12 lead surface ECGs as a surrogate of atrial electrical remodeling - a deep learning based approach

Abstract

Background and purpose:

Atrial fibrillation (AF), a common arrhythmia, is linked with atrial electrical and structural changes, notably low voltage areas (LVAs) which are associated with poor ablation outcomes and increased thromboembolic risk. This study aims to evaluate the efficacy of a deep learning model applied to 12-lead ECGs for non-invasively predicting the presence of LVAs, potentially guiding pre-ablation strategies and improving patient outcomes.

Methods:

A retrospective analysis was conducted on 204 AF patients, who underwent catheter ablation. Preprocedural sinus rhythm ECGs and electroanatomical maps (EAM) were utilized alongside demographic data to train a deep learning model combining Long Short-Term Memory networks and Convolutional Neural Networks with a cross-attention layer. Model performance was evaluated using a 5-fold cross-validation strategy.

Results:

The model effectively identified the presence of LVA on the examined atrial walls, achieving accuracies of 78 % for both the anterior and posterior walls, and 82 % for the LA roof. Moreover, it accurately predicted the global left atrial (LA) average voltage <0.7 mV, with an accuracy of 88 %.

Conclusion:

The study showcases the potential of deep learning applied to 12-lead ECGs to effectively predict regional LVAs and global LA voltage in AF patients non-invasively. This model offers a promising tool for the pre-ablation assessment of atrial substrate, facilitating personalized therapeutic strategies and potentially enhancing ablation success rates.

Introduction

Atrial fibrillation (AF), a prevalent cardiac arrhythmia, is associated with functional and electrical remodeling of the atria [1]. The characterization of the atrial electrical substrate has recently seen a growing interest, particularly concerning the identification of low voltage areas (LVAs) in the left atrium (LA). LVAs, identified through electroanatomical mapping (EAM), have been associated with procedural outcome and thromboembolic events in AF patients [2,3]. Their presence is linked to poor ablation outcomes and correlated with mitral regurgitation and LA stiffness [4,5], highlighting the significance of functional remodeling associated with LVAs [5]. Furthermore, recent studies underscore the clinical potential of targeting LVAs as a complementary strategy to conventional pulmonary vein isolation (PVI) in enhancing ablation success [6,7].

Currently, LVAs can only be assessed invasively using electroanatomic mapping (EAM) [8]. However, a non-invasive pre-ablation assessment of LVAs could significantly enhance clinical decision-making and help stratify AF patients. Various methods, including clinical scores and electrocardiography (ECG), have been investigated for predicting the LA substrate [9,10]. Earlier work by Schreiber et al. demonstrated a negative correlation between P-wave amplitude and the extent of LA LVAs, while more recent studies have shown that P-wave duration is associated with the presence of LVAs [11]. Despite these insights, the clinical utility of such findings has been limited. This is largely due to the reliance on conventional, manual P-wave analysis, which introduces examiner-dependent variability and reduces their practicality in routine clinical practice. Moreover, each of these studies identified different P-wave features—either amplitude or duration—as the most relevant marker of LVAs, suggesting that traditional methods of ECG analysis may be too narrow in scope by focusing on isolated features rather than capturing the full complexity of the signal.

Recent advances in artificial intelligence (AI) and deep learning have created new opportunities for non-invasive detection of various cardiac and non-cardiac abnormalities using ECG [12–14]. These developments suggest that more information can be extracted from the ECG waveform than previously recognized. In this study, we explore the application of deep learning to raw ECG signals to predict the presence of both regional and global LVAs.

Additionally, we plan to incorporate clinical parameters such as sex, age, and BMI, which are embedded in the digital ECG file, into our analysis. Previous studies have demonstrated that LVAs are more pronounced in older patients and females [15,16], while BMI significantly influences ECG signals specifically the P wave duration [17,18]. Therefore, we included these parameters in our model to enhance its predictive power with minimal additional effort in acquiring the data.

By utilizing this comprehensive approach, we aim to enable a non-invasive, accurate, and rapid assessment of the atrial electrical substrate from ECG. This could pave the way for a tool that can be seamlessly integrated into clinical practice, potentially improving patient outcomes by offering more efficient and reliable atrial substrate assessments.

