Artificial intelligence-driven diagnosis of acute thoracic aortic dissection: integrating imaging, biomarkers, and clinical workflows—a narrative review

Introduction

Thoracic aortic dissection (TAD) is a life-threatening medical emergency characterized by a tear in the aortic wall, which can rapidly lead to catastrophic complications such as severe internal bleeding, organ ischemia, or even sudden death if not promptly diagnosed and treated (1). If left untreated, the mortality rate of acute thoracic aortic dissection (ATAD) is 23.7%, with the risk increasing by approximately 0.5% per hour within the first 48 hours of presentation (2). This highlights the critical need for rapid diagnosis and immediate intervention to reduce the high risk of fatal outcomes.

Approximately 80% of ATAD patients present with chest pain in the emergency department (3). However, among all patients experiencing chest pain, only 1.14% are ultimately diagnosed with ATAD (4). Other chest pain-related emergencies, such as acute coronary syndrome (ACS), and acute pulmonary embolism (PE) are far more prevalent, often overshadowing ATAD. This disparity contributes to a high misdiagnosis rate of ATAD, delaying critical treatment and significantly increasing the risk of fatal outcomes. Previous studies have shown initial misdiagnosis rates ranging from 14% to 31% in patients later confirmed to have TAD (5,6).

Diagnosis of ATAD is often achieved by assessing pre-defined biomarkers and utilizing medical imaging modalities in patients with suspected cases. Given the urgent need for immediate intervention in ATAD, the European Society of Cardiology (ESC) has recommended several biomarkers to facilitate earlier confirmation of the correct diagnosis, complementing imaging findings. Imaging modalities, such as computed tomography (CT) and echocardiography, remain the gold standard for definitive diagnosis. However, their diagnostic accuracy can be influenced by variability in imaging interpretation (3).

With advancements in artificial intelligence (AI), several models have been developed to expedite the triage of patients presenting with chest pain upon admission to the emergency department and to reduce the misdiagnosis rate of conditions such as TAD. Various review articles have recently been published with respect to AI applications in managing aortic dissection (AD). In 2022, Mastrodicasa et al. published an article to summarise AI applications in AD imaging in screening, image segmentation and outcome prediction (7). One year after, Farina et al. summarized the use of AI-based tools applied to chest X-rays (CXRs) for diagnosing cardiovascular conditions and predicting outcomes, including AD (8). Similarly, Lum et al. examined radiogenomics studies on genetically triggered TAD, emphasizing the growing use of AI to extract high-throughput features and perform statistical analyses for prognostic modelling (9). Additionally, Lo et al. reviewed the performance and clinical utility of AI-powered models for measuring aortic diameter, discussing their integration into clinical workflows (10).

Despite these contributions, no review article has specifically summarised AI-powered diagnostic models for ATAD that incorporate multiple imaging modalities and biomarkers, with a focus on clinical applicability. This gap underscores the uniqueness of our review, which highlights models that integrate multiple imaging modalities, adopt a multimodal approach within the clinical workflow, and combine imaging data with biomarker analysis to enhance diagnostic precision and improve patient outcomes. We present this article in accordance with the Narrative Review reporting checklist (available at https://atm.amegroups.com/article/view/10.21037/atm-25-82/rc).

Methods

A comprehensive literature search was conducted using PubMed, Scopus, and Web of Science to identify relevant studies published between 2015 and 2025. The search utilized keywords such as (chest pain OR acute OR emergency) AND (aortic dissection) AND (screening OR triage OR diagnosis) AND (artificial intelligence OR deep learning OR machine learning OR algorithm). Inclusion criteria for this review included peer-reviewed full articles and non-peer-reviewed preprints that specifically addressed the differentiation of ATAD patients from other chest pain-related conditions in emergency settings, with the application of AI. Exclusion criteria were non-English studies (Table 1).

Traditional diagnostic paradigms in emergency setting

CT imaging in ATAD diagnosis

In emergency settings, CT angiography (CTA) is the gold standard for assessing ATAD, including its extension and branch involvement (3). A typical protocol involves plain CT followed by early- and late-phase contrast-enhanced CT for comprehensive vascular visualization (11).

Plain CT is useful for differentiating intramural hematoma (IMH) from other causes of aortic wall thickening upon hospital presentation, while CTA provides detailed imaging for aortic reconstruction. Thin-slice arterial phase CT scans allow full aortic assessment and aid in treatment planning (12). Although plain CT is accurate for detection of fresh thrombus in AD, it is less reliable in identifying intimal flaps or atherosclerotic changes (13). CTA bridges this gap but is contraindicated in patients with renal impairment or contrast allergies (12). In view of the pros and cons of plain CT and contrast-enhanced CT, plain CT serves as a triage and guidance to direct what further phase contrast-enhanced CT is needed.

The 2014 ESC guidelines highlight the use of the triple-rule-out approach for evaluating patients presenting with acute chest pain in the emergency department (3). This approach involves three CT phases: electrocardiogram (ECG)-gated CT for assessing coronary artery disease and ACS (14); the pulmonary arterial phase for detecting PE (15), and a CT aortogram for identifying ATAD (16). However, this method has been associated with significant drawbacks, including high radiation exposure for patients, prolonged diagnostic times, and high cost (16-18). As per the updated 2021 evaluation and diagnosis of chest pain guideline, imaging protocols should be customized to the most likely diagnosis rather than using the triple rule-out approach (19).

Role of CXR in screening ATAD

CXR can indicate TAD through specific findings such as displacement of intimal calcifications, presence of a double density, or a blurred aortic knob (20). According to the 2022 American College of Cardiology/American Heart Association (ACC/AHA) guideline, a widened mediastinum observed on CXR may suggest AD, which should be confirmed with subsequent CT imaging (12).

CXR is commonly utilized as a rapid screening tool to differentiate various causes of chest pain and as an initial test for ATAD. However, its diagnostic value for ATAD is limited, particularly when the abnormality involves the ascending aorta (3). The sensitivity and specificity of CXR for detecting ATAD through a widened aortic silhouette are only 67% and 70% respectively (21). The International Registry of Aortic Dissection (IRAD) reported that 11.3% of patients with ATAD showed no abnormalities on CXR (22). Therefore, while CXR is useful for screening ATAD, additional imaging, such as CT is necessary to confirm the diagnosis.

