Convolutional Neural Network for Real‑Time Localization of Ganglionated Plexi from Bipolar Intracardiac Electrograms

Convolutional Neural Network for Real‑Time Localization of Ganglionated Plexi from Bipolar Intracardiac Electrograms | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Convolutional Neural Network for Real‑Time Localization of Ganglionated Plexi from Bipolar Intracardiac Electrograms Tumer Erdem Guler, Metin Cagdas, Sukriye Ebru Onder, Serdar Bozyel, and 6 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7974349/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 28 Mar, 2026 Read the published version in Journal of Interventional Cardiac Electrophysiology → Version 1 posted You are reading this latest preprint version Abstract Background Precise localization of ganglionated plexi (GP) is critical for effective cardioneuroablation, yet current mapping relies on labour‑intensive stimulation and subjective electrogram (EGM) interpretation. Recent advancements in deep learning (DL) have shown the potential to automate and improve outcomes an atrial fibrillation by analyzing EGMs. We aimed to apply DL to raw bipolar EGMs in order to automate GP detection. Methods A total of 189 760 bipolar windows (18 left‑atrium and 15 right‑atrium maps, respectively) were collected from 18 patients. GP annotation was performed independently by two experienced electrophysiologists. Five atrial maps from three patients were withheld for external testing; the remaining 15 patients yielded 119 222 clean windows for model development (GP prevalence ≈ 3.5%). A lightweight one‑dimensional convolutional neural network (CNN) was implemented using PyTorch. Training used focal loss (α = 0.75, γ = 2.0) and class‑balanced sampling. Performance was assessed with ROC/PR curves, threshold sweeps and gradient‑weighted class activation mapping (GCAM) saliency mapping. Results On the validation set the model achieved 69.6% accuracy; GP precision, recall and F1‑score were 0.09, 0.85 and 0.17, respectively. External testing on 34 976 unseen windows produced ROC‑AUC = 0.870 and PR‑AUC = 0.349. A probability threshold of 0.70 captured 51% of reference GP sites while highlighting anatomically plausible “hot‑spots” (513/9 063 nodes). GCAM consistently focused on central waveform segments (indices 140–160), aligning with fractionated autonomic signatures and reinforcing model interpretability. Conclusions The proposed explainable one‑dimensional CNN detects GP substrates with high sensitivity despite pronounced class imbalance and generalizes to unseen atria. Its probability maps and saliency outputs provide intuitive visual guidance, supporting real‑time, physiology‑aware decision making in cardioneuroablation. ganglionated plexus intracardiac electrogram cardioneuroablation convolutional neural network class imbalance Grad‑CAM interpretability autonomic mapping deep learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Figure 8 Introduction The intrinsic cardiac autonomic nervous system, organized into discrete ganglionated plexi (GP) embedded within the atrial fat pads, exerts powerful control over sinus and atrioventricular node function (1–3). Over-activity of these neural clusters has been implicated in functional atrioventricular block, neurally mediated syncope, vagally mediated atrial fibrillation and sinus bradycardia (4–6). Cardioneuroablation (CNA)—targeted radio-frequency elimination of GP sites—has therefore emerged as a promising treatment when conventional pharmacology or pacing is ineffective in these clinical scenarios (7–15). Success, however, hinges on precise localization of autonomic tissue. Present-day mapping relies on labor-intensive high-frequency stimulation, voltage/fractionation surrogates and the operator’s subjective assessment of electrogram (EGM) complexity, all of which are time-consuming and prone to inter-observer variability (3, 16–18). Despite this progress, no published work has focused on the automatic localization of autonomic ganglia, a task made even more challenging by their sparse distribution within routine mapping datasets. Recently, deep learning methods have shown great potential in the healthcare and medical areas. Specifically, some pioneering work has shown success in using deep learning methods for atrial fibrillation detection (19). Convolutional neural network (CNN) is a type of deep learning that excels in processing 2D data, such as images. However, by considering signals as 1-dimensional (1D) data, studies have shown promising results using convolutions for signal processing (20, 21). Rectified linear unit (ReLU) activations is a popular activation functions used in neural networks, especially in deep learning models. We therefore developed a lightweight one-dimensional convolutional neural network (CNN) that ingests short, 150-ms bipolar windows and outputs the probability that a point harbors GP tissue. To counter the extreme class imbalance typical of CNA procedures (< 2% GP points), the model incorporates focal-loss optimization, balanced mini-batch sampling and early weight stabilization. Performance is rigorously evaluated on a large, patient-wise hold-out cohort and interpreted with gradient-weighted class-activation mapping (Grad-CAM) to ensure physiological plausibility. We hypothesized that this mathematically coherent pipeline— ReLU activations in hidden layers, sigmoid output, focal loss for class imbalance—would (i) discriminate GP from non-GP sites with an accuracy comparable to state-of-the-art arrhythmia classifiers, (ii) offer flexible operating thresholds for screening, balanced navigation and high-confidence confirmation, and (iii) provide interpretable saliency maps aligned with known electroanatomic features of autonomic tissue. The present study tests these hypotheses and positions deep learning as a viable decision-support tool for streamlining and objectifying CNA. Methods Data Acquisition and Annotation Intracardiac bipolar electrograms (iEGMs) were obtained from 18 patients who underwent CNA procedures at Kocaeli City Hospital between April 2023 and July 2025. Recordings were made using multielectrode catheters connected to an EnSite X™ mapping system (Abbott, USA), sampled at 2 kHz with 24‑bit resolution and a 30–300 Hz band‑pass plus notch filter. A total of 189,760 bipolar windows (each 150 ms, or 300 samples) were extracted across both atria. Each window was paired with metadata including three‑dimensional coordinates (x, y, z), peak‑to‑peak voltage, and a binary label indicating whether the point harbored GP tissue. Total 18 left‑atrium and 15 right‑atrium maps were included in the study (Fig. 1 ). Nine predefined GP regions were evaluated—LSGP (superior left atrial GP), RSGP‑Left (superior right atrial GP – left atrial access), RSGP‑Right (superior right atrial GP – right atrial access), RIGP (inferior right atrial GP), LIGP (inferior left atrial GP), MTGP (Marshall Tract GP), PMLGP‑Left (posteromedial left atrial GP– left atrial access), PMLGP‑Right (posteromedial left atrial GP– right atrial access), and SVC‑Ao GP (Superior Vena Cava–Aorta GP) (Fig. 2 ). GP annotation was performed independently by two senior electrophysiologists, each with more than a decade of invasive‑mapping experience and discrepancies were resolved by consensus. A mapping point was designated GP only when three conditions were simultaneously fulfilled. First, the bipolar electrogram displayed a highly fractionated morphology—multiple sharp deflections and complex high‑frequency content—characteristic of autonomic tissue. Second, the point’s three‑dimensional location lay within anatomical regions recognized to harbor major atrial GP on the electro‑anatomic map. Third, focal radio‑frequency application at that site provoked an immediate vagal response, such as pronounced sinus slowing or transient atrioventricular block. Mapping points that satisfied all three criteria were labelled GP; all remaining points were recorded as non‑GP. Pre‑processing iEGMs were collected from 18 patients. To evaluate final model performance, five atrial maps obtained from three patients were set aside as an independent test set, leaving 28 maps (15 left‑atrial and 13 right‑atrial) for model development. In total, 189,760 bipolar signal windows of 300 samples each were extracted. Raw amplitude strings were converted to 300‑element NumPy vectors and screened for completeness. After removing malformed or corrupted records, 119,222 windows remained for training and validation. An 80 : 20 split was applied on an atrium‑wise basis—that is, all windows from a given atrial map were assigned exclusively to either the training or validation fold to prevent patient‑level data leakage. Amplitude vectors were min–max normalized within each fold and reshaped to (N, 1, 300) tensors for input to the CNN. Corresponding binary labels (GP vs non‑GP) and three‑dimensional coordinate matrices were retained for subsequent interpretability analysis. The resulting training set comprised 95,871 windows, of which 3,355 (3.5%) were labelled as GP, yielding a pronounced class imbalance. To counter this skew, we employed focal loss during optimization and constructed mini‑batches with a class‑balanced sampler. Empirically derived class weights were 0.52 for non‑GP (label 0) and 14.29 for GP (label 1), ensuring that gradient updates remained sensitive to the minority class throughout training. Model Architecture A lightweight one‑dimensional convolutional neural network (1D‑CNN) was implemented using PyTorch. Using Optuna’s default Bayesian TPE sampler together with an ASHA pruner, we tuned dropout (0.10–0.50), Adam learning‑rate (1 × 10⁻⁵–1 × 10⁻³) and batch size (16/32/64) while keeping the full preprocessing pipeline fixed. The best configuration found dropout = 0.30, lr ≈ 8.6 × 10⁻⁵, batch = 32. The architecture comprised three convolutional blocks with kernel sizes of 7, 5, and 3, each followed by batch normalization, ReLU activation, and dropout layers (dropout rates 0.2, 0.3, 0.4 respectively). Temporal dimensionality was reduced via global average pooling, followed by two fully connected layers with 256 and 128 neurons. A sigmoid output neuron produced the final GP probability. The full model contained approximately 179,745 trainable parameters (Fig. 3 ). Training Procedure Training employed focal loss with hyper‑parameters α = 0.75 and γ = 2.0. The Adam optimizer was used (initial learning rate 1 × 10⁻³, weight decay 1 × 10⁻⁵), along with a ReduceLROnPlateau scheduler (patience = 8, factor = 0.5, minimum LR = 1 × 10⁻⁷). Early stopping was triggered if the validation loss failed to improve for 15 consecutive epochs (maximum 100 epochs). The mini‑batch size was set to 32. Class weights were automatically calculated from the training distribution and applied during loss computation. Evaluation Metrics and Threshold Optimization Model performance was evaluated at multiple thresholds between 0.05 and 0.95. For each threshold, metrics including sensitivity, specificity, precision, F1‑score, Youden’s J statistic, and accuracy were computed. The optimal threshold was selected based on application‑specific requirements: high sensitivity for screening, balanced Youden index for procedural navigation, or high precision for confirmation. Gradient‑weighted class activation mapping (Grad‑CAM) was applied to all validation samples to assess the physiological interpretability of network attention. Model Export and Deployment The trained model, normalization scalers, threshold logic, and Grad‑CAM outputs were serialized and stored for deployment. All evaluation metrics, confusion matrices, ROC and PR curves, and per‑window prediction probabilities were saved for clinical review and integration into intra‑procedural guidance tools. Results A total of 189 760 bipolar iEGM windows