Methods

Study design

This retrospective study involved 204 AF patients who underwent catheter ablation. Multimodal data, including preprocedural ECGs taken in sinus rhythm, clinical covariates - age, sex, body mass index (BMI), and EAM maps, were collected from all participants. The EAM maps were processed and utilized alongside the ECGs and clinical covariates to train the deep learning model. The study protocol adhered to the principles outlined in the Declaration of Helsinki and was approved by the local ethical committee. An overview of the study is summarized in Fig. 1.

Fig. 1.

Overview on the potential deployment of deep learning on ECG to detect Low Voltage Areas (LVA). The input consists of anonymized 12 lead ECG and BMI, age, and sex. The output consists of binary outcome as a yes/no for each predicted variable (presence of LVA on the anterior, posterior walls or the roof, or a total average atrial voltage of less than 0.7 mV). In this patient the output of the model stated that no LVA on either of the atrial walls is detected and the average LA voltage is >0.7 mV which correlated with the electroanatomical map.

12-Lead ECG and clinical covariate collection and processing

Twelve-lead ECGs in sinus rhythm were collected 1 to 7 days prior to ablation using Schiller CS 200 devices, and the digital signals were exported as anonymized DICOM files for analysis. All ECGs were recorded at a sampling frequency of 500 Hz, with a duration of 10 s, resulting in 5000 data points per ECG. The ECG-data was organized in a format of N x 12 × 5000, where N represents the number of patients, 12 corresponds to the ECG leads, and 5000 is the number of data points per lead. The pydicom package was employed to process the DICOM files and extract the raw ECG signals. These signals were filtered using a bandpass filter (cutoff frequencies - 3 Hz and 45 Hz). Alongside the ECG signals, clinical covariates such as age, sex, and BMI were collected. All the data was then scaled between 0 and 1. Samples with missing data were excluded, ensuring that only patients with complete data were included in the study. The processed ECG and clinical data were used to train the deep learning model.

Electroanatomical mapping process and analysis

EAM mapping was conducted using the CARTO^® 3 System while patients were in sinus rhythm. The mapping was performed using a multipolar mapping catheter, either PENTARAY (20 Electrodes, 2–6–2 mm, Biosense Webster) or LASSO (20 Electrodes, 4.5 mm, Biosense Webster), Bipolar voltage amplitudes from EAM points were recorded across the anterior and posterior walls and the LA roof. The average voltage amplitude for each region was calculated.

The presence of LVAs on each wall was defined by an average voltage amplitude of less than 0.5 mV across the wall, which is the clinically accepted cutoff for distinguishing between healthy and unhealthy atrial substrate [10,19]. Electrical remodeling in the atrium is a diffuse process with LVA presence being associated with a widespread reduction in voltage [20]. Hence, we additionally analyzed the global average LA voltage as a surrogate for the electrical remodeling in the atria with a threshold of 0.7 mV, as it is associated with the ablation outcome [21]. The deep learning model was trained to predict each of these binary outcomes (The general presence of LVA and the regional presence of LVA on each of the 3 examined atrial regions and the global average voltage <0.7 mV).

The multimodal deep learning network

Multiple architectures including LSTMs and CNNs have been used to analyze ECG waveforms. LSTMs have been used in analyzing time-varying signals as they can capture the global features from the signal. CNNs, on the other hand, have been a go-to deep learning model architecture since they are very successful in image classification and can capture local features very well. Our study uses a novel architecture that combines LSTM and CNN using a cross-attention layer. This ensures that the model captures both (local and global) features from the signal.

A 1D CNN with a kernel size of 3 and stride of 1 with padding set to 1 to preserve the signal length was used. The CNN was followed by a ReLU activation layer and a MaxPooling layer with a kernel size of 5, stride of 1, and padding of 2, ensuring constant signal length. The LSTM network was bidirectional in nature and dropout was added to ensure regularization. Both CNN and LSTM were trained concurrently. The 12-lead ECG waveform is passed to the LSTM layer and the CNN layer for feature extraction. The local and global extracted features are fused together using a cross-attention layer. These ECG features are then combined with the clinical covariates and passed to a Fully Connected Layer to predict the final binary outcomes like presence of LVA on the anterior wall, etc. Each binary variable was predicted using a different model – resulting in a total of 5 models for the 5 outcome variables. Fig. 2 illustrates the deep learning architecture used for each of the models in this study.

