Summary
I am interested in the intersection between machine learning, automatic control, signal processing, and its application in medicine.- Samples of my work on machine learning: ICLR (2026), NeurIPS (2025), ICML (2025), AISTATS (2025), NeurIPS (2024), ICML (2024), NeurIPS (2023), AISTATS (2020) .
- Samples of my work on system identification and signal processing: Automatica (2025), ICASSP (2025), IEEE Transactions on Signal Processing (2023), ICASSP (2021), Automatica (2020) .
- Samples of my work on medical applications: Nature Communications (2026), npj Digital Medicine (2026), npj Digital Medicine (2025), npj Digital Medicine (2025), npj Digital Medicine (2024), IEEE Transactions on Biomedical Engineering (2023), Circulation (2023), Nature Communications (2022), Nature Communications (2021), Nature Communications (2020) .
Full Publication List
See a curated publication list below. You can also download the PDF version.2026
- A multi-source domain fine-tuning framework for deep generalization performance in physiological time series analysisMachine Learning: Health (2026)
Abstract
A major challenge in translating medical AI systems into clinical practice is their limited generalization. In the field of physiological time series analysis, we propose a fine-tuning framework that leverages multiple small annotated datasets from diverse domains to improve out-of-distribution generalization performance (OOD-GP). Through an ablation study, we demonstrate the performance of our framework by evaluating the role of incorporating a greater number of independent datasets for tine-tuning to improve OOD-GP. Our experiments involve thirteen publicly available ECG and EEG datasets across four distinct tasks. In addition, we develop a method to measure the alignment of the latent space of target domains. We use this method to interpret our results, suggesting that multi-source domain training facilitates the learning of robust cross-domain features while minimizing learning of shortcut features. To support further research, we provide reproducible source code, establishing a framework and benchmark for studies on OOD-GP [URL provided upon publication.
- CODE-II: A large-scale dataset for artificial intelligence in ECG analysisnpj Digital Medicine (2026)
Abstract
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 open 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.
- Cardiac health assessment across scenarios and devices using a multimodal foundation model pretrained on data from 1.7 million individualsNature Machine Intelligence, 8, 220–233 (2026)Cover of the february issue
Abstract
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.
- A deep learning ECG model for localization of occlusion myocardial infarctionNature Communications, 17 (2026)
Abstract
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 trained and validated a deep learning model using 540,372 emergency ECGs paired with definitive catheterization outcomes. The model achieved a C-statistic of \textgreater0.95 for OMI and \textgreater0.87 for non-OMI infarctions and could localize culprit lesions in the left main/LAD, LCX, and RCA coronary branches, which can guide the angiographer. Performance was similar across age, sex, and ECG hardware subgroups. Obviating dependence on ST-elevations and troponins, this model for identification and localization of OMI has the potential to shorten 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.
- Explaining deep learning for ECG using time-localized clustersIEEE Transactions on Biomedical Engineering (2026)
Abstract
Deep learning has significantly advanced electrocardiogram (ECG) analysis, enabling automatic annotation, disease screening, and prognosis beyond traditional clinical capabilities. However, understanding these models remains a challenge, limiting interpretation and gaining knowledge from these developments. In this work, we propose a novel interpretability method for convolutional neural networks applied to ECG analysis. Our approach extracts time-localized clusters from the model’s internal representations, segmenting the ECG according to the learned characteristics while quantifying the uncertainty of these representations. This allows us to visualize how different waveform regions contribute to the model’s predictions and assess the certainty of its decisions. By providing a structured and interpretable view of deep learning models for ECG, our method enhances trust in AI-driven diagnostics and facilitates the discovery of clinically relevant electrophysiological patterns.
- On th Interaction of Compressibility and Adversarial RobustnessInternational Conference for Learning Representations (ICLR) (2026)
Abstract
Modern neural networks are expected to simultaneously satisfy a host of desirable properties: accurate fitting to training data, generalization to unseen inputs, parameter and computational efficiency, and robustness to adversarial perturbations. While compressibility and robustness have each been studied extensively, a unified understanding of their interaction still remains elusive. In this work, we develop a principled framework to analyze how different forms of compressibility - such as neuron-level sparsity and spectral compressibility - affect adversarial robustness. We show that these forms of compression can induce a small number of highly sensitive directions in the representation space, which adversaries can exploit to construct effective perturbations. Our analysis yields a simple yet instructive robustness bound, revealing how neuron and spectral compressibility impact Ł_}infty and Ł_2 robustness via their effects on the learned representations. Crucially, the vulnerabilities we identify arise irrespective of how compression is achieved - whether via regularization, architectural bias, or implicit learning dynamics. Through empirical evaluations across synthetic and realistic tasks, we confirm our theoretical predictions, and further demonstrate that these vulnerabilities persist under adversarial training and transfer learning, and contribute to the emergence of universal adversarial perturbations. Our findings show a fundamental tension between structured compressibility and robustness, and suggest new pathways for designing models that are both efficient and secure.
- A Robust Optimization Approach to Sparse Principal Component AnalysisTechnical Report (2026)
Abstract
While principal component analysis (PCA) is a fundamental tool for dimensionality reduction, its dense representations make it ill-suited for high-dimensional data. Existing methods address this by promoting sparsity through explicit \ell1-penalties, but these are not obvious to tune due to the unsupervised nature of the task. In contrast, we propose Adversarial PCA (AdvPCA), which leverages robust optimization to achieve sparsity by optimizing the reconstruction objective against bounded, worst-case latent space perturbations. We show that this formulation admits a closed-form reduction, leading to a practical iterative algorithm that alternates between adversarial linear regression-style updates for the sparse encoder and orthogonal updates for the decoder. By theoretically characterizing the solution, we derive a data-adaptive parameterization that allows the algorithm to perform effectively out of the box. We validate these claims through numerical experiments on synthetic and real-world genomics data.
- How Do Electrocardiogram Models Scale?Technical Report (2026)
Abstract
While scaling laws have established a fundamental framework for foundation models in natural language processing, their applicability to electrocardiogram (ECG) models remains poorly characterized. Indeed, recent studies do not always yield consistent downstream gains as one increases the model size or pre-training dataset size of ECG models, leaving the exact roles of architectural inductive biases, pre-training paradigms, and expected improvements with size largely unanswered. In this work, we systematically investigate neural and loss-to-loss scaling laws within the ECG domain. By pre-training over \120 models (ranging from \20\K to \200\M parameters) on the large-scale CODE dataset (\2.3\M records), we decouple the effects of model architecture (ResNet vs. Transformer) and pre-training paradigm, namely supervised learning (SL) versus self-supervised learning (SSL). We found that (i) SL models are data-bottlenecked in-distribution, whereas SSL models scale robustly across both model and data sizes; (ii) for out-of-distribution (OOD) generalization, ResNets are \1.3 to \2.5 times more parameter-efficient than Transformers, while SSL is up to \16 times more data-efficient and achieves up to \7.6 times higher transfer efficiency than SL on unseen clinical tasks; (iii) across the observed scales, ResNet-based models generally achieve the lowest OOD loss, with SSL dominating on unseen clinical tasks and self-supervised Transformers overtaking at very large model sizes. Our results suggest that the path to effective ECG foundation models lies in the strategic alignment of architecture and paradigm rather than brute-force scaling.
