Artificial Intelligence and Venous Thromboembolism: A Narrative Review of Applications, Benefits, and Limitations

Abstract

Background

Venous thromboembolism (VTE), including deep vein thrombosis and pulmonary embolism, remains a leading cause of cardiovascular morbidity and mortality. Artificial intelligence (AI) holds promise for potential improvement of risk stratification, diagnosis, and management of VTE.

Summary

This narrative review explores the applications, benefits, and limitations of AI in VTE management. AI models were shown to outperform conventional methods in identifying high-risk candidates for VTE prophylaxis treatments in several postsurgical settings. It has also been demonstrated to be efficient in the early detection of VTE events, particularly through point-of-care AI-guided sonography and computer tomography image processing. Data biases, model transparency, and the need for regulatory frameworks remain significant limitations in the full integration of AI into clinical practice.

Key Messages

AI has the potential to improve VTE care by enhancing risk stratification and diagnosis. The integration of AI-driven models into clinical workflows has the potential to reduce costs, streamline diagnostic processes, and ensure effective management of VTE. Safe and effective integration of AI into VTE care requires addressing its limitations, such as interpretability, privacy, and algorithmic bias.

Introduction

Venous thromboembolism (VTE), most commonly including deep vein thrombosis (DVT) and pulmonary embolism (PE), remains a leading cause of cardiovascular morbidity and mortality [1]. In Europe and North America, VTE incidence is estimated at 1 to 2 cases per 1,000 individuals annually [2]. This condition imposes a substantial healthcare burden, both acutely and in the long term, leading to complications such as post-thrombotic syndrome, chronic thromboembolic pulmonary hypertension, and portal hypertension [3–6].

To mitigate this burden, numerous risk assessment models and prediction algorithms have been developed to guide prevention and diagnosis in various clinical settings [7, 8]. Additionally, traditional clinical decision rules – exemplified by the Wells and Geneva scores – alongside laboratory tests like D-dimer and imaging modalities including ultrasonography, CT, and V/Q scanning, play pivotal roles in early detection and management [9–15]. Despite widespread use, these predictive models are limited by static evaluations, subjective interpretations, and reduced generalizability [12–15], with area under the curve (AUC) values typically around 0.7 [16–18].

The emergence of artificial intelligence (AI) holds promise for addressing the challenges inherent in VTE risk stratification, diagnosis, and management. By integrating data-driven machine learning (ML) algorithms, natural language processing, and imaging analytics, AI can potentially improve risk prediction accuracy, reduce diagnostic uncertainty, and optimize therapeutic decision-making [19]. AI has transformed medicine, as seen in its ability to enhance diagnostics in radiology [20], predict atrial fibrillation in cardiology [21], and advance precision medicine in cancer treatment [22]. These examples demonstrate AI’s potential to improve VTE risk prediction, diagnosis, and management. This review examines the evolving landscape of AI-driven solutions for VTE care, discussing their current applications and potential advantages over conventional tools.

Glossary of AI Terms and Models

General AI Terms

Artificial Intelligence

AI is a technology that enables machines to perform tasks usually requiring human intelligence. These tasks include decision-making, understanding and generating language, recognizing patterns, and learning from previous experiences [23, 24].

Machine Learning

ML is a form of AI where systems learn from data to enhance their performance over time without being explicitly programmed [25]. ML systems analyze large amounts of data, recognize patterns, and make decisions or predictions accordingly [26].

Natural Language Processing

Natural language processing is a field of AI concerned with the computational processing of natural language, including understanding, interpretation, and generation of text and speech [27].

Large Language Models

They are designed to understand and generate text that closely mimics human language. They achieve this by training on massive amounts of text data and employing a specialized architecture called transformers [23, 28]. Unlike traditional models that process data sequentially, transformers handle sequential data in parallel, using word interaction to simultaneously assess relationships between all words in a text, resulting in significantly improving MUDd speed and efficiency [29–31].

Basic Concepts of ML

Statistics primarily focuses on inferring relationships through probabilistic models, aiming to understand systems and quantify confidence in findings. In contrast, ML prioritizes prediction, identifying patterns in high-dimensional data with minimal assumptions.

