Related Experiment Video
Updated: Jul 12, 2026
Real-Time Cardiac Mapping with a Noninvasive Imageless Electrocardiographic Imaging System
Published on: April 11, 2025
Prediction of Left Ventricular Systolic Dysfunction from ICU Electrocardiograms Using Vision Transformer Embeddings
Jacopo Lenkowicz1, Nicoletta di Giorgi1
1Fondazione Policlinico Universitario A. Gemelli IRCCS.
This study developed a new method to identify heart pumping problems in intensive care patients by analyzing standard heart rhythm tests. By combining visual patterns learned from heart scans with traditional heart rhythm data, the researchers created a tool that can quickly screen for reduced heart function. This approach helps doctors identify patients who need urgent care even when specialized imaging is not immediately available.
Area of Science:
- Cardiovascular medicine and critical care research
- Vision Transformer embeddings in clinical diagnostics
Background:
No prior work had resolved the challenge of rapid heart function assessment for critically ill patients lacking immediate imaging access. Echocardiography and cardiac magnetic resonance imaging provide vital prognostic data but often face significant delays. That uncertainty drove the need for alternative screening methods using readily available diagnostic tools. Prior research has shown that standard heart rhythm recordings contain hidden information about cardiac structure. However, existing models struggle to extract these subtle features effectively from routine clinical data. This gap motivated the development of advanced computational techniques to bridge the divide between rhythm analysis and structural imaging. Investigators now seek to leverage deep learning to interpret these complex electrical signals. This study addresses the urgent requirement for timely identification of reduced pumping capacity in intensive care settings.
Purpose Of The Study:
The aim of this study was to develop a multimodal approach for predicting reduced left ventricular ejection fraction in critically ill patients. Researchers sought to overcome the delays associated with traditional imaging techniques like echocardiography. The project investigated whether structural information from cardiac magnetic resonance could be transferred to standard heart rhythm tests. This motivation stemmed from the frequent unavailability of advanced imaging in intensive care units. The team hypothesized that combining deep embeddings with quantitative features would improve diagnostic sensitivity. They addressed the challenge of identifying patients with pumping function at or below 40 percent. By focusing on 12-lead recordings, the study aimed to create a more accessible screening tool. This work specifically targets the need for timely clinical intervention in high-acuity settings.
Main Methods:
Review approach involved a retrospective analysis of approximately 900 intensive care unit admissions from a single tertiary center. The research team implemented three distinct classification algorithms to evaluate predictive performance. These included random forest, extreme gradient boosting, and a shallow multilayer perceptron. The multilayer perceptron utilized focal loss to address potential class imbalances within the dataset. Investigators compared these models using a stratified five-fold cross-validation framework to ensure statistical validity. The primary input combined deep structural embeddings with standard quantitative features extracted from 12-lead electrical recordings. This design enabled the direct prediction of reduced pumping function without immediate imaging. The approach focused on validating the transferability of imaging-derived insights to rhythm-based diagnostic tasks.
Main Results:
Key findings from the literature indicate that the multilayer perceptron achieved the highest predictive performance among the evaluated models. This specific architecture reached an area under the curve of 0.87. The model also demonstrated a recall of 0.68 and an F1 score of 0.65. These metrics confirm that the integration of visual embeddings enhances the detection of reduced pumping capacity. The results show that structural information from imaging successfully transfers to electrical signal analysis. The study confirms that 12-lead recordings can identify patients with a left ventricular ejection fraction of 40 percent or less. These findings suggest that the proposed multimodal approach outperforms traditional methods relying solely on rhythm features. The data support the utility of this technique for rapid screening in acute care environments.
Conclusions:
The researchers propose that cross-modal visual embeddings successfully transfer structural cardiac insights to electrical signal analysis. Synthesis and implications suggest that routine heart rhythm tests can serve as an early screening tool for ventricular dysfunction. The multilayer perceptron classifier outperformed other tested models in identifying reduced pumping capacity. These findings indicate that integrating deep learning with standard clinical data improves diagnostic accuracy. The study demonstrates that structural information from imaging can enhance the utility of simple diagnostic recordings. Authors suggest that this multimodal strategy offers a viable path for rapid patient assessment. The evidence supports the feasibility of deploying such models within intensive care environments to guide clinical decision-making. This work highlights the potential for artificial intelligence to overcome limitations in current diagnostic workflows.
Frequently Asked Questions
The multilayer perceptron model achieved an area under the curve of 0.87, a recall of 0.68, and an F1 score of 0.65. This performance indicates that the architecture effectively identifies reduced pumping capacity compared to the other tested classifiers.
The researchers utilized a Vision Transformer, which is a deep learning architecture pretrained on cardiac magnetic resonance data. This tool extracts structural embeddings that are then combined with quantitative features from electrocardiograms to improve predictive accuracy.
A stratified five-fold cross-validation approach was necessary to ensure the reliability of the model across the cohort. This technique prevents overfitting and provides a robust estimate of performance compared to a simple train-test split.
The study integrated deep embeddings from imaging data with conventional quantitative features derived from 12-lead electrocardiograms. This multimodal data structure allows the model to interpret electrical signals through the lens of cardiac anatomy.
The researchers measured the ability of the models to detect a left ventricular ejection fraction of 40 percent or lower. This threshold defines the clinical condition of reduced pumping function in the studied intensive care population.
The authors propose that their model facilitates early screening for ventricular dysfunction. This approach helps clinicians prioritize care for patients who might otherwise wait for delayed imaging results.