- Anytime-Valid Conformal Risk ControlTechnical Report (2026)
Abstract
Prediction sets provide a means of quantifying the uncertainty in predictive tasks. Using held out calibration data, conformal prediction and risk control can produce prediction sets that exhibit statistically valid error control in a computationally efficient manner. However, in the standard formulations, the error is only controlled on average over many possible calibration datasets of fixed size. In this paper, we extend the control to remain valid with high probability over a cumulatively growing calibration dataset at any time point. We derive such guarantees using quantile-based arguments and illustrate the applicability of the proposed framework to settings involving distribution shift. We further establish a matching lower bound and show that our guarantees are asymptotically tight. Finally, we demonstrate the practical performance of our methods through both simulations and real-world numerical examples.
2025
- The cost of explainability in artificial intelligence-enhanced electrocardiogram modelsnpj Digital Medicine, 8, 747 (2025)
Abstract
Artificial intelligence-enhanced electrocardiogram (AI-ECG) models have shown outstanding performance in diagnostic and prognostic tasks, yet their black-box nature hampers clinical adoption. Meanwhile, a growing demand for explainable AI in medicine underscores the need for transparent, trustworthy decision-making. Moving beyond post-hoc explainability techniques that have shown unreliable results, we focus on explicit representation learning using variational autoencoders (VAE) to capture inherently interpretable ECG features. While VAEs have demonstrated potential for ECG interpretability, the presumed performance-explainability trade-off remains underexplored, with many studies relying on complex, non-linear methods that obscure the morphological information of the features. In this work, we present a novel framework (VAE-SCAN) to model bi-directional, interpretable associations between ECG features and clinical factors. We also investigate how different representations affect ECG decoding performance across models with varying levels of explainability. Our findings demonstrate the cost introduced by intrinsic ECG interpretability, based on which we discuss potential implications and directions.
- A comparison of artificial intelligence–enhanced electrocardiography approaches for the prediction of time to mortality using electrocardiogram imagesEuropean Heart Journal - Digital Health, 6, 180–189 (2025)
Abstract
Abstract Aims Most artificial intelligence-enhanced electrocardiogram (AI-ECG) models used to predict adverse events including death require that the ECGs be stored digitally. However, the majority of clinical facilities worldwide store ECGs as images. Methods and results A total of 1 163 401 ECGs (189 539 patients) from a secondary care data set were available as both natively digital traces and PDF images. A digitization pipeline extracted signals from PDFs. Separate 1D convolutional neural network (CNN) models were trained on natively digital or digitized ECGs, with a discrete-time survival loss function to predict time to mortality. A 2D CNN model was trained on 310 \times 868 px ECG images. External validation was performed in 958 954 ECGs (645 373 patients) from a Brazilian primary care cohort and 1022 ECGs (1022 patients) from a Chagas disease cohort. The image 2D CNN model and digitized 1D CNN model performed comparably to natively digital 1D CNN model in internal [C-index 0.780 (0.779–0.781), 0.772 (0.771–0.774), and 0.775 (0.774–0.776), respectively] and external validation. Models trained on natively digital 1D ECGs had comparable performance when applied to digitized 1D ECGs [C-index 0.773 (0.771–0.774)]. Conclusion Both the image 2D CNN and digitized 1D CNN enable mortality prediction from ECG images, with comparable performance to natively digital 1D CNN. Models trained on natively digital 1D ECGs can also be applied to digitized 1D ECGs, without any significant loss in performance. This work allows AI-ECG mortality prediction to be applied in diverse global settings lacking digital ECG infrastructure.
- Prediction of Atrial Fibrillation From the ECG in the Community Using Deep Learning: A Multinational StudyCirculation: Arrhythmia and Electrophysiology (2025)
- Explainable AI associates ECG aging effects with increased cardiovascular risk in a longitudinal population studynpj Digital Medicine, 8 (2025)
Abstract
Aging affects the 12-lead electrocardiogram (ECG) and correlates with cardiovascular disease (CVD). AI-ECG models estimate aging effects as a novel biomarker but have only been evaluated on single ECGs—without utilizing longitudinal data. We validated an AI-ECG model, originally trained on Brazilian data, using a German cohort with over 20 years of follow-up, demonstrating similar performance (r2 = 0.70) to the original study (0.71). Incorporating longitudinal ECGs revealed a stronger association with cardiovascular risk, increasing the hazard ratio for mortality from 1.43 to 1.65. Moreover, aging effects were associated with higher odds ratios for atrial fibrillation, heart failure, and mortality. Using explainable AI methods revealed that the model aligns with clinical knowledge by focusing on ECG features known to reflect aging. Our study suggests that aging effects in longitudinal ECGs can be applied on population level as a novel biomarker to identify patients at risk early.
- Artificial intelligence-enhanced electrocardiography for the identification of a sex-related cardiovascular risk continuum: a retrospective cohort studyLancet Digit. Health, 7, e184–e194 (2025)
- Evaluation of AI ECG age in the prediction of cardiovascular diseases and risk factors: Exploratory data analysisJournal of Electrocardiology, 93, 154123 (2025)
Abstract
Cardiovascular diseases (CVDs) are the most widespread cause of death across the world, and this aspect requires a better risk stratification method. Effectively, this paper tests the potential of the artificial intelligence (AI)-predicted electrocardiogram (ECG) age as a novel CVD risk predictor biomarker. We used the CODE-15 % data (over 344,000 ECG records with clinical annotation) to create a logistic regression model that incorporated age predicted from ECG, as well as demographic and comorbidity variables, to evaluate cardiovascular outcomes. The model performed better with ECG-predicted age than using chronological age alone, with a 92 % receiver operating characteristic curve area (AUROC) for atrial fibrillation (AF) detection, compared to 88 % when chronological age was used alone. From the results, the ECG-predicted age showed a stronger correlation with AF (r = 0.17), bundle branch blocks (r = 0.14), and Chagas disease (r = 0.074) than chronological age. Sex-specific patterns were also observed; male sex had an increased risk of sinus bradycardia (odds ratio [OR]: 3.090), while female sex had a lower risk of having a left bundle branch block (LBBB) OR: 0.885) with the variable of ECG-predicted age. These results demonstrate how ECG-predicted age may be used as a biologically meaningful indicator for cardiovascular risk assessment. This study advances personalized medicine by offering a cost-effective and scalable method for detecting CVD, particularly in under-resourced regions that lack conventional biomarkers.
- Exploring the feasibility of olfactory brain–computer interfacesScientific Reports, 15, 1–13 (2025)
- Deep learning amplified early stopping bias: Overestimating performance on small datasetsInternational conference on acoustics, speech, and signal processing (ICASSP) (2025)
Abstract
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.
- Human-Aligned Image Models Improve Visual Decoding from the BrainInternational Conference on Machine Learning (ICML) (2025)
Abstract
Decoding visual images from brain activity has significant potential for advancing brain-computer interaction and enhancing the understanding of human perception. Recent approaches align the representation spaces of images and brain activity to enable visual decoding. In this paper, we introduce the use of human-aligned image encoders to map brain signals to images. We hypothesize that these models more effectively capture perceptual attributes associated with the rapid visual stimuli presentations commonly used in visual brain data recording experiments. Our empirical results support this hypothesis, demonstrating that this simple modification improves image retrieval accuracy by up to 21% compared to state-of-the-art methods. Comprehensive experiments confirm consistent performance improvements across diverse EEG architectures, image encoders, alignment methods, participants, and brain imaging modalities.