ML models learn relationships between input features (e.g., signals and images) and target variables, known as labels [26]. For example, given clinical and demographic data for healthy individuals and individuals with a disease, the model learns to map these features to the corresponding health status label (healthy or diseased) [32]. Often, these models learn this connection based on the training data and further validate their prediction ability on the test data [33].

Machine Learning Models

Linear regression is a method used to find the optimal weighted sum of input features that best predicts the label [34].

Generalized linear regression (GLR) extends regular linear regression to enable nonlinear relationships between features and labels. This makes GLR models suitable for a wider range of prediction tasks compared to linear regression [35].

Decision trees (DTs) use data features to predict the value of a label by splitting the data into branches based on specific feature values. Each branch represents a decision rule, and the leaves show the predicted label [36].

Random forest (RF) improves the accuracy and robustness of predictions by combining multiple DTs, each trained on different parts of the data. The final prediction combines the results from all the trees, using a majority vote for classification tasks or averaging the tree predictions for numerical labels. This process reduces errors and makes the model more reliable and accurate compared to a single tree [37].

Artificial neural networks (ANNs) are models for classifying data and identifying patterns within it. They are built of interconnected layers of units often termed “neurons” that transfer information to each other. Input data are processed by these neurons, and the network learns by adjusting the connections between them to minimize the difference between its output and the correct answers [25, 38, 39].

Convolutional neural networks (CNNs) are powerful neural networks designed to learn spatial hierarchies of features, making them ideal for image-based tasks. They use convolution layers, pooling layers, and fully connected layers to achieve high accuracy in classification problems. While ANNs are preferred for limited datasets and nonimage inputs, CNNs dominate computer vision and image-dependent ML applications [40].

Metric Definitions

ROC analysis is a popular method for evaluating the accuracy of medical diagnostic systems. In clinical epidemiology, it quantifies how well a test distinguishes between 2 patient states: positive (typically “diseased”) and negative (typically “non-diseased”). The ROC curve provides a graphical representation of diagnostic performance, allowing for the assessment and comparison of multiple tests. It plots the true-positive rate (sensitivity) against the false-positive rate (specificity), illustrating the balance between detecting true-positive cases and minimizing false-positive cases [41, 42].

The AUC quantifies a model’s ability to distinguish between positive and negative cases, representing its overall discriminatory power. An AUC of 1.0 indicates perfect discrimination, whereas an AUC of 0.5 suggests performance equivalent to random selection [41, 42].

Methods

A search of the published literature was conducted in November 2024 using PubMed, Web of Science, and Scopus. Our search targeted original studies at the intersection of AI and VTE, employing a set of search terms relevant to both fields. AI-related search terms included various ML models, while VTE-related search terms included DVT and PE.

We included peer-reviewed, original, English-language articles that were pivotal or seminal in addressing the phenomenon of interest, along with other relevant manuscripts. Articles that were not related to AI applications in VTE or were not original research were excluded.

Using AI for VTE Prediction and Primary Prevention

ML models offer a new approach to identifying at-risk patients and guiding more precise and adaptable prophylactic measures [28]. Several studies have explored the potential of AI models for VTE prediction and primary prevention (Table 1).

Table 1.

A summary of the reviewed articles about using AI for VTE prediction and primary prevention

First author	Year of publication	Journal	Aim	Participants, n	Primary outcome
Chen [43]	2024	Arteriosclerosis, Thrombosis, and Vascular Biology	Examine the performance of ML model in the prediction of 1-year risk of VTE	159,000	AUC for 1-year risk prediction-ML model: 0.64–0.78
					Padua: 0.54–0.65
Chen [44]	2024	Medicine	Examine the performance of RF, ANN, and GLR, in the prediction of post-gynecological laparoscopy VTE	489	AUC predictive performance for DVT-RF: 0.862
					ANN: 0.813
					GLR: 0.709
Lin [45]	2024	Journal of Obstetrics and Gynecology Research	DT and LR models were developed to predict the occurrence of VTE after hysterectomy for gynecological malignancies	1,087	AUC predictive performance for DVT-DT: 0.950
					LR: 0.722
Zhou [46]	2024	Scientific Reports	A multivariate LR model was developed to predict DVT in gastric cancer patients after undergoing radical gastrectomy	693	AUC predictive performance for DVT: 0.875
Katiyar [47]	2023	Clinical Spine Surgery	Testing 6 ML models in predicting VTE risk in spine surgery patients	63	VTE predictive accuracy rate
					RF: 88.89%
					Simple logistic: 84.13%
Huang [48]	2023	BMJ Open Quality	Evaluate the effectiveness of an AI-CDSS designed to assist healthcare professionals in clinical decision-making in reducing HA-VTE	19,785	The group which utilized an AI-CDSS in addition to standard prevention measures showed a significant 46% reduction in HA-VTE incidence compared to the group which followed conventional VTE prevention protocols

