This study adheres to PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses; Checklist 1) guidelines for conducting systematic reviews [18]. The study protocol was registered on PROSPERO (International Prospective Register of Systematic Reviews; registration number: CRD42024625842).

A comprehensive search was conducted in both Chinese and English databases, including Web of Science, Embase, Scopus, CNKI, Wanfang, Vipro, and PubMed, from the time of their formation until December 15, 2024. The following search terms were applied: “joint arthroplasty,” “venous thromboembolism,” “machine learning,” “risk prediction,” “prediction models,” and their synonyms and variant terms. A manual search of reference lists containing studies to find related publications was conducted. For detailed search strategies, see Multimedia Appendix 1, which includes keyword combinations and database settings.

The following were the inclusion criteria: (1) studies on patients after joint replacement; (2) observational studies, case-control studies, cohort studies, and randomized controlled trials; (3) postoperative VTE risk was predicted using an ML model; (4) studies that provide performance evaluation metrics (eg, accuracy, sensitivity, specificity, area under the curve [AUC]) for the model and give clear statistical analysis results; and (5) studies published in English or Chinese.

The following were the exclusion criteria: (1) review articles, case reports, opinion pieces, commentaries, conference abstracts, and other nonoriginal research; (2) studies that do not mention particular kinds of ML models or algorithms; (3) animal experiments, simulation studies, and other nonclinical research; and (4) studies for which the entire text was not retrievable.

The retrieved results were imported into EndNote 21, and duplicates were eliminated. Two researchers trained in systematic evaluation techniques independently conducted the study selection process based on preestablished eligibility criteria. First, the titles and abstracts of the literature were screened, and the studies failing to meet the eligibility criteria were initially excluded. Second, it was filtered through full-text reading, and if 2 researchers disagreed, a third researcher stepped in to reach an agreement.

The data extraction process was also performed independently by 2 researchers following CHARMS (Critical Appraisal and Data Extraction for Systematic Reviews of Prediction) guidelines [19], and a third researcher settled any discrepancies. There are 2 groups into which the data taken from the chosen studies are separated: (1) basic information: author, year of publication, country, study design, participants, data source, clinical outcome, and sample size and (2) model information: missing data handling, candidate predictors, screening variable methods, prediction timepoint, prediction horizon, model development, internal verification, external verification, model performance, number of variates, calibration, and method of visualization.

The included studies’ applicability and bias risk were evaluated using the PROBAST (Prediction Model Risk of Bias Assessment Tool) tool. The assessment has 20 questions, including 4 basic areas: participants, predictors, outcomes, and analysis. There are answers to the questions in each domain: “yes/probably” (low risk of bias), “no/probably not” (high risk of bias), and “no information” (unclear). A domain was deemed to have a high risk of bias if its answer contained “no/probably not.” If every question had a “yes/probably yes” response, the domain was deemed to have a low risk of bias. When there is “no information,” it is deemed to be highly biased if it cannot be explained by the original or additional material. If each domain was considered low risk, then the total bias risk was considered low. Furthermore, the total risk of bias was deemed high if any domains were deemed to be at high risk.

The bias risk in each model is assessed independently by 2 researchers, and if there are differences, they will be discussed to reach an agreement. Due to the majority of research using various models or having several outcomes, the bias risk is evaluated independently for each model and outcome. Finally, since there are no differences in the risk of bias between models in a single study, each study only shows the result once.

The study selection process is detailed in the PRISMA flow diagram (Figure 1). A total of 4344 records were initially identified from 6 electronic databases (PubMed, n=2001; Embase, n=661; Web of Science, n=1415; Wanfang, n=112; CNKI, n=73; and Wipro, n=82). After removing 1466 duplicates, 2878 unique records remained for screening. During the title and abstract screening phase, 2863 records were excluded because they were either unrelated to the topic (n=2,641) or were nonoriginal research, such as reviews and newspaper articles. This left 15 reports to be sought for full-text retrieval. The subsequent critical appraisal excluded 6 studies based on predefined criteria: 3 did not use an ML approach, 2 had incomplete outcome data, and 1 did not involve a model development process. Ultimately, 9 studies met the full inclusion criteria and were included in this systematic review.