were recorded from 18 patients. Of these, 4 345 windows (2.3%) satisfied all electrophysiological criteria for GP activity. After removing corrupted or incomplete segments, 119 222 clean windows were retained for model development. To preserve patient‑level independence, the data were divided on an atrium‑wise basis into 95 871 training samples and 23 351 validation samples, maintaining the native class ratio (GP prevalence ≈ 3.5% in training, 3.6% in validation). The training fold contained 3 355 GP windows, whereas the validation fold contained 833. Model training converged after 16 epochs, with early stopping triggered by the minimum validation loss. The restored best‑epoch weights yielded an overall validation accuracy of 69.6%. For the minority GP class, the network achieved a precision of 0.09, recall of 0.85, and an F1‑score of 0.17; for the majority non‑GP class, precision, recall, and F1‑score were 0.99, 0.69, and 0.81, respectively. These values correspond to a macro‑averaged F1‑score of 0.49 and a weighted F1‑score of 0.79. The confusion matrix showed 705 true‑positive and 128 false‑negative GP detections, alongside 6 978 false positives, confirming high sensitivity but modest precision in the face of severe class imbalance. Comprehensive threshold analysis is summarized in Fig. 4 . The receiver‑operating‑characteristic curve demonstrated an area under the curve (AUC) of 0.870, and the precision–recall plot yielded an average precision of 0.349, indicating robust discrimination despite the 30‑to‑1 class skew. Operating points were compared across accuracy, F1‑score, Youden’s J statistic, and the sensitivity–specificity trade‑off; the optimum threshold for balanced intra‑procedural guidance lay near 0.55. Re‑evaluation on an independent test cohort reproduced an ROC‑AUC of 0.870 and a PR‑AUC of 0.349, confirming excellent generalizability. Network interpretability was explored with gradient‑weighted class‑activation mapping. Figure 5 A depicts the distribution of peak Grad‑CAM indices, revealing a dominant cluster between sample indices 140–160. The fraction of windows exceeding an importance score of 0.7 mirrored this concentration, supporting physiologically plausible localization of salient features (Fig. 5 B). Time‑domain overlays highlighted a sharply rising biphasic deflection at the waveform centre for GP sites (red) versus a broader, less asymmetric profile for non‑GP sites (blue) (Fig. 5 C). Class‑specific means confirmed these distinctions, and the corresponding Grad‑CAM traces showed consistently higher activations in GP windows, peaking in the same temporal segment (Fig. 6 ). The mean Grad‑CAM profile averaged over all windows is shown in Fig. 7 A, further illustrating the network’s focus on the 150‑sample region. Aggregated waveforms for GP and non‑GP classes are provided in Figs. 7 B and C, respectively. Finally, three‑dimensional probability maps projected model outputs onto patient‑specific atrial shells (Supplemental Fig. 1–3). In the independent right‑atrial example, 33 of 65 validated GP sites (51%) resided within regions scoring ≥ 0.70, while left‑atrial renderings demonstrated anatomically coherent clusters along known autonomic corridors in both left‑anterior‑oblique and postero‑anterior views. These visualizations illustrate how the model integrates spatial continuity with point‑wise probabilistic predictions to highlight clinically actionable “hot‑spots.” To support the interpretability and reliability of the spatial probability maps, we further analyzed the training dynamics and overall development process of the model. As shown in Supplemental Fig. 4, training and validation loss curves demonstrated rapid convergence and overall stability, with early stopping applied at epoch 16 to prevent overfitting. The full development pipeline—including data acquisition, preprocessing, model architecture, validation strategy, and external testing—is summarized in Fig. 8 . This systematic workflow underscores the robustness of the model and contextualizes the clinical validity of its output visualizations. All trained parameters, normalization scalers, operating thresholds, Grad‑CAM heat‑maps, and complete metric reports have been archived to ensure full reproducibility and to facilitate future integration into real‑time intra‑procedural guidance systems. Discussion Deep-learning–based classification of GP regions from intracardiac electrograms (iEGMs) is emerging as a powerful adjunct to autonomic-substrate mapping. Unlike rule-based algorithms—often confounded by the dynamic, fractionated and patient-specific morphology of electrograms—convolutional neural networks (CNNs) learn hierarchical spatial-temporal features directly from raw waveforms and can, in principle, generalize across a spectrum of clinical presentations. Yet several domain-specific hurdles must be overcome for such models to achieve reliable clinical adoption. Class imbalance is paramount: GP windows typically constitute < 5% of any mapping dataset, so naive optimization yields high specificity but unacceptably low sensitivity. Loss re-weighting, class-balanced resampling and focal loss are indispensable to force the network to attend to scarce positives. Subtle and overlapping signal features further complicate learning; in atrial remodeling or fibrosis, both GP and non-GP sites may exhibit fractionation or high-frequency deflections, blurring class boundaries and fostering over-fitting to noise. Inter- and intra-patient variability—arising from anatomy, autonomic tone, catheter orientation and contact—requires aggressive data augmentation and regularization to maintain generalizability. Reliable ground truth is equally problematic. Annotation noise and ambiguity result from subjective criteria (vagal reflexes, spectral markers) that can both miss true GP sites and mislabel equivocal regions, diminishing effective signal-to-noise ratio during training. Interpretability and trust present another barrier: even with Grad-CAM visualization, clinicians remain wary of black-box tools, especially if saliency occasionally highlights artefacts or baseline shifts. A purely temporal signal is also an imperfect proxy for a three-dimensional problem; without anatomical context, temporal–spatial resolution mismatches can undermine localization accuracy. Electrogram quality is further threatened by noise and preprocessing artefacts—over-filtering may erase genuine autonomic signatures, whereas under-filtering leaves residual noise that degrades feature extraction. Procedural bias is a final concern: models trained at a single center risk memorizing institution-specific artefacts (catheter type, stimulation protocol), limiting transportability. Ultimately, clinical integration demands real-time inference with strict latency and false-alarm constraints; excessive false positives prolong procedures, whereas false negatives risk incomplete denervation. Our network architecture addresses many of these issues. Stacked 1D convolutions treat each 150-ms iEGM segment as a time-series image; kernels learn autonomic motifs such as high-frequency bursts or sharp biphasic deflections. Convolutional feature maps are bias-corrected for baseline shifts, ReLU-activated to suppress noise, and down-sampled with max pooling to curb over-fitting and computational load. Global-max pooling condenses each map to its most informative scalar; fully connected layers then assign class probabilities via a sigmoid output. End-to-end optimization with focal loss concentrates gradient updates on the minority GP class, mitigating severe imbalance, while label stratification, Gaussian noise, gain scaling and temporal shifts bolster robustness. Grad-CAM saliency maps consistently localized attention to the central 140–160-sample region, mirroring expert recognition of autonomic signatures. Recent studies corroborate the value of such approaches. Rodrigo et al (22) reported CNNs that distinguish AF from atrial tachycardia (AT) with AUCs of 0.97 (unipolar) and 0.92 (bipolar). Liao et al (23) achieved an AUC of ≈ 0.923 for focal AF-trigger detection, while Chen et al (24) improved local-activation-time annotation to 81% accuracy with a CNN–LSTM hybrid versus 67% for decision trees. Extending these findings to autonomic mapping, our model maintained a recall of 0.85 despite a GP prevalence of only 2.3% across 18 patients and achieved a weighted F1‑score of 0.79. Precision remained modest (≈ 0.09 at the balanced‑threshold operating point), but threshold sweeps demonstrated tunable modes—ranging from high‑sensitivity screening to higher‑précision confirmation—allowing operators to trade off false‑positive burden against the risk of missed denervation sites. At the balanced-threshold operating point, the model’s precision was approximately 0.09, primarily due to morphological similarities between GP and non-GP sites, particularly in atrial regions affected by remodeling. In the procedural setting, this trade-off is acceptable for initial screening, as false positives mainly prompt additional mapping or stimulation rather than unnecessary ablation, thereby reducing the risk of missed GP targets. A recent study by Wang et al (25) utilized 46 hand-crafted electrogram features as input to a machine learning model for substrate localization. While this approach achieved promising accuracy, it relied on predefined feature engineering, which can inadvertently omit subtle morphological or temporal patterns not captured by the selected metrics. In contrast, our end-to-end CNN framework ingests minimally pre-processed raw bipolar iEGM segments without manual feature extraction, allowing the network to learn discriminative spatiotemporal representations directly from the signal. This design reduces bias from feature selection, enables discovery of novel electrophysiological patterns, and facilitates generalization across variable signal morphologies. The pipeline’s modular design—standardized preprocessing, stratified sampling, threshold calibration and interpretable outputs—facilitates intra-procedural deployment and regulatory auditing. Archived weights, scalers, prediction logs and saliency maps provide a transparent foundation for multicenter validation. Future researches should enlarge datasets across institutions, incorporate three-dimensional catheter-localization data, and explore transformer or hybrid CNN–transformer architectures capable of modelling long-range dependencies. Prospective trials correlating AI-guided ablation with autonomic endpoints (e.g., vagal reflex elimination) and long-term rhythm outcomes will be essential to confirm real-world benefit. In summary, a compact, interpretable CNN can exploit fine-grained temporal structure in bipolar iEGMs to detect autonomic substrates with high sensitivity, offering a viable decision-support tool for targeted CNA. When integrated with conventional mapping strategies, such models have the potential to shorten procedures, enhance lesion precision and improve patient outcomes. Declarations Conflict of interest The authors have no conflicts of interest to disclose. Ethical approval The study was approved by the institutional ethics committee. Informed consent Informed consent was waived due to retrospective nature of the study. Funding information The authors have no funding sources to disclose. Author Contribution T.E.G., S.E.Ö., S. B, and S.N.D. wrote the main manuscript text A.S, A.