Artificial Intelligence Techniques in Cardiac Neuromodulation: Mechanisms, Applications, and Pathways to Clinical Translation

ABSTRACT

Cardiac neuromodulation includes various methods, such as vagus nerve stimulation, baroreflex activation therapy, renal denervation, and stellate ganglion intervention, and targets the autonomic imbalance contributing to the pathophysiology of many cardiovascular diseases. Despite promising mechanistic evidence, several landmark trials, including INOVATE‐HF, NECTAR‐HF, and SYMPLICITY HTN‐3, did not meet their primary clinical outcomes, with substantial numbers of non‐responders observed across therapies. Variation in patient response is attributed to several unresolved issues, including insufficient stimulation dosing, off‐target or non‐selective fiber activation, and differences in autonomic phenotypes between patients. Both problems highlight the need for individualized approaches to patient selection, therapy delivery, and monitoring. Artificial intelligence (AI) offers tools to address these problems. In this narrative review, we describe seven families of AI techniques relevant to cardiac neuromodulation: supervised machine learning, deep learning, representation learning, reinforcement learning, multimodal fusion, digital twins with physics‐informed AI, and explainable AI with federated learning. For each family, we summarize how the method works, the cardiac neuromodulation problem it addresses, and the available evidence in the field of cardiac electrophysiology. We then map these techniques to the three core problems of patient selection, real‐time stimulation control, and longitudinal response monitoring. The strongest evidence to date supports representation learning for VNS responder identification, reinforcement learning for closed‐loop VNS control, and digital twins for in silico testing of stimulation protocols. The opportunity for the field is to translate these methods, most of which were developed in adjacent fields, into prospective cardiac neuromodulation trials.

Artificial intelligence could improve patient selection, stimulation control, and response monitoring in cardiac neuromodulation, yet direct clinical evidence remains scarce. We map available techniques onto these three challenges and argue that safe, adaptive closed‐loop control and established therapy efficacy are prerequisites for translation.

1. Introduction

The autonomic nervous system controls heart rate, myocardial contractility, vascular tone, and conduction, and an imbalance in this system is thought to play an etiological role in most prevalent cardiovascular diseases [1]. Cardiac neuromodulation describes a group of device‐based therapies that target this imbalance directly, including cervical and transcutaneous vagus nerve stimulation (cVNS, tVNS), baroreflex activation therapy (BAT), renal denervation (RDN), and stellate ganglion block (SGB) [2, 3]. Despite having encouraging mechanistic data, various crucial trials such as INOVATE‐HF, NECTAR‐HF, and SYMPLICITY HTN‐3 missed their primary endpoints, and a high proportion of non‐responders has been observed across therapies [4, 5, 6]. In post hoc analysis of INOVATE‐HF, only about 30% of patients met the responder criterion [7]. This response heterogeneity is most often attributed to under‐dosing of stimulation or to autonomic phenotype heterogeneity within enrolled cohorts, both of which point to the need for individualized decisions about treatment, delivery, and monitoring [7, 8, 9]. Other mechanisms may also contribute. Cervical vagus stimulation is not fully selective and can activate afferent fibers. This afferent activation may raise sympathetic tone and offset the intended vagal effect. The timing of stimulation within the disease course may matter as well. These factors are candidate predictors, not just confounders. Fiber selectivity, reflex responses, and intervention timing could be encoded as input features in responder and risk models.

Artificial intelligence (AI) provides tools to address this need. Machine learning (ML) and deep learning (DL) extract patterns from high‐dimensional and multimodal data without predefined feature sets [10]. Reinforcement learning (RL) supports closed‐loop control by adjusting stimulation in response to physiological feedback [11]. Representation learning compresses raw signals into interpretable autonomic phenotypes [12]. Digital twin and physics‐informed approaches embed mechanistic knowledge of cardiac autonomic physiology within data‐driven models [13, 14]. Each of these techniques has been demonstrated in cardiovascular medicine, and several have early proof‐of‐concept applications in cardiac neuromodulation [15, 16]. Existing reviews address AI in cardiology broadly [17], AI in cardiac electrophysiology [18], or AI‐guided neuromodulation in heart failure as a single clinical area [19]. A recent review by Varzideh et al. addressed AI for decoding brain–heart network interactions, focused on diagnostic and risk‐prediction applications such as arrhythmia detection, stroke‐related autonomic dysfunction, and stress phenotyping [20]. The application of AI to cardiac neuromodulation as a therapeutic discipline, including stimulation control, dose adaptation, and response monitoring, has not been reviewed as a therapeutic discipline.

