This article used a targeted literature search using keywords such as “artificial intelligence”, “cardiovascular stents”, “drug-coated balloons”, “computational modeling”, and “intravascular imaging” through PubMed, Scopus, and Web of Science databases between 2020 and 2025. Peer-reviewed studies reporting original data, validated platforms, or ready-to-implement tools were emphasized. Studies were included based on relevance, originality, and clinical utility to AI applications in DCCDs. Reviews, preclinical studies, and conference proceedings were excluded unless they presented novel, validated models or datasets.

The key contributions of this comprehensive narrative review are as follows: (i) summarizing the clinical development and current limitations of DCCDs, including DESs and DCBs, to demonstrate the need for AI integration, (ii) systematically examining the latest AI models applicable to device design, intravascular imaging, and post-procedural monitoring, (iii) exploring how AI contributes to the development of novel DCCDs, including predictive drug delivery modeling and stent optimization, (iv) discussing the evaluation of the regulatory, technical, and ethical issues involved in translating AI-integrated DCCDs into clinical use, (v) identifying research gaps and suggesting future directions, such as adaptive AI tools for real-time procedural feedback.

The rest of this article is organized as follows: Section 2 provides background information on the current state of drug-coated cardiovascular devices, setting the clinical context for AI applications. Section 3 examines the role of AI in cardiovascular interventions by addressing pre-procedure planning, real-time guidance, and post-procedure monitoring. Section 4 focuses on AI applications in DCCD development, including device design optimization and drug release kinetics modeling. Section 5 discusses the various challenges associated with AI applications, including ethical, economic, and regulatory considerations. Section 6 acknowledges the limitations of this review, and Section 7 concludes with future perspectives and actionable recommendations for advancing the field.

Current DCCD platforms primarily utilize two drug classes: the limus family of compounds (sirolimus, everolimus, zotarolimus), which inhibit mammalian target of rapamycin (mTOR) signaling and prevent smooth muscle cell proliferation, and paclitaxel, which disrupts microtubule dynamics to inhibit cell division and migration [26, 27, 28]. Emerging therapeutic approaches are exploring anti-inflammatory agents such as tacrolimus and corticosteroids to address the inflammatory components of restenosis [29, 30].

Despite significant technological advances, there are several challenges that still require intelligence-driven solutions.

The first challenge is related to optimizing device performance. Current device selection is primarily based on angiographic measurements and operator judgment, potentially missing opportunities for patient-specific optimization based on lesion characteristics, vascular biology, and individual risk factors [8].

Another challenge is risk of ISR, which affects 5–15% of the patients with modern DES. This variability reflects the complex interactions between patient factors (genetic polymorphisms, diabetes, inflammatory status), lesion characteristics (calcification, vessel size, plaque composition), and device characteristics (polymer biocompatibility, drug release kinetics, and support design) [8, 31]. Late stent thrombosis (LST), while rare, remains a significant concern, necessitating prolonged antiplatelet therapy and careful patient evaluation [7]. Achieving optimal drug concentrations in target tissues while minimizing systemic exposure remains challenging, particularly in calcified lesions or complex anatomies where drug penetration may be compromised [6, 32].

Recent innovations have explored dual-drug delivery systems (e.g., paclitaxel-sirolimus combinations), polymer-free coatings, and feedback-controlled pharmacokinetics to enhance vascular healing while minimizing systemic exposure [32, 33, 34]. However, the widespread clinical adoption of these advanced systems is limited by the lack of patient-specific optimization and real-time decision support, creating an area where artificial intelligence has transformative potential.

The complexity of optimizing DCCD performance across diverse patient populations and clinical scenarios creates an ideal environment for AI-driven innovation. The multidimensional nature of cardiovascular interventions, encompassing imaging data, procedural variables, patient characteristics, and device features, aligns with AI’s strengths in pattern recognition, predictive modeling, and decision support.

Modern interventional cardiology generates vast datasets through intravascular imaging [intravascular ultrasound (IVUS) and optical coherence tomography (OCT)], physiological assessments [instantaneous wave free ratio (iFR), fractional flow reserve (FFR)], angiographic analysis, and electronic health records [35]. These rich data sources, combined with evolving computational capabilities, enable AI applications that can improve every aspect of DCCD use, from initial device design to long-term patient follow-up.