- Efficient Optimization Algorithms for Linear Adversarial TrainingInternational Conference on Artificial Intelligence and Statistics (AISTATS) (2025)
Abstract
Adversarial training can be used to learn models that are robust against perturbations. For linear models, it can be formulated as a convex optimization problem. Compared to methods proposed in the context of deep learning, leveraging the optimization structure allows significantly faster convergence rates. Still, the use of generic convex solvers can be inefficient for large-scale problems. Here, we propose tailored optimization algorithms for the adversarial training of linear models, which render large-scale regression and classification problems more tractable. For regression problems, we propose a family of solvers based on iterative ridge regression and, for classification, a family of solvers based on projected gradient descent. The methods are based on extended variable reformulations of the original problem. We illustrate their efficiency in numerical examples.
- Detection of Chagas Disease from the ECG: The George B. Moody PhysioNet Challenge 2025Computing in Cardiology (CinC) (2025)
Abstract
Objective: Chagas disease is a parasitic infection that is endemic to South America, Central America, and, more recently, the U.S., primarily transmitted by insects. Chronic Chagas disease can cause cardiovascular diseases and digestive problems. Serological testing capacities for Chagas disease are limited, but Chagas cardiomyopathy often manifests in ECGs, providing an opportunity to prioritize patients for testing and treatment. Approach: The George B. Moody PhysioNet Challenge 2025 invites teams to develop algorithmic approaches for identifying Chagas disease from electrocardiograms (ECGs). Main results: This Challenge provides multiple innovations. First, we leveraged several datasets with labels from patient reports and serological testing, provided a large dataset with weak labels and smaller datasets with strong labels. Second, we augmented the data to support model robustness and generalizability to unseen data sources. Third, we applied an evaluation metric that captured the local serological testing capacity for Chagas disease to frame the machine learning problem as a triage task. Significance: Over 630 participants from 111 teams submitted over 1300 entries during the Challenge, representing diverse approaches from academia and industry worldwide.
- Contrasting by Augmented Patient Electrocardiograms to Learn Representations for a Foundation ModelInternational Conference on Artificial Intelligence in Medicine, 15735, 202–206 (2025)
- Single-Lead, Single-Beat: Smartwatch ECG Risk Assessment in Decompensated Heart Failure Patients via Domain-Guided AttentionTechnical Report (2025)
Abstract
Deep learning models have achieved remarkable diagnostic performance in electrocardiography signals but are often criticized for lacking physiological interpretability and robustness in consumer-grade devices. In this work, we propose a novel, domain-guided attention framework by constraining learning to specific intervals (P-wave, QRS complex, full heartbeat) to investigate whether arrhythmia-specific information can be localized efficiently in a single heart beat in a single lead. We evaluate the suitability by performing electrocardiography classification on public 12-lead databases and predicting clinical events from 1-lead smartwatch data of cardiac patients (NCT06819618). First, models were trained on 12-lead data, i.e. the Brazilian CODE dataset, internally validated on CODE15, and externally tested on PTB-XL, Georgia, and CPSC-2018. Being trained on single heartbeats and lead I only, our approach achieved comparable performance across four arrhythmia compared to full ECGs. Constraining attention does not degrade performance and Explainable AI analysis confirmed that attention aligns with physiology. Second, we applied the model to smartwatch data of N=32 patients with decompensated heart failure with reduced ejection fraction, showing that beat-level arrhythmia burden predicts rehospitalization and mortality in 3-month follow-up. Our results demonstrate that domain guidance enables physiology-aware classification and risk assessment-in single heart beats measured via consumer-grade smartwatches.
2024
- Prognostic Significance and Associations of Neural Network–Derived Electrocardiographic FeaturesCirculation: Cardiovascular Quality and Outcomes, 0 (2024)
- Genetic and phenotypic architecture of human myocardial trabeculationNature Cardiovascular Research (2024)
Abstract
Cardiac trabeculae form a network of muscular strands that line the inner surfaces of the heart. Their development depends on multiscale morphogenetic processes and, while highly conserved across vertebrate evolution, their role in the pathophysiology of the mature heart is not fully understood. Here we report variant associations across the allele frequency spectrum for trabecular morphology in 47,803 participants of the UK Biobank using fractal dimension analysis of cardiac imaging. We identified an association between trabeculation and rare variants in 56 genes that regulate myocardial contractility and ventricular development. Genome-wide association studies identified 68 loci in pathways that regulate sarcomeric function, differentiation of the conduction system and cell fate determination. We found that trabeculation-associated variants were modifiers of cardiomyopathy phenotypes with opposing effects in hypertrophic and dilated cardiomyopathy. Together, these data provide insights into mechanisms that regulate trabecular development and plasticity, and identify a potential role in modifying monogenic disease expression.
- Generalization challenges in electrocardiogram deep learning: insights from dataset characteristics and attention mechanismFuture Cardiology, 0 (2024)
Abstract
Aim: Deep learning’s widespread use prompts heightened scrutiny, particularly in the biomedical fields, with a specific focus on model generalizability. This study delves into the influence of training data characteristics on the generalization performance of models, specifically in cardiac abnormality detection. Materials & methods: Leveraging diverse electrocardiogram datasets, models are trained on subsets with varying characteristics and subsequently compared for performance. Additionally, the introduction of the attention mechanism aims to improve generalizability. Results: Experiments reveal that using a balanced dataset, just 1% of a large dataset, leads to equal performance in generalization tasks, notably in detecting cardiology abnormalities. Conclusion: This balanced training data notably enhances model generalizability, while the integration of the attention mechanism further refines the model’s ability to generalize effectively. This study tackles a common problem for deep learning models: they often struggle when faced with new, unfamiliar data that they have not been trained on. This phenomenon is also known as performance drop in out-of-distribution generalization. This reduced performance on out-of-distribution generalization is a key focus of the research, aiming to improve the models’ ability to handle diverse data sets beyond their training data. The study examines how the characteristics of the dataset used to train deep learning models affect their ability to detect abnormal heart activities when applied to new, unseen data. Researchers trained these models using various sets of electrocardiogram (ECG) data and then evaluated their performance in identifying abnormalities. They also introduced an attention mechanism to enhance the models’ learning capabilities. The attention mechanism in deep learning is like a spotlight that helps the model focus on important information while ignoring less relevant details. The findings were particularly noteworthy. Despite being trained on a small, well-balanced subset of a larger dataset, the models excelled in detecting heart abnormalities in new, unfamiliar data. This training method significantly improved the models’ generalization and performance with unseen data. Furthermore, integrating the attention mechanism substantially enhanced the models’ ability to generalize effectively on new information. Investigate the impact of training data characteristics and attention mechanism on deep learning model generalizability in cardiac abnormality detection. Balanced dataset (1% of the total) improves model performance in generalization tasks, especially in detecting cardiology abnormalities. The attention mechanism further enhances the model’s capacity to comprehend and utilize out-of-distribution data effectively. Utilized multiple electrocardiogram datasets for the study. Trained models on subsets with varying characteristics and evaluated performance. Added attention mechanism to enhance learning capabilities. Balanced training data significantly enhances model generalizability. Attention mechanism improves the model’s ability to generalize on out-of-distribution data. Lack of clinical user information in datasets due to privacy and ethical considerations. Future research may consider patient-specific models for improved generalization in biomedical machine learning. Balanced and curated datasets are crucial for training high-performing models in cardiac abnormality detection using deep learning. Attention mechanisms show promise in enhancing model accuracy and generalization.