Three features of cardiac neuromodulation make it particularly well suited to AI methods. First, the data that define an autonomic phenotype are high‐dimensional and multimodal. A single patient considered for vagus nerve stimulation or renal denervation generates a 12‐lead ECG, ambulatory HRV indices, echocardiographic strain, cardiac magnetic resonance, ambulatory blood pressure, and increasingly omics‐derived biomarkers. Traditional clinical frameworks struggle to integrate the vast and diverse datasets generated in cardiac neuromodulation [9], whereas AI methods are specifically designed to recognize patterns across such complex information. Second, wearable and implanted devices continuously monitor the autonomic state and generate massive real‐time data streams. As a result, treatments cannot depend only on infrequent clinic reviews and may require more adaptive, real‐time decision systems [11]. Third, cardiac neuromodulation trials often involve relatively small patient populations [4, 5, 6], which makes self‐supervised pretraining [21], transfer from adjacent fields such as vagus nerve stimulation for epilepsy [12], federated learning [22], and physics‐informed modeling [13] particularly attractive, because each is designed to extract value from limited labeled data.

In this narrative review, we describe seven families of AI techniques relevant to cardiac neuromodulation. For each family, we explain how the method works, the type of problem it can address, and the available evidence in cardiac neuromodulation or closely adjacent fields. We then map these techniques to the core problems of patient selection, real‐time stimulation control, and longitudinal response monitoring, and close with the limitations of the current evidence base and the priorities for clinical translation.

2. Methods

We searched PubMed, IEEE Xplore, arXiv, and ClinicalTrials.gov from January 2015 to May 2026. We included peer‐reviewed studies and recent preprints reporting development or application of an AI method to a cardiac neuromodulation problem, or to a closely adjacent problem in cardiology or autonomic neuroscience. Studies were prioritized on the basis of methodological relevance and direct applicability to cardiac neuromodulation.

3. Artificial Intelligence Techniques Relevant to Cardiac Neuromodulation

3.1. Supervised Machine Learning

Supervised machine learning trains an algorithm on input–output pairs, so the model learns to map clinical, imaging, electrophysiologic, or procedural features to a defined outcome [18]. The family includes least‐squares and regularized regression, support vector machines (SVMs), random forests, gradient‐boosted trees, and shallow neural networks [18, 23]. In cardiology, supervised models predict atrial fibrillation recurrence after ablation [24], estimate ablation lesion size and ablation index targets from procedural parameters [25], and stratify sudden cardiac death risk in dilated cardiomyopathy using cardiovascular coupling between heart rate and blood pressure series [26]. Random forests and SVMs have predicted response and mortality after cardiac resynchronization therapy more accurately than guideline‐based selection [27, 28, 29].

However, direct evidence in cardiac neuromodulation is limited. In renal denervation, Zhang et al. developed a ridge regression model in 69 patients with resistant hypertension and reported a mean absolute error of 6.40 mmHg for systolic blood pressure change at 6 months [30]. The PREDICT‐RDN trial (NCT06845579) is recruiting 90 patients to develop a multivariable prediction model for blood pressure response [31]. A related feasibility example comes from van Es et al. They used supervised models to predict vagal‐tone HRV metrics in 10 patients with heart failure during daily activity [32]. This study did not involve neuromodulation. It shows that supervised models can estimate autonomic indices from ambulatory wearable data, which is the same prediction step needed before or during stimulation. In vagus nerve stimulation, Fang et al. used HRV‐based supervised classifiers to discriminate likely responders from non‐responders in drug‐resistant epilepsy [33], and Chen et al. used an SVM on EEG functional connectivity features in pediatric pharmacoresistant epilepsy with an AUC of 0.88 [34]. Both serve as methodologic precedents rather than cardiac evidence.

Supervised ML offers a direct route to several unsolved problems in cardiac neuromodulation. The 30% non‐response rate in INOVATE‐HF, the variable blood pressure reduction after renal denervation, and the heterogeneous outcomes of baroreflex activation therapy reflect a selection problem that current criteria cannot resolve [7, 35, 36]. Models trained on baseline phenotype, autonomic biomarkers, imaging, and procedural parameters could identify likely responders before implantation, in the same way HRV and EEG features have predicted VNS response in epilepsy cohorts [33, 34]. Computational modeling of human cervical VNS suggests that supervised models could also help predict the stimulation dose window that activates therapeutic fibers without provoking laryngeal side effects [37]. Outside RDN, no published supervised ML study has directly addressed responder prediction in cardiac VNS, BAT, or stellate ganglion block.