The following sections examine how AI technologies are being integrated throughout the DCCD lifecycle to address these persistent challenges and unlock new possibilities for personalized cardiovascular care.

AI encompasses multiple computational approaches that have demonstrated significant potential in cardiovascular medicine [9]. The technical foundation of AI applications in DCCDs rests on several core methodologies, each optimized for specific clinical tasks.

Among AI methodologies, ML and its subfields are central to cardiovascular device applications [36]. Supervised learning approaches, including support vector machines (SVM), random forests, and gradient boosting methods (XGBoost, LightGBM), excel at risk prediction and outcome classification tasks. These algorithms achieve superior performance in predicting in-stent restenosis (area under the curve (AUC) 0.82–0.89) and bleeding complications (AUC 0.76–0.84) compared to traditional clinical scoring systems [32, 37].

DL, a subfield of machine learning based on artificial neural networks, has shown remarkable potential in processing high-dimensional imaging data. Convolutional neural networks (CNNs), particularly U-Net and ResNet architectures, are particularly effective in analyzing IVUS, OCT, and coronary computed tomography angiography (CCTA) and can automate plaque characterization, lumen measurement, and stent sizing [35]. For intravascular imaging, CNN-based segmentation algorithms achieve 92–95% accuracy in vessel boundary detection and 88–92% accuracy in plaque characterization [38, 39]. Recurrent neural networks (RNNs) and Long Short-Term Memory (LSTM) networks provide excellent results in temporal data analysis, enabling real-time monitoring of operational parameters and outcome prediction.

Reinforcement learning (RL), which learns optimal strategies through reward-based feedback, is emerging in interventional cardiology for procedural planning, such as selecting stent type and length based on lesion characteristics and simulated outcomes. Unsupervised learning techniques, including clustering and dimensionality reduction algorithms, are used to discover new patient subtypes or treatment response profiles from electronic health records and registries [40].

Computer vision (CV)-based advanced image processing techniques such as semantic segmentation, object detection (YOLO, R-CNN), and image registration algorithms enable automated analysis of angiographic and intravascular imaging data. These methods provide human-level performance in stenosis detection (sensitivity 94%, specificity 91%) and vessel measurement accuracy within 0.1 mm of expert judgment [35].

Natural language processing (NLP) techniques, including transformer-based models (BERT, GPT variants), facilitate the extraction of clinical insights from unstructured electronic health records, transactional reports, and literature analysis, supporting clinical decision-making and research synthesis [41].

AI systems in cardiovascular interventions are evaluated using established metrics [42, 43]. Important metrics for classification tasks include area AUC, sensitivity, specificity, and positive/negative predictive values. Mean absolute error (MAE), root mean square error (RMSE), and R-squared values are prominent in regression tasks. Apart from these, metrics such as the dice similarity coefficient and Hausdorff distance used for image analysis; C-index, net reclassification improvement (NRI), and integrated discrimination improvement (IDI) for survival analysis used in clinical validation are widely used.

AI technologies are increasingly being adopted in cardiovascular medicine to support diagnostic and procedural decision-making [9]. Broadly, AI simulates human cognitive functions, including learning, reasoning, and problem solving, using various algorithmic approaches. A key subset of AI is ML, which enables systems to learn patterns from data without the need for explicit programming. DL, a branch of ML, uses multilayered neural networks to model complex relationships in big data. DL has demonstrated particularly strong performance in cardiovascular imaging tasks due to its ability to automate feature extraction and improve classification accuracy [44].

NLP is increasingly being used to extract meaningful clinical information from unstructured data such as electronic health records or research literature. CNNs, a type of DL architecture, have proven particularly effective for image-based tasks and are widely used in computer vision, including automated analysis of coronary angiograms and IVUS [38, 39].