İ.C and A.B.D prepared figures 1-5. H.D.H. prepared remaining figures. T.A and M.C. did trial design and statistical evaluation. All authors reviewed the manuscript. References Armour JA, Murphy DA, Yuan BX, Macdonald S, Hopkins DA. Gross and microscopic anatomy of the human intrinsic cardiac nervous system. Anat Rec 1997;247:289‐98. Pauza DH, Skripka V, Pauziene N, Stropus R. Morphology, distribution, and variability of the epicardiac neural ganglionated subplexuses in the human heart. Anat Rec 2000;259:353-82. Aksu T, Gopinathannair R, Gupta D, Pauza DH. Intrinsic cardiac autonomic nervous system: What do clinical electrophysiologists need to know about the "heart brain"? J Cardiovasc Electrophysiol 2021;32:1737-47. Ajijola OA, Aksu T, Arora R, Biaggioni I, Chen PS, De Ferrari G, Dusi V, Fudim M, Goldberger JJ, Green AL, Herring N, Khalsa SS, Kumar R, Lakatta E, Mehra R, Meyer C, Po S, Stavrakis S, Somers VK, Tan AY, Valderrabano M, Shivkumar K. Clinical neurocardiology: defining the value of neuroscience-based cardiovascular therapeutics - 2024 update. J Physiol. 2025 Mar;603(7):1781-1839. Brignole M, Moya A, de Lange FJ, Deharo JC, Elliott PM, Fanciulli A, et al. 2018 ESC Guidelines for the diagnosis and management of syncope. Eur Heart J 2018;39:1883–1948. Glikson M, Nielsen JC, Kronborg MB, Michowitz Y, Auricchio A, Barbash IM, Barrabés JA, Boriani G, Braunschweig F, Brignole M, Burri H, Coats AJS, Deharo JC, Delgado V, Diller GP, Israel CW, Keren A, Knops RE, Kotecha D, Leclercq C, Merkely B, Starck C, Thylén I, Tolosana JM; ESC Scientific Document Group. 2021 ESC Guidelines on cardiac pacing and cardiac resynchronization therapy. Eur Heart J. 2021 Sep 14;42(35):3427-3520. doi: 10.1093/eurheartj/ehab364. Aksu T, Brignole M, Calo L, Debruyne P, Di Biase L, Deharo JC, Fanciulli A, Fedorowski A, Kulakowski P, Morillo C, Moya A, Piotrowski R, Stec S, Sutton R, van Dijk JG, Wichterle D, Tse HF, Yao Y, Sheldon RS, Vaseghi M, Pachon JC, Scanavacca M, Meyer C, Amin R, Gupta D, Magnano M, Malik V, Schauerte P, Shen WK, Acosta JCZ. Cardioneuroablation for the treatment of reflex syncope and functional bradyarrhythmias: A Scientific Statement of the European Heart Rhythm Association (EHRA) of the ESC, the Heart Rhythm Society (HRS), the Asia Pacific Heart Rhythm Society (APHRS) and the Latin American Heart Rhythm Society (LAHRS). Europace. 2024 Aug 3;26(8):euae206. doi: 10.1093/europace/euae206. Erratum in: Europace. 2025 Feb 5;27(2):euaf023. doi: 10.1093/europace/euaf023. Aksu T, Piotrowski R, Tung R, De Potter T, Markman TM, du Fay de Lavallaz J, Rekvava R, Alyesh D, Joza JE, Badertscher P, Do DH, Bradfield JS, Upadhyay G, Sood N, Sharma PS, Guler TE, Gul EE, Kumar V, Koektuerk B, Dal Forno ARJ, Woods CE, Rav-Acha M, Valeriano C, Enriquez A, Sundaram S, Glikson M, d'Avila A, Shivkumar K, Kulakowski P, Huang HD. Procedural and Intermediate-term Results of the Electroanatomical-guided Cardioneuroablation for the Treatment of Supra-Hisian Second- or Advanced-degree Atrioventricular Block: the PIRECNA multicentre registry. Europace. 2024 Jul 2;26(7):euae164. doi: 10.1093/europace/euae164. Gopinathannair R, Olshansky B, Turagam MK, Gautam S, Futyma P, Akella K, Tanboga HI, Bozyel S, Yalin K, Padmanabhan D, Shenthar J, Lakkireddy D, Aksu T. Permanent pacing versus cardioneuroablation for cardioinhibitory vasovagal syncope. J Interv Card Electrophysiol. 2025 Mar;68(2):203-210. doi: 10.1007/s10840-022-01456-x. Vandenberk B, Lei LY, Ballantyne B, Vickers D, Liang Z, Sheldon RS, Chew DS, Aksu T, Raj SR, Morillo CA. Cardioneuroablation for vasovagal syncope: A systematic review and meta-analysis. Heart Rhythm. 2022 Nov;19(11):1804-1812. Pachon JC, Pachon EI, Pachon JC, Lobo TJ, Pachon MZ, Vargas RN, Jatene AD. "Cardioneuroablation"--new treatment for neurocardiogenic syncope, functional AV block and sinus dysfunction using catheter RF-ablation. Europace. 2005 Jan;7(1):1-13. doi: 10.1016/j.eupc.2004.10.003. Francia P, Viveros D, Falasconi G, Penela D, Soto-Iglesias D, Martí-Almor J, Alderete J, Saglietto A, Bellido AF, Franco-Ocaña P, Zaraket F, Matiello M, Fernández-Armenta J, San Antonio R, Berruezo A. Clinical impact of aging on outcomes of cardioneuroablation for reflex syncope or functional bradycardia: Results from the cardionEuroabLation: patiEnt selection, imaGe integrAtioN and outComEs-The ELEGANCE multicenter study. Heart Rhythm. 2023 Sep;20(9):1279-1286. doi: 10.1016/j.hrthm.2023.06.007. Rivarola EWR, Hachul D, Wu TC, Pisani C, Scarioti VD, Hardy C, Darrieux F, Scanavacca M. Long-Term Outcome of Cardiac Denervation Procedures: The Anatomically Guided Septal Approach. JACC Clin Electrophysiol. 2023 Aug;9(8 Pt 1):1344-1353. Debruyne P, Rossenbacker T, Janssens L, Collienne C, Ector J, Haemers P, et al. Durable Physiological Changes and Decreased Syncope Burden 12 Months After Unifocal Right-Sided Ablation Under Computed Tomographic Guidance in Patients With Neurally Mediated Syncope or Functional Sinus Node Dysfunction. Circ Arrhythm Electrophysiol 2021; 14:e009747. Wileczek A, Stodolkiewicz-Nowarska E, Reichert A, Kustron A, Sledz J, Biernikiewicz W, Orlik B, Lipka M, Kutarski A, Hering D, Zając M, Stec S. Reevaluation of indications for permanent pacemaker implantation after cardioneuroablation. Kardiol Pol. 2023;81(12):1272-1275. Onder SE, Guler TE, Bozyel S, Cagdas M, Dalgic SN, Sipal A, Gecer S, Kılıc E, Santangeli P, Aksu T. Integration of automated peak frequency annotation with voltage mapping for identifying ventricular tachycardia ablation sites. J Interv Card Electrophysiol. 2025 Nov;68(8):1573-1583. doi: 10.1007/s10840-025-02045-4. Aksu T, Skeete JR, Huang HH. Ganglionic Plexus Ablation: A Step-by-step Guide for Electrophysiologists and Review of Modalities for Neuromodulation for the Management of Atrial Fibrillation. Arrhythm Electrophysiol Rev. 2023 Jan;12:e02. doi: 10.15420/aer.2022.37. Kim MY, Coyle C, Tomlinson DR, Sikkel MB, Sohaib A, Luther V, Leong KM, Malcolme-Lawes L, Low B, Sandler B, Lim E, Todd M, Fudge M, Wright IJ, Koa-Wing M, Ng FS, Qureshi NA, Whinnett ZI, Peters NS, Newcomb D, Wood C, Dhillon G, Hunter RJ, Lim PB, Linton NWF, Kanagaratnam P. Ectopy-triggering ganglionated plexuses ablation to prevent atrial fibrillation: GANGLIA-AF study. Heart Rhythm. 2022 Apr;19(4):516-524. doi: 10.1016/j.hrthm.2021.12.010. Cai W, Chen Y, Guo J, Han B, Shi Y, Ji L, Wang J, Zhang G, Luo J. Accurate detection of atrial fibrillation from 12-lead ECG using deep neural network. Comput Biol Med. 2020;116:103378. Huang M-L, Wu Y-S. Classification of atrial fibrillation and normal sinus rhythm based on convolutional neural network. Biomed Eng Lett. 2020;10(2):183–93. Tutuko B, Nurmaini S, Tondas AE, Rachmatullah MN, Darmawahyuni A, Esafri R, Firdaus F, Sapitri AI. AFibNet: an implementation of atrial fibrillation detection with convolutional neural network. BMC Med Inform Decis Mak. 2021 Jul 14;21(1):216. Rodrigo M, Alhusseini MI, Rogers AJ, Krittanawong C, Thakur S, Feng R, Ganesan P, Narayan SM. Atrial fibrillation signatures on intracardiac electrograms identified by deep learning. Comput Bill Med. 2022; 145: 105451. Liao S, Ragot D, Nayyar S, Suszko A, Zhang Z, Wang B, Chauhan V. Deep Learning Classification of Unipolar Electrograms in Human Atrial Fibrillation: Application in Focal Source Mapping. Front Physiol. 2021 Jul 30:12:704122. Chen X, Cheng Z, Wang S, Lu G, Xv G, Liu Q, Zhu X. Atrial fibrillation detection based on multi-feature extraction and convolutional neural network for processing ECG signals. Comput Bill Med. 2021; 202: 106009. Wang X, Dennis A, Hesselkilde EM, et al. Machine learning approach for automated localization of ventricular tachycardia ablation targets from substrate maps: development and validation in a porcine model. Eur Heart J Digit Health. 2025;6(6):645-655. Additional Declarations No competing interests reported. Supplementary Files SupplementalFigure1A.png SupplementalFigure1B.png SupplementalFigure2B.png SupplementalFigure3B.png SupplementalFigure2A.png SupplementalFigure3A.png SupplementalFigure4.png SupplementalFigures.docx Cite Share Download PDF Status: Published Journal Publication published 28 Mar, 2026 Read the published version in Journal of Interventional Cardiac Electrophysiology → Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7974349","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":543538098,"identity":"65efa283-7b03-4bd9-92b8-ccd027e677a2","order_by":0,"name":"Tumer Erdem Guler","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA30lEQVRIiWNgGAWjYNCCAhDBfABISMgQoZyZsYHBAMRgSwBp4SFFCw+EJKhBt73/+IMfBnbRBufPfH51o8aCh4H98NEN+LSYnTnM2NhjkJy74UbuNuucY0CH8aSl3cCr5UYyYwOPATNQC+824xw2oBYJHjP8Wu4/Zmz8Y1Cfu+H8mWfGOf+I0XKDmbGZx+Bw7oYDOcyPc9uI0XIm2XC2jMHx3Jk30syYc/skeNgI+uX4wQcf31RU5/adP/z4c863Ojl+9sPH8GpBBmwSYJJY5SDA/IEU1aNgFIyCUTByAAAurUskFetk5gAAAABJRU5ErkJggg==","orcid":"","institution":"Kocaeli State Hospital","correspondingAuthor":true,"prefix":"","firstName":"Tumer","middleName":"Erdem","lastName":"Guler","suffix":""},{"id":543538099,"identity":"c757b88c-4e2b-4591-9ba9-20e4202a49cb","order_by":1,"name":"Metin Cagdas","email":"","orcid":"","institution":"Kocaeli State Hospital","correspondingAuthor":false,"prefix":"","firstName":"Metin","middleName":"","lastName":"Cagdas","suffix":""},{"id":543538100,"identity":"647f8e30-c3d0-4674-8590-5a7730bdb0b9","order_by":2,"name":"Sukriye Ebru Onder","email":"","orcid":"","institution":"Kocaeli State Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sukriye","middleName":"Ebru","lastName":"Onder","suffix":""},{"id":543538101,"identity":"4289f269-5dbe-437b-a502-a22ff85890c0","order_by":3,"name":"Serdar Bozyel","email":"","orcid":"","institution":"Kocaeli State Hospital","correspondingAuthor":false,"prefix":"","firstName":"Serdar","middleName":"","lastName":"Bozyel","suffix":""},{"id":543538102,"identity":"354c87c9-9bd7-4b68-95ea-9941081a8ff8","order_by":4,"name":"Sadiye Nur Dalgic","email":"","orcid":"","institution":"Kocaeli State Hospital","correspondingAuthor":false,"prefix":"","firstName":"Sadiye","middleName":"Nur","lastName":"Dalgic","suffix":""},{"id":543538103,"identity":"008318d1-f0f5-49c2-a436-a7272fd23ece","order_by":5,"name":"Abdulcebbar Sipal","email":"","orcid":"","institution":"Kocaeli State Hospital","correspondingAuthor":false,"prefix":"","firstName":"Abdulcebbar","middleName":"","lastName":"Sipal","suffix":""},{"id":543538104,"identity":"7000205d-8063-4d92-80ad-375302e4c779","order_by":6,"name":"Aziz İnan Celik","email":"","orcid":"","institution":"Gebze Fatih State Hospital","correspondingAuthor":false,"prefix":"","firstName":"Aziz","middleName":"İnan","lastName":"Celik","suffix":""},{"id":543538105,"identity":"a5cc62e6-f67e-46ad-92f5-d219a14f314b","order_by":7,"name":"Ahmet Berk Duman","email":"","orcid":"","institution":"Kocaeli State Hospital","correspondingAuthor":false,"prefix":"","firstName":"Ahmet","middleName":"Berk","lastName":"Duman","suffix":""},{"id":543538106,"identity":"809a4feb-37cf-4260-be2d-346ecb9aa624","order_by":8,"name":"Henry D. Huang","email":"","orcid":"","institution":"University of Chicago","correspondingAuthor":false,"prefix":"","firstName":"Henry","middleName":"D.","lastName":"Huang","suffix":""},{"id":543538107,"identity":"008db54d-e888-44be-a5a0-7d28622b5f2c","order_by":9,"name":"Tolga Aksu","email":"","orcid":"","institution":"Istanbul Aydin University","correspondingAuthor":false,"prefix":"","firstName":"Tolga","middleName":"","lastName":"Aksu","suffix":""}],"badges":[],"createdAt":"2025-10-29 01:08:25","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7974349/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7974349/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1007/s10840-026-02307-9","type":"published","date":"2026-03-28T16:12:25+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":96244651,"identity":"be2ea7e0-f92f-4c9f-9a8e-8d1292bbbebf","added_by":"auto","created_at":"2025-11-19 