3.2. Deep Learning

Deep learning (DL) is a subfield of machine learning that uses multilayered artificial neural networks [38]. Unlike conventional models, DL learns features directly from raw data such as waveforms, images, and signals [38]. In cardiovascular medicine, convolutional neural networks (CNNs) applied to the 12‐lead electrocardiogram (ECG) have detected left ventricular dysfunction and atrial fibrillation from otherwise normal‐appearing tracings [39, 40]. Similar models have classified echocardiographic views, segmented cardiac magnetic resonance images, and predicted incident atrial fibrillation from ECG and clinical risk factors [41, 42, 43]. These applications show that DL can extend diagnostic reach beyond what trained professionals achieve.

In cardiomodulation, DL has been applied to identify candidates and predict response to device‐based and autonomic interventions. Wouters et al. trained a variational autoencoder on 1.1 million ECGs and applied the resulting model to cardiac resynchronization therapy (CRT) candidates [44]. The deep‐learning model improved outcome prediction beyond QRS area and current guideline criteria [44]. In the autonomic domain, a deep CNN applied to R‐R interval data has distinguished active from sham transcutaneous auricular vagus nerve stimulation (taVNS) [16]. This supports the use of DL for objective monitoring of stimulation effects. Recurrent neural network controllers with long short‐term memory layers have titrated vagus nerve stimulation (VNS) in silico to track target heart rate and blood pressure trajectories [45]. For renal denervation, machine learning models, including emerging DL approaches, have begun to predict blood pressure response from pre‐procedural ECG and ambulatory blood pressure data [46]. In atrial fibrillation (AF) ablation, a CNN trained on patient‐specific atrial tissue models has predicted optimal ablation strategy [47]. These early reports suggest that DL can guide neuromodulatory therapy beyond the population averages on which current selection criteria rest.

Several open problems in cardiomodulation may benefit from DL. Patient selection remains imperfect for VNS, baroreflex activation therapy, and renal denervation, and DL applied to multimodal data may improve responder prediction [48]. Real‐time decoding of autonomic state from wearable signals could enable closed‐loop devices that titrate stimulation to physiologic demand [11]. DL applied to high‐resolution microscopy has begun to map the microscopic anatomy of the human vagus nerve, supporting computational models of neuromodulation therapy [49]. Finally, generative and computational models may allow in silico testing of stimulation protocols before clinical use, reducing risk and accelerating translation [47].

3.3. Representation Learning and Unsupervised Phenotyping

Representation learning is a class of methods in which a neural network learns its own internal description of the data. Instead of relying on hand‐crafted features, the model is trained to compress a raw input into a low‐dimensional latent space and then to reconstruct or contrast it [50]. Because the training requires no labels, large unlabeled datasets can be used. The resulting representation captures structural features that may be invisible to standard measurements such as QRS duration or RR interval. Common architectures include autoencoders, variational autoencoders (VAEs), and contrastive self‐supervised models [50]. In cardiology, VAEs applied to UK Biobank ECGs have yielded latent spaces that recapitulate conventional parameters and reveal 170 genetic loci not previously linked to electrocardiographic phenotypes [21].

In cardiac neuromodulation, representation learning is just beginning to inform patient selection. Suresh and colleagues trained a deep representation learning model on preoperative T1‐weighted MRI and outperformed clinical predictors in identifying responders to vagus nerve stimulation (VNS) [12]. Although developed in pediatric epilepsy, the framework offers a transferable approach to cardiac VNS, where responder identification remains a central problem. The transfer is not straightforward. Epilepsy VNS acts mainly through vagal afferents to the brain, whereas cardiac VNS also depends on cardiac remodeling, myocardial fibrosis, and the fascicular anatomy of the cervical vagus nerve. These differences must be addressed before an imaging‐based responder model can be applied to the heart. In the future, similar approaches applied to multimodal data could surface autonomic phenotypes that current selection criteria miss. Inputs may include ECG, heart rate variability, neural electrograms, and cardiac imaging. Self‐supervised pretraining on large unlabeled ECG and wearable datasets could also yield foundation models that adapt to small cardiomodulation cohorts, where labeled data is scarce.