ECG as preliminary workup for ATAD

ECG plays a crucial role in the initial assessment and management of patients presenting with chest pain as it is simple, inexpensive, non-invasive, readily available, and rapidly interpretable at the time of presentation. A study reported that nearly 60% of patients with ATAD presented with abnormal ECG (23). Among ATAD patients, ischemic ST-T changes such as ST-segment elevation or depression or negative T waves are frequently observed on presentation ECG, especially when the ATAD is associated with malperfusion to the coronary arteries (24). However, these ECG patterns are no unique to ATAD and may indicate other cardiac conditions, such as myocardial infarction (MI) (23). Further examination is required to confirm diagnosis.

CTA is often incorporated with ECG. ECG-gated CT is the imaging modality used in “Triple-rule out” approach, to evaluate patients with chest pain for three potential causes, ATAD, PE and ACS (3). ECG-gated CT minimizes motion artifacts caused by cardiac activity, especially in detecting ascending AD and intimal tears (13). Contrast dose can also be reduced with image quality guaranteed (25). However, ECG gating is not widely adopted in emergency settings, due to the additional time and resources required for setup. In leading centers in the US, only 30% of CT scans for aortic diseases are ECG-gated (26).

Biomarkers in aiding diagnosis of ATAD

The 2024 European Association for Cardio-Thoracic Surgery/Society of Thoracic Surgeons (EACTS/STS) guidelines emphasize that measuring biomarkers, together with ECG and CXR, can serve as first-line examination in patients with acute chest pain to rule out alternative diagnosis (1). Common biomarkers used for screening ATAD include proteins, ribonucleic acids (RNAs), and deoxyribonucleic acids (DNAs) (27). Recently, blood-based diagnostic biomarkers have gained attention because of being non-invasive and cost-effective nature. Rapid and non-invasive blood tests can help differentiate ATAD from other causes of chest pain in clinical settings, particularly in rural clinics in China where CTA and experienced radiologists are unavailable (28). These biomarkers are regarded as reliable tools to compensate for the limitations of imaging examinations (29).

Among the various biomarkers, D-dimer stands out for its high diagnostic accuracy. Despite the fact that no single biomarker is currently considered definitive and imaging remains the gold standard for diagnosis, D-dimer, when combined with the Aortic Dissection Detection Risk Score (ADD-RS), is suggested as a valuable method for identifying suspected ATAD patients who require imaging examinations (30).

AI models driven by different modalities

In recent years, there has been a rise in novel algorithms utilizing AI for the diagnosis of ATAD (7). While CTA remains the preferred imaging modality, the diagnostic roles of CXRs and ECGs are expanding. AI-driven models integrate multiple imaging modalities or incorporate biomarkers have emerged to enhance diagnostic accuracy. The potential role of AI-powered models in supporting clinical decision making is summarized in Figure 1. These advancements in AI-based diagnostic tools for ATAD in emergency settings enable prompt and reliable diagnosis of this life-threatening condition while significantly reducing misdiagnosis rates. Table 2 summarises the findings from recent literatures, with a focus in comparing their performances of screening and diagnosis of ATAD in emergency settings.

Figure 1 A flow diagram illustrating the integration of AI-powered models in clinical decision-making for acute thoracic aortic dissection. The AI system evaluates a comprehensive set of patient data, including blood tests, CXRs, ECGs, plain CT scans, and CT aortograms. By analyzing these multimodal inputs, the AI assists clinicians in achieving more precise diagnoses and enhancing the overall decision-making process. ACS, acute coronary syndrome; AD, aortic dissection; AI, artificial intelligence; CT, computed tomography; CTA, computed tomography angiogram; CXR, chest X-ray; ECG, electrocardiogram; PE, pulmonary embolism.

AI-powered triage using CTA

CT, particularly CTA, is considered the gold standard and the definitive imaging modality for diagnosing ATAD. As a result, most AI-powered models are designed around CT to aid emergency physicians and junior radiologists in identifying ATAD. Since these models serve as the safety net in diagnosing ATAD, ensuring their classification performance is of paramount importance.

Sensitivity and specificity are standard metrics used to evaluate classification performance in binary outcome scenarios, representing the true positive and false positive cases, respectively (49). Models powered by CTA achieve desirable classification performance. In Liu et al.’s study, their model achieved an impressive accuracy of 0.98, with the capability to classify both Stanford type A and type B AD. The overall sensitivity and specificity for Type A AD were 0.969 and 0.982, respectively, while for Type B AD, they were 0.946 and 0.996. However, patients with thrombus within the aorta were excluded from their study, as the presence of a thrombus could obscure or compromise the visualization of the intimal flap (31). On the other hand, Chang et al. and Laletin et al. incorporated CTA scans with various conditions (32,33). Chang et al. created a diagnosis model to differentiate between AD, IMH and penetrating atherosclerotic ulcer (PAU). As it is difficult to identify subtle manifestations of IMH and PAU from AD, the model was validated using external dataset comprised with other aortic abnormalities such as atherosclerotic disease, aortic aneurysm, and arterial dissection. Their model achieves sensitivity of 0.94 and specificity of 0.93 (32). Laletin et al. incorporated CTA studies from more than 200 sites with various confounding conditions, such as thoracic or abdominal aneurysms, IMH, calcifications, and post-surgery cases, to enhance the diversity of their training dataset. Despite the inclusion of such a varied dataset, their model demonstrated a sensitivity of 94.2% and a specificity of 97.3% (33).

Image pre-processing plays a crucial role in improving classification performance. Wada et al. developed two models using CTA scans from patients with aortic emergencies. Model A, utilized cropped images of the aorta demonstrated significantly better classification performance compared to model B that used the original, unprocessed images. Notably, the specificity of model A was 0.879, which was much higher than that of model B (0.349). This improvement can be attributed to the larger proportion of the aorta being used in model A, effectively addressing the challenges posed by convolutional central network (CNN) models in detecting small objects against complex backgrounds (34).

Cotena et al. conducted the only study to simulate their model in a real-world emergency setting (35). Initially, CTAs were assessed by radiologists following the traditional clinical workflow of First In, First Out (FIFO). After a one-month washout period, the same batch of CTAs was reassessed with suspected cases flagged by AI. The model achieved a sensitivity of 94.29% and a specificity of 100%, while significantly reducing the time required for diagnosis by 26.8%. These results highlight the model’s standalone effectiveness in high-pressure, high-volume, and time-sensitive clinical environments (35).