07:19:00","extension":"png","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2368521,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/b8662d9fbd20f9fdccb7d0a5.png"},{"id":95909939,"identity":"4724601d-467a-4d43-a937-7786076bef91","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":51198,"visible":true,"origin":"","legend":"","description":"","filename":"MainText.docx","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/e5b3b342195444121010f534.docx"},{"id":95909955,"identity":"7c71e421-ff09-4330-a1f9-1de56ad16deb","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14125262,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/8cfcc5bf2f7389bd0b213e88.png"},{"id":96243554,"identity":"3adecd79-226a-4ace-bb98-21a694be6764","added_by":"auto","created_at":"2025-11-19 07:16:39","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11374868,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/2bb95399e5a20bdda03b207b.png"},{"id":96244535,"identity":"a84b8e70-dca3-4630-9a6f-152cfbd4a115","added_by":"auto","created_at":"2025-11-19 07:18:47","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1612285,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/90078fdc457ca80e99042fb9.png"},{"id":96242308,"identity":"503643e1-239d-4ddf-9674-6ed96aa0a3f3","added_by":"auto","created_at":"2025-11-19 07:12:35","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1418532,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/837a769fad2d80d71fdf9750.png"},{"id":96243562,"identity":"97372376-58f5-4b90-8307-8210d87a52e6","added_by":"auto","created_at":"2025-11-19 07:16:40","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1439084,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/63f4bbacfc3bb4d96f7dbce4.png"},{"id":96242634,"identity":"a701c690-7023-46e6-8e37-cbe44b6bd755","added_by":"auto","created_at":"2025-11-19 07:13:53","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1362956,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/4e5c04ccb954c7fd5682ea66.png"},{"id":95909964,"identity":"197e673e-f5b2-42f3-9c6b-aa2443ea0852","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":244041,"visible":true,"origin":"","legend":"","description":"","filename":"Figure8.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/00feec3bf91763dda941df59.png"},{"id":95909960,"identity":"81245edd-3907-4400-8df4-86257d099fd1","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"json","order_by":9,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":10675,"visible":true,"origin":"","legend":"","description":"","filename":"6281eae2f7fe48c68d587dbb14e72c3d.json","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/149a1427612c9f80a4de5bf9.json"},{"id":95909962,"identity":"6dc5dd6b-e5e7-492c-8c87-174526f9d6de","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":10,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":406393,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure1A.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/92edf8528b50c5837c724fb4.png"},{"id":95909972,"identity":"f3d0fa6a-2766-4054-8ef0-fb57c59203d0","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":11,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":512863,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure1B.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/268a503fb5e7e98d98b0ee67.png"},{"id":96243675,"identity":"753b7ad9-00ce-46fb-9beb-d851e3ccf433","added_by":"auto","created_at":"2025-11-19 07:16:50","extension":"png","order_by":12,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":789049,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure2A.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/32f6018e1339e531d703a02a.png"},{"id":96242511,"identity":"45cc44ba-f881-45bd-adb9-f9b5e8c6c380","added_by":"auto","created_at":"2025-11-19 07:13:17","extension":"png","order_by":13,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":683925,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure2B.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/538cabdb18b66e2fa18ec1d1.png"},{"id":96243008,"identity":"349735f7-e49d-4908-b61e-1f81a8bca3cb","added_by":"auto","created_at":"2025-11-19 07:15:10","extension":"png","order_by":14,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":602997,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure3A.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/17ae6e9e6c6b88a77f9341b8.png"},{"id":95909970,"identity":"f8c6342d-64b0-461e-8a7c-60236c1c928d","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":15,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":762604,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure3B.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/00b21ca76d9b80cdfc22bc6b.png"},{"id":95909984,"identity":"8630eef9-47ae-442f-8b26-1b640f300cee","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":16,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":183332,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/f6decc1a435f89cdd4634129.png"},{"id":96243238,"identity":"b3c7629e-13b5-4959-bbc1-fed03b938489","added_by":"auto","created_at":"2025-11-19 07:15:53","extension":"xml","order_by":17,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":114008,"visible":true,"origin":"","legend":"","description":"","filename":"6281eae2f7fe48c68d587dbb14e72c3d1enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/b86da388091e59001c8bb271.xml"},{"id":96242919,"identity":"0fe658db-35c7-4d09-ae60-5f41f67977c7","added_by":"auto","created_at":"2025-11-19 07:14:51","extension":"png","order_by":18,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":2368521,"visible":true,"origin":"","legend":"","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/f89fde2a7e8f3c17341bf673.png"},{"id":96242929,"identity":"c599129b-af9f-491b-968b-c96180dcc568","added_by":"auto","created_at":"2025-11-19 07:14:53","extension":"png","order_by":19,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":14125262,"visible":true,"origin":"","legend":"","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/b8b256f37a589945626856a0.png"},{"id":96242289,"identity":"c4c2901f-78e6-4177-a892-4a3776e071a5","added_by":"auto","created_at":"2025-11-19 07:12:34","extension":"png","order_by":20,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":11374868,"visible":true,"origin":"","legend":"","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/696bcab9534af4191f803e09.png"},{"id":96243804,"identity":"94028513-fd3e-44e7-826c-bbb737d9714b","added_by":"auto","created_at":"2025-11-19 07:17:04","extension":"png","order_by":21,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1612285,"visible":true,"origin":"","legend":"","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/a79ba554002f319ba9997041.png"},{"id":96243591,"identity":"0eaaac30-cc50-4db6-ad56-796b05ce44e9","added_by":"auto","created_at":"2025-11-19 07:16:41","extension":"png","order_by":22,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1418532,"visible":true,"origin":"","legend":"","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/fd2d056d7dc6d4292b2e1bcd.png"},{"id":95909982,"identity":"f67af112-07fc-468c-8adf-fa3fef3b9b05","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":23,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1439084,"visible":true,"origin":"","legend":"","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/95ba5465b9bbc320501745ea.png"},{"id":95909977,"identity":"be8f2bc7-6bcc-439b-ae77-962881a8b3eb","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":24,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":1362956,"visible":true,"origin":"","legend":"","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/41b5de88f08509d84be77f8e.png"},{"id":96243274,"identity":"ac75a276-e6e3-48ef-95b4-25b8584d0fbf","added_by":"auto","created_at":"2025-11-19 07:15:58","extension":"png","order_by":25,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":244041,"visible":true,"origin":"","legend":"","description":"","filename":"Figure8.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/a73fb2b117c096572ff7e1c6.png"},{"id":96242925,"identity":"ca094a35-0b2f-4ede-9ef6-3ea3a930042b","added_by":"auto","created_at":"2025-11-19 07:14:53","extension":"png","order_by":26,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":256573,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/4f1bd2db613e357131d4e6c6.png"},{"id":95909968,"identity":"a33c1c0f-befe-467d-909e-7134eacd3fba","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":27,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":907270,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/6ff96f6b904d5b6343701b80.png"},{"id":96243254,"identity":"66ee975f-6958-42e0-abca-5e912b76bf3d","added_by":"auto","created_at":"2025-11-19 07:15:56","extension":"png","order_by":28,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":471402,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/1ad6e59ad4596774f7762405.png"},{"id":96243131,"identity":"53da1852-3e22-4672-8afc-b82c9a2a8777","added_by":"auto","created_at":"2025-11-19 07:15:38","extension":"png","order_by":29,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":262897,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/e21286fe801cf0d0c8c02469.png"},{"id":95909980,"identity":"8dce4997-33fb-467f-acc1-b115aee2a2b4","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":30,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":175819,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/066114e24d6c93b460bd252b.png"},{"id":96243229,"identity":"55e21ec8-9120-42fd-b737-d09b3cca5d9c","added_by":"auto","created_at":"2025-11-19 07:15:52","extension":"png","order_by":31,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":210224,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/7a8fadb7b8fdea27c7317e0e.png"},{"id":95909986,"identity":"dcf259ae-f009-4f60-a6f6-277f0e63aa04","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":32,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":273841,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/c910c78de7af562ad718c9bf.png"},{"id":95909990,"identity":"3c7ff8c6-ce89-4dcb-aac0-f47a602dd700","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":33,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64117,"visible":true,"origin":"","legend":"","description":"","filename":"OnlineFigure8.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/f0cc4babc87d2235ada0b514.png"},{"id":95909975,"identity":"6c93c0de-f4da-4e7c-a8d4-178e758040de","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"xml","order_by":34,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":110617,"visible":true,"origin":"","legend":"","description":"","filename":"6281eae2f7fe48c68d587dbb14e72c3d1structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/23b40635c821a2322b246708.xml"},{"id":96243220,"identity":"105a972c-19e3-47be-9fba-9471b28ef360","added_by":"auto","created_at":"2025-11-19 07:15:51","extension":"html","order_by":35,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":126107,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/5cfeeb4a4f9a6befc6896740.html"},{"id":96242541,"identity":"a4b26c27-65de-40ae-8b65-bd896382d247","added_by":"auto","created_at":"2025-11-19 07:13:25","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":2368521,"visible":true,"origin":"","legend":"\u003cp\u003eIllustration of input signal (A), structured data format with spatial and voltage features (B), and criteria used for GP annotation (C).\u003c/p\u003e","description":"","filename":"Figure1.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/f21e714224934af699504550.png"},{"id":95909961,"identity":"814978b9-5075-46ff-abbe-4e1e0920ef06","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":14125262,"visible":true,"origin":"","legend":"\u003cp\u003eAnatomical distribution of predefined ganglionated plexus (GP) sites targeted for annotation and classification.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/f25a211c1930903b682acf71.png"},{"id":96242705,"identity":"65e5f4a2-76ef-4eb3-9ca4-fd69bf18de83","added_by":"auto","created_at":"2025-11-19 07:14:06","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":11374868,"visible":true,"origin":"","legend":"\u003cp\u003eArchitecture of the 1D convolutional neural network (1D-CNN) implemented for GP classification.