These innovations have led to the development of AI-integrated intravascular imaging platforms that support vessel segmentation, lesion detection, and real-time device selection. These tools provide rapid and standardized imaging interpretations, reducing operator variability and improving processing precision. For example, many commercially available AI tools now offer automatic vessel size estimation and lesion characterization, enabling personalized stent selection and optimization of percutaneous coronary intervention (PCI) strategies [10, 11].

Among the leading commercial platforms, the AVVIGO+ Multimodality Guidance System (Boston Scientific) exemplifies the integration of AI into intravascular imaging. It combines high-resolution IVUS with physiological data and uses a machine learning algorithm based on U-Net CNNs for automatic segmentation of the vessel lumen and its boundaries. A distinctive feature of the AI-powered AVVIGO+ system is the ability to automatically determine lesion length, proximal and distal reference segments, and the location of the minimal lumen area in the longitudinal view. This capability can significantly improve accurate and rapid device sizing without manual input. In a study by Matsumura et al. [35], this automation achieved 92.4% agreement with expert-based sizing for balloon selection, highlighting its clinical utility and sensitivity.

Another leading platform, the Ultreon 2.0 software (Version: 2.0, Abbott Vascular, Santa Clara, CA, USA), applies AI to OCT imaging. Ultreon 2.0 provides automatic detection and quantification of calcified plaque, providing key measurements such as total calcium arc, maximum thickness, and plaque length, which are necessary to assess the need for calcium modification before stenting. The system also detects the external elastic lamina (EEL) to assist with stent sizing and landing site selection. By automating these complex assessments, Ultreon 2.0 improves pre- and post-stenting decisions, increases the reproducibility of image interpretation, and reduces reliance on operator expertise [45]. Table 1 summarizes the key features of AI-enabled intravascular imaging software from selected commercially available examples.

Taken together, these AI technologies underpin current and emerging applications in the development and implementation of DCCD. The following sections discuss how these tools support planning, real-time intervention, post-procedure prediction, and device design.

AI technologies have played a significant role in improving real-time decision-making during PCI by assisting operators with dynamic image interpretation, procedural navigation, and device positioning. Modern AI-enabled intravascular imaging platforms support intraprocedural adjustments and optimize procedural outcomes in real time.

The design and material composition of DCCDs significantly affect their performance. AI algorithms are widely used in stent and balloon design, particularly for optimizing structural geometry, mechanical behavior, and drug delivery efficiency. For example, supervised learning models trained on past stent deployment results have achieved over 90% predictive accuracy in predicting strut malposition and mechanical deformation [54]. Gaussian process regression models and Bayesian optimization techniques explore large design spaces to identify optimal configurations that balance mechanical performance, hemodynamic properties, and drug delivery characteristics. These approaches reduce design iteration cycles from months to weeks while simultaneously examining thousands of design variants [63, 64, 65].

ML and DL models are used to predict how design parameters (e.g., support geometry, polymer composition, drug-polymer interactions) affect device behavior. For example, supervised learning algorithms trained on clinical, and materials datasets have been used to optimize scaffold thickness and expansion profiles for novel stents [54, 64]. Samant et al. [66] demonstrated that AI-based modeling significantly improved the mechanical performance and drug distribution profile of everolimus-eluting stents specifically designed for left main coronary interventions. Similarly, Poletti et al. [64] developed an image-based, patient-specific digital twin approach to virtually model and test stent deployment strategies. These tools allow for iterative design testing under physiological conditions without physical prototyping, accelerating the development cycle and reducing costs.

In recent years, in silico studies based on finite element analysis (FEA) have been conducted to consider the geometric and material properties of intravascular devices such as stents and balloons, their different deployment strategies, and the patient’s vascular anatomical characteristics. FEA frameworks, when combined with RL, further support iterative modeling of intravascular mechanics under pulsating flow conditions. These frameworks enable rapid testing of structural changes and device positioning strategies without the need for physical prototyping [54, 64].