- Decoding 2.3 Million ECGs: Interpretable Deep Learning for Advancing Cardiovascular Diagnosis and Mortality Risk StratificationEuropean Heart Journal - Digital Health, ztae014 (2024)
Abstract
Electrocardiogram (ECG) is widely considered the primary test for evaluating cardiovascular diseases. However, the use of artificial intelligence to advance these medical practices and learn new clinical insights from ECGs remains largely unexplored. Utilising a dataset of 2,322,513 ECGs collected from 1,558,772 patients with 7 years of follow-up, we developed a deep learning model with state-of-the-art granularity for the interpretable diagnosis of cardiac abnormalities, gender identification, and hyper- tension screening solely from ECGs, which are then used to stratify the risk of mortality. The model achieved the area under the receiver operating characteristic curve (AUC) scores of 0.998 (95% confidence interval (CI), 0.995-0.999), 0.964 (0.963-0.965), and 0.839 (0.837-0.841) for the three diagnostic tasks separately. Using ECG-predicted results, we find high risks of mortality for subjects with sinus tachycardia (adjusted hazard ratio (HR) of 2.24, 1.96-2.57), and atrial fibrillation (adjusted HR of 2.22, 1.99-2.48). We further use salient morphologies produced by the deep learning model to identify key ECG leads that achieved similar performance for the three diagnoses, and we find that the V1 ECG lead is important for hypertension screening and mortality risk stratification of hypertensive cohorts, with an AUC of 0.816 (0.814-0.818) and a univariate HR of 1.70 (1.61-1.79) for the two tasks separately. Using ECGs alone, our developed model showed cardiologist-level accuracy in interpretable cardiac diagnosis, and the advancement in mortality risk stratification; In addition, the potential to facilitate clinical knowledge discovery for gender and hypertension detection which are not readily available.
- Artificial intelligence–enabled electrocardiogram for mortality and cardiovascular risk estimation: An actionable, explainable and biologically plausible platformnpj Digital Medicine, 7 (2024)
- AI-ECG and prediction of new atrial fibrillation: when the heart tells the ageEuropean Heart Journal, ehae809 (2024)
Abstract
This editorial refers to ‘Artificial intelligence–derived electrocardiographic aging and risk of atrial fibrillation: a multi-national study’, by S. Cho et al., https://doi.org/10.1093/eurheartj/ehae790.Ageing is a complex, multifaceted process encompassing biological, physiological, and psychosocial changes occurring over time. At the cellular and molecular levels, ageing is characterized by accumulating damage in cells and tissues, leading to a progressive decline in the body’s ability to maintain homeostasis and repair itself. This biological deterioration is not uniform across individuals or even within different organs of the same individual. While some organs may exhibit accelerated ageing, others might remain relatively preserved, highlighting the heterogeneity of the ageing process across the human body.1 The concept of ‘biological age’ vs. ‘chronological age’ is pivotal in understanding ageing. Chronological age is a simple measure of the years a person has lived, whereas biological age reflects the functional status and health of an individual’s organs and systems. Individuals of the same chronological age may have vastly different biological ages due to genetic predisposition, lifestyle factors, and environmental exposures. This variation can significantly impact disease risk and mortality. Understanding and measuring organ-specific ageing can open up avenues for targeted interventions that could decelerate ageing in specific organs, thereby improving overall health and extending a healthy life span.2
- Artificial intelligence-enabled electrocardiogram for mortality and cardiovascular risk estimation: a model development and validation studyThe Lancet Digital Health, 6, e791–e802 (2024)
Abstract
Summary Background Artificial intelligence (AI)-enabled electrocardiography (ECG) can be used to predict risk of future disease and mortality but has not yet been adopted into clinical practice. Existing model predictions do not have actionability at an individual patient level, explainability, or biological plausibi. We sought to address these limitations of previous AI-ECG approaches by developing the AI-ECG risk estimator (AIRE) platform. Methods The AIRE platform was developed in a secondary care dataset (Beth Israel Deaconess Medical Center [BIDMC]) of 1 163 401 ECGs from 189 539 patients with deep learning and a discrete-time survival model to create a patient-specific survival curve with a single ECG. Therefore, AIRE predicts not only risk of mortality, but also time-to-mortality. AIRE was validated in five diverse, transnational cohorts from the USA, Brazil, and the UK (UK Biobank [UKB]), including volunteers, primary care patients, and secondary care patients. Findings AIRE accurately predicts risk of all-cause mortality (BIDMC C-index 0\textperiodcentered775, 95% CI 0\textperiodcentered773–0\textperiodcentered776; C-indices on external validation datasets 0\textperiodcentered638–0\textperiodcentered773), future ventricular arrhythmia (BIDMC C-index 0\textperiodcentered760, 95% CI 0\textperiodcentered756–0\textperiodcentered763; UKB C-index 0\textperiodcentered719, 95% CI 0\textperiodcentered635–0\textperiodcentered803), future atherosclerotic cardiovascular disease (0\textperiodcentered696, 0\textperiodcentered694–0\textperiodcentered698; 0\textperiodcentered643, 0\textperiodcentered624–0\textperiodcentered662), and future heart failure (0\textperiodcentered787, 0\textperiodcentered785–0\textperiodcentered789; 0\textperiodcentered768, 0\textperiodcentered733–0\textperiodcentered802). Through phenome-wide and genome-wide association studies, we identified candidate biological pathways for the prediction of increased risk, including changes in cardiac structure and function, and genes associated with cardiac structure, biological ageing, and metabolic syndrome. Interpretation AIRE is an actionable, explainable, and biologically plausible AI-ECG risk estimation platform that has the potential for use worldwide across a wide range of clinical contexts for short-term and long-term risk estimation. Funding British Heart Foundation, National Institute for Health and Care Research, and Medical Research Council.
- Evaluating regression and probabilistic methods for ECG-based electrolyte predictionScientific Reports, 14, 15273 (2024)
Abstract
Imbalances in electrolyte concentrations can have severe consequences, but accurate and accessible measurements could improve patient outcomes. The current measurement method based on blood tests is accurate but invasive and time-consuming and is often unavailable for example in remote locations or an ambulance setting. In this paper, we explore the use of deep neural networks (DNNs) for regression tasks to accurately predict continuous electrolyte concentrations from electrocardiograms (ECGs), a quick and widely adopted tool. We analyze our DNN models on a novel dataset of over 290,000 ECGs across four major electrolytes and compare their performance with traditional machine learning models. For improved understanding, we also study the full spectrum from continuous predictions to a binary classification of extreme concentration levels. Finally, we investigate probabilistic regression approaches and explore uncertainty estimates for enhanced clinical usefulness. Our results show that DNNs outperform traditional models but model performance varies significantly across different electrolytes. While discretization leads to good classification performance, it does not address the original problem of continuous concentration level prediction. Probabilistic regression has practical potential, but our uncertainty estimates are not perfectly calibrated. Our study is therefore a first step towards developing an accurate and reliable ECG-based method for electrolyte concentration level prediction—a method with high potential impact within multiple clinical scenarios.