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 3 Legend \u003c/strong\u003eThe model comprises three successive convolutional blocks with kernel sizes of 7, 5, and 3, respectively. Each block includes batch normalization, ReLU activation, and dropout (rates 0.2, 0.3, 0.4). Feature maps are temporally reduced by global average pooling, followed by two fully connected layers (256 and 128 units). A sigmoid output neuron generates the GP classification probability. The final model contains approximately 179,745 trainable parameters.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/2e3580a3a4c31bccb02dae8b.png"},{"id":95909947,"identity":"adecac59-aa07-4c8b-a73c-d7488d10d2b8","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":1612285,"visible":true,"origin":"","legend":"\u003cp\u003eComprehensive threshold analysis of the CNA‑CNN.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 4 Legend \u003c/strong\u003eEight diagnostic panels summaries performance across decision thresholds from 0.05 to 0.95. The ROC curve (upper left) shows an AUC of 0.870, while the precision–recall (PR) curve (upper middle) yields an average precision of 0.349. Accuracy and F1‑score peak at thresholds of 0.55 and 0.75, respectively (upper right). Sensitivity and specificity intersect near 0.55 (center left), and GP‑specific precision rises monotonically with stricter thresholds (center right). Youden’s J statistic is maximal at 0.55 (lower left). The class‑distribution bar plot (lower right) highlights the 30 : 1 imbalance between non‑GP and GP windows in the validation set. Abbreviations: AUC, area under the curve; CNN, convolutional neural network; GP, ganglionated plexus; PR, precision–recall; ROC, receiver‑operating characteristic.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/d93d651ce9522362cb4e9bd0.png"},{"id":95909948,"identity":"ef852476-8735-4bb6-9698-ac25470164f9","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":1418532,"visible":true,"origin":"","legend":"\u003cp\u003e(A)\u003cstrong\u003e \u003c/strong\u003eHistogram of Grad‑CAM peak‑importance locations, (B) Fraction of windows exceeding a Grad‑CAM importance threshold of 0.7, and (C) Overlay of raw, time‑aligned iEGM windows\u003cstrong\u003e\u003cbr\u003e\n Figure 5 Legend\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) For every validation window the sample index exhibiting maximal Grad‑CAM intensity is recorded. Peaks cluster sharply between indices 140 and 160, consistent with the time window where fractionated autonomic potentials typically occur. \u003cstrong\u003e(B) \u003c/strong\u003eAcross the 300‑sample window, the proportion of traces with high salience (\u0026gt; 0.7) mirrors the peak‑index histogram, with \u0026gt; 45 % of all windows showing maximal salience between indices 150–155. \u003cstrong\u003e(C) \u003c/strong\u003eScaled signals from GP sites (red) and non‑GP sites (blue) are superimposed. GP traces exhibit a narrow, high‑amplitude biphasic deflection centered on index 150, whereas non‑GP traces display broader, lower‑frequency morphology.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eGrad‑CAM, gradient‑weighted class‑activation mapping; GP, ganglionated plexus.\u003c/em\u003e\u003cstrong\u003e \u003c/strong\u003e\u003cem\u003eGP, ganglionated plexus; iEGM, intracardiac electrogram.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/fe4ecf7bf8c91389f2139148.png"},{"id":95909959,"identity":"9c04320d-25d7-4674-b88c-cfc2d6937f99","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1439084,"visible":true,"origin":"","legend":"\u003cp\u003e(A) Average Grad-CAM importance in GP sites, (B) Average Grad-CAM importance in non-GP sites, and (C) Class-spesific average Grad-CAM importance\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 6 Legend\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) The mean time‑domain waveform for GP windows (upper panel, red) shows a sharp positive spike followed by a negative undershoot; the non‑GP average (lower panel, blue) is wider and less asymmetric, underscoring morphological differences leveraged by the CNN. (B) Mean Grad‑CAM traces for GP (red) and non‑GP (blue) classes highlight markedly higher salience in the GP ensemble, peaking at index 150. This alignment with physiological “hot‑spots” supports the plausibility of the network’s decision strategy. (C) The mean saliency curve peaks sharply at indices 150–155, indicating that, on average, the network concentrates its attention on the mid‑segment of each 300‑sample iEGM window—precisely where fractionated autonomic signatures typically arise. \u0026nbsp;\u003cem\u003eCNN, convolutional neural network; Grad‑CAM, gradient‑weighted class‑activation mapping; GP, ganglionated plexus; iEGM, intracardiac electrogram.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure6.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/29dffcdc8e58cf3d9f2f3a22.png"},{"id":96242686,"identity":"a63ac2c1-348a-456d-aa75-d18c5465bbf3","added_by":"auto","created_at":"2025-11-19 07:14:04","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":1362956,"visible":true,"origin":"","legend":"\u003cp\u003e(A) Average Grad‑CAM importance across all validation windows, (B) Aggregated time‑domain waveform for GP‑labelled sites, and (C) Aggregated time‑domain waveform for non‑GP sites\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 7 Legend\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(A) The mean saliency curve peaks sharply at indices 150–155, indicating that, on average, the network concentrates its attention on the mid‑segment of each 300‑sample iEGM window—precisely where fractionated autonomic signatures typically arise. (B) After min–max scaling and temporal alignment, the class‑mean signal (red) displays a narrow, high‑amplitude biphasic spike at index ≈ 150, preceded by a shallow pre‑potential and followed by a brief negative undershoot—morphology consistent with autonomic‑rich tissue. (C) The class‑mean trace for non‑GP windows (blue) shows a broader, less asymmetric complex at index ≈ 150, accompanied by a deeper pre‑potential trough, reflecting the smoother conduction properties of ordinary atrial myocardium. \u003cem\u003eGrad‑CAM, gradient‑weighted class‑activation mapping; GP, ganglionated plexus.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure7.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/fb24e620546dd199295d49ad.png"},{"id":95909969,"identity":"d866b896-4bc8-4f3a-a73e-649717bee5ae","added_by":"auto","created_at":"2025-11-14 10:09:11","extension":"png","order_by":8,"title":"Figure 8","display":"","copyAsset":false,"role":"figure","size":244041,"visible":true,"origin":"","legend":"\u003cp\u003eFive‑stage workflow for deep‑learning–based localisation of ganglionated plexus (GP) sites from intracardiac electrograms.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFigure 8 Legend\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e(1) Data acquisition and pre‑processing: bipolar iEGM windows were collected from 18 patients (33 atria); after cleaning and min–max normalisation, 119 222 windows remained. (2) CNN architecture and training: a three‑block 1‑D CNN with focal‑loss optimization was trained on 95 871 windows drawn from 15 patients (30 atria) using class‑balanced sampling. (3) Validation and early stopping: 23 351 patient‑wise windows monitored convergence, with the best weights recovered at epoch 16. (4) External testing: generalizability was confirmed on 34 976 unseen windows from three additional patients (five atria), yielding a ROC‑AUC of ≈ 0.870. (5) Visual outputs: model probabilities were projected onto 3‑D atrial shells and interrogated with Grad‑CAM to highlight physiologically plausible regions of autonomic activity.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003e3‑D, three‑dimensional; CNN, convolutional neural network; dev, development set; Grad‑CAM, gradient‑weighted class‑activation mapping; iEGM, intracardiac electrogram; ROC‑AUC, area under the receiver‑operating characteristic curve.\u003c/em\u003e\u003c/p\u003e","description":"","filename":"Figure8.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/865db71ad60081dba00f8ac2.png"},{"id":105904041,"identity":"c3bdc40b-52ee-4a94-9fa1-9207290a728f","added_by":"auto","created_at":"2026-04-01 10:02:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":36751567,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/1a2f48e3-553a-4130-a510-576ef3f89bff.pdf"},{"id":96242603,"identity":"1567c044-64b7-4546-a2c0-1e82b57d40f8","added_by":"auto","created_at":"2025-11-19 07:13:38","extension":"png","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":406393,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure1A.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/f5f32ce2d428bcddb623abd4.png"},{"id":95909940,"identity":"833399ba-8035-4440-b444-1550e43184f5","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":512863,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure1B.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/3c37628842c56fd5613c05e9.png"},{"id":95909942,"identity":"87e01846-2434-4d6a-a595-d29cf601445c","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":683925,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure2B.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/d482bea277078edc6a365f59.png"},{"id":95909946,"identity":"d70d9ed5-c41a-46b4-9659-c84ff0a9efe5","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":762604,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure3B.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/5473df52efb69505905ab0fa.png"},{"id":96242565,"identity":"3f6c4514-2196-4625-b4f4-afcd4e642b90","added_by":"auto","created_at":"2025-11-19 07:13:30","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":789049,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure2A.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/76c0064b4e74b501532b2b2c.png"},{"id":95909952,"identity":"2afc2b18-43d1-475e-9577-26a2f74669ad","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"supplement","size":602997,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure3A.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/ad4dee834db3f4822ad4ee09.png"},{"id":96242904,"identity":"7e6ceebb-f42c-411b-ae35-9a435565558e","added_by":"auto","created_at":"2025-11-19 07:14:50","extension":"png","order_by":7,"title":"","display":"","copyAsset":false,"role":"supplement","size":183332,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/1b93d129f3a8ab05750f4dca.png"},{"id":95909965,"identity":"99d026a6-b914-44eb-8330-b9418d1659a9","added_by":"auto","created_at":"2025-11-14 10:09:10","extension":"docx","order_by":8,"title":"","display":"","copyAsset":false,"role":"supplement","size":16865,"visible":true,"origin":"","legend":"","description":"","filename":"SupplementalFigures.docx","url":"https://assets-eu.researchsquare.com/files/rs-7974349/v1/cfd14e9f47c29dc5b9046bc0.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Convolutional Neural Network for Real‑Time Localization of Ganglionated Plexi from Bipolar Intracardiac Electrograms","fulltext":[{"header":"Introduction","content":"\u003cp\u003eThe intrinsic cardiac autonomic nervous system, organized into discrete ganglionated plexi (GP) embedded within the atrial fat pads, exerts powerful control over sinus and atrioventricular node function (1\u0026ndash;3). Over-activity of these neural clusters has been implicated in functional atrioventricular block, neurally mediated syncope, vagally mediated atrial fibrillation and sinus bradycardia (4\u0026ndash;6). Cardioneuroablation (CNA)\u0026mdash;targeted radio-frequency elimination of GP sites\u0026mdash;has therefore emerged as a promising treatment when conventional pharmacology or pacing is ineffective in these clinical scenarios (7\u0026ndash;15). Success, however, hinges on precise localization of autonomic tissue. Present-day mapping relies on labor-intensive high-frequency stimulation, voltage/fractionation surrogates and the operator\u0026rsquo;s subjective assessment of electrogram (EGM) complexity, all of which are time-consuming and prone to inter-observer variability (3, 16\u0026ndash;18).