AI-CDSS, AI Clinical Decision Support System; HA-VTE, hospital-associated VTE.

An ML model trained and validated in the Mount Sinai Data Warehouse on over 159,000 participants to predict the 1-year risk of VTE outperformed traditional risk assessment tools, such as the Padua score. The models demonstrated robust performance when tested on external datasets of over 500,000 participants, with AUC scores ranging from 0.64 to 0.78 for 1-year risk prediction, while Padua AUC scores ranged from 0.54 to 0.65. Additionally, the ML models showed consistent performance across different patient subgroups, including hospitalized patients, surgical patients, patients of different ethnicities, and patients with cancer [43].

Hospitalized and, in particular, surgical patients are at increased risk of VTE [49]. Several studies have explored the potential of AI models in improving the assessment of VTE risk following surgical gynecological procedures. Chen et al. [44] examined the performance of 3 ML models in the prediction of post-gynecological laparoscopy VTE using the RF, ANN, and GLR. The RF model demonstrated the highest predictive performance for DVT, with an AUC ranging from 0.851 to 0.862. Lin et al. [45] included 1,087 patients who underwent hysterectomy for gynecological malignancies. In this study, a univariate logistic regression (LR) analysis was first used to identify risk factors associated with VTE occurrence within 30 days after the surgery. Next, DT and LR models were developed to predict the occurrence of postoperative VTE. The DT model demonstrated the highest performance in predicting VTE, achieving an AUC of 0.950.

ML applications were demonstrated to be efficiently applied in the prediction of DVT in other surgical fields as well. A retrospective case-control study of 693 gastric cancer patients undergoing radical gastrectomy analyzed 49 clinical indicators across baseline, preoperative, surgical, and pathological data. Following the identification of 14 potential risk factors for DVT, a predictive model for postoperative DVT was developed using multivariate LR, achieving high accuracy with an AUC of 0.875 in the test set [46]. In spine surgery patients, a retrospective cohort study analyzed 113 features from 63 individuals, testing 6 ML models in predicting VTE risk. The RF model demonstrated the highest accuracy, correctly classifying 88.89% of instances as either having or not having VTE. It also demonstrated a high VTE predictive ability of 93.75% [47].

The effectiveness of an AI Clinical Decision Support System, a big data governance system designed to assist healthcare professionals in clinical decision-making, was evaluated for its effectiveness in reducing hospital-associated VTE. Almost 20,000 adult hospitalized patients were randomly assigned to either the control or intervention group. The control group followed conventional VTE prevention protocols, while the intervention group utilized an AI Clinical Decision Support System in addition to standard prevention measures. The intervention group showed a significant 46% reduction in hospital-associated VTE incidence compared to the control group [48].

In summary, ML demonstrates significant potential for improving VTE prediction and guiding prophylactic measures. The studies highlight common important features, including age, BMI, immobility, prior VTE, malignancy, recent surgery, and D-dimer levels. These factors consistently contribute to model performance, emphasizing their relevance in future AI risk stratification efforts. While ML models have shown promising performance in various settings, including outperforming traditional risk scores and achieving high accuracy in surgical populations, further research is needed to validate these findings in larger, more diverse cohorts and to assess the impact of AI-driven interventions on clinical outcomes. Nevertheless, these advancements suggest that AI holds considerable promise for improving VTE prevention.