- Transferability and Adversarial Training in Automatic Classification of the Electrocardiogram with Deep LearningComputing in Cardiology (CinC) (2024)
- No Double Descent in Principal Component Regression: A High-Dimensional AnalysisInternational Conference on Machine Learning (ICML) (2024)
2023
- Overparameterized Linear Regression under Adversarial AttacksIEEE Transactions on Signal Processing (2023)
Abstract
As machine learning models start to be used in critical applications, their vulnerabilities and brittleness become a pressing concern. Adversarial attacks are a popular framework for studying these vulnerabilities. In this work, we study the error of linear regression in the face of adversarial attacks. We provide bounds of the error in terms of the traditional risk and the parameter norm and show how these bounds can be leveraged and make it possible to use analysis from non-adversarial setups to study the adversarial risk. The usefulness of these results is illustrated by shedding light on whether or not overparameterized linear models can be adversarially robust. We show that adding features to linear models might be either a source of additional robustness or brittleness. We show that these differences appear due to scaling and how the {}ell_1 and {}ell_2 norms of random projections concentrate. We also show how the reformulation we propose allows for solving adversarial training as a convex optimization problem. This is then used as a tool to study how adversarial training and other regularization methods might affect the robustness of the estimated models.
- On Merging Feature Engineering and Deep Learning for Diagnosis, Risk-Prediction and Age Estimation Based on the 12-Lead ECGIEEE Transactions on Biomedical Engineering (2023)
Abstract
Objective: Over the past few years, deep learning (DL) has been used extensively in research for 12-lead electrocardiogram (ECG) analysis. However, it is unclear whether the explicit or implicit claims made on DL superiority to the more classical feature engineering (FE) approaches, based on domain knowledge, hold. In addition, it remains unclear whether combining DL with FE may improve performance over a single modality. Methods: To address these research gaps and in-line with recent major experiments, we revisited three tasks: cardiac arrhythmia diagnosis (multiclass-multilabel classification), atrial fibrillation risk prediction (binary classification), and age estimation (regression). We used an overall dataset of 2.3M 12-lead ECG recordings to train the following models for each task: i) a random forest taking FE as input; ii) an end-to-end DL model; and iii) a merged model of FE+DL. Results: FE yielded comparable results to DL while necessitating significantly less data for the two classification tasks. DL outperformed FE for the regression task. For all tasks, merging FE with DL did not improve performance over DL alone. These findings were confirmed on the additional PTB-XL dataset. Conclusion: We found that for traditional 12-lead ECG based diagnosis tasks, DL did not yield a meaningful improvement over FE, while it improved significantly the nontraditional regression task. We also found that combining FE with DL did not improve over DL alone, which suggests that the FE were redundant with the features learned by DL. Significance: Our findings provides important recommendations on 12-lead ECG based machine learning strategy and data regime to choose for a given task. When looking at maximizing performance as the end goal, if the task is nontraditional and a large dataset is available then DL is preferable. If the task is a classical one and/or a small dataset is available then a FE approach may be the better choice.
- Invertible Kernel PCA with Random Fourier FeaturesIEEE Signal Processing Letters (2023)
Abstract
Kernel principal component analysis (kPCA) is a widely studied method to construct a low-dimensional data representation after a nonlinear transformation. The prevailing method to reconstruct the original input signal from kPCA—an important task for denoising—requires us to solve a supervised learning problem. In this paper, we present an alternative method where the reconstruction follows naturally from the compression step. We first approximate the kernel with random Fourier features. Then, we exploit the fact that the nonlinear transformation is invertible in a certain subdomain. Hence, the name invertible kernel PCA (ikPCA) . We experiment with different data modalities and show that ikPCA performs similarly to kPCA with supervised reconstruction on denoising tasks, making it a strong alternative.
- Heart age gap by explainable advanced electrocardiography is associated with cardiovascular risk factors and survivalEuropean Heart Journal - Digital Health (2023)
- End-to-end risk prediction of atrial fibrillation from the 12-Lead ECG by deep neural networksJournal of Electrocardiology, 81, 193–200 (2023)
Abstract
BACKGROUND: Atrial fibrillation (AF) is one of the most common cardiac arrhythmias that affects millions of people each year worldwide and it is closely linked to increased risk of cardiovas- cular diseases such as stroke and heart failure. Machine learning methods have shown promising results in evaluating the risk of developing atrial fibrillation from the electrocardiogram. We aim to develop and evaluate one such algorithm on a large CODE dataset collected in Brazil. METHODS: We used the CODE cohort to develop and test a model for AF risk prediction for individual patients from the raw ECG recordings without the use of additional digital biomarkers. The cohort is a collection of ECG recordings and annotations by the Telehealth Network of Minas Gerais, in Brazil. A convolutional neural network based on a residual network architecture was implemented to produce class probabilities for the classification of AF. The probabilities were used to develop a Cox proportional hazards model and a Kaplan-Meier model to carry out survival analysis. Hence, our model is able to perform risk prediction for the development of AF in patients without the condition. RESULTS: The deep neural network model identified patients without indication of AF in the presented ECG but who will develop AF in the future with an AUC score of 0.845. From our survival model, we obtain that patients in the high-risk group (i.e. with the probability of a future AF case being \textquestiondown0.7) are 50% more likely to develop AF within 40 weeks, while patients belonging to the minimal-risk group (i.e. with the probability of a future AF case being less than or equal to 0.1) have \textquestiondown85% chance of remaining AF free up until after seven years. CONCLUSION: We developed and validated a model for AF risk prediction. If applied in clinical practice, the model possesses the potential of providing valuable and useful information in decision- making and patient management processes.
- Electrocardiographic Age Predicts Cardiovascular Events in Community: The Framingham Heart StudyCirculation: Cardiovascular Quality and Outcomes (2023)
- Detection of Left Ventricular Systolic Dysfunction from Electrocardiographic ImagesCirculation (2023)
Abstract
BACKGROUND: Left ventricular (LV) systolic dysfunction is associated with a \textgreater8-fold increased risk of heart failure and a 2-fold risk of premature death. The use of ECG signals in screening for LV systolic dysfunction is limited by their availability to clinicians. We developed a novel deep learning–based approach that can use ECG images for the screening of LV systolic dysfunction. METHODS: Using 12-lead ECGs plotted in multiple different formats, and corresponding echocardiographic data recorded within 15 days from the Yale New Haven Hospital between 2015 and 2021, we developed a convolutional neural network algorithm to detect an LV ejection fraction \textless40%. The model was validated within clinical settings at Yale New Haven Hospital and externally on ECG images from Cedars Sinai Medical Center in Los Angeles, CA; Lake Regional Hospital in Osage Beach, MO; Memorial Hermann Southeast Hospital in Houston, TX; and Methodist Cardiology Clinic of San Antonio, TX. In addition, it was validated in the prospective Brazilian Longitudinal Study of Adult Health. Gradient-weighted class activation mapping was used to localize class-discriminating signals in ECG images. RESULTS: Overall, 385 601 ECGs with paired echocardiograms were used for model development. The model demonstrated high discrimination power across various ECG image formats and calibrations in internal validation (area under receiving operation characteristics [AUROCs], 0.91; area under precision-recall curve [AUPRC], 0.55); and external sets of ECG images from Cedars Sinai (AUROC, 0.90 and AUPRC, 0.53), outpatient Yale New Haven Hospital clinics (AUROC, 0.94 and AUPRC, 0.77), Lake Regional Hospital (AUROC, 0.90 and AUPRC, 0.88), Memorial Hermann Southeast Hospital (AUROC, 0.91 and AUPRC 0.88), Methodist Cardiology Clinic (AUROC, 0.90 and AUPRC, 0.74), and Brazilian Longitudinal Study of Adult Health cohort (AUROC, 0.95 and AUPRC, 0.45). An ECG suggestive of LV systolic dysfunction portended a \textgreater27-fold higher odds of LV systolic dysfunction on transthoracic echocardiogram (odds ratio, 27.5 [95% CI, 22.3–33.9] in the held-out set). Class-discriminative patterns localized to the anterior and anteroseptal leads (V2 to V3), corresponding to the left ventricle regardless of the ECG layout. A positive ECG screen in individuals with an LV ejection fraction \geq40% at the time of initial assessment was associated with a 3.9-fold increased risk of developing incident LV systolic dysfunction in the future (hazard ratio, 3.9 [95% CI, 3.3–4.7]; median follow-up, 3.2 years). CONCLUSIONS: We developed and externally validated a deep learning model that identifies LV systolic dysfunction from ECG images. This approach represents an automated and accessible screening strategy for LV systolic dysfunction, particularly in low-resource settings.