Plain CT on the rise

Compared to CTA, diagnosing ATAD using plain CT is significantly more challenging, as distinguishing features are less apparent. Even experienced radiologists often achieve low diagnostic sensitivity (50). Radiomics helps address this challenge by offering a quantitative approach to medical imaging. It extracts a large number of predefined high-throughput features from images and applies statistical methods to filter relevant features. When combined with AI, radiomics can support clinical decision-making by automating repetitive, routine tasks (51).

Two studies have explored the use of radiomics for diagnosing ATAD on plain CT. Rather than directly identifying the tear or intimal flap, Guo et al. analyzed blood heterogeneity to distinguish thoracic AD from other aortic conditions using radiomics. Turbulent blood flow—caused by different flow patterns—was quantified as a distinctive feature of AD patients on plain CT (36). In Ma et al.’s study, nine radiomics features were identified to differentiate acute aortic syndromes (AAS), including ATAD. Their model demonstrated remarkable accuracy, achieving 0.991 in distinguishing AAS even when significant features were absent on plain CT (37).

Additionally, Kim et al. developed a model for simultaneously detecting AD and IMH on plain CT. In this study, plain CT images were paired with post-contrast images during model training (38). However, plain CT scans from a single patient typically consist of only 70–100 images, which increases the risk of missing subtle but critical pathological features, such as the intimal flap (52).

Advancements in AI for CXR-based screening of ATAD

CXRs have been utilized as a screening tool for ATAD due to their low cost, non-invasive nature, and widespread availability (53). However, the sensitivity of CXRs for detecting ATAD is only 70%, which increases the risk of misdiagnosis (21). With the advancements in AI, subtle features that may go undetected by the human eye can now be identified, and the relationships between imaging patterns and clinical metadata can be explored.

Two multicenter, retrospective studies have been conducted to develop AI-powered screening tools using CXRs. CNN was used in both studies. In the study by Lin et al., the performances of various CNN architectures on CXRs were compared, with Inception-v3 demonstrating strong diagnostic capability for ATAD, achieving a precision of 85%. However, the model’s sensitivity for identifying AAS, which includes ATAD and thoracic aortic aneurysm, was only 68% (39). In contrast, Lee et al. reported that their model achieved a sensitivity of 94.44% and an accuracy of 90.2%, which are comparable to the diagnostic performance of CT. This can be attributed to the model identifying subtle abnormalities in the diameters of both the ascending and descending aortas (40).

It is important to consider sample diversity when evaluating model performance in real-world settings. Lee et al.’s study included only TAD and normal cases during model training, whereas Lin et al.’s study included a broader range of conditions under AAS, such as penetrating atherosclerotic aortic ulcer, IMH, and ATAD (39,40). This difference may explain the variations in model performance. In the study by Kolossváry et al., patients exhibiting symptoms of acute chest pain were presented with external devices, such as ECG electrodes and wires. Despite this, their model was still able to extract information from the heart and lungs that correlated with patient outcomes, which is crucial given the diverse clinical presentations (41).

In Kolossváry et al. study, three logistic regression model were developed and compared to simulate clinical value—model 1: age and sex; model 2: model 1 and biomarker positivity and model 3: model 2 and deep learning (DL) model. Model 3 outperformed other models, in terms of deferrable rate at sensitivity threshold of 99% (14% vs. 2%) (41). It proves that model 3 performed better in differential diagnosis of ATAD, by integrating CXR, biomarkers and patient demographics.

The role of biomarkers in machine learning (ML)-driven ATAD diagnosis

The 2024 EACTS/STS guidelines suggested that more research should be done to identify diagnostic biomarkers for confirming ATAD (1). In addition to identifying novel diagnostic biomarkers, Huo et al., Liu et al. and Tavafi et al. proposed ML classification models to assist in the diagnosis of ATAD using routine blood test results and patient demographics (42-44). Huo et al. and Tavafi et al. selected 13 and 26 attributes derived from the lipid profile, blood routine examination, and patient admission data respectively, while Liu et al. utilized 76 features based on the 2014 ESC guidelines. Both studies compared the performance of various ML models (42-44). In Huo et al.’s study, the Bayesian Network achieved the best performance, with a precision of 85.54% and an area under the curve (AUC) of 0.8136 (42). In contrast, Liu et al. identified the SmoteBagging and EasyEnsemble models as the most effective, reporting recall rates of 78.1% and 7.8%, and specificities of 79.2% and 79.3%, respectively (43). However, regarding sample selection, Huo et al. included positive misdiagnosed cases instead of screening all cases comprehensively (42). Among the three studies reviewed, Tavafi et al. reported that combination of Relief F feature selection and a random forest classifier achieved the best performance, with AUC of 1.0 and sensitivity of 99.3% (44).

With the increasing digitalization of health records, it is now possible to integrate patient demographics with blood test results for analysis ML models, enabling rapid generation of screening results. However, missing values in training datasets are a common challenge, as highlighted in both studies. Huo et al. reported that their AD dataset contained a mix of nominal and numeric attributes with missing values (42). To address this, Liu et al. employed a stratified random sampling method to fill the missing data by sampling values within the positive and negative categories (43). However, their AD datasets exhibited significant overlap between AD samples (positive class) and non-AD samples (negative class). Since the stratified random sampling method is sensitive to overlapping data, this overlap caused a skew toward the negative class, introducing biases into the results (54). Tavafi et al. adopted Multiple Imputation by Chained Equations (MICE) to handle missing data. Their dataset exhibited a significant class imbalance, with 129 positive cases of AAS compared to 148,578 non-AAS cases. To mitigate this imbalance and reduce model bias, they applied the Synthetic Minority Oversampling Technique (SMOTE) to generate synthetic samples of the positive cases (44).

Multimodality approaches—CXR, ECG and biomarker

Several studies combined CXR, biomarker with ECG to optimize diagnostic performance. Patients with ATAD frequently present with ECG abnormalities at the time of admission. While ECG is a widely available and non-invasive screening tool, its utility in diagnosing ATAD is limited. Nonspecific ST-segment changes observed on ECG can indicate various conditions, including ACS, acute MI and ATAD. Misdiagnosing ATAD as ACS or MI can result in the use of thrombolytic or antiplatelet therapies, which significantly increase the risk of severe bleeding and mortality (55,56). Therefore, in emergency settings, ECG is typically combined with CXR and biomarker analysis to aid in differential diagnosis. With the integration of AI, models have been emerged to improve the diagnostic accuracy of AD by analyzing ECG results in conjunction with CXR or biomarker data.