Table 1 shows the basic characteristics of the 9 studies. These studies were published from 2018 to 2024 and included 1 prospective study [3] and 8 retrospective studies [20-27]. Among them, 4 studies [3,20,21,25] were conducted in China; 3 studies [22-24] were from the United States; 1 was conducted in the Netherlands; and 1 transnational study used data from the Netherlands, the United Kingdom, and Denmark. In terms of study population, 2 studies [3,20] focused on patients with THA, 1 study [21] targeted patients with TKA, and the remaining studies [22-27] focused on both THA and TKA. In terms of clinical outcomes, 2 studies [20,21] concentrated solely on DVT, while the remaining 7 studies [3,22-27] evaluated both DVT and PE. In addition, the studies’ sample sizes ranged from 60 to 3,92,661. Cases of VTE ranged from 24 to 4042, with an incidence of 0.9% to 40%.

Table 2 provides a detailed summary of the methodological design and data characteristics of the included studies. Table 3 presents a summary of the performance of the model development, validation, and inclusion of the studies. Among the 9 studies included, 5 studies [3,20,23-25] evaluated multiple ML algorithms, and 4 studies [21,22,26,27] used a single algorithm. Across these studies, 15 kinds of ML algorithms were used to develop 34 predictive models, and each study developed 1-5 models. Logistic regression (LR) and XGB were the most common algorithms, appearing in 5 studies (5/9, 56%) each. This is followed by RF (3/9, 33%) [23-25] and SVM (3/9, 33%) [20,24,25]. In particular, the XGB model showed strong performance in predicting DVT and PE, with an AUC range of 0.71‐0.982. Figure 2 details the frequency of the model used.

The included studies used a variety of metrics to evaluate ML models, of which AUC was the most common (8/9, 89%). The predicted AUC range for DVT and PE was 0.579‐0.982. One study did not report AUC but reported C-statistic [22]. XGB, popularized by Chen and Guestrin [28] in 2016, is a widely adopted ML algorithm on the basis of gradient boosting [29,30]. The included XGB studies [3,20,21,24,25] reported AUC ranging from 0.71 to 0.982, demonstrating strong model performance. LR estimates the association of one or more independent variables with binary dependent variables [31,32]. The included LR studies [3,20,25-27] reported AUC ranging from 0.579 to 0.944. RF is a widely used ML algorithm that helps mitigate overfitting and improves the model’s stability and accuracy by aggregating the outcomes from various decision trees [33]. It functions as a modeling algorithm as well as a variable selection technique. Three studies [23-25] used RF, with reported AUC ranging from 0.698 to 0.907. In addition, calibration plots were provided in 3 studies [22,24,27]; Brier scores were provided in 1 study [20]; and 1 study [26] also provided Hosmer-Lemeshow, calibration plot, and Brier score. The other 4 studies did not include data on calibration performance. Additional reported metrics encompass accuracy, sensitivity, and specificity.

These 9 studies differ in terms of the optimal timing for risk assessment. Six studies have created preoperative prediction models [3,22,23,25-27] to identify high-risk individuals before surgery. The last 3 studies focused on postoperative prediction, assessing immediate risks using postoperative data [20,21,24]. Additionally, there is diversity in the risk assessment’s predictive horizon. Its duration ranges from a brief period following the procedure, such as “within 3‐5 days after surgery” or “during the postoperative hospital stay” [3,20], to a lengthy risk assessment that might continue for up to a year following the procedure [26]. Common time horizons include 30 days [22,23] and 90 days [24,27].

From a low of 5 [26] to a high of 66 [25], the number of predictor variables that ultimately made it into the model varied widely. The descriptive analysis of the predictive factors used in each study (as shown in Figure 3) shows that a constant core set of variables is regularly utilized, despite the wide variations in the number of predictive factors used in each investigation. These variables can be broadly classified into four categories: (1) demographic and biological characteristics of the patients (such as age, gender, BMI); (2) major comorbidities (such as hypertension, diabetes, coronary heart disease, and malignant tumors); (3) specific VTE risk histories (such as previous VTE); and (4) key laboratory indicators (such as D-dimer).

All of the models in the included studies underwent internal validation; however, only 2 studies [22,27] used external validation and reported the AUC value of external validation. The main methods used for internal validation were k-fold cross-validation and random-split validation. External validation plays a key role in confirming whether a model can maintain good performance in different populations and clinical settings. The absence of external validation will limit the generalization and clinical application of these models.