- Association of lifestyle with deep-learning based ECG-ageFrontiers in Cardiovascular Medicine (2023)
- Screening for Chagas disease from the electrocardiogram using a deep neural networkPLOS Neglected Tropical Diseases, 17, e0011118 (2023)
Abstract
Background Worldwide, it is estimated that over 6 million people are infected with Chagas disease (ChD). It is a neglected disease that can lead to severe heart conditions in its chronic phase. While early treatment can avoid complications, the early-stage detection rate is low. We explore the use of deep neural networks to detect ChD from electrocardiograms (ECGs) to aid in the early detection of the disease. Methods We employ a convolutional neural network model that uses 12-lead ECG data to compute the probability of a ChD diagnosis. Our model is developed using two datasets which jointly comprise over two million entries from Brazilian patients: The SaMi-Trop study focusing on ChD patients, enriched with data from the CODE study from the general population. The model’s performance is evaluated on two external datasets: the REDS-II, a study focused on ChD with 631 patients, and the ELSA-Brasil study, with 13,739 civil servant patients. Findings Evaluating our model, we obtain an AUC-ROC of 0.80 (CI 95% 0.79-0.82) for the validation set (samples from CODE and SaMi-Trop), and in external validation datasets: 0.68 (CI 95% 0.63-0.71) for REDS-II and 0.59 (CI 95% 0.56-0.63) for ELSA-Brasil. In the latter, we report a sensitivity of 0.52 (CI 95% 0.47-0.57) and 0.36 (CI 95% 0.30-0.42) and a specificity of 0.77 (CI 95% 0.72-0.81) and 0.76 (CI 95% 0.75-0.77), respectively. Additionally, when considering only patients with Chagas cardiomyopathy as positive, the model achieved an AUC-ROC of 0.82 (CI 95% 0.77-0.86) for REDS-II and 0.77 (CI 95% 0.68-0.85) for ELSA-Brasil. Interpretation The neural network detects chronic Chagas cardiomyopathy (CCC) from ECG—with weaker performance for early-stage cases. Future work should focus on curating large higher-quality datasets. The CODE dataset, our largest development dataset includes self-reported and therefore less reliable labels, limiting performance for non-CCC patients. Our findings can improve ChD detection and treatment, particularly in high-prevalence areas.
2022
- Development and validation of deep learning ECG-based prediction of myocardial infarction in emergency department patientsScientific reports, 12, 19615 (2022)Equal contribution: S. Gustafsson and D. Gedon
Abstract
Myocardial infarction diagnosis is a common challenge in the emergency department. In managed settings, deep learning-based models and especially convolutional deep models have shown promise in electrocardiogram (ECG) classification, but there is a lack of high-performing models for the diagnosis of myocardial infarction in real-world scenarios. We aimed to train and validate a deep learning model using ECGs to predict myocardial infarction in real-world emergency department patients. We studied emergency department patients in the Stockholm region between 2007 and 2016 that had an ECG obtained because of their presenting complaint. We developed a deep neural network based on convolutional layers similar to a residual network. Inputs to the model were ECG tracing, age, and sex; and outputs were the probabilities of three mutually exclusive classes: non-ST-elevation myocardial infarction (NSTEMI), ST-elevation myocardial infarction (STEMI), and control status, as registered in the SWEDEHEART and other registries. We used an ensemble of five models. Among 492,226 ECGs in 214,250 patients, 5,416 were recorded with an NSTEMI, 1,818 a STEMI, and 485,207 without a myocardial infarction. In a random test set, our model could discriminate STEMIs/NSTEMIs from controls with a C-statistic of 0.991/0.832 and had a Brier score of 0.001/0.008. The model obtained a similar performance in a temporally separated test set of the study sample, and achieved a C-statistic of 0.985 and a Brier score of 0.002 in discriminating STEMIs from controls in an external test set. We developed and validated a deep learning model with excellent performance in discriminating between control, STEMI, and NSTEMI on the presenting ECG of a real-world sample of the important population of all-comers to the emergency department. Hence, deep learning models for ECG decision support could be valuable in the emergency department.
- Automated multilabel diagnosis on electrocardiographic images and signalsNature Communications, 13, 1583 (2022)
2021
- Electrocardiographic Predictors of Mortality: Data from a Primary Care Tele-Electrocardiography Cohort of Brazilian PatientsHearts, 2, 449–458 (2021)
Abstract
Computerized electrocardiography (ECG) has been widely used and allows linkage to electronic medical records. The present study describes the development and clinical applications of an electronic cohort derived from a digital ECG database obtained by the Telehealth Network of Minas Gerais, Brazil, for the period 2010–2017, linked to the mortality data from the national information system, the Clinical Outcomes in Digital Electrocardiography (CODE) dataset. From 2,470,424 ECGs, 1,773,689 patients were identified. A total of 1,666,778 (94%) underwent a valid ECG recording for the period 2010 to 2017, with 1,558,421 patients over 16 years old; 40.2% were men, with a mean age of 51.7 [SD 17.6] years. During a mean follow-up of 3.7 years, the mortality rate was 3.3%. ECG abnormalities assessed were: atrial fibrillation (AF), right bundle branch block (RBBB), left bundle branch block (LBBB), atrioventricular block (AVB), and ventricular pre-excitation. Most ECG abnormalities (AF: Hazard ratio [HR] 2.10; 95% CI 2.03–2.17; RBBB: HR 1.32; 95%CI 1.27–1.36; LBBB: HR 1.69; 95% CI 1.62–1.76; first degree AVB: Relative survival [RS]: 0.76; 95% CI0.71–0.81; 2:1 AVB: RS 0.21 95% CI0.09–0.52; and RS 0.36; third degree AVB: 95% CI 0.26–0.49) were predictors of overall mortality, except for ventricular pre-excitation (HR 1.41; 95% CI 0.56–3.57) and Mobitz I AVB (RS 0.65; 95% CI 0.34–1.24). In conclusion, a large ECG database established by a telehealth network can be a useful tool for facilitating new advances in the fields of digital electrocardiography, clinical cardiology and cardiovascular epidemiology.