Liu et al.’s study developed a DL model for diagnosing ATAD by integrating CXR, ECG, and biomarkers. The inclusion of D-dimer significantly enhanced the model’s performance, achieving a sensitivity of 100% and specificity of 87%. Although ST-segment elevation was treated as a non-AD feature, the model was still able to correctly identify AD using ECG data. However, the mechanisms linking ECG features to AD remain unclear, creating a “black box” issue that limits interpretability and may hinder the clinical implementation of such AI models (45). In contrast, the study by Arita et al. focused its CNN using ECG and biomarkers to analyze the QRS and ST-T segments of ECG data, resulting in a higher sensitivity (77%) compared to Liu et al.’s model (44%) (45,46). Despite this, the Liu et al. study underwent prospective validation, making its findings more reliable in real-world clinical settings, whereas the Arita et al. model lacks such validation (45,46).

Similar to the findings by Arita et al., the studies by Wang et al. and Zhou et al. also highlighted the significance of distinctive information from ECG, which surpasses most biochemical indicators in importance (47,48). Wang et al. focused on distinguishing acute MI from ATAD. Their model demonstrated an AUC that exceeded the performance of three experienced human readers, showcasing its superior ability to differentiate between MI and ATAD. By identifying complex patterns in ECG data, the model was able to detect subtle functional changes crucial for distinguishing acute MI from ATAD (47). Similarly, Zhou et al. supported these findings, reporting that abnormal ECG features, particularly ST-segment abnormalities, were the most strongly associated with the model’s predictions (48).

Study design

With a surge of AI models in clinical application, various reporting guidelines on AI in diagnosis and decision making have been published. Table 3 summarizes the relevant reporting guidelines. Key features related to the literatures are discussed below.

The importance of dataset size, diversity, and representativeness

According to TRIPOD-AI and MINIMAR, it is important to describe how the data were separated for development and evaluation of model performance (57,58). In view of all searched literatures, number of samples divided for training, testing and evaluation have been clearly outlined. Adequate dataset size is essential to prevent overfitting and ensure robust performance on unseen data. Notably, over 50,000 CXRs and patient data were included in Kolossváry et al. study, and over 148,000 patient data were collected in Tavafi et al. study (41,44). When the sample size is too small, cross-validation can be utilized to evaluate the model and enhance generalizability (59). Cross-validation involves splitting the dataset into multiple folds for iterative training and testing, with various methods available to define these folds depending on the task requirements (60). Commonly, 5-fold or 7-fold cross-validation is employed, as reported in the literatures (34,40,42,43,45,46,48).

However, an excessively large sample size is not always necessary, as data diversity also matters (60). Data diversity is particularly critical in emergency settings, as patients admit with heterogenous conditions. Lee et al.’s research focused solely on TAD and normal cases during model development (40), whereas Lin et al.’s work encompassed a wider spectrum of AAS conditions, including penetrating atherosclerotic aortic ulcers, IMH, and ATAD (39). Meanwhile, Kolossváry et al. reported that patients displaying symptoms of acute chest pain were equipped with external devices like ECG electrodes and wires (41).

Future-AI emphasizes the importance of data representativeness, as capturing the essential characteristics of the population is necessary to achieve the specific goals of an AI model (61). For example, Zhou et al. included non-AD patients with chest pain as a control group to validate their model’s ability to differentiate between AD and non-AD patients presenting with chest pain (48). TRIPOD+AI also highlights that datasets used for model evaluation must be representative of the target population to ensure unbiased and accurate predictions (57).

Clinical implementation

Ideally, a multi-center, prospective approach should be adopted to enhance the generalizability of a model. However, as stated by DECIDE-AI, due to challenges in accessing data, many studies are constrained to small, single-center datasets or lack external validation (62). This can result in dataset shift, where models behave differently when applied to external data that differs slightly from the training and validation data (60). Some models, however, have been validated across multiple sites (32,33,35,39,40) For instance, the Laletin et al. model was validated in over 200 cities across the U.S. and Europe. All U.S. and European data were de-identified to comply with the HIPAA Privacy Rule and the General Data Protection Regulation (GDPR) in Europe (33). By adhering to the privacy regulations, patient-sensitive data can be analyzed in a safe and private manner.

Only three studies have undergone prospective evaluation to ensure robust validation in real-world settings (45-47). In Liu et al.’s study, their model notified physicians of the positive results and recommended further CXR or D-dimer tests, after ECG examination (45). However, prospective validation may not be feasible in clinical setting as it requires clinician participation and may disrupt routine clinical workflows. To test model’s clinical applicability, Cotena et al. simulated two distinct clinical workflows to compare the traditional FIFO approach without AI assistance to an AI-driven method that prioritized tasks, with a 1-month washout period separating the two workflows. Model predictions can also be compared to actual patient outcomes to enhance generalizability (35).

Building clinicians’ trust towards AI

The “black-box” nature of many AI/ML algorithms poses a significant challenge to their interpretability and clinician trust (60). Some literatures reported that when decision-making processes are not transparent and results are difficult to verify, physicians may hesitate utilize these tools (39,41,45,48). This issue is particular common in ECG studies as mechanisms linking ECG features to AD remain unclear, which may make the algorithms less explainable. In order to boost model interpretability, Zhou et al. adopted a quantitative approach by discerning strong AD signals on ECG to different aortic risk levels (48).

AI has the potential to significantly reduce clinical workloads while delivering reliable results, prompting physicians to focus on comparisons between AI algorithms and expert readers (63). It is crucial to provide a detailed description of the reference standard, as variability in observer assessments of ground truth can result in label inconsistencies and an underestimation of the ML model’s performance (60). In the study by Guo et al., readers with over 5 and 10 years of experience manually segmented the aorta, and their results were compared with the AI (36). Similarly, Ma et al. utilized two radiologists with more than 8 and 10 years of experience to assess the model’s reliability (37). Cotena et al. evaluated time-reporting efficiency and performance parameters in the context of routine clinical workflows (35). Therefore, human factors evaluations play a key role in boosting reproducibility of study outcomes.