3.4. Reinforcement Learning and Closed‐Loop Control

Reinforcement learning (RL) is a class of methods in which an agent learns a control policy by interacting with an environment and receiving rewards for desirable outcomes [51]. Unlike supervised learning, RL requires no labeled examples; the agent discovers good actions through trial and error against a feedback signal. In cardiology, RL has been proposed for adaptive programming of cardiac implantable electronic devices, where policies are learned to tune pacing parameters in response to physiologic state [52]. The framework is well suited to settings in which therapy must be adjusted continuously and the optimal action depends on patient‐specific physiology.

In cardiac neuromodulation, RL has been applied most directly to closed‐loop vagus nerve stimulation (VNS). Sarikhani and colleagues used RL to automate the design of closed‐loop VNS control systems in a computational study of cardiovascular regulation [11]. The agent learned to titrate stimulation parameters to track target heart rate and blood pressure trajectories in silico, without manual tuning. Earlier work by Branen and colleagues used long short‐term memory networks for predictive control of VNS in a similar setting [45]. These reports show that RL‐based controllers can adapt VNS in a way that fixed‐parameter open‐loop devices cannot. However, both studies remain entirely in silico; no closed‐loop RL controller has been tested in a cardiac patient.

Several control problems in cardiac neuromodulation align well with the RL framework. Stimulation parameters that work in one patient often fail in another, and an RL agent could learn patient‐specific policies from continuous physiologic feedback. Autonomic state also drifts over time, and a closed‐loop RL controller could re‐tune stimulation as the underlying disease progresses. A distinct problem is that the response to a fixed stimulation protocol is itself unstable over time. Repeated or continuous stimulation can produce habituation or compensatory autonomic adjustment, so identical parameters may yield a weaker or altered response on later delivery. This makes the reward signal non‐stationary. An RL controller must therefore track response against a drifting baseline rather than a fixed target, and separate genuine therapeutic effect from adaptation. Drift also arises on the device side. Electrode encapsulation, tissue remodeling, and changes in stimulation efficiency can alter the response to a fixed protocol over months to years. A controller that monitors the input–output relationship could in principle detect this decay and flag or compensate for it, in the same way it adapts to biological change. Safe exploration remains the central concern, since trial‐and‐error learning in a patient is not acceptable. Constrained or offline RL methods will be needed before clinical translation [52]. Safety is a particular concern in the cardiac setting, since autonomic stimulation can itself be proarrhythmic: excessive vagal drive can provoke bradycardia, sinus pauses, or atrioventricular block, while sympathetic modulation can shorten refractoriness and favor tachyarrhythmia. Any RL or digital‐twin controller acting in real time therefore requires hard safety constraints rather than learned caution alone, including a bounded action space that caps stimulation intensity and rate, watchdog monitoring that halts stimulation when a bradyarrhythmia or hemodynamic compromise is detected, and a fail‐safe default to a previously validated fixed program when a fault occurs. A digital twin can further screen a candidate action for proarrhythmic risk before it is delivered to the patient [13].

3.5. Multimodal Fusion Models

Multimodal fusion models combine inputs from different data sources into a single neural network that learns a joint representation [53]. Common inputs in cardiology include the 12‐lead ECG, cardiac imaging, electronic health records (EHR), polygenic risk scores (PRS), and wearable signals. Fusion can be implemented at the input (early fusion), at an intermediate representation layer, or by combining modality‐specific predictions (late fusion) [54]. In cardiology, these models have most clearly improved coronary artery disease (CAD) risk prediction. A recent synthesis of 39 empirical studies reported a median 6.4% improvement in predictive accuracy over the best single‐modality baselines [54]. Pezel and colleagues combined coronary CT angiography, cardiac magnetic resonance, ECG, and clinical variables, achieving an externally validated AUC of 0.86 for cardiovascular events in obstructive CAD [55].

In cardiomodulation, multimodal fusion has been applied to two complementary problems: identifying responders to device‐based therapy and detecting target engagement during autonomic stimulation. Puyol‐Antón et al. developed the first multimodal deep learning model for cardiac resynchronization therapy (CRT) response prediction [56]. The model combined segmentations of 2D echocardiography and cardiac magnetic resonance (CMR) through a joint latent space and achieved 77% accuracy in classifying responders. In autonomic neuromodulation, Gurel and colleagues fused features from the ECG and the photoplethysmogram (PPG) to detect target engagement of transcutaneous cervical vagus nerve stimulation (tcVNS). Their model used 88 fused ECG and PPG features and outperformed unimodal classifiers under traumatic stress triggers. These early applications show that multimodal fusion can improve both responder prediction and stimulation monitoring for cardiomodulation therapies.