According to the PROBAST evaluation, all models had a high risk of bias (Table 4). The assessment results showed that the main sources of bias were concentrated in the “participants” and “analysis” domains. In the participants domain, 8 studies had a high risk of bias, mainly because these studies relied on retrospective data. In the analysis domain, all models had a high risk of bias. The PROBAST assessment pointed out several specific methodological issues: 4 studies [21,23-25] did not perform model calibration, 4 studies [3,21,23,25] did not report the handling of missing data, and 2 studies [21,23] were insufficient in measures to prevent model overfitting. In contrast, the outcome domain was rated as having a low bias risk level in all studies. In terms of applicability assessment, only 1 study [25] had poor applicability because the model was overly complex and required 66 predictor variables.

VTE is a common and serious postoperative complication that is preventable [10]. Risk assessment and management-based strategies can help prevent or reduce the occurrence of VTE [34]. Therefore, selecting an appropriate predictive model for VTE is crucial. This systematic review examined the methodological quality and clinical applicability of ML methods in predicting VTE after joint replacement surgery. Among the reviewed studies, 15 ML algorithms were utilized to construct 34 predictive models. XGB and LR were the most used model types, followed by RF and SVM. The AUC values of the models ranged from 0.579 to 0.982, indicating that the discrimination ability of the models varied greatly, and some of the models showed excellent performance. Two research [1,6] reports demonstrated extremely high discrimination ability, with their AUC values exceeding 0.9. Such high performance often raises concerns about overfitting, where the model learns specific statistical noise specific to the training data rather than the generalizable underlying biological signals, leading to an optimistically biased performance estimate. Moreover, these models have not yet undergone external validation. This deficiency is a major limitation. In the clinical environment of joint replacement, there may be significant differences among patients, and surgical techniques as well as VTE prevention protocols may also vary among different medical institutions. Models trained based on data from a single hospital are likely to learn the specific treatment patterns of that hospital rather than a universally applicable VTE prediction signature. This may lead to unreliable prediction results when the model is applied to different patient groups. The risk of overfitting and the lack of reliable external validation seriously undermine the clinical value of these highly performing models. These models can only be considered for practical application after undergoing extensive external validation.

Another finding is that the XGB model demonstrates superior predictive performance compared to other models. For instance, 2 studies [3,20] compared the XGB and SVM models with the LR models, and the results showed that the XGB and SVM models outperformed the LR model. This supremacy is probably due to the smart way the algorithm was made. First, XGB has strong built-in features (such as regularization) that keep the model from overfitting when working with small datasets, making it stable even when the data are noisy. Second, XGB is very good at identifying the most important risk factors on its own. It can also capture complicated nonlinear interaction relationships, which typical linear models cannot do. Furthermore, the fact that XGB functions so effectively most likely indicates that the research team spent a great deal of time and energy optimizing the hyperparameters and determining the ideal configurations for their particular dataset. However, research has indicated that LR models may be more successful and easier to apply in some situations, whereas XGB and SVM models may not necessarily perform better. LR’s fundamental structure makes it simple to comprehend and analyze. Additionally, it produces dependable findings and has strong convergence, particularly when used for small datasets [35]. The Caprini score is a clinically prevalent instrument utilized to evaluate VTE risk by analyzing many clinical risk factors, including age, surgical history, malignancy, and obesity [36]. One study [25] compared the Caprini Score to ML models and showed that the ML models were better at predicting VTE risk. Although the predictive performance of most models ranged from moderate to good, the risk of bias in all studies restricted the applicability of these models in clinical practice.