Future directions

Optimize dataset size and variety

Apart from cross validation of data, data augmentation is a widely used method to deal with insufficient amount of training data and uneven class balance within the datasets. Classical image transformations include rotating, cropping, zooming of images (64). Generative adversarial network (GAN) is a relatively new method in image generation. However, proper quality control is needed to avoid reinforcing biases, including artifacts and overfitting to synthetic patterns (65).

Synthetic data generation is another way to expand dataset. Medghalchi et al. proposed a novel pipeline called Medical Dataset Enhancement via Diversified Augmentation Pipeline (MEDDAP), a diffusion model to automatically generate new informative labeled samples by augmenting existing small datasets (66). Hosseini and Serag demonstrate reliability of synthetic CXR in disease classification, with critical radiological biomarkers preserved (67).

Clinical validation

ML tools may perform differently on real-world data that does not align with the algorithm’s design compared to their performance during internal or external validation. Consequently, prospective studies that compare the predictions of ML algorithms with those of clinicians are crucial for evaluating their effectiveness in clinical practice (49). Furthermore, a single-center evaluation restricts the generalizability of the model’s results to a wider range of clinical scenarios or diverse patient populations.

Silent deployment is a strategy used to test models in real time without direct interaction with end-users (49). Before formally implementing a model, this approach allows the generated results to be compared with actual patient outcomes. Silent deployment enables the model to run alongside routine clinical practice without disrupting real-world clinical workflows, which can make it easier to engage sites and clinicians in prospective validation efforts (68).

Security and privacy considerations are critical when implementing AI models in clinical settings, as substantial amounts of sensitive data are often required for training and validation (60). Data anonymization may not always be sufficient to fully de-identify privacy-sensitive information (69), and this concern becomes even more significant in multi-center studies involving large datasets. Federated learning offers a decentralized alternative by keeping training data distributed across local devices while collaboratively building a shared model. In this approach, each user computes updates to the global model managed by a central server without uploading their local training data, thereby enhancing data privacy and security (70).

Makes AI explainable and generalizable

An explainable model allows end users to interpret the AI system and its outputs, gain a clear understanding of the tool’s capabilities and limitations, and intervene when necessary (61). The development of explainable AI or ML can help reveal what these tools are doing “under the hood”, making their operations more transparent (60). By increasing the visibility of AI algorithms, it is possible to identify potential biases and reinforce discriminatory practices tied to desirable features (71).

The FUTURE-AI framework suggests two approaches to enhance the explainability of AI models. First, during the design phase, clinicians should be able to understand which features most significantly influence the model’s diagnosis and reach a consensus (61). For instance, the model should identify features such as the intimal flap and false lumen on CT scans based on the agreed-upon standards (13). The explainable AI method should first be assessed quantitatively through computational techniques, followed by a qualitative evaluation involving end users to determine its clinical performance (61). Formative evaluation, which involves generating and sharing data with the research team and target clinicians at various stages of implementation, enables the implementation team to identify and address challenges and adapt the solution for better integration into clinical workflows (72). This can be achieved through human reader evaluation. In this context, the interaction between AI systems, clinicians, and implementation environments is a critical factor in optimizing the clinical decision-making process. Clinician participation in deploying AI models in real-world settings is essential for ensuring their generalizability.

AI-powered echocardiography application

Although is a vital tool for diagnosing AD in emergency settings, no AI study has been done using echocardiography as an imaging modality. Trans-thoracic echocardiogram is recommended as the first-line imaging modality while CT is being prepared (73).

In recent years, pocket echocardiography has been recommended by the European Association of Cardiovascular Imaging as a quick screening tool and a supplement to physical examination (74). Pocket-size mobile echocardiography has demonstrated promising results, achieving a true-positive value of 92.1% in the screening of thoracic aortic aneurysm among hypertensive patients (75). Nishigami proposed a screening protocol using point-of-care echocardiography for the evaluation of AD, PE, and ACS in patients presenting with chest pain (76). In the future, it would be feasible to combine AI with pocket echocardiography for ATAD screening.

Growing role of AI in TAD diagnosis

AI is revolutionizing the field of medicine, with large language models (LLMs) leveraging DL techniques to provide language-based diagnostic support. For instance, Goyal et al. evaluated ten AD cases using ChatGPT 4.0 to compare its diagnostic outputs, achieving promising results (77). This underscores the potential of LLMs in aiding clinical diagnoses. Meanwhile, generative AI is gaining momentum in automating radiology report generation, particularly for CXR (78,79). Additionally, risk stratification has become a growing focus in medical AI. Zhou et al. introduced the AI-aortic-dissection-ECG (AADE) score, to aids clinicians in assessing ATAD severity and predicting mortality (48). Al-Alusi et al. developed a digital biomarker powered by ECG data to assess the risk of cardiovascular diseases, including AD, in individuals with hypertension (80). These advancements indicate a future where AI applications play a significant role in AD diagnosis and risk assessment.

Conclusions

In conclusion, the integration of AI, imaging modalities, and biomarkers represents a significant advancement in the diagnostic approach ATAD. Beyond traditional diagnostic methods like CTA, multimodal models incorporating CXR, ECG, and biomarkers have been explored. While these models show promising results, their generalizability in real-world clinical settings must be carefully evaluated. To support the clinical implementation of AI models, several reporting guidelines have been developed. It is evident that AI will play an increasingly important role in the screening and diagnosis of ATAD. However, the involvement of clinicians remains essential to ensure the seamless integration of AI into clinical practice.

Acknowledgments

None.

Footnote

Reporting Checklist: The authors have completed the Narrative Review reporting checklist. Available at https://atm.amegroups.com/article/view/10.21037/atm-25-82/rc

Peer Review File: Available at https://atm.amegroups.com/article/view/10.21037/atm-25-82/prf

Funding: None.