Several challenges in cardiac neuromodulation align with what multimodal fusion is designed to address. Responder identification for vagus nerve stimulation, baroreflex activation therapy, and renal denervation likely depends on signals spanning the ECG, autonomic indices, cardiac imaging, and circulating biomarkers. No single modality is sufficient on its own. Multimodal models trained on these combined inputs could yield more accurate selection criteria than current scoring systems. In closed‐loop settings, fusion of ECG and wearable autonomic signals could give a controller a richer state representation, supporting more stable stimulation. A recent framework for VNS evaluation has called for AI‐based multimodal fusion of imaging, electrophysiological, and behavioral measures to support precision therapy [57]. These applications remain to be tested in cardiac neuromodulation cohorts.

3.6. Digital Twins and Physics‐Informed AI

A digital twin is a computational replica of a patient or organ that integrates mechanistic physiological models with patient‐specific data to enable in silico simulation [58]. In cardiology, digital twins of the heart have been used to plan ablation in atrial fibrillation, predict ventricular tachycardia inducibility, and personalize device therapy [59]. Physics‐informed neural networks (PINNs) are a related class of models that constrain a deep learning system to obey governing physical equations, allowing the network to interpolate from sparse data while preserving physiological plausibility [60]. In cardiovascular hemodynamics, physics‐informed approaches have been used to infer pressure, flow, and wall shear stress from imaging data that would otherwise require invasive measurement [61]. Together, digital twins and physics‐informed AI bridge mechanistic modeling and data‐driven inference within the same framework.

In cardiac neuromodulation, digital twins have been developed to capture the connection between the autonomic nervous system and the heart. Yang and colleagues built the first multiscale neurocardiac digital twin under the NIH SPARC program [13]. The model linked atomistic simulations of β‐adrenergic receptor–noradrenaline binding to integrate‐and‐fire models of the sympathetic and parasympathetic networks, single‐cell electrophysiology of the sinoatrial node and ventricular myocardium, and tissue‐level pseudo‐electrocardiograms [13]. The twin reproduced experimentally measured firing patterns and predicted that parasympathetic stimulation could suppress sympathetically driven triggered activity in both healthy and heart‐failure tissue [13]. A complementary line of work has built virtual brain twins for epilepsy that personalize stimulation strategies, and the same conceptual approach is now being explored for neurostimulation targets that modulate cardiac function [62]. These models show how digital twins can serve as a sandbox to test stimulation protocols before applying them to patients.

Several emerging directions could expand the role of digital twins in cardiac neuromodulation. Patient‐specific calibration of twin parameters from non‐invasive data, for example from echocardiogram videos or wearable ECG, using physics‐informed self‐supervised learning could remove the need for invasive measurement [63]. Generative AI applied to digital twins could enable in silico trials of novel stimulation regimens, exploring parameter spaces too large or too risky to test clinically [64]. Closed‐loop systems that pair a twin with a real‐time stimulator could allow controllers to plan stimulation by querying the twin and selecting actions predicted to restore autonomic balance. These applications remain early, and their translation will depend on careful validation against patient outcomes.

3.7. Explainable AI and Federated Learning

Explainable AI (XAI) refers to techniques that make machine learning predictions interpretable to clinicians [65]. Common methods attribute the output to specific input features or regions using SHapley Additive exPlanations (SHAP), gradient‐based saliency, or attention‐weight visualization [65]. Federated learning (FL) is a complementary framework in which model training is distributed across institutions; only model updates, not raw patient data, are exchanged [66]. This preserves privacy while allowing models to learn from multi‐institutional cohorts. In cardiology, XAI and FL have been applied most actively to heart disease prediction from the 12‐lead ECG. Sorayaie Azar and colleagues built a federated random forest with SHAP attribution for coronary artery disease prediction, achieving 83% accuracy while preserving privacy through k‐anonymity [67]. Zeleke and Bochicchio combined federated training with an attention‐based temporal convolutional network for arrhythmia detection, reaching 93.5% balanced accuracy across decentralized clients [68]. A recent systematic review identified explainability and federated training as central methodological gaps for clinical translation of cardiovascular AI [69].