AI Use for VTE Diagnosis

Advances in AI offer the potential to enhance DVT early detection by mitigating traditional diagnostic methods’ limitations, such as reducing false positives and negatives, radiation exposure, contrast-related risks, and dependence on operator expertise [12–15]. Table 2 provides a summary of the reviewed articles about using AI for VTE diagnosis.

Table 2.

A summary of the reviewed articles about using AI for VTE diagnosis

First author	Year of publication	Journal	Aim	Participants, n	Primary outcome
Nothnagel [50]	2024	BJGP Open	Assess AI-assisted POCUS for DVT diagnosis using a guidance app to help nonspecialists perform handheld ultrasound exams	91	75% of AI-guided POCUS scans achieved adequate image quality
					Diagnosis by remote experts based on ultrasound images captured by nonspecialists
					Sensitivity: 100%
					Specificity: 91%
Seo [51]	2023	Scientific Reports	CNN-based AI models were assessed for detecting and localizing iliofemoral DVT lesions in CTA of the lower extremities	Training: 114	Identifying DVT
				Testing: 38	Sensitivity: 80%
					Precision: 65%
Ayobi [52]	2024	Clinical Imaging	Evaluate the performance of an AI model for detecting PE in CT pulmonary angiography scans	1,204	Sensitivity: 93.9%
					Specificity: 94.8%
Valente Silva [53]	2023	Revista Portuguesa de Cardiologia	Evaluate deep learning model performance to diagnose acute PE by using only 12-lead ECG data	Training: 1,014	Sensitivity: 50%
					Specificity: 100%

A recent study evaluated the potential of AI-assisted point-of-care ultrasound (POCUS) for diagnosing DVT, comparing its effectiveness to traditional ultrasound performed by specialists. The researchers used a guidance app, which assists nonspecialist healthcare providers through the ultrasound procedure using a handheld ultrasound probe. Participants with suspected DVT underwent AI-guided POCUS, followed by standard ultrasound scans performed by trained professionals for comparison. The quality of the ultrasound images captured by the nonspecialists was evaluated by remote experts who provided opinions regarding the technical performance achieved. Among 91 participants, 75% of AI-guided POCUS scans achieved adequate image quality. Remote diagnosis enabled 100% diagnostical sensitivity (correctly identifying DVT cases) and 91% specificity (correctly ruling out DVT when it was not present) [50].

AI was investigated for detecting and localizing iliofemoral DVT lesions in computed tomography angiography (CTA) of the lower extremities. CNN-based AI models were trained on CT scans from 114 patients and tested on CT scans from 38 other patients. The models demonstrated high sensitivity, correctly identifying over 80% of DVT cases, but exhibited lower precision, with approximately 65% of the cases identified as DVT by the model being true positives [51].

A study also evaluated the performance of an AI model for detecting PE in CT pulmonary angiography scans. The diagnosis was determined by three expert radiologists, and the AI’s interpretations were compared against these findings. The AI achieved a high sensitivity of 93.9% and specificity of 94.8% per scan when compared to the experts’ opinion [52]. Another study described a deep learning model aimed to diagnose acute PE by using only 12-lead electrocardiogram (ECG) data. The model was trained by using ECG data from 1,014 emergency department patients suspected of having PE. When tested, the model demonstrated 100% specificity, correctly identifying all patients who did not have acute PE but moderate sensitivity (50%), correctly identifying only half of the patients who were eventually diagnosed with PE [53].

In conclusion, AI demonstrates promising potential for enhancing the early detection of VTE, including both DVT and PE. Studies utilizing AI-assisted POCUS, CTA analysis, and even ECG interpretation have shown varying degrees of success. Imaging-based methods, such as AI-assisted POCUS and CTA analysis, have generally demonstrated high sensitivity and specificity. However, the AI model for ECG-based detection of PE, while achieving 100% specificity, had a moderate sensitivity of only 50%. Notably, most AI applications for VTE diagnosis have relied on CNN-based models. While vision transformers have shown success in other medical imaging tasks [54, 55], their use in VTE detection remains limited and warrants further exploration.