AI-driven computational modeling allows the simulation of various design parameters through in silico testing and enables virtual models to predict the physical and biological behavior of devices under various conditions. For example, stent designs can be tested in virtual environments for durability, hemodynamic performance, and drug release properties. This provides significant advantages in optimizing stent designs and accelerating the time from inception to clinical use [63]. RL techniques have also been applied to simulate catheter-based deployment strategies, providing adaptive guidance to minimize misplacement and vessel wall stress [54, 55]. Such simulation environments, when integrated with patient-specific imaging data, form the basis for AI-assisted design processes tailored to individualized anatomy and lesion morphology.

Modeling drug diffusion and pharmacokinetics in complex vascular environments remains a fundamental challenge in DCCD development. By analyzing large datasets of material properties and clinical outcomes, AI can also help select optimal polymer coatings and drug combinations. This approach could significantly contribute to the development of DCCDs with improved biocompatibility and targeted drug delivery capabilities [32, 67].

Although DCCDs are widely used to reduce the rate of restenosis, ISR remains a challenge. The development of new DCCDs requires observation of the pharmacodynamic and pharmacokinetic properties of the loaded drugs, as well as testing for efficacy and safety. This process is time-consuming and costly [22, 68]. In the context of drug kinetic modeling, AI can make a powerful contribution to overcoming these limitations. AI can play an important role in optimizing various aspects, from drug release from the stent or balloon to drug transport in the blood stream and passage through the vascular wall. For example, AI-assisted algorithms can significantly contribute to the design of polymeric biocompatible cardiovascular devices with tunable mechanical and release properties, as well as optimize process parameters for coating technologies to achieve desired thickness, uniformity, and adhesion [69].

AI algorithms have enabled the modeling of complex pharmacokinetic and pharmacodynamic processes. For example, supervised CNNs have been trained on in vitro and animal model data to predict arterial wall drug concentration gradients with an average absolute error of less than 10% [69]. Graph neural networks (GNNs) have demonstrated potential for modeling drug diffusion pathways across non-uniform vascular tissue structures. Gracia’s computational framework modeled stent-based delivery systems using CNN-assisted simulations and validated outputs against physical measurements in porcine artery models [69]. These AI-assisted simulators accurately replicate in vivo behavior, reducing reliance on animal testing.

Beyond passive diffusion, AI models are now contributing to the design of feedback-controlled delivery systems that integrate biosensor inputs (e.g., endothelial response, flow shear stress). These systems can adapt drug release rates in real time using reinforcement learning controllers [33, 34].

AI models also play a key role in the development of patient-specific DCCDs. Vascular properties, such as calcification in the lesion composition, have a significant impact on drug absorption and distribution. This underscores the importance of modeling binding-diffusion kinetics and developing strategies for lesion-specific DCCDs [6]. AI algorithms are also transforming the personalization of antiplatelet therapy by customizing the dosage for each the patient. Traditional fixed dual antiplatelet therapy (DAPT) may not be optimal for all patients in balancing the risks of ischemia and bleeding. In this regard, ML models have been created to tailor the duration of DAPT to individual patient profiles [70]. Specific AI models such as XGBoost have been developed for this purpose [60]. These approaches hold the promise of personalizing not only device configurations but also pharmacotherapy regimens.

Table 2 summarizes the main types of AI models applied at various stages of the DCCD development process, including their core functions, advantages, and limitations.

AI-powered integration of imaging modalities (e.g., OCT, IVUS, CT) into design workflows enables real-time feedback loops between anatomy, device design, and delivery simulations. Advanced segmentation tools based on U-Net and DenseNet architectures achieve Dice similarity coefficients $>$0.9 for vessel and lesion segmentation, enabling precise simulation of device-tissue interactions [35, 53] These imaging algorithm pipelines are essential for reducing geographic mismatch and customizing drug delivery profiles based on lesion morphology and calcification.

In the manufacturing arena, computer vision models, including generative adversarial networks (GANs), have been used to detect defects in polymer coatings and structural alignment. Real-time optical inspection systems have demonstrated over 95% accuracy in identifying microstructural inconsistencies, improving both quality control and manufacturing efficiency [42, 43].

To improve the interpretability of complex workflows, Fig. 2 summarizes the AI-assisted pipeline for DCCD development, including design optimization, pharmacokinetic modeling, and simulation-guided evaluation.