\u003c/p\u003e\u003cp\u003eDespite this progress, no published work has focused on the automatic localization of autonomic ganglia, a task made even more challenging by their sparse distribution within routine mapping datasets. Recently, deep learning methods have shown great potential in the healthcare and medical areas. Specifically, some pioneering work has shown success in using deep learning methods for atrial fibrillation detection (19). Convolutional neural network (CNN) is a type of deep learning that excels in processing 2D data, such as images. However, by considering signals as 1-dimensional (1D) data, studies have shown promising results using convolutions for signal processing (20, 21). Rectified linear unit (ReLU) activations is a popular activation functions used in neural networks, especially in deep learning models.\u003c/p\u003e\u003cp\u003eWe therefore developed a lightweight one-dimensional convolutional neural network (CNN) that ingests short, 150-ms bipolar windows and outputs the probability that a point harbors GP tissue. To counter the extreme class imbalance typical of CNA procedures (\u0026lt;\u0026thinsp;2% GP points), the model incorporates focal-loss optimization, balanced mini-batch sampling and early weight stabilization. Performance is rigorously evaluated on a large, patient-wise hold-out cohort and interpreted with gradient-weighted class-activation mapping (Grad-CAM) to ensure physiological plausibility.\u003c/p\u003e\u003cp\u003eWe hypothesized that this mathematically coherent pipeline\u0026mdash; ReLU activations in hidden layers, sigmoid output, focal loss for class imbalance\u0026mdash;would (i) discriminate GP from non-GP sites with an accuracy comparable to state-of-the-art arrhythmia classifiers, (ii) offer flexible operating thresholds for screening, balanced navigation and high-confidence confirmation, and (iii) provide interpretable saliency maps aligned with known electroanatomic features of autonomic tissue. The present study tests these hypotheses and positions deep learning as a viable decision-support tool for streamlining and objectifying CNA.\u003c/p\u003e"},{"header":"Methods","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eData Acquisition and Annotation\u003c/h2\u003e\u003cp\u003eIntracardiac bipolar electrograms (iEGMs) were obtained from 18 patients who underwent CNA procedures at Kocaeli City Hospital between April 2023 and July 2025. Recordings were made using multielectrode catheters connected to an EnSite X\u0026trade; mapping system (Abbott, USA), sampled at 2 kHz with 24‑bit resolution and a 30\u0026ndash;300 Hz band‑pass plus notch filter. A total of 189,760 bipolar windows (each 150 ms, or 300 samples) were extracted across both atria. Each window was paired with metadata including three‑dimensional coordinates (x, y, z), peak‑to‑peak voltage, and a binary label indicating whether the point harbored GP tissue. Total 18 left‑atrium and 15 right‑atrium maps were included in the study (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eNine predefined GP regions were evaluated\u0026mdash;LSGP (superior left atrial GP), RSGP‑Left (superior right atrial GP \u0026ndash; left atrial access), RSGP‑Right (superior right atrial GP \u0026ndash; right atrial access), RIGP (inferior right atrial GP), LIGP (inferior left atrial GP), MTGP (Marshall Tract GP), PMLGP‑Left (posteromedial left atrial GP\u0026ndash; left atrial access), PMLGP‑Right (posteromedial left atrial GP\u0026ndash; right atrial access), and SVC‑Ao GP (Superior Vena Cava\u0026ndash;Aorta GP) (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). GP annotation was performed independently by two senior electrophysiologists, each with more than a decade of invasive‑mapping experience and discrepancies were resolved by consensus. A mapping point was designated GP only when three conditions were simultaneously fulfilled. First, the bipolar electrogram displayed a highly fractionated morphology\u0026mdash;multiple sharp deflections and complex high‑frequency content\u0026mdash;characteristic of autonomic tissue. Second, the point\u0026rsquo;s three‑dimensional location lay within anatomical regions recognized to harbor major atrial GP on the electro‑anatomic map. Third, focal radio‑frequency application at that site provoked an immediate vagal response, such as pronounced sinus slowing or transient atrioventricular block. Mapping points that satisfied all three criteria were labelled GP; all remaining points were recorded as non‑GP.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\n\u003ch3\u003ePre‑processing\u003c/h3\u003e\n\u003cp\u003eiEGMs were collected from 18 patients. To evaluate final model performance, five atrial maps obtained from three patients were set aside as an independent test set, leaving 28 maps (15 left‑atrial and 13 right‑atrial) for model development. In total, 189,760 bipolar signal windows of 300 samples each were extracted. Raw amplitude strings were converted to 300‑element NumPy vectors and screened for completeness. After removing malformed or corrupted records, 119,222 windows remained for training and validation. An 80 : 20 split was applied on an atrium‑wise basis\u0026mdash;that is, all windows from a given atrial map were assigned exclusively to either the training or validation fold to prevent patient‑level data leakage. Amplitude vectors were min\u0026ndash;max normalized within each fold and reshaped to (N, 1, 300) tensors for input to the CNN. Corresponding binary labels (GP vs non‑GP) and three‑dimensional coordinate matrices were retained for subsequent interpretability analysis.\u003c/p\u003e\u003cp\u003eThe resulting training set comprised 95,871 windows, of which 3,355 (3.5%) were labelled as GP, yielding a pronounced class imbalance. To counter this skew, we employed focal loss during optimization and constructed mini‑batches with a class‑balanced sampler. Empirically derived class weights were 0.52 for non‑GP (label 0) and 14.29 for GP (label 1), ensuring that gradient updates remained sensitive to the minority class throughout training.\u003c/p\u003e\n\u003ch3\u003eModel Architecture\u003c/h3\u003e\n\u003cp\u003eA lightweight one‑dimensional convolutional neural network (1D‑CNN) was implemented using PyTorch. Using Optuna\u0026rsquo;s default Bayesian TPE sampler together with an ASHA pruner, we tuned dropout (0.10\u0026ndash;0.50), Adam learning‑rate (1 \u0026times; 10⁻⁵\u0026ndash;1 \u0026times; 10⁻\u0026sup3;) and batch size (16/32/64) while keeping the full preprocessing pipeline fixed. The best configuration found dropout\u0026thinsp;=\u0026thinsp;0.30, lr\u0026thinsp;\u0026asymp;\u0026thinsp;8.6 \u0026times; 10⁻⁵, batch\u0026thinsp;=\u0026thinsp;32. The architecture comprised three convolutional blocks with kernel sizes of 7, 5, and 3, each followed by batch normalization, ReLU activation, and dropout layers (dropout rates 0.2, 0.3, 0.4 respectively). Temporal dimensionality was reduced via global average pooling, followed by two fully connected layers with 256 and 128 neurons. A sigmoid output neuron produced the final GP probability. The full model contained approximately 179,745 trainable parameters (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\n\u003ch3\u003eTraining Procedure\u003c/h3\u003e\n\u003cp\u003eTraining employed focal loss with hyper‑parameters α\u0026thinsp;=\u0026thinsp;0.75 and γ\u0026thinsp;=\u0026thinsp;2.0. The Adam optimizer was used (initial learning rate 1 \u0026times; 10⁻\u0026sup3;, weight decay 1 \u0026times; 10⁻⁵), along with a ReduceLROnPlateau scheduler (patience\u0026thinsp;=\u0026thinsp;8, factor\u0026thinsp;=\u0026thinsp;0.5, minimum LR\u0026thinsp;=\u0026thinsp;1 \u0026times; 10⁻⁷). Early stopping was triggered if the validation loss failed to improve for 15 consecutive epochs (maximum 100 epochs). The mini‑batch size was set to 32. Class weights were automatically calculated from the training distribution and applied during loss computation.\u003c/p\u003e\n\u003ch3\u003eEvaluation Metrics and Threshold Optimization\u003c/h3\u003e\n\u003cp\u003eModel performance was evaluated at multiple thresholds between 0.05 and 0.95. For each threshold, metrics including sensitivity, specificity, precision, F1‑score, Youden\u0026rsquo;s J statistic, and accuracy were computed. The optimal threshold was selected based on application‑specific requirements: high sensitivity for screening, balanced Youden index for procedural navigation, or high precision for confirmation. Gradient‑weighted class activation mapping (Grad‑CAM) was applied to all validation samples to assess the physiological interpretability of network attention.\u003c/p\u003e\u003cdiv id=\"Sec8\" class=\"Section2\"\u003e\u003ch2\u003eModel Export and Deployment\u003c/h2\u003e\u003cp\u003eThe trained model, normalization scalers, threshold logic, and Grad‑CAM outputs were serialized and stored for deployment. All evaluation metrics, confusion matrices, ROC and PR curves, and per‑window prediction probabilities were saved for clinical review and integration into intra‑procedural guidance tools.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003eA total of 189 760 bipolar iEGM windows were recorded from 18 patients. Of these, 4 345 windows (2.3%) satisfied all electrophysiological criteria for GP activity. After removing corrupted or incomplete segments, 119 222 clean windows were retained for model development. To preserve patient‑level independence, the data were divided on an atrium‑wise basis into 95 871 training samples and 23 351 validation samples, maintaining the native class ratio (GP prevalence\u0026thinsp;\u0026asymp;\u0026thinsp;3.5% in training, 3.6% in validation). The training fold contained 3 355 GP windows, whereas the validation fold contained 833.\u003c/p\u003e\u003cp\u003eModel training converged after 16 epochs, with early stopping triggered by the minimum validation loss. The restored best‑epoch weights yielded an overall validation accuracy of 69.6%. For the minority GP class, the network achieved a precision of 0.09, recall of 0.85, and an F1‑score of 0.17; for the majority non‑GP class, precision, recall, and F1‑score were 0.99, 0.69, and 0.81, respectively. These values correspond to a macro‑averaged F1‑score of 0.49 and a weighted F1‑score of 0.79. The confusion matrix showed 705 true‑positive and 128 false‑negative GP detections, alongside 6 978 false positives, confirming high sensitivity but modest precision in the face of severe class imbalance.\u003c/p\u003e\u003cp\u003eComprehensive threshold analysis is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig6\" class=\"InternalRef\"\u003e4\u003c/span\u003e. The receiver‑operating‑characteristic curve demonstrated an area under the curve (AUC) of 0.870, and the precision\u0026ndash;recall plot yielded an average precision of 0.349, indicating robust discrimination despite the 30‑to‑1 class skew. Operating points were compared across accuracy, F1‑score, Youden\u0026rsquo;s J statistic, and the sensitivity\u0026ndash;specificity trade‑off; the optimum threshold for balanced intra‑procedural guidance lay near 0.55. Re‑evaluation on an independent test cohort reproduced an ROC‑AUC of 0.870 and a PR‑AUC of 0.349, confirming excellent generalizability.