1 sample of negative individuals were invited for serology (index test-dependent sampling), the reference standard. Weighted analyses (inverse probability) accounted for this sampling design. Diagnostic metrics included sensitivity, specificity, predictive values, likelihood ratios, area under the receiver operating characteristics curve, and area under the precision recall curve. The AI-ECG-EPI model was compared with both an epidemiological questions model and an AI ECG-only model. RESULTS: Among 75 779 eligible ECGs, the AI-ECG-EPI model was available for 59 154 individuals; 6812 (11.5%) screened positive. A total of 7804 selected were invited for serology, and 3509 completed testing (2517 positive; 992 negative). ChD was confirmed in 36.8% of positive and 9.3% of negative individuals. The weighted prevalence of ChD was 12.2%. Sensitivity was 34.3%, specificity 91.6%, harmonic mean of precision and recall score 0.52, and diagnostic odds ratio 5.69. The AI-ECG-EPI model showed higher discrimination than both the epidemiological questions model and AI ECG-only model (area under the receiver operating characteristics curve, 77.6% versus 70.0% and 64.5%, respectively; P<0.001). Precision recall curves showed higher precision across most recall levels. Exploratory analysis suggested improved risk reclassification (net reclassification improvement, 0.503). Performance was higher in women and in individuals with Chagas cardiomyopathy. CONCLUSIONS: In this community-based study, an AI-ECG-EPI model integrated into a public telecardiology network enabled feasible opportunistic screening for ChD in primary care. It showed acceptable diagnostic accuracy, with higher discrimination than epidemiological questions and AI ECG-only models, and identified many previously undiagnosed patients in endemic regions, supporting its potential role in scalable screening strategies.REGISTRATION: URL: ; Unique identifier: NCT02646943.

Background

Females are typically underserved in cardiovascular medicine. The use of sex as a dichotomous variable for risk stratification fails to capture the heterogeneity of risk within each sex. We aimed to develop an artificial intelligence-enhanced electrocardiography (AI-ECG) model to investigate sex-specific cardiovascular risk.

Methods

In this retrospective cohort study, we trained a convolutional neural network to classify sex using the 12-lead electrocardiogram (ECG). The Beth Israel Deaconess Medical Center (BIDMC) secondary care dataset, comprising data from individuals who had clinically indicated ECGs performed in a hospital setting in Boston, MA, USA collected between May, 2000, and March, 2023, was the derivation cohort (1 163 401 ECGs). 50% of this dataset was used for model training, 10% for validation, and 40% for testing. External validation was performed using the UK Biobank cohort, comprising data from volunteers aged 40–69 years at the time of enrolment in 2006–10 (42 386 ECGs). We examined the difference between AI-ECG-predicted sex (continuous) and biological sex (dichotomous), termed sex discordance score.

Findings

AI-ECG accurately identified sex (area under the receiver operating characteristic 0·943 [95% CI 0·942–0·943] for BIDMC and 0·971 [0·969–0·972] for the UK Biobank). In BIDMC outpatients with normal ECGs, an increased sex discordance score was associated with covariate-adjusted increased risk of cardiovascular death in females (hazard ratio [HR] 1·78 [95% CI 1·18–2·70], p=0·006) but not males (1·00 [0·63–1·58], p=0·996). In the UK Biobank cohort, the same pattern was seen (HR 1·33 [95% CI 1·06–1·68] for females, p=0·015; 0·98 [0·80–1·20] for males, p=0·854). Females with a higher sex discordance score were more likely to have future heart failure or myocardial infarction in the BIDMC cohort and had more male cardiac (increased left ventricular mass and chamber volumes) and non-cardiac phenotypes (increased muscle mass and reduced body fat percentage) in both cohorts.

Interpretation

Sex discordance score is a novel AI-ECG biomarker capable of identifying females with disproportionately elevated cardiovascular risk. AI-ECG has the potential to identify female patients who could benefit from enhanced risk factor modification or surveillance.

Electrocardiograms (ECGs) record the heart's electrical activity, offering valuable diagnostic and prognostic insights. We introduce Contrasting by Augmented Patient Electrocardiograms (CAPE), a foundation model trained using patient contrastive learning enhanced with temporal augmentations, on over six million unlabeled ECGs from diverse cohorts across three continents. We demonstrate CAPE's effectiveness to generate a meaningful generic representation and validate its performance on an open-source external cohort, where it surpasses existing pretraining methods and fully supervised deep networks.

Cross-validation is commonly used to estimate machine learning model performance on new samples. However, using it for both hyperparameter selection and error estimation can lead to overestimating model performance, especially with extensive hyperparameter searches that overly tailor models to validation data. We demonstrate that deep learning further amplifies this bias, with even minor model adjustments causing significant overestimation. Our extensive experiments on simulated and real data focus on the bias from early stopping during cross-validation. We find that overestimation intensifies with network depth and is especially severe in small datasets, which are common in physiological signal processing applications. Selecting the early stopping point during cross-validation can result in ROC-AUC estimates exceeding 90% on random data, and this effect persists across various sample sizes, architectures, and network sizes.