In cardiac neuromodulation, XAI and FL remain at an early stage, with no published study applying either framework specifically to autonomic neuromodulation cohorts. Patient numbers in this field are small and outcome heterogeneity is high, conditions that make federated and explainable methods particularly well suited. Federated training could enable multi‐center pooling of vagus nerve stimulation, baroreflex activation therapy, and renal denervation cohorts without sharing raw data [22]. This would address the small‐sample limitation that currently slows responder prediction. Explainable AI could expose which physiological features (autonomic indices, ECG morphology, imaging biomarkers) drive a model's response prediction, supporting clinician trust in stimulation decisions. Attention‐based architectures, already validated for arrhythmia detection [68], could be applied to HRV and neural electrogram data during stimulation. They could both classify response and identify the signal components most informative to that decision. These directions remain to be tested in cardiac neuromodulation trials.

Table 1 summarizes representative studies applying these AI methods to cardiac neuromodulation, together with their interventions, datasets, input signals, and reported performance, and highlights the technique families for which no such study has yet been published.

TABLE 1.

Studies applying AI methods to cardiac neuromodulation.

AI family	Study (ref)	Intervention	Cohort/dataset	Input data	AI method	Primary outcome	Performance
Supervised ML	Zhang [30]	Renal denervation	69 patients, resistant HTN	Clinical features	Ridge regression	6‐month SBP change	MAE 6.40 mmHg
Deep learning	Tarasenko [16]	taVNS	Healthy volunteers	R‐R intervals	Deep CNN	Active versus sham stimulation	Significant separation
Deep learning	Branen [45]	VNS (cardiovascular)	In silico	Cardiovascular signals	LSTM controller	HR/BP tracking	Trajectory tracking achieved
Representation learning	No published study in cardiac neuromodulation	—	—	—	—	—	—
Reinforcement learning	Sarikhani [11]	VNS (cardiovascular)	In silico cardiovascular model	Cardiovascular signals	RL controller	Closed‐loop HR/BP control	Automated parameter design
Multimodal fusion	Gurel [70]	tcVNS	Healthy + traumatic‐stress paradigm	ECG + PPG (88 features)	Multimodal classifier	Target engagement detection	Outperformed unimodal classifiers
Digital twin/PINN	Yang [13]	Autonomic neuromodulation of the heart	In silico, multiscale	Atomic → tissue scales	Multiscale neurocardiac digital twin	SNS/PNS effect prediction on SAN and ventricle	Reproduced experimental data
Explainable AI/Federated learning	No published study in cardiac neuromodulation	—	—	—	—	—	—

Abbreviations: BP, blood pressure; CNN, convolutional neural network; ECG, electrocardiogram; HR, heart rate; HTN, hypertension; LSTM, long short‐term memory; MAE, mean absolute error; PINN, physics‐informed neural network; PNS, parasympathetic nervous system; PPG, photoplethysmography; RL, reinforcement learning; SAN, sinoatrial node; SBP, systolic blood pressure; SNS, sympathetic nervous system; taVNS, transcutaneous auricular vagus nerve stimulation; tcVNS, transcutaneous cervical vagus nerve stimulation; VNS, vagus nerve stimulation.

4. How AI Techniques Map to Cardiac Neuromodulation Problems

The three clinical problems that recur across vagus nerve stimulation, baroreflex activation therapy, renal denervation, and stellate ganglion intervention are patient selection, real‐time stimulation control, and longitudinal response monitoring. Each AI family aligns with one or more of these problems, and the alignment is not arbitrary. Figure 1 summarizes this mapping.

FIGURE 1.

Conceptual roadmap of AI‐enabled cardiac neuromodulation for precision cardiovascular therapy. The left panel illustrates the cardiac autonomic network and major neuromodulation targets, including vagus nerve stimulation (VNS), baroreflex activation therapy (BAT), stellate ganglion block (SGB), and renal denervation (RDN), with sympathetic and parasympathetic afferent/efferent pathways. The middle panel summarizes multimodal physiological data streams integrated into the AI framework, while the right panel depicts machine learning–based approaches for patient selection, adaptive stimulation control, response monitoring.

