Deep Learning for The Management of Deep Vein Thrombosis
Welcome to Health Connect, the podcast for health professionals where we will share the latest news and information on science and technology in the medical field. In this episode, we will talk about how artificial intelligence is changing the way professionals can predict the risk and complications of venous thromboembolism and major bleeding.
Did you know that, in the United States alone, more than 540,000 total knee arthroplasties and 230,000 total hip arthroplasties are performed annually?1 These numbers are projected to increase substantially by the year 2030.2
Venous thromboembolism, or VTE, and major bleeding are feared complications in total joint arthroplasty and are influenced by numerous patient- and surgical-related factors. VTE is associated with immense morbidity and increased cost of care and, therefore, risk stratification and prophylaxis are crucial. Knowing the individual risk factors associated with VTE is necessary; however, that alone is not enough for clinical decision-making and overall risk stratification. Besides, risk stratification for major bleeding remains lacking in total joint arthroplasty.1
To that end, a recent study by Shohat and collaborators used a contemporary large institutional database with granular data to develop and validate an algorithm suitable for everyday use in clinical practice that could predict the probability of developing VTE and major bleeding in total joint arthroplasty patients. The algorithm would take into account many variables, including sex, age, race, body mass index, and patient-reported past medical history, as well as Charlson Comorbidity and Elixhauser Comorbidity indexes.1
This retrospective cohort study collected the medical records of 37,948 patients who underwent either primary or revision total hip or knee arthroplasty between January 2009 and October 2020. Of those, 35,963 patients were included in the study. Out of which, 308 developed VTE, and 293 developed major bleeding, with significant differences in patient demographics, characteristics, comorbidities, anticoagulation medications, and operative factors between patients who developed the complications and those who didn’t.1
Throughout the study, two distinct outcomes were evaluated. The first was the occurrence of symptomatic deep venous thrombosis (or DVT) or pulmonary embolism (PE) within 90 days of surgery. To avoid including superficial clots that were not clinically significant, only patients with a documented diagnosis, confirmatory study, and treatment for VTE were considered to have met the primary endpoint. The second main outcome was the occurrence of major bleeding events as defined by the Scientific and Standardization Committee of the International Society on Thrombosis and Haemostasis.1
Before running the predictive algorithms, a set of descriptive statistics were performed to understand the data distributions. Here, patients with VTE and major bleeding were compared to controls who did not have VTE or major bleeding. Following the descriptive breakdown, various machine-learning methods were applied to determine specific variables that increased the chance of DVT, pulmonary embolism, or major bleeding. The four models tested were Random Forest, LASSO, Gradient Boosting Trees, and Support Vector Machines. However, due to the nature of the imbalance in the data, some of the models were also performed using down-sampling.1
Then, each model analyzed used repeated cross-validation techniques. To determine the best models from training, areas under the curve and precision curves were calculated. Once the “best” model was selected, the remaining data was tested on it to ensure it had the proper performance.1
Models were created for all VTE grouped together, as well as for PE and DVT individually. One important finding was that, while PE and DVT have many risk factors in common, there’s an advantage to using different models to predict each one. So, separate models were developed for DVT and PE prediction and were tested using repeated cross-validation. Gradient boosting trees had the highest performance in the prediction of both conditions and were chosen for algorithm development.1
Gradient boosting trees analysis showed that the 10 most important factors associated with pulmonary embolism were, by order of importance: active cancer, hyper-coagulopathy, blood transfusion, warfarin for VTE prophylaxis, older age, operative duration, revision surgery, history of VTE, atrial fibrillation, and underlying fracture.1
While for DVT, the ten most important factors were: hyper-coagulopathy, older age, allogenic blood transfusions, revision surgery, warfarin prophylaxis, simultaneous bilateral surgery, active cancer, active or former smoking, underlying fracture, and male sex.1
On the other hand, Lasso analysis for the entire cohort, which had the highest area under the curve for major bleeding events, showed that the 10 most important factors associated with major bleeding were: revision surgery, chronic use of warfarin preoperatively, operative duration, general anesthesia, peptic ulcer disease, allogenic blood transfusions, older age, knee joint procedure, varicose veins, and current or past smoking.1
Modifiable factors such as tranexamic acid, type of anesthesia, and type of prophylaxis are other relevant variables that previous studies or algorithms did not consider. Also, previous works assumed that when patients have multiple risk factors and calculated high risk, they could benefit from more potent chemoprophylaxis; however, recent studies have shown that is not the case. It is therefore interesting to note that the influence of these variables could be tested in real-time to examine whether they can change the individual patient’s risk.1