Fig. 2.

Deep Learning Model Architecture. The input consists of a 10-s ECG sample. A convolutional neural network (CNN) layer and a long short-term memory (LSTM) network were used independently for extracting short-term and long-term features. These extracted features were then combined in a cross-attention layer. The clinical parameters (age, sex, BMI) were incorporated in a fully connected layer. The figure shows a general structure of the model with two neurons in the output layer since independent models were developed to predict different variables in the data.

The data was split into 80–20 train-test split. A 5-fold cross-validation approach was used to ensure model robustness and generalizability. The train set in each fold was further split in 80–20 fashion to create the validation set for each fold ensuring the test set remained untouched during the model development process. The effective dataset splits were 64–16–20 train-validation-test splits. Hyperparameter tuning was performed during cross-validation. The hyperparameter space consisted of the number of CNN and LSTM layers from 1 to 3, and the number of fully connected layers from 1 to 3. A constraint on the number of CNN and LSTM layers to be the same was enforced. The hidden dimensions in the LSTM and attention layers were chosen from (32, 64, 128). CNN kernels were calculated based on number of layers i.e. the last CNN layer would always have 32 kernels. For example, in a 3-layer CNN – the neurons would be 128, 64 and 32 in the respective layers. The bidirectional nature of the LSTM was also considered as a tunable parameter. Finally, the dropout rate was also tuned which ranged from 0.1 to 0.3 with a step size of 0.1. All hyperparameter tuning was performed using only the training and validation sets.

The training spanned 100 epochs, with early stopping if no performance improvement was observed over 10 epochs. The learning rate was dynamically adjusted using an exponential scheduler, starting at 0.01 and reducing with a decay parameter of 0.9.

The analysis was performed in Python using PyTorch for deep learning, along with NumPy and SciPy for signal processing, and Matplotlib for visualization. An in-house Nvidia RTX A4500 GPUs with 20Gb VRAM was used for training all models.

Model interpretability

To interpret the deep learning model, Integrated Gradients (IG) method was used [22]. In the IG definition, a function F represents the deep learning model, input x ∈ Rⁿ, and the baseline xʹ ∈ Rⁿ. A straight-line path between x and x’ is assumed and gradients are computed along that path in m interpolation steps. IGs were computed using the captum package [23].

Data demographics statistical methods

Data are presented as median with interquartile range (IQR). The model’s accuracy, sensitivity, specificity and positive predicted value (PPV) along with the corresponding confidence intervals were calculated [24]. Statistical computations were performed using R Statistical Software (v4.2.2; R Core Team, 2021, R: A language and environment for statistical computing, R Foundation for Statistical Computing, Vienna, Austria).

Results

In this study, 204 patients were included for analysis. The baseline characteristics of the patients are summarized in Table 1.

Table 1.

Baseline clinical characteristics.

	n = 204
Age, years	64.0 [56.0–72.0]
Sex, female, n (%)	61 (29.9)
Body mass index, kg/m2	26.5 [24.4–29.7]
Diabetes mellitus, n (%)	15 (7.4)
Hypertension, n (%)	126 (61.8)
Persistent AF, n (%)	103 (50.5)
CHA2DS2VASc-Score	2.6 [1.9–3.6]
EF%	55.0 [55.0–60.0]
Patient with no LVA, n (%)	141 (69.1)
Patient with LVA on the anterior wall, n (%)	32 (15.7)
Patient with LVA on the posterior wall, n (%)	22(10.7)
Patient with LVA on the LA roof, n (%)	47 (23)
Patient with avg LA Voltage of < 0.7 mV, n (%)	43 (21.1)

Continuous variables are presented as median [interquartile range]. Categorical variables are presented as n (%).

AF: Atrial fibrillation, EF: Ejection fraction, LVA: Low Voltage Areas, avg.: Average.

Predictive capacity of deep learning ECG model for the identification of global LA average voltage and the regional presence of LVAs

Performance metrics for our model, detailed in Table 2, reflect the number of positive samples in the test set for each predicted variable, the model’s accuracy, sensitivity, specificity and PPV with their confidence intervals in identifying the presence of LVAs across the examined atrial regions and in predicting the average LA voltage <0.7 mV. The model performance was better in predicting the global average voltage in the LA compared to the regional presence of LVA on the atrial walls.