- Deep neural network estimated electrocardiographic-age as a mortality predictorNature Communications, 12 (2021)Equal contribution: E. M. Lima, A. H. Ribeiro, G. M. M. Paixao
Abstract
The electrocardiogram (ECG) is the most commonly used exam for the screening and evaluation of cardiovascular diseases. Here we propose that the age predicted by artificial intelligence (AI) from the raw ECG tracing (ECG-age) can be a measure of cardiovascular health and provide prognostic information. A deep convolutional neural network was trained to predict a patient9s age from the 12-lead ECG using data from patients that underwent an ECG from 2010 to 2017 - the CODE study cohort (n=1,558,415 patients). On the 15% hold-out CODE test split, patients with ECG-age more than 8 years greater than chronological age had a higher mortality rate (hazard ratio (HR) 1.79, p<0.001) in a mean follow-up of 3.67 years, whereas those with ECG-age more than 8 years less than chronological age had a lower mortality rate (HR 0.78, p<0.001). Similar results were obtained in the external cohorts ELSA-Brasil (n=14,236) and SaMi-Trop (n=1,631). The ability to predict mortality from the ECG predicted age remains even when we adjust the model for cardiovascular risk factors. Moreover, even for apparent normal ECGs, having a predicted ECG-age 8 or more years greater than chronological age remained a statistically significant predictor of risk (HR 1.53, p<0.001 in CODE 15% test split). These results show that AI-enabled analysis of the ECG can add prognostic information to the interpretation of the 12-lead ECGs.
- Atrial fibrillation risk prediction from the 12-lead ECG using digital biomarkers and deep representation learningEuropean Heart Journal - Digital Health (2021)
Abstract
This study aims to assess whether information derived from the raw 12-lead electrocardiogram (ECG) combined with clinical information is predictive of atrial fibrillation (AF) development.We use a subset of the Telehealth Network of Minas Gerais (TNMG) database consisting of patients that had repeated 12-lead ECG measurements between 2010-2017 that is 1,130,404 recordings from 415,389 unique patients. Median and interquartile of age for the recordings were 58 (46-69) and 38% of the patients were males. Recordings were assigned to train-validation and test sets in an 80:20% split which was stratified by class, age and gender. A random forest classifier was trained to predict, for a given recording, the risk of AF development within 5-years. We use features obtained from different modalities, namely demographics, clinical information, engineered features, and features from deep representation learning.The best model performance on the test set was obtained for the model combining features from all modalities with an AUROC=0.909 against the best single modality model which had an AUROC=0.839.Our study has important clinical implications for AF management. It is the first study integrating feature engineering, deep learning and EMR metadata to create a risk prediction tool for the management of patients at risk of AF. The best model that includes features from all modalities demonstrates that human knowledge in electrophysiology combined with deep learning outperforms any single modality approach. The high performance obtained suggest that structural changes in the 12-lead ECG are associated with existing or impending AF.
- How convolutional neural networks deal with aliasingIEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2755–2759 (2021)
- First Steps Towards Self-Supervised Pretraining of the 12-Lead ECGComputing in Cardiology (CinC), 48, 1–4 (2021)
Abstract
Self-supervised learning is a paradigm that extracts general features which describe the input space by artificially generating labels from the input without the need for explicit annotations. The learned features can then be used by transfer learning to boost the performance on a downstream task. Such methods have recently produced state of the art results in natural language processing and computer vision. Here, we propose a self-supervised learning method for 12-lead electrocardiograms (ECGs). For pretraining the model we design a task to mask out subsegements of all channels of the input signals and try to predict the actual values. As the model architecture, we use a U-ResNet containing an encoder-decoder structure. We test our method by self-supervised pretraining on the CODE dataset and then transfer the learnt features by finetuning on the PTB-XL and CPSC benchmarks to evaluate the effect of our method in the classification of 12-leads ECGs. The method does provide modest improvements in performance when compared to not using pretraining. In future work we will make use of these ideas in smaller dataset, where we believe it can lead to larger performance gains.
- Deep Energy-Based NARX ModelsIFAC Symposium on System Identification (SYSID), 54, 505–510 (2021)
Abstract
This paper is directed towards the problem of learning nonlinear ARX models based on system input–output data. In particular, our interest is in learning a conditional distribution of the current output based on a finite window of past inputs and outputs. To achieve this, we consider the use of so-called energy-based models, which have been developed in allied fields for learning unknown distributions based on data. This energy-based model relies on a general function to describe the distribution, and here we consider a deep neural network for this purpose. The primary benefit of this approach is that it is capable of learning both simple and highly complex noise models, which we demonstrate on simulated and experimental data.
2020
- SciPy 1.0–Fundamental Algorithms for Scientific Computing in PythonNature Methods, 17, 261–272 (2020)
Abstract
SciPy is an open source scientific computing library for the Python programming language. SciPy 1.0 was released in late 2017, about 16 years after the original version 0.1 release. SciPy has become a de facto standard for leveraging scientific algorithms in the Python programming language, with more than 600 unique code contributors, thousands of dependent packages, over 100,000 dependent repositories, and millions of downloads per year. This includes usage of SciPy in almost half of all machine learning projects on GitHub, and usage by high profile projects including LIGO gravitational wave analysis and creation of the first-ever image of a black hole (M87). The library includes functionality spanning clustering, Fourier transforms, integration, interpolation, file I/O, linear algebra, image processing, orthogonal distance regression, minimization algorithms, signal processing, sparse matrix handling, computational geometry, and statistics. In this work, we provide an overview of the capabilities and development practices of the SciPy library and highlight some recent technical developments.
- Evaluation of Mortality in Atrial Fibrillation: Clinical Outcomes in Digital Electrocardiography (CODE) StudyGlobal Heart, 15 (2020)
Abstract
Methods: This observational retrospective study of primary care patients was developed with the digital ECG database from the Telehealth Network of Minas Gerais, Brazil. ECGs performed from 2010 to 2017 were interpreted by cardiologists and the University of Glasgow automated analysis software. An electronic cohort was obtained linking data from ECG exams and those from a national mortality information system, using standard probabilistic linkage methods. We considered only the first ECG of each patient. Patients under 16 years were excluded. Hazard ratios (HR) for mortality were adjusted for demographic and self-reported clinical factors and estimated with Cox regression. Results: From a dataset of 1,773,689 patients, 1,558,421 were included, mean age 51.6 years; 40.2% male. There were 3.34% deaths from all causes in 3.68 years of median follow up. The prevalence of AF was 1.33%. AF was an independent risk factor for all-cause mortality (HR 2.10, 95%CI 2.03–2.17) and cardiovascular mortality (HR 2.06, 95%CI 1.86–2.29). Females with AF had a higher risk of overall and cardiovascular mortality compared with males (p \textless 0.001). Conclusions: AF was a strong predictor of cardiovascular and all-cause mortality in a primary care population, with increased risk in women.
- Contextualized Interpretable Machine Learning for Medical DiagnosisCommunications of the ACM (2020)
- On the smoothness of nonlinear system identificationAutomatica, 121, 109158 (2020)
Abstract
We shed new light on the smoothness of optimization problems arising in prediction error parameter estimation of linear and nonlinear systems. We show that for regions of the parameter space where the model is not contractive, the Lipschitz constant and beta-smoothness of the objective function might blow up exponentially with the simulation length, making it hard to numerically find minima within those regions or, even, to escape from them. In addition to providing theoretical understanding of this problem, this paper also proposes the use of multiple shooting as a viable solution. The proposed method minimizes the error between a prediction model and the observed values. Rather than running the prediction model over the entire dataset, multiple shooting splits the data into smaller subsets and runs the prediction model over each subset, making the simulation length a design parameter and making it possible to solve problems that would be infeasible using a standard approach. The equivalence to the original problem is obtained by including constraints in the optimization. The new method is illustrated by estimating the parameters of nonlinear systems with chaotic or unstable behavior, as well as neural networks. We also present a comparative analysis of the proposed method with multi-step-ahead prediction error minimization.