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eNetwork interpretability was explored with gradient‑weighted class‑activation mapping. Figure\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003eA depicts the distribution of peak Grad‑CAM indices, revealing a dominant cluster between sample indices 140\u0026ndash;160. The fraction of windows exceeding an importance score of 0.7 mirrored this concentration, supporting physiologically plausible localization of salient features (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003eB). Time‑domain overlays highlighted a sharply rising biphasic deflection at the waveform centre for GP sites (red) versus a broader, less asymmetric profile for non‑GP sites (blue) (Fig.\u0026nbsp;\u003cspan refid=\"Fig7\" class=\"InternalRef\"\u003e5\u003c/span\u003eC). Class‑specific means confirmed these distinctions, and the corresponding Grad‑CAM traces showed consistently higher activations in GP windows, peaking in the same temporal segment (Fig.\u0026nbsp;\u003cspan refid=\"Fig9\" class=\"InternalRef\"\u003e6\u003c/span\u003e). The mean Grad‑CAM profile averaged over all windows is shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e7\u003c/span\u003eA, further illustrating the network\u0026rsquo;s focus on the 150‑sample region. Aggregated waveforms for GP and non‑GP classes are provided in Figs.\u0026nbsp;\u003cspan refid=\"Fig11\" class=\"InternalRef\"\u003e7\u003c/span\u003eB and C, respectively.\u003c/p\u003e\u003cp\u003eFinally, three‑dimensional probability maps projected model outputs onto patient‑specific atrial shells (Supplemental Fig.\u0026nbsp;1\u0026ndash;3). In the independent right‑atrial example, 33 of 65 validated GP sites (51%) resided within regions scoring\u0026thinsp;\u0026ge;\u0026thinsp;0.70, while left‑atrial renderings demonstrated anatomically coherent clusters along known autonomic corridors in both left‑anterior‑oblique and postero‑anterior views. These visualizations illustrate how the model integrates spatial continuity with point‑wise probabilistic predictions to highlight clinically actionable \u0026ldquo;hot‑spots.\u0026rdquo;\u003c/p\u003e\u003cp\u003eTo support the interpretability and reliability of the spatial probability maps, we further analyzed the training dynamics and overall development process of the model. As shown in Supplemental Fig.\u0026nbsp;4, training and validation loss curves demonstrated rapid convergence and overall stability, with early stopping applied at epoch 16 to prevent overfitting.\u003c/p\u003e\u003cp\u003eThe full development pipeline\u0026mdash;including data acquisition, preprocessing, model architecture, validation strategy, and external testing\u0026mdash;is summarized in Fig.\u0026nbsp;\u003cspan refid=\"Fig13\" class=\"InternalRef\"\u003e8\u003c/span\u003e. This systematic workflow underscores the robustness of the model and contextualizes the clinical validity of its output visualizations.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eAll trained parameters, normalization scalers, operating thresholds, Grad‑CAM heat‑maps, and complete metric reports have been archived to ensure full reproducibility and to facilitate future integration into real‑time intra‑procedural guidance systems.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eDeep-learning\u0026ndash;based classification of GP regions from intracardiac electrograms (iEGMs) is emerging as a powerful adjunct to autonomic-substrate mapping. Unlike rule-based algorithms\u0026mdash;often confounded by the dynamic, fractionated and patient-specific morphology of electrograms\u0026mdash;convolutional neural networks (CNNs) learn hierarchical spatial-temporal features directly from raw waveforms and can, in principle, generalize across a spectrum of clinical presentations.\u003c/p\u003e\u003cp\u003eYet several domain-specific hurdles must be overcome for such models to achieve reliable clinical adoption. Class imbalance is paramount: GP windows typically constitute\u0026thinsp;\u0026lt;\u0026thinsp;5% of any mapping dataset, so naive optimization yields high specificity but unacceptably low sensitivity. Loss re-weighting, class-balanced resampling and focal loss are indispensable to force the network to attend to scarce positives. Subtle and overlapping signal features further complicate learning; in atrial remodeling or fibrosis, both GP and non-GP sites may exhibit fractionation or high-frequency deflections, blurring class boundaries and fostering over-fitting to noise. Inter- and intra-patient variability\u0026mdash;arising from anatomy, autonomic tone, catheter orientation and contact\u0026mdash;requires aggressive data augmentation and regularization to maintain generalizability.\u003c/p\u003e\u003cp\u003eReliable ground truth is equally problematic. Annotation noise and ambiguity result from subjective criteria (vagal reflexes, spectral markers) that can both miss true GP sites and mislabel equivocal regions, diminishing effective signal-to-noise ratio during training. Interpretability and trust present another barrier: even with Grad-CAM visualization, clinicians remain wary of black-box tools, especially if saliency occasionally highlights artefacts or baseline shifts. A purely temporal signal is also an imperfect proxy for a three-dimensional problem; without anatomical context, temporal\u0026ndash;spatial resolution mismatches can undermine localization accuracy. Electrogram quality is further threatened by noise and preprocessing artefacts\u0026mdash;over-filtering may erase genuine autonomic signatures, whereas under-filtering leaves residual noise that degrades feature extraction. Procedural bias is a final concern: models trained at a single center risk memorizing institution-specific artefacts (catheter type, stimulation protocol), limiting transportability. Ultimately, clinical integration demands real-time inference with strict latency and false-alarm constraints; excessive false positives prolong procedures, whereas false negatives risk incomplete denervation.\u003c/p\u003e\u003cp\u003eOur network architecture addresses many of these issues. Stacked 1D convolutions treat each 150-ms iEGM segment as a time-series image; kernels learn autonomic motifs such as high-frequency bursts or sharp biphasic deflections. Convolutional feature maps are bias-corrected for baseline shifts, ReLU-activated to suppress noise, and down-sampled with max pooling to curb over-fitting and computational load. Global-max pooling condenses each map to its most informative scalar; fully connected layers then assign class probabilities via a sigmoid output. End-to-end optimization with focal loss concentrates gradient updates on the minority GP class, mitigating severe imbalance, while label stratification, Gaussian noise, gain scaling and temporal shifts bolster robustness. Grad-CAM saliency maps consistently localized attention to the central 140\u0026ndash;160-sample region, mirroring expert recognition of autonomic signatures.\u003c/p\u003e\u003cp\u003eRecent studies corroborate the value of such approaches. Rodrigo et al (22) reported CNNs that distinguish AF from atrial tachycardia (AT) with AUCs of 0.97 (unipolar) and 0.92 (bipolar). Liao et al (23) achieved an AUC of \u0026asymp;\u0026thinsp;0.923 for focal AF-trigger detection, while Chen et al (24) improved local-activation-time annotation to 81% accuracy with a CNN\u0026ndash;LSTM hybrid versus 67% for decision trees. Extending these findings to autonomic mapping, our model maintained a recall of 0.85 despite a GP prevalence of only 2.3% across 18 patients and achieved a weighted F1‑score of 0.79. Precision remained modest (\u0026asymp;\u0026thinsp;0.09 at the balanced‑threshold operating point), but threshold sweeps demonstrated tunable modes\u0026mdash;ranging from high‑sensitivity screening to higher‑pr\u0026eacute;cision confirmation\u0026mdash;allowing operators to trade off false‑positive burden against the risk of missed denervation sites. At the balanced-threshold operating point, the model\u0026rsquo;s precision was approximately 0.09, primarily due to morphological similarities between GP and non-GP sites, particularly in atrial regions affected by remodeling. In the procedural setting, this trade-off is acceptable for initial screening, as false positives mainly prompt additional mapping or stimulation rather than unnecessary ablation, thereby reducing the risk of missed GP targets.\u003c/p\u003e\u003cp\u003eA recent study by Wang et al (25) utilized 46 hand-crafted electrogram features as input to a machine learning model for substrate localization. While this approach achieved promising accuracy, it relied on predefined feature engineering, which can inadvertently omit subtle morphological or temporal patterns not captured by the selected metrics. In contrast, our end-to-end CNN framework ingests minimally pre-processed raw bipolar iEGM segments without manual feature extraction, allowing the network to learn discriminative spatiotemporal representations directly from the signal. This design reduces bias from feature selection, enables discovery of novel electrophysiological patterns, and facilitates generalization across variable signal morphologies.\u003c/p\u003e\u003cp\u003eThe pipeline\u0026rsquo;s modular design\u0026mdash;standardized preprocessing, stratified sampling, threshold calibration and interpretable outputs\u0026mdash;facilitates intra-procedural deployment and regulatory auditing. Archived weights, scalers, prediction logs and saliency maps provide a transparent foundation for multicenter validation.\u003c/p\u003e\u003cp\u003eFuture researches should enlarge datasets across institutions, incorporate three-dimensional catheter-localization data, and explore transformer or hybrid CNN\u0026ndash;transformer architectures capable of modelling long-range dependencies. Prospective trials correlating AI-guided ablation with autonomic endpoints (e.g., vagal reflex elimination) and long-term rhythm outcomes will be essential to confirm real-world benefit.\u003c/p\u003e\u003cp\u003eIn summary, a compact, interpretable CNN can exploit fine-grained temporal structure in bipolar iEGMs to detect autonomic substrates with high sensitivity, offering a viable decision-support tool for targeted CNA. When integrated with conventional mapping strategies, such models have the potential to shorten procedures, enhance lesion precision and improve patient outcomes.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eConflict of interest\u003c/h2\u003e\u003cp\u003eThe authors have no conflicts of interest to disclose.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eEthical approval\u003c/strong\u003e\u003cp\u003e The study was approved by the institutional ethics committee.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eInformed consent\u003c/strong\u003e\u003cp\u003e Informed consent was waived due to retrospective nature of the study.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding information\u003c/h2\u003e\u003cp\u003eThe authors have no funding sources to disclose.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eT.E.G., S.E.\u0026Ouml;., S. B, and S.N.D. wrote the main manuscript text A.S, A.İ.C and A.B.D prepared figures 1-5. H.D.H. prepared remaining figures. T.A and M.C. did trial design and statistical evaluation. All authors reviewed the manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eArmour JA, Murphy DA, Yuan BX, Macdonald S, Hopkins DA. Gross and microscopic anatomy of the human intrinsic cardiac nervous system. Anat Rec 1997;247:289‐98.\u003c/li\u003e\n\u003cli\u003ePauza DH, Skripka V, Pauziene N, Stropus R. Morphology, distribution, and variability of the epicardiac neural ganglionated subplexuses in the human heart. Anat Rec 2000;259:353-82.