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://atm.amegroups.com/article/view/10.21037/atm-25-82/coif). R.H.L.W. serves as an unpaid editorial board member of Annals of Translational Medicine from September 2024 to August 2026. Both E.M.K.L. and R.H.L.W. serve as reviewers of Annals of Translational Medicine. The other author has no conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Czerny M, Grabenwöger M, Berger T, et al. EACTS/STS Guidelines for diagnosing and treating acute and chronic syndromes of the aortic organ. Eur J Cardiothorac Surg 2024;65:ezad426. [Crossref] [PubMed]
2. Harris KM, Nienaber CA, Peterson MD, et al. Early Mortality in Type A Acute Aortic Dissection: Insights From the International Registry of Acute Aortic Dissection. JAMA Cardiol 2022;7:1009-15. [Crossref] [PubMed]
3. Erbel R, Aboyans V, Boileau C, et al. 2014 ESC Guidelines on the diagnosis and treatment of aortic diseases: Document covering acute and chronic aortic diseases of the thoracic and abdominal aorta of the adult. The Task Force for the Diagnosis and Treatment of Aortic Diseases of the European Society of Cardiology (ESC). Eur Heart J 2014;35:2873-926. [Crossref] [PubMed]
4. China Heart House, Chest Pain Professional Committee of National Clinical Research Center for Interventional Medicine. Summary of quality control report of Chest Pain Center in 2019. Chinese Journal of Interventional Cardiology 2020;421-4.
5. Zhan S, Hong S, Shan-Shan L, et al. Misdiagnosis of aortic dissection: experience of 361 patients. J Clin Hypertens (Greenwich) 2012;14:256-60. [Crossref] [PubMed]
6. Asouhidou I, Asteri T. Acute aortic dissection: be aware of misdiagnosis. BMC Res Notes 2009;2:25. [Crossref] [PubMed]
7. Mastrodicasa D, Codari M, Bäumler K, et al. Artificial Intelligence Applications in Aortic Dissection Imaging. Semin Roentgenol 2022;57:357-63. [Crossref] [PubMed]
8. Farina JM, Pereyra M, Mahmoud AK, et al. Artificial Intelligence-Based Prediction of Cardiovascular Diseases from Chest Radiography. J Imaging 2023;9:236. [Crossref] [PubMed]
9. Lum RTW, Wang X, Zhang M, et al. Emerging role of radiogenomics in genetically triggered thoracic aortic aneurysm and dissection: a narrative review. Ann Transl Med 2023;11:356. [Crossref] [PubMed]
10. Lo EMK, Chen S, Ng KHL, et al. Artificial intelligence-powered solutions for automated aortic diameter measurement in computed tomography: a narrative review. Ann Transl Med 2024;12:116. [Crossref] [PubMed]
11. Baliyan V, Shaqdan K, Hedgire S, et al. Vascular computed tomography angiography technique and indications. Cardiovasc Diagn Ther 2019;9:S14-27. [Crossref] [PubMed]
12. Isselbacher EM, Preventza O, Hamilton Black J 3rd, et al. 2022 ACC/AHA Guideline for the Diagnosis and Management of Aortic Disease: A Report of the American Heart Association/American College of Cardiology Joint Committee on Clinical Practice Guidelines. Circulation 2022;146:e334-e482. [Crossref] [PubMed]
13. Tanaka R, Yoshioka K, Abiko A. Updates on Computed Tomography Imaging in Aortic Aneurysms and Dissection. Ann Vasc Dis 2020;13:23-7. [Crossref] [PubMed]
14. Dobra M, Bordi L, Nyulas T, et al. Clinical update. Computed Tomography — an Emerging Tool for Triple Rule-Out in the Emergency Department. A Review. J Cardiovasc Emergencies 2017;3:36-40.
15. Soliman HH. Value of triple rule-out CT in the emergency department. Egypt J Radiol Nucl Med 2015;46:621-7.
16. Esterson YB, Esterson AY, Jankowska D, et al. Incidental myocardial perfusion defect detected on ECG-gated CT aortogram. Clin Imaging 2019;54:148-52. [Crossref] [PubMed]
17. Madder RD, Raff GL, Hickman L, et al. Comparative diagnostic yield and 3-month outcomes of "triple rule-out" and standard protocol coronary CT angiography in the evaluation of acute chest pain. J Cardiovasc Comput Tomogr 2011;5:165-71. [Crossref] [PubMed]
18. Takakuwa KM, Halpern EJ, Shofer FS. A time and imaging cost analysis of low-risk ED observation patients: a conservative 64-section computed tomography coronary angiography "triple rule-out" compared to nuclear stress test strategy. Am J Emerg Med 2011;29:187-95. [Crossref] [PubMed]
19. Gulati M, Levy PD, Mukherjee D, et al. 2021 AHA/ACC/ASE/CHEST/SAEM/SCCT/SCMR Guideline for the Evaluation and Diagnosis of Chest Pain: A Report of the American College of Cardiology/American Heart Association Joint Committee on Clinical Practice Guidelines. Circulation 2021;144:e368-e454. [Crossref] [PubMed]
20. Damberg A, Ziganshin BA, Elefteriades JA. Chapter 12 - Chest X-Ray in Aortic Disease. In: New Approaches to Aortic Diseases from Valve to Abdominal Bifurcation. Elsevier; 2018:129-31.
21. von Kodolitsch Y, Nienaber CA, Dieckmann C, et al. Chest radiography for the diagnosis of acute aortic syndrome. Am J Med 2004;116:73-7. [Crossref] [PubMed]
22. Hagan PG, Nienaber CA, Isselbacher EM, et al. The International Registry of Acute Aortic Dissection (IRAD): new insights into an old disease. JAMA 2000;283:897-903. [Crossref] [PubMed]
23. Costin NI, Korach A, Loor G, et al. Patients With Type A Acute Aortic Dissection Presenting With an Abnormal Electrocardiogram. Ann Thorac Surg 2018;105:92-9. [Crossref] [PubMed]
24. Kosuge M, Kimura K, Uchida K, et al. Clinical Implications of Electrocardiograms for Patients With Type A Acute Aortic Dissection. Circ J 2017;81:1254-60. [Crossref] [PubMed]
25. Hou QR, Gao W, Sun AM, et al. A prospective evaluation of contrast and radiation dose and image quality in cardiac CT in children with complex congenital heart disease using low-concentration iodinated contrast agent and low tube voltage and current. Br J Radiol 2017;90:20160669. [Crossref] [PubMed]