Large Language Model Roles in VTE Management

Large language models (LLMs), such as OpenAI’s ChatGPT and Google’s Gemini, are relatively new ML models that have gained significant popularity since the introduction of ChatGPT in late 2022 [56]. However, since LLMs are not specifically designed for medical purposes, their widespread adoption in healthcare raises concerns about their reliability [23]. Table 3 offers a summary of the reviewed articles about using LLMs for VTE management.

Table 3.

A summary of the reviewed articles about using LLMs for VTE management

First author	Year of publication	Journal	Aim	Sample size	Primary outcome
Duey [57]	2023	The Spine Journal	Evaluate the ability of GPT models to improve thromboembolic prophylaxis in spine surgery by assessing their concordance with clinical guidelines for antithrombotic therapies	12 questions	Accuracy rate
					ChatGPT-3.5: 33%
					GPT-4.0: 92%
Rosenzveig [58]	2024	Journal of the American Heart Association	Evaluate ChatGPT‐4.0 and Google Geminis’ ability to answer patients’ questions about PE	15 questions	Accuracy rate
					Gemini: 67%
					GPT-4.0: 93.3%
Sarangi [59]	2024	The Indian Journal of Radiology and Imaging	Evaluate the accuracy and reliability of LLMs in providing clinical decision support for initial imaging in suspected PE cases	---	Accuracy rate
					Open-ended questions: 0.58–0.83
					Select all that apply questions: 0.56–0.96

GPT models were evaluated on their ability to improve thromboembolic prophylaxis in spine surgery. Questions from the 2009 North American Spine Society (NASS) clinical guidelines for antithrombotic therapies were introduced to ChatGPT and evaluated for concordance with the guidelines. ChatGPT responses were considered accurate if they were consistent with the clinical guidelines. GPT-4.0 demonstrated high performance, achieving an accuracy rate of 92% [57].

LLMs were also shown to be effective in delivering medical information about PE to patients. A recent study evaluated different LLMs’ ability to answer patients’ questions about PE. The researchers created 15 commonly searched patient questions using Google Trends data. Responses from the models were reviewed by a panel of physicians specializing in vascular medicine, cardiology, and hematology. ChatGPT-4.0 achieved a 93.3% accuracy rate, as evaluated by the expert group [58].

To evaluate the accuracy and reliability of LLMs in providing clinical decision support for initial imaging in suspected PE cases, questions based on PE case scenarios were presented to the models. The responses were then evaluated by comparing them to the American College of Radiology Appropriateness Criteria. The accuracy rate for open-ended questions ranged from 0.58 to 0.83, while the accuracy rate for select all that apply questions ranged from 0.56 to 0.96 [59].

A recent manuscript showed that LLMs can also engage in detailed conversations about DVT, offering accessible health information. ChatGPT interacted in a conversation and explained DVT’s causes, symptoms, treatments, and preventive measures in simple terms. The ChatGPT responses emphasized risk factors such as immobility, surgery, and lifestyle choices while suggesting clinical interventions such as anticoagulants and thrombolytic therapy [60]. While LLMs have demonstrated high accuracy in some areas, their performance in more complex clinical scenarios, like imaging selection in suspected PE, requires further investigation.

Discussion

This review has explored the field of AI in VTE management, encompassing prediction, primary prevention, diagnosis, and the emerging role of LLMs. The reviewed studies demonstrate AI’s considerable potential to enhance various aspects of VTE care but also highlight important limitations that must be addressed for safe and effective clinical integration.

Limitations of AI in VTE

Despite its promising potential, the current application of AI in medical care has many limitations. First, it is difficult for clinicians to trace the AI models’ decision-making process, as the models often function as “black boxes,” meaning they do not provide an explicit understanding of their internal workings. This can limit clinicians’ trust and adoption of the models [61, 62].

Additionally, some models are trained and validated on patients’ EHR, raising concerns about data privacy and security [63]. De-identification techniques can reduce privacy risks, but they can also limit the training and test data, which can reduce model accuracy [64].

AI models are often trained on biased datasets in which certain population groups, such as minorities, are underrepresented. As a result, AI algorithms may be biased and fail to provide accurate predictions for these groups [65, 66]. Finally, as physicians’ reliance on AI increases, they may become less involved in patient management, which could harm their ability to explain medical plans to patients and reduce clinician-patient communication [67].