The algorithm presented by this team not only accurately predicts VTE and major bleeding, but it can also guide how different prophylactic measures may mitigate an individual’s risk, so it could be used by clinicians in practice to improve decision-making and patient counseling. All in all, the results of this study show that the use of these models can help in decision-making before surgery and in patient-specific management.1
CÁPSULA
Deep vein thrombosis is a common postoperative complication of knee or hip arthroplasty with rates higher than 70% in the absence of appropriate preventive measures. But even with standard preventive measures, the incidence rates of DVT still remain high after these procedures.2
The evaluation of DVT risk before surgery can help in the early identification of high-risk populations and the subsequent adoption of effective preventive measures, both pharmacological and non-pharmacological, to reduce the risk of complications, accelerate patient recovery, and promote a better utilization of healthcare resources. Machine learning can help with the prediction of high-risk populations and the individualization of preventive measures.2
FIN DE CÁPSULA
With improvements in intraoperative strategies and the adoption of effective and standardized perioperative preventive measures, such as early postoperative out-of-bed activities, physical prevention, and pharmacological prevention, DVT incidence has decreased significantly, yet it is still an important postoperative complication.2
Predicting the occurrence of DVT after total joint arthroplasty is essential, since severe DVT may cause subsequent pulmonary complications, such as PE, and is one of the major causes of death in the perioperative period. In addition, medical prophylaxis, may increase the risk of local ecchymoses, bleeding from wounds, anemia, and intracerebral hemorrhage.2
Currently, the methods most commonly used in clinical practice for predicting DVT include measurement scales developed based on statistical analyses of clinical data, like the Autar DVT scale, JFK Medical Center DVT risk assessment tool, Padua Prediction Score, the Risk Assessment Profile for Thromboembolism Scale, and Caprini Score for Venous Thromboembolism.2
Although these measurement scales are easy to use and score, they also present several shortcomings, such as low prediction accuracy, consideration of influencing factors, the inability to provide dynamic real-time early warnings, and the considerable time and effort required for scoring. Recently, however, with the advancing medical technology and the continuous accumulation of data, the clinical application of artificial intelligence has revolutionized diagnostics and treatment paradigms. AI-based prediction methods are already capable of effectively predicting the risk of peripherally inserted central venous catheter-related thrombosis, as well as thrombosis following total shoulder arthroplasty.2
Nonetheless, the feasibility of AI-based methods for predicting lower extremity DVT risk after knee or hip arthroplasty has not been fully studied. Therefore, a recent retrospective study by Wang and collaborators aimed to investigate the use of AI methods for predicting DVT after total joint arthroplasty using electronic health records and compare those results with traditional measurement scales represented by Caprini scores.2
The sample included patients who underwent knee or hip arthroplasty between January 2017 and December 2021, and underwent bilateral lower extremity venous ultrasonography examination during hospital stay after surgery.2
Input variables included information on demographics, clinical characteristics, laboratory test measurements, medication use, examinations, and comorbidities. All variables of the electronic health records, from hospital admission to the postoperative bilateral lower extremity venous ultrasonography examination, were extracted to form the feature dataset. The cohort was composed of 6,897 joint arthroplasty cases: 5,214 for knee and 1,683 for hip. Among them, 1,161 were positive DVT cases and 5,736 were negative DVT cases.2
Patient data was randomly assigned to two datasets: 80% for the training dataset; and the remaining 20% for the test dataset for model performance evaluation. The proportions of positive cases in the training and test datasets were identical. Each model was then subjected to five-fold cross-validation for the selection of the best parameter combinations from the candidate pools to obtain reliable and stable models. The cross-validation involved the random division of experimental data into five groups; four groups were used for model training and the other for model testing. After repeating the process five times, the team obtained five models and their corresponding evaluation results. The final model performance was determined by calculating the average values of the performance metrics of the five models.2
Different models were compared, those being: eXtreme gradient boosting, random forest, support vector machines, logistic regression, and backpropagation neural network. Since each machine learning algorithm solves the problems in a slightly distinct way, different algorithms may give different answers to the same problem. Thus, the investigators adopted an ensemble model by weighted voting, an ensemble of all the models, except the backpropagation neural network; with the highest probability among the four model predictions defined as the final prediction result.2
All models had an area under the curve above 0.88 except for the Caprini score. The Ensembled Model showed the highest area under the curve in prediction performance, which was 0.92. Sensitivity was the highest for the Caprini score, with 0.89, followed by the support vector machine model, with 0.84. All models had a specificity greater than 0.80, except for the Caprini score, the backpropagation neural network being the highest.2
