Prediction of Left Ventricular Ejection Fraction from Lead-II ECG monitoring waveforms (EMW) Using Deep Learning Algorithms

Background： Left ventricular ejection fraction (EF) is a key marker of cardiac systolic function and plays a central role in hemodynamic management of critically ill patients. Although echocardiography remains the standard for EF assessment, its routine use in intensive care units (ICUs) is limited by equipment availability, operator dependency, and lack of capability for continuous monitoring. This study aimed to develop and validate a deep learning model using lead-II electrocardiogram (ECG) monitoring waveforms (EMW) to estimate EF continuously and non-invasively, providing a practical alternative for real-time cardiac function monitoring in the ICU. Methods This prospective, single-center observational study was conducted in the Emergency Intensive Care Unit (EICU) of the First Affiliated Hospital of the University of Science and Technology of China between September 2024 and March 2025. Patients underwent 120-second lead-II EMW recordings at 500 Hz, synchronized with M-mode echocardiography. EF was quantified using the Teichholz method and verified by three experienced echocardiographers. EMW signals were processed using wavelet-based denoising, R-peak detection, and heartbeat segmentation. Two datasets were developed: the Single-beat EF Dataset (SEF), with beat-level EF labels, and the Averaged EF Dataset (AveEF), with patient-level average EF labels. A convolutional neural network combined with a long short-term memory architecture (CNN-LSTM) was trained for continuous EF regression. Performance was evaluated using mean squared error (MSE) and mean absolute error (MAE), as illustrated in Figure 1. Results： A total of 42 patients were enrolled, generating 6,948 EMW-EF paired samples. In the SEF Dataset, the CNN-LSTM model achieved a mean MAE<4 and worst MAE<10, outperforming models including CNN, LSTM, and Transformer architectures (R²= 0.9521). In the AveEF Dataset, the model achieved improved stability (R² =0.9596), with > 95% of samples presenting MAE<4. Subgroup analysis stratified by EF levels showed the highest predictive accuracy among patients with EF <40%, 93.1% of predictions had MAE <4. Prospective validation in three new patients confirmed generalizability, with all samples showing MAE <10. Conclusions: A CNN-LSTM model using short-duration lead-II EMW enables accurate, continuous, and non-invasive EF estimation. These findings support future development of waveform-based monitoring and wearable technologies for dynamic cardiac assessment in critical care.