How Artificial Intelligence Applied to the ECG Can Estimate Left Ventricular Ejection Fraction

Artificial intelligence applied to the ECG (AI-ECG) has opened a new clinical possibility: estimating the probability of reduced left ventricular ejection fraction (LVEF) from a test that has traditionally never measured systolic function.

The conventional ECG does not calculate ejection fraction. It records electrical activity. However, ventricular remodeling, fibrosis, conduction abnormalities, and repolarization heterogeneity leave diffuse electrical signatures. Deep learning models can capture these signatures in an integrated way, identifying patterns that escape traditional human interpretation.

The conceptual shift is important: we are not turning the ECG into an echocardiogram, but using advanced statistical learning to estimate the probability of systolic dysfunction.

How AI Infers Ejection Fraction From the ECG

AI-ECG models generally use convolutional neural networks trained on thousands or millions of ECG-echocardiogram pairs. During training, the algorithm learns to map complex electrical patterns to the presence or absence of reduced EF.

Unlike classic analysis based on isolated criteria, the model does not look only for a wide QRS, left bundle branch block, or low voltage. It analyzes the raw tracing as a continuous signal, including:

• Complete morphology of all 12 leads
• Subtle temporal relationships between waves and segments
• Microvariations in repolarization
• Spatial distribution of the electrical vectors
• Global patterns consistent with ventricular remodeling

The algorithm’s output is not a numerical EF value. What it provides is a probability of systolic dysfunction (for example, EF < 40% or ≤ 35%, depending on the threshold adopted in the study).

Therefore, AI-ECG does not replace the echocardiogram. It acts as a probabilistic screening tool, capable of flagging patients with a higher likelihood of left ventricular dysfunction.

Clinical Evidence: Beyond the Pioneering Study

The pioneering study by Attia et al. showed that a neural network applied to the 12-lead ECG was able to detect EF ≤ 35% with excellent diagnostic performance.

Main results of the original study:

• Area under the curve (AUC) ~0.93
• Sensitivity ~86%
• Specificity ~85%

These numbers placed AI-ECG on a level comparable to many diagnostic tests used in clinical practice.

After this initial publication, several investigations expanded and refined the concept:

• External validations in different populations, maintaining AUCs between 0.85 and 0.95.
• Prospective studies showing the ability of AI-ECG to identify asymptomatic ventricular dysfunction before the formal clinical diagnosis of heart failure.
• Applications in outpatient and primary care settings, with preserved performance even when the ECG was considered “normal” by conventional interpretation.

Beyond diagnostic accuracy, subsequent analyses showed that the continuous score generated by the algorithm may have incremental prognostic value, being associated with heart failure hospitalization and cardiovascular mortality.

These findings suggest that AI-ECG not only detects established reduced EF, but may also work as an early electrical marker of ventricular remodeling.

Potential Clinical Applications

From a care-delivery standpoint, the main role of AI-ECG is to expand screening capacity. In settings with high patient volume and limited access to imaging, this tool can help prioritize those who truly need an echocardiogram.

Possible applications include:

• Opportunistic screening of outpatients who undergo an ECG for any reason
• Early identification of asymptomatic ventricular dysfunction
• Selection of patients for priority echocardiography
• Population screening in regions with limited access to cardiac imaging

Studies suggest that serial use of AI-ECG may detect a progressive rise in risk before heart failure becomes clinically evident. This opens the door to earlier intervention strategies.

PMcardio and the Commercial Translation of AI-ECG

Among the commercial platforms that incorporate artificial intelligence into the electrocardiogram, PMcardio is one of the best-known examples in Europe. Developed by Powerful Medical, it uses proprietary deep learning algorithms applied to the 12-lead ECG, including a module designed to estimate the probability of reduced ejection fraction (LVsense™).

The goal is to bring the academic concept of AI-ECG to the point of care, generating a risk alert for systolic dysfunction within seconds. Data released by the company report an AUC above 0.90 for detecting reduced EF, although part of this information comes from technical reports and sponsored studies.

The platform holds CE marking as a medical device in certain jurisdictions, which indicates regulatory compliance but does not replace independent clinical validation and performance assessment in the local population.

The main impact of these solutions lies not only in the accuracy of the algorithm, but in its integration into the care workflow: AI-ECG can act as a trigger to order an echocardiogram or further risk stratification. Rational adoption requires a critical analysis of performance, local disease prevalence, and the potential impact of false positives.

Technical Limitations and Challenges

Despite robust performance in early studies, AI-ECG has relevant limitations.

• Generalizability: models trained in tertiary centers may perform worse in populations with different characteristics.
• Dataset bias: a high prevalence of reduced EF in the training cohort may overestimate performance when applied to low-risk populations.
• Interpretability: many models work as a “black box”, making it difficult to understand the mechanism behind the result.
• Dependence on imaging confirmation: the definitive diagnosis of reduced ejection fraction is still made by echocardiography or other imaging methods.

Therefore, AI-ECG should be understood as a decision-support tool, not as a substitute for the reference method.

Conclusion

Artificial intelligence applied to the ECG represents a relevant conceptual shift: using a simple electrical test to infer the probability of left ventricular systolic dysfunction.

AI-ECG can estimate the risk of reduced EF with consistent performance across multiple studies and external validations. Its greatest value lies in screening and early detection, especially in settings where echocardiography is not readily available.

Echocardiography remains the standard for measuring ejection fraction. In this scenario, AI expands the clinical ability to suspect earlier and refer better.