- Explaining end-to-end ECG automated diagnosis using contextual featuresEuropean Conference on Machine Learning and Principles and Practice of Knowledge Discovery in Databases (ECML-PKDD) (2020)
- Explaining black-box automated electrocardiogram classification to cardiologistsComputing in Cardiology (CinC), 47 (2020)
Abstract
In this work, we present a method to explain “end-toend” electrocardiogram (ECG) signal classifiers, where the explanations were built along with seniors cardiologist to provide meaningful features to the final users. Our method focuses exclusively on automated ECG diagnosis and analyzes the explanation in terms of clinical accuracy for interpretability and robustness. The proposed method uses a noise-insertion strategy to quantify the impact of intervals and segments of the ECG signals on the automated classification outcome. An ECG segmentation method was applied to ECG tracings, to obtain: (1) Intervals, Segments and Axis; (2) Rate, and (3) Rhythm. Noise was added to the signal to disturb the ECG features in a realistic way. The method was tested using Monte Carlo simulation and the feature impact is estimated by the change in the model prediction averaged over 499 executions and a feature is defined as important if its mean value changes the result of the classifier. We demonstrate our method by explaining diagnoses generated by a deep convolutional neural network. The proposed method is particularly effective and useful for modern deep learning models that take raw data as input.
- Automatic 12-lead ECG classification using a convolutional network ensembleComputing in Cardiology (CinC) (2020)
Abstract
The 12-lead electrocardiogram (ECG) is a major diagnostic test for cardiovascular diseases and enhanced automated analysis tools might lead to more reliable diagnosis and improved clinical practice. Deep neural networks are models composed of stacked transformations that learn tasks by examples. Inspired by the success of these models in computer vision, we propose an end-to-end approach for the task at hand. We trained deep convolutional neural network models in the heterogeneous dataset provided in the Physionet 2020 Challenge and used an ensemble of seven of these convolutional models for the classification of abnormalities present in the ECG records. Ensembles use the output of multiple models to generate a combined prediction and are known to improve performance and generalization when compared to the individual models. In our submission, we use an ensemble of neural networks with the architecture similar to the one described in Nat Commun 11, 1760 (2020) for 12-lead ECGs classification. On the partially hidden test dataset from the challenge, the best-scored entry for our team (the “Code Team”) had a performance of 0.657, which place us in the 7-th place team-wise in the challenge leaderboard.
2019
- Evaluation of mortality in bundle branch block patients from an electronic cohort: Clinical Outcomes in Digital Electrocardiography (CODE) studyJournal of Electrocardiology (2019)
- Tele-electrocardiography and bigdata: The CODE (Clinical Outcomes in Digital Electrocardiography) studyJournal of Electrocardiology (2019)
Abstract
Digital electrocardiographs are now widely available and a large number of digital electrocardiograms (ECGs) have been recorded and stored. The present study describes the development and clinical applications of a large database of such digital ECGs, namely the CODE (Clinical Outcomes in Digital Electrocardiology) study. ECGs obtained by the Telehealth Network of Minas Gerais, Brazil, from 2010 to 17, were organized in a structured database. A hierarchical free-text machine learning algorithm recognized specific ECG diagnoses from cardiologist reports. The Glasgow ECG Analysis Program provided Minnesota Codes and automatic diagnostic statements. The presence of a specific ECG abnormality was considered when both automatic and medical diagnosis were concordant; cases of discordance were decided using heuristisc rules and manual review. The ECG database was linked to the national mortality information system using probabilistic linkage methods. From 2,470,424 ECGs, 1,773,689 patients were identified. After excluding the ECGs with technical problems and patients \textless16 years-old, 1,558,415 patients were studied. High performance measures were obtained using an end-to-end deep neural network trained to detect 6 types of ECG abnormalities, with F1 scores \textgreater80% and specificity \textgreater99% in an independent test dataset. We also evaluated the risk of mortality associated with the presence of atrial fibrillation (AF), which showed that AF was a strong predictor of cardiovascular mortality and mortality for all causes, with increased risk in women. In conclusion, a large database that comprises all ECGs performed by a large telehealth network can be useful for further developments in the field of digital electrocardiography, clinical cardiology and cardiovascular epidemiology.
- Deep Convolutional Networks in System IdentificationIEEE Conference on Decision and Control (CDC), 3670–3676 (2019)Equal contribution C. Andersson and A. H. Ribeiro
Abstract
Recent developments within deep learning are relevant for nonlinear system identification problems. In this paper, we establish connections between the deep learning and the system identification communities. It has recently been shown that convolutional architectures are at least as capable as recurrent architectures when it comes to sequence modeling tasks. Inspired by these results we explore the explicit relationships between the recently proposed temporal convolutional network (TCN) and two classic system identification model structures; Volterra series and block-oriented models. We end the paper with an experimental study where we provide results on two real-world problems, the well-known Silverbox dataset and a newer dataset originating from ground vibration experiments on an F-16 fighter aircraft.
2018
- ”Parallel Training Considered Harmful?”: Comparing series-parallel and parallel feedforward network trainingNeurocomputing, 316, 222–231 (2018)
Abstract
Neural network models for dynamic systems can be trained either in parallel or in series-parallel configurations. Influenced by early arguments, several papers justify the choice of series-parallel rather than parallel configuration claiming it has a lower computational cost, better stability properties during training and provides more accurate results. Other published results, on the other hand, defend parallel training as being more robust and capable of yielding more accurate long-term predictions. The main contribution of this paper is to present a study comparing both methods under the same unified framework with special attention to three aspects: (i) robustness of the estimation in the presence of noise; (ii) computational cost; and, (iii) convergence. A unifying mathematical framework and simulation studies show situations where each training method provides superior validation results and suggest that parallel training is generally better in more realistic scenarios. An example using measured data seems to reinforce such a claim. Complexity analysis and numerical examples show that both methods have similar computational cost although series-parallel training is more amenable to parallelization. Some informal discussion about stability and convergence properties is presented and explored in the examples.
- Lasso Regularization Paths for NARMAX Models via Coordinate DescentAmerican Control Conference (ACC), 5268–5273 (2018)
2017
- Shooting Methods for Parameter Estimation of Output Error ModelsIFAC World Congress, 50, 13998–14003 (2017)
Abstract
This paper studies parameter estimation of output error (OE) models. The commonly used approach of minimizing the free-run simulation error is called single shooting in contrast with the new multiple shooting approach proposed in this paper, for which the free-run simulation error of sub-datasets is minimized subject to equality constraints. The names “single shooting” and “multiple shooting” are used due to the similarities with techniques for estimating ODE (ordinary differential equation) parameters. Examples with nonlinear polynomial models illustrate the advantages of OE models as well as the capability of the multiple shooting approach to avoid undesirable local minima.
2015
- Selecting transients automatically for the identification of models for an oil wellIFAC Workshop on Automatic Control in Offshore Oil and Gas Production, 48, 154–158 (2015)