Applications of AI in healthcare, including its integration into cardiovascular interventions, raise several ethical challenges that need to be addressed. One of these challenges is data privacy and security. AI systems often rely on large datasets containing sensitive personal health information. Risks such as data breaches, unauthorized use, and cyberattacks can lead to discrimination or harm [71]. Regulatory bodies, such as the Health Insurance Portability and Accountability Act (HIPAA) in the United States and the General Data Protection Regulation (GDPR) in the European Union, are working to reduce these risks [72, 73]. For example, model inversion is a type of attack that exploits the final output of an AI model. It can reconstruct or extract training data from the outputs of a machine learning model, potentially exposing sensitive information. Attackers use the model’s behavior to “reverse engineer” the data it was trained on, even without direct access to the original dataset. These attacks pose significant privacy risks, especially when dealing with sensitive data like medical records or personal information [74]. Advanced privacy-preserving technologies offer robust solutions to these challenges. Differential privacy, which provides quantifiable privacy ($\epsilon$ $\le$ 1) guarantees for cardiovascular risk models, can offer a mathematical framework for privacy protection [75].

Algorithmic bias and fairness are another important challenge. AI models trained on biased or non-diverse datasets often struggle when applied to populations different from those on which they were trained [76, 77]. Developing AI applications that address specific health needs at the local community or individual level can hinder the creation of robust AI models due to limited data availability or variability. However, a comprehensive bias reduction framework addresses these challenges through a variety of approaches. Bias-detection algorithms, such as fairness-aware machine learning and statistical equity tests, can accurately detect unfair model behavior. Regular fairness audit protocols using standardized fairness metrics ensure ongoing compliance with equity standards.

The third challenge is explainability and transparency. Many AI models, especially those using deep learning, operate as “black boxes”. Their opaque nature hinders trust and accountability. Transparent or federated AI systems and explainable AI (XAI) approaches are encouraged to increase user trust [65, 66]. Physicians must trust and understand algorithmic outputs, especially when considering factors affecting stent selection, drug-coating parameters, or treatment duration. Therefore, explainability should be a non-negotiable requirement for AI applications in interventional cardiology. XAI methods such as SHAP (SHapley Additive Explanations), LIME (Local Interpretable Model-Independent Explanations), and attention heatmaps are increasingly being used to clarify model rationale for individual predictions. These tools help uncover hidden biases and increase user confidence. However, considering the difficulties that explainability brings with it trade-offs in terms of model complexity, latency, and interpretation accuracy, hybrid strategies should not be ignored [39, 73, 78].

The fourth challenge is accountability and responsibility. Defining the roles among developers, clinicians, and institutions in AI-driven clinical decisions remains a challenge. Unfortunately, current legal systems have difficulty assigning responsibility for this issue [17]. Finally, global inequalities and accessibility to services can be addressed. AI tools may not be accessible to the desired level in low-resource environments due to reasons such as cost, infrastructure, or local barriers. Minimizing this requires adapting AI to local health needs and encouraging international regulatory cooperation [39, 79, 80].

Furthermore, traditional informed consent models are inadequate for continuously learning and evolving AI systems, creating ongoing consent challenges [81]. Blockchain-based systems and natural language processing systems could provide significant improvements [82, 83].

Techniques such as federative learning and adversarial bias removal have been proposed to increase data privacy and fairness in AI models for cardiovascular applications [17]. These methods can minimize demographic bias and improve generalizability.

Healthcare applications, including AI-based models, require initial setup, training, infrastructure, and other investments requirements. This not only presents a technical hurdle for AI implementation but also poses significant economic burdens. Economic impact metrics are critical in calculating these costs [84]. Traditional health economics models struggle to capture the dynamic benefits and costs of AI, especially for systems that improve over time [85]. Advanced economic modeling, such as dynamic cost-effectiveness models (e.g., Markov models), value-based pricing, total cost of ownership, and budget impact analysis, can address these challenges [84]. Integrating AI with these strategies has the potential to significantly reduce healthcare costs while improving diagnostic accuracy and treatment efficiency. For example, AI-assisted screening for diabetic retinopathy has been shown to lead to better health outcomes at lower costs than traditional methods [47]. Therefore, it is crucial to address the economic impact of AI adoption in healthcare.