The study found that machine learning models based on electronic health records can predict the risk of DVT after total joint arthroplasty with a better prediction accuracy than the traditional Caprini score. Also, high prediction accuracy can be achieved despite the lack of prior knowledge about relevant risk factors by including all variables of the electronic health records in model training. By combining the results of multiple machine learning models, a model with better-generalized performance and prediction effects can be obtained. All of the aforementioned models had better diagnostic efficiency than the traditional Caprini score.2
Since these machine learning models can provide a good prediction for the risk of DVT after knee or hip arthroplasty, they can be beneficial for the early identification of high-risk populations and the subsequent adoption of effective preventive measures, reducing the risk of postoperative complications, such as pulmonary embolism. Additionally, machine learning-based prediction tools can be embedded in clinical decision support systems within electronic health records to accomplish automatic hospitalization information extraction and real-time early warnings of DVT, which cannot be achieved with conventional risk assessment scales.2
In the field of thrombosis detection, current imaging modalities for DVT diagnosis include ultrasonography, MRI, catheter venography, and computed tomography angiography of the bilateral lower extremities, better known as LECTA. Among these, LECTA is superior because it provides more objective images than ultrasonography. Also, it provides information about extravascular tissues in the bilateral lower extremities and abdominopelvic region. In the clinical setting, to accurately decide on DVT in LECTA, adjacent slices should be considered rather than one CT slice with the suspicious existence of the lesion.3
However, to overcome the limitations and drawbacks of manual analysis, convolutional neural network-based artificial intelligence algorithms have been used in the medical imaging field as computer-aided diagnosis system tools. Some studies have been conducted using various imaging modalities and have shown potential and efficiency of AI for DVT diagnosis.3
One of these studies was done by Seo and colleagues. The team aimed to evaluate the performance of an AI algorithm in the detection of iliofemoral deep venous thrombosis on computed tomography angiography of the lower extremities, LECTA. To investigate the effectiveness of this method, they created synthesized images to consider practical diagnostic procedures and applied them to the convolutional neural network-based RetinaNet model. RetinaNet, a deep learning-based one-stage detection model that uses a focal loss function was employed because it has the advantages of time efficiency and high accuracy based on its function and structure. The model has demonstrated strong performance in addressing the foreground-background class imbalance, which is the main drawback of one-stage object detectors.3
RetinaNet has a feature pyramid network combined with the ResNet backbone, which has been applied and used in many detection models in medical imaging because it exhibits a high level of detection performance with minimal resource requirements for computation. Also, its structure has two distinct subnetworks, one performs regression for localization to the bounding box of the target object task, while the other performs object classification.3
Employing the synthesized images mentioned, the model outperformed the one-slice images in detecting iliofemoral DVT on LECTA. This AI algorithm could allow radiologists to achieve more accuracy by presenting them with a probable location with the existence of DVT. From there, the radiologists can confirm the locations suggested by the AI model in CT volume data, comprised of numerous slices images, which will, in turn, improve the reading efficiency and reduce the burden on these professionals.3
DVT commonly develops in the lower extremities and can cause complications, such as recurrent pulmonary embolism and venous thromboembolism, which increase mortality and decrease the quality of life. However, the treatment and long-term prognosis depend on an accurate and timely diagnosis, which might be delayed due to the absence of a radiologist on duty. Because the clinical symptoms and signs have low specificity for diagnosing DVT, an imaging workup is necessary to confirm or exclude the diagnosis. AI models can improve the diagnostic performance and reduce the burden of clinicians.3
As we’ve discovered, recent developments in machine learning have facilitated a more comprehensive, accurate, and user-friendly platform that may help clinicians in decision-making.1 This includes using machine learning models based on electronic health records, which provide a valid tool for predicting the risk of DVT after total joint arthroplasty.2
Thanks for joining us on this episode of Health Connect. Don't miss out on our next episode.
Discover more medical news on Viatris Connect.
Referências:
- 1. Shohat N, Ludwick L, Sherman MB, Fillingham Y, Parvizi J. Using machine learning to predict venous thromboembolism and major bleeding events following total joint arthroplasty. Sci Rep. 2023;13(1):2197. Available at: Using machine learning to predict venous thromboembolism and major bleeding events following total joint arthroplasty
- 2. Wang X, Xi H, Geng X, Li Y, Zhao M, Li F, et al. Artificial intelligence-based prediction of lower extremity deep vein thrombosis risk after knee/hip arthroplasty. Clin Appl Thromb Hemost. 2023;29:10760296221139264. Available at: http://dx.doi.org/10.1177/10760296221139263
- 3. Seo JW, Park S, Kim YJ, Hwang JH, Yu SH, Kim JH, et al. Artificial intelligence-based iliofemoral deep venous thrombosis detection using a clinical approach. Sci Rep. 2023;13(1):967. Available at: Artificial intelligence-based iliofemoral deep venous thrombosis detection using a clinical approach
Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.
NON-2023-9979
NON-2023-2512