Table 2.

Performance metrics for identifying LVA presence on individual atria regions and the global LA average voltage.

Predicted Variable	N-positive	Accuracy (95 % CI)	Sensitivity (95 % CI)	Specificity (95 % CI)	PPV (95 % CI)
LVA Presence on the anterior wall	11	78 (74–82)	55 (25–84)	86 (74–99)	60 (30–90)
LVA Presence on the posterior wall	6	78 (74–82)	67 (29–104)	80 (66–93)	37 (32–40)
LVA Presence on the LA roof	8	83 (80–86)	75 (45–105)	84 (72–97)	55 (7–64)
Global LA average Voltage of <0.7 mV	9	88 (85–91)	78 (50–105)	90 (80–100)	70 (42–98)
LVA presence in LA	11	80 (77–83)	36 (8–65)	97 (90–103)	80 (45–115)

N-positive: number of positive samples in the test set for each predicted variable, LVA: Low Voltage Areas, LA: Left atrium.

Interpretability analysis

The interpretability analysis is presented in Fig. 3. The analysis demonstrates that the model relies on the entire ECG signal to generate predictions.

Fig. 3.

Interpretability analysis of the deep learning model. The ECG waveforms are color-coded, ranging from blue to red, where red signifies areas of high importance to the model’s predictions, and blue indicates lower importance. In the ECG panels, key segments of the waveforms in specific leads (Leads V1 and V3 in Panel A, and Leads V1 and aVL in Panel B) are marked in red, showing that these regions were crucial for the model’s predictions. (For interpretation of the references to color in this figure legend, the reader is referred to the web version of this article.)

Discussion

In this study, we developed a deep learning model that uses preprocedural 12-lead ECGs taken in sinus rhythm, along with clinical covariates (age, sex, and BMI), to predict the presence of LA LVAs in patients with AF.

The key findings from our research are summarized as follows:

• Deep learning applied to ECG data showed high accuracy in identifying the average LA voltage <0.7 mV.

• Our deep learning model effectively determined the general and regional presence of LVA on each of the examined atrial walls.

The model performed better in predicting the global average voltage in the LA compared to detecting the regional presence of LVAs on the atrial walls. This enhanced accuracy can be attributed to the global average encompassing a wider spectrum of electrical remodeling across the entire atria, which is more readily detected on surface ECG compared to localized changes.

Unlike traditional ECG analysis, which focuses solely on the P-wave to assess the atrial substrate, our approach utilizes the entire ECG waveform, including complete 10-s digital ECG recordings.

Our deep learning model facilitates the detection of atrial electrical remodeling using non-invasive, cost-effective, and accessible tools like the ECG. This approach is crucial due to the established association between LVAs and AF recurrence post-ablation, atrial functional mitral regurgitation, and thromboembolic events [2,3,25]. With increasing evidence supporting the benefits of additional substrate modification alongside PVI in patients with LVAs [6,7], the predictive capability of our model to identify LVAs prior to intervention can enable patient-specific ablation approaches and may guide the selection of optimal ablation techniques.

Upon examining the explainability analysis, we observed that the deep learning model leverages the entire ECG waveform. By using the full 10-s ECG rather than a median P-wave beat, the model captures dynamic aspects of the heart’s electrical activity, such as heart rate variability, which could influence the detection of LVAs. However, the specific contributions of different waveform components still require further investigation. This approach distinguishes the deep learning model from conventional ECG analysis, which typically focuses solely on the P-wave to assess the atrial substrate. Furthermore, the deep learning model offers a more streamlined and adaptable method for clinical use, reducing examiner-dependent variability and enhancing its practicality. Integrating such models into the decision-making process for elective procedures could significantly improve clinical outcomes.

Limitations

This study has several limitations. It relies on a relatively small sample size, which may impact the generalizability of the findings. Additionally, variations in electrode placement and patient anatomy can introduce variability into the data, potentially influencing the results. Furthermore, there is a need for external validation and a larger dataset to robustly establish these findings.

Conclusion

This study highlights the potential of deep learning applied to 12-lead ECGs for assessing the left atrial substrate in AF patients. Our model effectively identifies the presence and location of LVAs with high accuracy, offering a novel, non-invasive method for evaluating the LA substrate. This underscores the potential of the ECG in personalized AF management.