On the other hand, the high initial costs of AI systems and the need for ongoing maintenance and updates pose significant implementation challenges. Furthermore, the lack of standardized methodologies makes it difficult to analyze the economics of AI interventions and understand their value. This challenge is particularly evident in regions with health inequalities [79, 86].

Rigorous validation of AI algorithms is critical for the development and clinical deployment of DCCDs. Validation strategies typically involve multistage testing. These include internal validation (e.g., cross-validation within the training set), external validation using independent datasets to assess generalizability, and prospective validation within clinical workflows [10, 16, 63].

Unlike static software tools, AI systems (especially those using deep learning) learn from large, high-dimensional datasets and can improve over time. Therefore, traditional validation metrics need to be expanded to assess both model performance and robustness under real-world variability. Performance metrics such as accuracy, AUC, sensitivity, specificity, and calibration should be tailored to each use case. Applications for DCCDs include predicting ISR, optimizing drug release kinetics, or customizing dual antiplatelet therapy duration [36, 60, 70].

Digital twin-based simulation frameworks and high-fidelity virtual experiments are emerging as complementary validation methods, reducing reliance on animal studies, and accelerating preclinical evaluation [64, 69]. These systems have the potential to play a critical role in modeling stent performance, drug elution dynamics, and vascular tissue interactions under physiological flow conditions.

It is crucial to establish validated and approved comprehensive regulatory frameworks to govern the application of these technologies. This is necessary to protect patients from misdiagnosis, misuse of personal data, and biases embedded in algorithms [87, 88]. Despite the intense efforts of regulatory bodies in some countries, there is still no global regulatory framework for AI applications in healthcare. Currently, the global regulatory environment primarily regulates the use of AI in healthcare for medical devices, specifically Software as a Medical Device (SaMD) [89]. Regulatory agencies such as the US Food and Drug Administration (FDA), the European Commission, and the UK Medicines and Healthcare products Regulatory Agency (MHRA) are developing frameworks to address the inherent risks of AI/ML-driven systems [18, 19, 88, 90].

The FDA’s Predetermined Change Control Plan (PCCP) framework is particularly important for machine learning-enabled devices, enabling post-approval algorithm updates while maintaining safety and performance assurance [18]. Similarly, the proposed EU Artificial Intelligence Act mandates risk stratification, documentation, and human oversight for high-risk AI applications in healthcare [19]. National organizations such as National institute for Health and Care Excellence (NICE (UK)) and the World Health Organization (WHO) have published evidence standards frameworks for AI tools that emphasize clinical effectiveness, economic value, and patient safety. These are crucial for deploying AI in DCCD personalization, where inaccurate predictions can have disastrous consequences [20, 80, 86].

AI/ML models continuously learn and improve over time, requiring adaptive algorithms and presenting unique regulatory challenges. To address this issue, the FDA introduced the “Proposed Regulatory Framework for Changes to AI/ML-based SaMD” in April 2019. This framework adopts a “total product life cycle (TPLC)” approach, acknowledging that AI/ML-based software frequently evolves as it learns from new data [90]. Regulatory fragmentation across different regions poses significant obstacles to global AI deployment. Harmonization initiatives such as the International Medical Device Regulators Forum (IMDRF) and ISO/IEC Standards (including ISO 14155 and ISO 13485) are addressing these challenges through various international efforts [91, 92].

Ensuring that AI algorithms perform reliably across diverse patient populations and clinical settings is essential. Regulators require robust clinical evidence for AI-based medical devices, but traditional clinical trial designs may be insufficient for adaptive AI systems [74]. Adaptive trials utilize Bayesian designs, allowing for protocol changes based on accumulated evidence and significantly increasing efficiency. Real-world evidence obtained through post-market surveillance using electronic health records and registry data enables continuous safety and efficacy monitoring. Digital biomarkers enable continuous monitoring using wearable devices and smartphone sensors, providing high-frequency outcome data [93, 94, 95].