Future Developments and Applications

The initial AI applications were limited to basic pattern recognition tasks [68]. As data volumes and computational power have increased, ML models have advanced to support complex diagnostics, personalized treatment plans, and predictive analytics [69]. In the context of VTE, AI can play a critical role in improving diagnostic precision, enhancing clinical efficiency, and enabling earlier interventions for acute cases [70].

In addition, AI can potentially reduce healthcare costs by improving diagnostic accuracy and optimizing workflows [71]. According to the National Bureau of Economic Research, wider adoption of AI could lead to savings of 5 to 10 percent in US healthcare spending, roughly USD 200 billion to USD 360 billion annually [72]. VTE poses a substantial burden on the US healthcare system, with medical costs estimated at USD 5–10 billion per year [73]. Implementing appropriate thromboprophylaxis is not only a life-saving intervention but also a cost-effective strategy that could further reduce healthcare expenditures [48, 74].

AI’s performance in VTE has mainly been studied in PE and DVT, as well as the general patient population, with limited focus on unique clinical scenarios or specific subpopulations. There are less common forms of VTE, such as renal vein thrombosis and portal vein thrombosis, that have not yet been adequately studied using AI. Further research is also needed in specific subpopulations (e.g., pregnant women, children).

To integrate ML models into clinical practice for VTE, future research should focus on enhancing the transparency and accountability of these models [75]. Increasing attention is being directed toward explainability tools such as SHapley Additive Explanations (SHAP) and Local Interpretable Model-agnostic Explanations (LIME), which can clarify the predictions made by complex AI models. These tools enhance model transparency by providing insights into how individual input features influence the model’s decisions. Incorporating explainability techniques is crucial for fostering clinician trust, facilitating the validation of AI outputs, and ensuring safe adoption of AI models in VTE management [76, 77]. These tools can clarify how key VTE risk factors, such as D-dimer levels, immobility, or recent surgery, drive AI predictions. Future research should prioritize integrating explainability methods into AI systems to promote their clinical acceptance and reliability.

In addition, robust security measures are essential to ensure the privacy and confidentiality of patient data accessed and analyzed by AI models within EHR systems [78]. It is also important to ensure that the models are continuously monitored and updated, reflecting the latest advancements in medical research, clinical guidelines, and treatment protocols [79]. VTE management is a field with frequent updates to clinical guidelines and new evidence emerging regularly. If AI models are not updated, outdated recommendations could lead to suboptimal treatment decisions.

Finally, legal regulations must be established to ensure its safe and ethical use. These regulations can help protect patient privacy, ensure data quality, and maintain the transparency of the ML models [80]. An example of recent regulation is the European Union (EU) AI Act, which was approved in March 2024. This regulation addresses several key areas, including ensuring the safety of AI systems and protecting fundamental rights [81]. The World Health Organization (WHO) and the International Telecommunication Union (ITU) also collaborated in 2023 to promote the safe development and use of AI in healthcare through a Focus Group. The group has explored key regulatory considerations, providing an overview covering areas such as AI transparency, data quality, and patient privacy [82]. As AI continues to evolve, the need for more extensive legal regulation becomes critical.

In conclusion, the use of AI in VTE management has the potential to improve risk prediction and VTE prevention, diagnostic accuracy, and support in treatment decision-making. AI models were demonstrated in several studies to perform better than conventional methods in identifying high-risk patients who will benefit from prophylaxis treatments for VTE. Some models demonstrated impressive performance in their ability to early detect VTE events, enabling early diagnosis and treatment. However, before implementing AI tools in clinical practice, some limitations must be carefully addressed to ensure the models are transparent, unbiased, and safe, while preserving clinician autonomy and patient trust.

Conflict of Interest Statement

The authors have no conflicts of interest to declare.

Funding Sources

This study was not supported by any sponsor or funder.

Author Contributions

A.M. and O.E. conceived and designed the review. A.M. conducted the literature search, organized the material, and drafted the manuscript. O.E. supervised the project and provided critical revisions to the manuscript’s content. Both authors approved the final manuscript.

Funding Statement

This study was not supported by any sponsor or funder.