4.1. Patient Selection

Selection is fundamentally a classification problem on multimodal baseline data: who is likely to benefit, and at what cost. Supervised machine learning has the most direct evidence here, with Feeny [28], Tokodi [29], and Kalscheur [27] showing that random forests and gradient‐boosted trees outperform guideline criteria for CRT response. In autonomic neuromodulation, Zhang's ridge regression for renal denervation [30] and Fang's HRV‐based classifier for VNS in epilepsy [33] serve as proofs of concept. Representation learning extends this approach to settings where the relevant features are not known in advance; the Suresh imaging model for VNS responders is the marquee example [12]. Multimodal fusion is the natural next step because no single modality predicts autonomic response on its own.

4.2. Real‐Time Stimulation Control

Control is a sequential decision problem with feedback. Open‐loop devices fail in this setting because the optimal action depends on the patient's current state. Reinforcement learning is designed for exactly this problem; Sarikhani 2024 [11] and Branen 2022 [45] demonstrate in silico closed‐loop VNS controllers that adapt stimulation to heart rate and blood pressure trajectories. Deep learning on raw signals supplies the controller's state representation because the controller needs to estimate autonomic state from short windows of ECG, HRV, and respiration [16, 43]. Digital twins close the loop by allowing the controller to query a simulated version of the patient before committing to an action, reducing the risk of unsafe exploration [13].

A further challenge is the timing of the response. Vagus nerve stimulation does not change heart rate and blood pressure at the same speed, and each follows its own delay and time constant. A controller that ignores these differences may act too late or too strongly, provoking tachycardia or hypotension. Closed‐loop designs therefore need to model response latency, not only the target values.

4.3. Longitudinal Response Monitoring

Monitoring is a change‐detection problem on time‐series data. Supervised classifiers trained on baseline can drift; representation learning and self‐supervised pretraining give a more stable substrate because they learn what is invariant across time [21, 50]. Multimodal fusion of ECG, HRV, and wearable signals supports detection of slow drifts in autonomic balance that no single signal would reveal [57, 70]. Explainable AI is not optional in this setting, because clinicians need to know what the model is responding to before they will act on a monitoring alert [69]. For instance, a controller that flags rising sympathetic tone should also surface the underlying HRV or ECG features driving the alert, not just the score. This reframes what it means to identify a responder. If a therapy's benefit is unproven and the response to a fixed protocol drifts over time, responder status is neither certain nor fixed. Response is therefore better treated as a trajectory to be tracked than a category to be assigned once. In this view, monitoring is not a final step after selection but a continuous process that selection feeds into.

Two cross‐cutting requirements run through all three problems. The first is data scale. Cardiac neuromodulation trials are small, and labeled multicentre cohorts of CRT scale do not yet exist for VNS, BAT, or RDN. Federated learning addresses this directly by allowing centres to train shared models without exchanging raw data [22, 66]; this matters more for cardiac neuromodulation than for almost any other cardiology subfield. The second is trust. Stimulation decisions carry direct physiological risk, and a model that cannot explain why it predicts response will not be adopted. Explainable AI is therefore a precondition for clinical use rather than an afterthought [65, 69].

The mapping is asymmetric. Patient selection has the most direct cardiac evidence, real‐time control has the most exciting computational work, and monitoring has the least mature literature. That asymmetry is a reasonable signal of where the field's next 5 years of effort should go.

5. Limitations, Ethics, Path Forward

Several limitations of the current evidence base should temper any near‐term clinical claims. AI can optimize neuromodulation therapy only when a real treatment effect exists. Its value therefore depends first on establishing the efficacy of the therapy itself, which is settled for cardiac resynchronization therapy but not yet for VNS, BAT, or RDN [4, 5, 6].

The further limitation is the current evidence base. Most published applications of AI in cardiac neuromodulation involve tens to a few hundred patients, often from a single centre. The Zhang renal‐denervation model used 70 patients [30] and the van Es heart‐failure feasibility study used 10 [32]. By contrast, machine‐learning models for cardiac resynchronization therapy were developed on multicentre cohorts of thousands [27, 28, 29]. Many of the most‐discussed contributions remain in silico. Sarikhani's reinforcement‐learning controller [11] and Yang's neurocardiac digital twin [13] are landmark conceptual advances, but in silico performance does not guarantee clinical performance. Bench‐to‐patient validation through instrumented preclinical models and then small first‐in‐human studies is the missing translational step for almost every method discussed in this review. Federated, multicentre datasets and pragmatic device trials will be needed before these models can be expected to generalize.