\u003c/li\u003e\n\u003cli\u003eAksu T, Gopinathannair R, Gupta D, Pauza DH. Intrinsic cardiac autonomic nervous system: What do clinical electrophysiologists need to know about the \u0026quot;heart brain\u0026quot;? J Cardiovasc Electrophysiol 2021;32:1737-47.\u003c/li\u003e\n\u003cli\u003eAjijola OA, Aksu T, Arora R, Biaggioni I, Chen PS, De Ferrari G, Dusi V, Fudim M, Goldberger JJ, Green AL, Herring N, Khalsa SS, Kumar R, Lakatta E, Mehra R, Meyer C, Po S, Stavrakis S, Somers VK, Tan AY, Valderrabano M, Shivkumar K. Clinical neurocardiology: defining the value of neuroscience-based cardiovascular therapeutics - 2024 update. J Physiol. 2025 Mar;603(7):1781-1839.\u003c/li\u003e\n\u003cli\u003eBrignole M, Moya A, de Lange FJ, Deharo JC, Elliott PM, Fanciulli A, et al. 2018 ESC Guidelines for the diagnosis and management of syncope. Eur Heart J 2018;39:1883\u0026ndash;1948.\u003c/li\u003e\n\u003cli\u003eGlikson M, Nielsen JC, Kronborg MB, Michowitz Y, Auricchio A, Barbash IM, Barrab\u0026eacute;s JA, Boriani G, Braunschweig F, Brignole M, Burri H, Coats AJS, Deharo JC, Delgado V, Diller GP, Israel CW, Keren A, Knops RE, Kotecha D, Leclercq C, Merkely B, Starck C, Thyl\u0026eacute;n I, Tolosana JM; ESC Scientific Document Group. 2021 ESC Guidelines on cardiac pacing and cardiac resynchronization therapy. Eur Heart J. 2021 Sep 14;42(35):3427-3520. doi: 10.1093/eurheartj/ehab364.\u003c/li\u003e\n\u003cli\u003eAksu T, Brignole M, Calo L, Debruyne P, Di Biase L, Deharo JC, Fanciulli A, Fedorowski A, Kulakowski P, Morillo C, Moya A, Piotrowski R, Stec S, Sutton R, van Dijk JG, Wichterle D, Tse HF, Yao Y, Sheldon RS, Vaseghi M, Pachon JC, Scanavacca M, Meyer C, Amin R, Gupta D, Magnano M, Malik V, Schauerte P, Shen WK, Acosta JCZ. Cardioneuroablation for the treatment of reflex syncope and functional bradyarrhythmias: A Scientific Statement of the European Heart Rhythm Association (EHRA) of the ESC, the Heart Rhythm Society (HRS), the Asia Pacific Heart Rhythm Society (APHRS) and the Latin American Heart Rhythm Society (LAHRS). Europace. 2024 Aug 3;26(8):euae206. doi: 10.1093/europace/euae206. Erratum in: Europace. 2025 Feb 5;27(2):euaf023. doi: 10.1093/europace/euaf023.\u003c/li\u003e\n\u003cli\u003eAksu T, Piotrowski R, Tung R, De Potter T, Markman TM, du Fay de Lavallaz J, Rekvava R, Alyesh D, Joza JE, Badertscher P, Do DH, Bradfield JS, Upadhyay G, Sood N, Sharma PS, Guler TE, Gul EE, Kumar V, Koektuerk B, Dal Forno ARJ, Woods CE, Rav-Acha M, Valeriano C, Enriquez A, Sundaram S, Glikson M, d\u0026apos;Avila A, Shivkumar K, Kulakowski P, Huang HD. Procedural and Intermediate-term Results of the Electroanatomical-guided Cardioneuroablation for the Treatment of Supra-Hisian Second- or Advanced-degree Atrioventricular Block: the PIRECNA multicentre registry. Europace. 2024 Jul 2;26(7):euae164. doi: 10.1093/europace/euae164.\u003c/li\u003e\n\u003cli\u003eGopinathannair R, Olshansky B, Turagam MK, Gautam S, Futyma P, Akella K, Tanboga HI, Bozyel S, Yalin K, Padmanabhan D, Shenthar J, Lakkireddy D, Aksu T. Permanent pacing versus cardioneuroablation for cardioinhibitory vasovagal syncope. J Interv Card Electrophysiol. 2025 Mar;68(2):203-210. doi: 10.1007/s10840-022-01456-x. \u003c/li\u003e\n\u003cli\u003eVandenberk B, Lei LY, Ballantyne B, Vickers D, Liang Z, Sheldon RS, Chew DS, Aksu T, Raj SR, Morillo CA. Cardioneuroablation for vasovagal syncope: A systematic review and meta-analysis. Heart Rhythm. 2022 Nov;19(11):1804-1812.\u003c/li\u003e\n\u003cli\u003ePachon JC, Pachon EI, Pachon JC, Lobo TJ, Pachon MZ, Vargas RN, Jatene AD. \u0026quot;Cardioneuroablation\u0026quot;--new treatment for neurocardiogenic syncope, functional AV block and sinus dysfunction using catheter RF-ablation. Europace. 2005 Jan;7(1):1-13. doi: 10.1016/j.eupc.2004.10.003.\u003c/li\u003e\n\u003cli\u003eFrancia P, Viveros D, Falasconi G, Penela D, Soto-Iglesias D, Mart\u0026iacute;-Almor J, Alderete J, Saglietto A, Bellido AF, Franco-Oca\u0026ntilde;a P, Zaraket F, Matiello M, Fern\u0026aacute;ndez-Armenta J, San Antonio R, Berruezo A. Clinical impact of aging on outcomes of cardioneuroablation for reflex syncope or functional bradycardia: Results from the cardionEuroabLation: patiEnt selection, imaGe integrAtioN and outComEs-The ELEGANCE multicenter study. Heart Rhythm. 2023 Sep;20(9):1279-1286. doi: 10.1016/j.hrthm.2023.06.007. \u003c/li\u003e\n\u003cli\u003eRivarola EWR, Hachul D, Wu TC, Pisani C, Scarioti VD, Hardy C, Darrieux F, Scanavacca M. Long-Term Outcome of Cardiac Denervation Procedures: The Anatomically Guided Septal Approach. JACC Clin Electrophysiol. 2023 Aug;9(8 Pt 1):1344-1353.\u003c/li\u003e\n\u003cli\u003eDebruyne P, Rossenbacker T, Janssens L, Collienne C, Ector J, Haemers P, et al. Durable Physiological Changes and Decreased Syncope Burden 12 Months After Unifocal Right-Sided Ablation Under Computed Tomographic Guidance in Patients With Neurally Mediated Syncope or Functional Sinus Node Dysfunction. Circ Arrhythm Electrophysiol 2021; 14:e009747.\u003c/li\u003e\n\u003cli\u003eWileczek A, Stodolkiewicz-Nowarska E, Reichert A, Kustron A, Sledz J, Biernikiewicz W, Orlik B, Lipka M, Kutarski A, Hering D, Zając M, Stec S. Reevaluation of indications for permanent pacemaker implantation after cardioneuroablation. Kardiol Pol. 2023;81(12):1272-1275.\u003c/li\u003e\n\u003cli\u003eOnder SE, Guler TE, Bozyel S, Cagdas M, Dalgic SN, Sipal A, Gecer S, Kılıc E, Santangeli P, Aksu T. Integration of automated peak frequency annotation with voltage mapping for identifying ventricular tachycardia ablation sites. J Interv Card Electrophysiol. 2025 Nov;68(8):1573-1583. doi: 10.1007/s10840-025-02045-4.\u003c/li\u003e\n\u003cli\u003eAksu T, Skeete JR, Huang HH. Ganglionic Plexus Ablation: A Step-by-step Guide for Electrophysiologists and Review of Modalities for Neuromodulation for the Management of Atrial Fibrillation. Arrhythm Electrophysiol Rev. 2023 Jan;12:e02. doi: 10.15420/aer.2022.37.\u003c/li\u003e\n\u003cli\u003eKim MY, Coyle C, Tomlinson DR, Sikkel MB, Sohaib A, Luther V, Leong KM, Malcolme-Lawes L, Low B, Sandler B, Lim E, Todd M, Fudge M, Wright IJ, Koa-Wing M, Ng FS, Qureshi NA, Whinnett ZI, Peters NS, Newcomb D, Wood C, Dhillon G, Hunter RJ, Lim PB, Linton NWF, Kanagaratnam P. Ectopy-triggering ganglionated plexuses ablation to prevent atrial fibrillation: GANGLIA-AF study. Heart Rhythm. 2022 Apr;19(4):516-524. doi: 10.1016/j.hrthm.2021.12.010.\u003c/li\u003e\n\u003cli\u003eCai W, Chen Y, Guo J, Han B, Shi Y, Ji L, Wang J, Zhang G, Luo J. Accurate detection of atrial fibrillation from 12-lead ECG using deep neural network. Comput Biol Med. 2020;116:103378.\u003c/li\u003e\n\u003cli\u003eHuang M-L, Wu Y-S. Classification of atrial fibrillation and normal sinus rhythm based on convolutional neural network. Biomed Eng Lett. 2020;10(2):183\u0026ndash;93.\u003c/li\u003e\n\u003cli\u003eTutuko B, Nurmaini S, Tondas AE, Rachmatullah MN, Darmawahyuni A, Esafri R, Firdaus F, Sapitri AI. AFibNet: an implementation of atrial fibrillation detection with convolutional neural network. BMC Med Inform Decis Mak. 2021 Jul 14;21(1):216.\u003c/li\u003e\n\u003cli\u003eRodrigo M, Alhusseini MI, Rogers AJ, Krittanawong C, Thakur S, Feng R, Ganesan P, Narayan SM. Atrial fibrillation signatures on intracardiac electrograms identified by deep learning. Comput Bill Med. 2022; 145: 105451.\u003c/li\u003e\n\u003cli\u003eLiao S, Ragot D, Nayyar S, Suszko A, Zhang Z, Wang B, Chauhan V. Deep Learning Classification of Unipolar Electrograms in Human Atrial Fibrillation: Application in Focal Source Mapping. Front Physiol. 2021 Jul 30:12:704122.\u003c/li\u003e\n\u003cli\u003eChen X, Cheng Z, Wang S, Lu G, Xv G, Liu Q, Zhu X. Atrial fibrillation detection based on multi-feature extraction and convolutional neural network for processing ECG signals. Comput Bill Med. 2021; 202: 106009.\u003c/li\u003e\n\u003cli\u003eWang X, Dennis A, Hesselkilde EM, et al. Machine learning approach for automated localization of ventricular tachycardia ablation targets from substrate maps: development and validation in a porcine model. Eur Heart J Digit Health. 2025;6(6):645-655.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"ganglionated plexus, intracardiac electrogram, cardioneuroablation, convolutional neural network, class imbalance, Grad‑CAM interpretability, autonomic mapping, deep learning","lastPublishedDoi":"10.21203/rs.3.rs-7974349/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7974349/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003ePrecise localization of ganglionated plexi (GP) is critical for effective cardioneuroablation, yet current mapping relies on labour‑intensive stimulation and subjective electrogram (EGM) interpretation. Recent advancements in deep learning (DL) have shown the potential to automate and improve outcomes an atrial fibrillation by analyzing EGMs. We aimed to apply DL to raw bipolar EGMs in order to automate GP detection.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eA total of 189 760 bipolar windows (18 left‑atrium and 15 right‑atrium maps, respectively) were collected from 18 patients. GP annotation was performed independently by two experienced electrophysiologists. Five atrial maps from three patients were withheld for external testing; the remaining 15 patients yielded 119 222 clean windows for model development (GP prevalence\u0026thinsp;\u0026asymp;\u0026thinsp;3.5%). A lightweight one‑dimensional convolutional neural network (CNN) was implemented using PyTorch. Training used focal loss (α\u0026thinsp;=\u0026thinsp;0.75, γ\u0026thinsp;=\u0026thinsp;2.0) and class‑balanced sampling. Performance was assessed with ROC/PR curves, threshold sweeps and gradient‑weighted class activation mapping (GCAM) saliency mapping.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eOn the validation set the model achieved 69.6% accuracy; GP precision, recall and F1‑score were 0.09, 0.85 and 0.17, respectively. External testing on 34 976 unseen windows produced ROC‑AUC\u0026thinsp;=\u0026thinsp;0.870 and PR‑AUC\u0026thinsp;=\u0026thinsp;0.349. A probability threshold of 0.70 captured 51% of reference GP sites while highlighting anatomically plausible \u0026ldquo;hot‑spots\u0026rdquo; (513/9 063 nodes). GCAM consistently focused on central waveform segments (indices 140\u0026ndash;160), aligning with fractionated autonomic signatures and reinforcing model interpretability.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eThe proposed explainable one‑dimensional CNN detects GP substrates with high sensitivity despite pronounced class imbalance and generalizes to unseen atria. Its probability maps and saliency outputs provide intuitive visual guidance, supporting real‑time, physiology‑aware decision making in cardioneuroablation.\u003c/p\u003e","manuscriptTitle":"Convolutional Neural Network for Real‑Time Localization of Ganglionated Plexi from Bipolar Intracardiac Electrograms","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-11-14 10:09:05","doi":"10.21203/rs.3.rs-7974349/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"d6657ca1-430e-45d2-892b-5ac371821bf9","owner":[],"postedDate":"November 14th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-30T16:23:26+00:00","versionOfRecord":{"articleIdentity":"rs-7974349","link":"https://doi.org/10.1007/s10840-026-02307-9","journal":{"identity":"journal-of-interventional-cardiac-electrophysiology","isVorOnly":false,"title":"Journal of Interventional Cardiac Electrophysiology"},"publishedOn":"2026-03-28 16:12:25","publishedOnDateReadable":"March 28th, 2026"},"versionCreatedAt":"2025-11-14 10:09:05","video":"","vorDoi":"10.1007/s10840-026-02307-9","vorDoiUrl":"https://doi.org/10.1007/s10840-026-02307-9","workflowStages":[]},"version":"v1","identity":"rs-7974349","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7974349","identity":"rs-7974349","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}