26. Asch FM, Yuriditsky E, Prakash SK, et al. The Need for Standardized Methods for Measuring the Aorta: Multimodality Core Lab Experience From the GenTAC Registry. JACC Cardiovasc Imaging 2016;9:219-26. [Crossref] [PubMed]
27. Zhao Y, Fu W, Wang L. Biomarkers in aortic dissection: Diagnostic and prognostic value from clinical research. Chin Med J (Engl) 2024;137:257-69. [Crossref] [PubMed]
28. Qiu P, Yang M, Pu H, et al. Potential Clinical Value of Biomarker-Guided Emergency Triage for Thoracic Aortic Dissection. Front Cardiovasc Med 2021;8:777327. [Crossref] [PubMed]
29. Chen H, Li Y, Li Z, et al. Diagnostic biomarkers and aortic dissection: a systematic review and meta-analysis. BMC Cardiovasc Disord 2023;23:497. [Crossref] [PubMed]
30. Ren S, Essat M, Pandor A, et al. Diagnostic accuracy of the aortic dissection detection risk score alone or with D-dimer for acute aortic syndromes: Systematic review and meta-analysis. PLoS One 2024;19:e0304401. [Crossref] [PubMed]
31. Liu HH, Chang CB, Chen YS, et al. Automated Detection and Differentiation of Stanford Type A and Type B Aortic Dissections in CTA Scans Using Deep Learning. Diagnostics (Basel) 2024;15:12. [Crossref] [PubMed]
32. Chang J, Lee KJ, Wang TH, et al. Multi-Stage Cascaded Deep Learning-Based Model for Acute Aortic Syndrome Detection: A Multisite Validation Study. J Clin Med 2025;14:4797. [Crossref] [PubMed]
33. Laletin V, Ayobi A, Chang PD, et al. Diagnostic Performance of a Deep Learning-Powered Application for Aortic Dissection Triage Prioritization and Classification. Diagnostics (Basel) 2024;14:1877. [Crossref] [PubMed]
34. Wada T, Takahashi M, Matsunaga H, et al. An automated screening model for aortic emergencies using convolutional neural networks and cropped computed tomography angiography images of the aorta. Int J Comput Assist Radiol Surg 2023;18:2253-60. [Crossref] [PubMed]
35. Cotena M, Ayobi A, Zuchowski C, et al. Enhancing Radiologist Efficiency with AI: A Multi-Reader Multi-Case Study on Aortic Dissection Detection and Prioritization. Diagnostics (Basel) 2024;14:2689. [Crossref] [PubMed]
36. Guo Y, Chen X, Lin X, et al. Non-contrast CT-based radiomic signature for screening thoracic aortic dissections: a multicenter study. Eur Radiol 2021;31:7067-76. [Crossref] [PubMed]
37. Ma Z, Jin L, Zhang L, et al. Diagnosis of Acute Aortic Syndromes on Non-Contrast CT Images with Radiomics-Based Machine Learning. Biology (Basel) 2023;12:337. [Crossref] [PubMed]
38. Kim YS, Kim JG, Choi HY, et al. Detection of Aortic Dissection and Intramural Hematoma in Non-Contrast Chest Computed Tomography Using a You Only Look Once-Based Deep Learning Model. J Clin Med 2024;13:6868. [Crossref] [PubMed]
39. Lin YT, Wang BC, Chung JY. Identifying Acute Aortic Syndrome and Thoracic Aortic Aneurysm from Chest Radiography in the Emergency Department Using Convolutional Neural Network Models. Diagnostics (Basel) 2024;14:1646. [Crossref] [PubMed]
40. Lee DK, Kim JH, Oh J, et al. Detection of acute thoracic aortic dissection based on plain chest radiography and a residual neural network (Resnet). Sci Rep 2022;12:21884. [Crossref] [PubMed]
41. Kolossváry M, Raghu VK, Nagurney JT, et al. Deep Learning Analysis of Chest Radiographs to Triage Patients with Acute Chest Pain Syndrome. Radiology 2023;306:e221926. [Crossref] [PubMed]
42. Huo D, Kou B, Zhou Z, et al. A machine learning model to classify aortic dissection patients in the early diagnosis phase. Sci Rep 2019;9:2701. [Crossref] [PubMed]
43. Liu L, Zhang C, Zhang G, et al. A study of aortic dissection screening method based on multiple machine learning models. J Thorac Dis 2020;12:605-14. [Crossref] [PubMed]
44. Tavafi M, Passi K, Ohle R. Early Risk Prediction in Acute Aortic Syndrome on Clinical Data Using Machine Learning. Algorithms 2025;18:257.
45. Liu WT, Lin CS, Tsao TP, et al. A Deep-Learning Algorithm-Enhanced System Integrating Electrocardiograms and Chest X-rays for Diagnosing Aortic Dissection. Can J Cardiol 2022;38:160-8. [Crossref] [PubMed]
46. Arita T, Suzuki S, Motogi J, et al. Identifying patients with acute aortic dissection using an electrocardiogram with convolutional neural network. Int J Cardiol Heart Vasc 2024;51:101389. [Crossref] [PubMed]
47. Wang L, Wu H, Wu C, et al. A deep-learning system integrating electrocardiograms and laboratory indicators for diagnosing acute aortic dissection and acute myocardial infarction. Int J Cardiol 2025;423:133008. [Crossref] [PubMed]
48. Zhou M, Lei L, Chen W, et al. Deep learning-based diagnosis of aortic dissection using an electrocardiogram: Development, validation, and clinical implications of the AADE score. Kardiol Pol 2024;82:63-71. [Crossref] [PubMed]
49. Antoniou T, Mamdani M. Evaluation of machine learning solutions in medicine. CMAJ 2021;193:E1425-9. [Crossref] [PubMed]
50. Kurabayashi M, Okishige K, Ueshima D, et al. Diagnostic utility of unenhanced computed tomography for acute aortic syndrome. Circ J 2014;78:1928-34. [Crossref] [PubMed]
51. van Timmeren JE, Cester D, Tanadini-Lang S, et al. Radiomics in medical imaging-"how-to" guide and critical reflection. Insights Imaging 2020;11:91. [Crossref] [PubMed]
52. McMahon MA, Squirrell CA. Multidetector CT of Aortic Dissection: A Pictorial Review. Radiographics 2010;30:445-60. [Crossref] [PubMed]
53. Ueda D, Matsumoto T, Ehara S, et al. Artificial intelligence-based model to classify cardiac functions from chest radiographs: a multi-institutional, retrospective model development and validation study. Lancet Digit Health 2023;5:e525-33. [Crossref] [PubMed]