Another prominent limitation is Black‐box concerns and clinician trust. Deep learning, representation learning, and reinforcement learning produce outputs that are difficult for clinicians to inspect. Without explainability, models will not pass institutional review for stimulation decisions, and patients will not consent. SHAP attributions, attention maps, and counterfactual explanations should be planned at the model‐design stage rather than added after the fact [65]. Attribution alone may not be enough. SHAP values and attention maps show which inputs a model weighted, but not why those inputs matter or how they act. For a control system that adjusts stimulation, this distinction is important. A model may flag a feature as predictive without revealing the mechanism that links it to the response. True interpretability requires understanding that mechanism, not only the feature ranking. Closing this gap is a harder problem than attribution, and it remains open for control‐based AI in cardiac neuromodulation.

A further practical constraint is signal quality. ECG and HRV data from wearable and implantable devices carry substantial motion artifacts and noise during daily activity, and a controller acting on corrupted input can misread the autonomic state. Robust preprocessing, artifact rejection, and signal quality indices that gate low‐confidence data are therefore prerequisites for any real‐time system, not optional refinements.

A final caveat concerns causal inference. Because direct literature in cardiac neuromodulation is sparse, an observed association between stimulation and a physiological change should not be read as evidence that the modulated neurons directly control the target, since they may act upstream, downstream, or in parallel to the intended mechanism.

Regulation has not yet caught up with adaptive AI in implantable cardiac neuromodulation devices. Closed‐loop controllers, by definition, change behavior over time, and existing device‐approval pathways assume a fixed algorithm. The FDA's predetermined change control plan framework points in the right direction but has not yet been tested on a cardiac neuromodulation device. Regulatory science is now part of the AI translation problem rather than an afterthought to it.

Another major limitation of AI is ethical considerations. Autonomic neuromodulation can affect mood, alertness, and cognition. Closed‐loop devices that change stimulation without the patient's knowledge raise consent questions that do not arise with fixed‐parameter devices. Federated learning addresses privacy at the data level [66] but does not address consent at the patient level.

The near‐term priorities for the field are, in our view, the following: build labeled multicentre cohorts for VNS, BAT, and RDN, federated where direct sharing is not feasible; require explainability as a model‐design constraint rather than a post hoc add‐on; move RL controllers from in silico to instrumented preclinical models with realistic safety constraints; calibrate digital twins to patient‐specific data using physics‐informed self‐supervised methods; and work with regulators on adaptive‐AI frameworks that allow closed‐loop devices to update without losing approval.

6. Conclusion

Cardiac neuromodulation has shown strong mechanistic promise but inconsistent clinical benefit. This gap reflects three key challenges: identifying which patients will respond, determining the optimal dose, and adjusting therapy over time. Current selection criteria, fixed‐parameter devices, and clinic‐based follow‐up are not designed to address these challenges effectively. Each of these problems aligns with a specific family of AI techniques. Supervised and representation learning sharpen patient selection. Reinforcement learning and digital twins enable closed‐loop stimulation. Multimodal fusion supports longitudinal monitoring across signals that no single device captures.

The evidence base in cardiac neuromodulation remains uneven. Renal denervation and CRT have working machine learning models. Vagus nerve stimulation has compelling in silico work in reinforcement learning and digital twins, with limited clinical validation. Baroreflex activation therapy, stellate ganglion block, and cardiac sympathetic interventions have very little AI literature at all. Federated and explainable learning have almost no direct evidence in this field, even though both are particularly well suited to it.

We expect the next phase of progress to come from three sources: multicenter federated cohorts that overcome small‐sample limits; closed‐loop reinforcement learning controllers paired with safety‐focused digital twins; and explainable multimodal models that make response prediction interpretable to clinicians. AI will not replace the clinical reasoning that defines neurocardiology. Instead, it offers tools that are well suited to the field's key challenges. The methodological groundwork already exists. The priority for the next five years is to take three specific steps: [1] launch federated multicenter VNS and RDN cohorts large enough to train responder models, [2] translate the Sarikhani‐style RL controller from in silico to a large‐animal closed‐loop preparation, and [3] develop the first explainable digital twin paired with a clinical stimulator.

Author Contributions

The authors confirm their contribution to the paper as follows: study conception and design: K.P., P.P.; data collection: S.K. and A.P.; analysis and interpretation of results: K.P. and P.P.; draft manuscript: K.P. All authors reviewed and approved the final version of the manuscript.

Funding

The authors have nothing to report.

Consent

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

Data sharing not applicable to this article as no datasets were generated or analyzed during the current study.