54. Sapsford R, Jupp V. Data Collection and Analysis. SAGE 2006;358.
55. Luo JL, Wu CK, Lin YH, et al. Type A aortic dissection manifesting as acute myocardial infarction: still a lesson to learn. Acta Cardiol 2009;64:499-504. [Crossref] [PubMed]
56. Xue Y, Chong H, Zhu X, et al. Aortic dissection patients mimic acute coronary syndrome with preoperative antiplatelet therapy. J Thorac Dis 2019;11:3385-90. [Crossref] [PubMed]
57. Collins GS, Moons KGM, Dhiman P, et al. TRIPOD+AI statement: updated guidance for reporting clinical prediction models that use regression or machine learning methods. BMJ 2024;385:e078378. [Crossref] [PubMed]
58. Hernandez-Boussard T, Bozkurt S, Ioannidis JPA, et al. MINIMAR (MINimum Information for Medical AI Reporting): Developing reporting standards for artificial intelligence in health care. J Am Med Inform Assoc 2020;27:2011-5. [Crossref] [PubMed]
59. Norgeot B, Quer G, Beaulieu-Jones BK, et al. Minimum information about clinical artificial intelligence modeling: the MI-CLAIM checklist. Nat Med 2020;26:1320-4. [Crossref] [PubMed]
60. Williams MC, Weir-McCall JR, Baldassarre LA, et al. Artificial Intelligence and Machine Learning for Cardiovascular Computed Tomography (CCT): A White Paper of the Society of Cardiovascular Computed Tomography (SCCT). J Cardiovasc Comput Tomogr 2024;18:519-32. [Crossref] [PubMed]
61. Lekadir K, Frangi AF, Porras AR, et al. FUTURE-AI: international consensus guideline for trustworthy and deployable artificial intelligence in healthcare. BMJ 2025;388:e081554. [Crossref] [PubMed]
62. Vasey B, Nagendran M, Campbell B, et al. Reporting guideline for the early-stage clinical evaluation of decision support systems driven by artificial intelligence: DECIDE-AI. Nat Med 2022;28:924-33. [Crossref] [PubMed]
63. Bluemke DA, Moy L, Bredella MA, et al. Assessing Radiology Research on Artificial Intelligence: A Brief Guide for Authors, Reviewers, and Readers-From the Radiology Editorial Board. Radiology 2020;294:487-9. [Crossref] [PubMed]
64. Mikolajczyk A, Grochowski M. Data augmentation for improving deep learning in image classification problem. In: 2018 International Interdisciplinary PhD Workshop (IIPhDW) [Internet]. Swinoujście: IEEE; [cited 2025 Apr 28]. 2018:117-22. Available from: https://ieeexplore.ieee.org/document/8388338/
65. Shorten C, Khoshgoftaar TM. A survey on Image Data Augmentation for Deep Learning. J Big Data 2019;6:1-48. [Crossref] [PubMed]
66. Medghalchi Y, Zakariaei N, Rahmim A, et al. MEDDAP: Medical Dataset Enhancement via Diversified Augmentation Pipeline. arXiv 2024. arXiv:2403.16335.
67. Hosseini A, Serag A. Is synthetic data generation effective in maintaining clinical biomarkers? Investigating diffusion models across diverse imaging modalities. Front Artif Intell 2024;7:1454441. [Crossref] [PubMed]
68. Verma AA, Murray J, Greiner R, et al. Implementing machine learning in medicine. CMAJ 2021;193:E1351-7. [Crossref] [PubMed]
69. Sweeney L. Simple Demographics Often Identify People Uniquely. Carnegie Mellon University, DataPrivacy Working Paper 3. Pittsburgh; 2000.
70. McMahan HB, Moore E, Ramage D, et al. Communication-Efficient Learning of Deep Networks from Decentralized Data. arXiv 2023. arXiv:1602.05629.
71. He J, Baxter SL, Xu J, et al. The practical implementation of artificial intelligence technologies in medicine. Nat Med 2019;25:30-6. [Crossref] [PubMed]
72. Elwy AR, Wasan AD, Gillman AG, et al. Using formative evaluation methods to improve clinical implementation efforts: Description and an example. Psychiatry Res 2020;283:112532. [Crossref] [PubMed]
73. Meredith EL, Masani ND. Echocardiography in the emergency assessment of acute aortic syndromes. Eur J Echocardiogr 2009;10:i31-9. [Crossref] [PubMed]
74. Cardim N, Dalen H, Voigt JU, et al. The use of handheld ultrasound devices: a position statement of the European Association of Cardiovascular Imaging (2018 update). Eur Heart J Cardiovasc Imaging 2019;20:245-52. [Crossref] [PubMed]
75. Wong RH, Yang F, Fujikawa T, et al. Pocket-Size Mobile Echocardiographic Screening of Thoracic Aortic Aneurysms in Hypertensive Patients. Ann Thorac Surg 2021;111:1554-9. [Crossref] [PubMed]
76. Nishigami K. Point-of-care echocardiography for aortic dissection, pulmonary embolism and acute coronary syndrome in patients with killer chest pain: EASY screening focused on the assessment of effusion, aorta, ventricular size and shape and ventricular asynergy. J Echocardiogr 2015;13:141-4. [Crossref] [PubMed]
77. Goyal A, Tariq MD, Ahsan A, et al. Can ChatGPT 4.0 Diagnose Acute Aortic Dissection? Integrating Artificial Intelligence into Medical Diagnostics. Am J Cardiol 2025;239:90-2. [Crossref] [PubMed]
78. Pan Y, Li K, Wu J, et al. Sequence-Integrated Radiology Report Generation Leveraging Positional Encoding for Text and Images. In: 2024 IEEE Cyber Science and Technology Congress (CyberSciTech) [Internet]. Boracay Island, Philippines: IEEE; 2024 [cited 2025 Apr 19]. p. 480–4. Available online: https://ieeexplore.ieee.org/document/10795493/
79. Hong EK, Roh B, Park B, et al. Value of Using a Generative AI Model in Chest Radiography Reporting: A Reader Study. Radiology 2025;314:e241646. [Crossref] [PubMed]
80. Al-Alusi MA, Friedman SF, Kany S, et al. A deep learning digital biomarker to detect hypertension and stratify cardiovascular risk from the electrocardiogram. NPJ Digit Med 2025;8:120. [Crossref] [PubMed]