The PROBAST assessment results indicate that all the models have a high risk of bias, suggesting that their predictive performance in practical applications may be lower than the reported levels. This widespread bias is mainly due to the flaws in the research design and the statistical analysis methods adopted by the models. In terms of the research design, most studies rely on retrospective data, which increases the risk of selecting an unrepresentative patient population, thereby weakening the generalization ability of the models beyond the original dataset. Additionally, there are numerous flaws in the analysis aspects of the model construction and testing processes. The main issues are that some studies did not check the calibration process. This makes it impossible for us to determine whether the “30% risk of venous thrombosis” predicted by the model truly means a 30% probability in reality, thus preventing its use in clinical decision-making. Some studies also failed to provide enough protections against the “overfitting” problem, in which the model just remembered the noise in the training data instead of identifying the true risk patterns, while others did not provide solutions for handling missing patient data. These analytical flaws make these models likely to perform poorly when applied to new patients. Furthermore, the actual usability is also an issue. One model used an excessive 66 predictive variables, making it overly complex and unsuitable for clinical use. To raise the accuracy and clinical application of predictive models, future studies should apply more rigorous methodological methods. Moreover, the significant bias risk indicates inconsistency in the development and validation of current models, which could influence their widespread application. Future studies should give priority to enhancements in a number of important domains, including the research design (ie, cohort or case-control study design), the number of events, the method for predictor selection, the processing of complicated data, and the processes of model calibration and fitting.

One serious and widespread flaw in these studies is the absence of model calibration reports. Although most of the included studies reported high discrimination metrics such as AUC, the discrimination does not guarantee that the predicted probabilities are accurate. Calibration addresses a more fundamental clinical question: Does the predicted “30% risk of venous thrombosis” really mean that there will be a 30% incidence rate among a group of similar patients? Inaccurate calibration might have very harmful consequences. For example, a model that consistently overestimates risks may incorrectly identify patients with low risks as having high risks [37]. This may lead to the improper use of potent anticoagulant drugs, thereby exposing patients to a serious risk of bleeding, a risk that could have been avoided. However, if the model underestimates the risks, it may prevent doctors from taking necessary preventive measures to ensure the safety of patients with high risk. This could lead to potentially fatal thrombosis incidents [38]. Consequently, owing to the absence of calibration data, most of these models are unreliable for informing individual patient decisions, irrespective of their quoted AUC values. Future research must prioritize calibration as the principal measure of model performance. We strongly urge that studies include both visual calibration plots and quantitative indicators (such as the Brier Score) at the same time to make sure their predictions are accurate in the actual world. Additionally, most of the included studies did not provide reports on the sensitivity and specificity of their models. This affects the overall assessment of model performance, as the reported AUC values may not fully reflect its clinical applicability. Future research should follow standardized reporting guidelines to evaluate its model performance.

External validation is critical to confirm the generalizability of the model’s performance and its applicability to different populations [39,40]. However, all the studies used internal validation methods, such as K-fold cross-validation or random split validation, and only 2 studies conducted external validation. Most of these models lack external validation, further increasing the challenge of clinical implementation. In such a sensitive sector as health care, where predictive models are finally meant to be employed in real-world situations, these ML models may not accurately reflect their applicability and dependability without outside validation [41]. Future studies should incorporate external validation where possible, such as in different hospitals or using publicly accessible databases, to optimize the model generalization through data synthesis and larger datasets [42]. Additionally, we advise temporal validation to assess the model’s resilience at different points in time.

The ML model’s predictive variables found in this study have a number of important effects on clinical practice. First, the predictive factors frequently used in these models play a very important role in current clinical practice and future research. For instance, D-dimer is a well-known clinically validated biomarker for VTE. Studies have shown that it is highly sensitive in predicting the risk of VTE. Prothrombin time, activated partial thromboplastin time, and international normalized ratio are a few other markers that may assist physicians better understand the risk of blood clots and quickly identify and treat patients with high risk [43]. Another established risk factor for VTE is age. Research has indicated that the risk of VTE rises with age. People over the age of 60 years account for around 70% of occurrences of VTE [44]. Considering age while evaluating patients before surgery may enhance outcomes and diminish complication rates. Common chronic disorders, such as diabetes and hypertension, were also among recurring factors. Research has shown that these disorders can raise VTE incidence. A personalized anticoagulant treatment strategy might help those with past histories of these disorders lower their VTE risk even more [45]. Second, the results of the 9 studies on the predictors can guide the development of future predictive models and contribute to later studies on related risk elements. For instance, 1 study [20] included bone cement prosthesis as a predictor in the model, although further study is required to confirm its association with VTE. However, variance in the choosing and reporting of predictors in these studies may affect the comparability and practical utility of the models in different clinical settings.

To our knowledge, this is the first study to evaluate ML in predicting VTE in patients undergoing joint replacement surgery. Nevertheless, the study had certain limitations. First, most of the included studies have been conducted in China and the United States. This geographical concentration restricts the general applicability of the model to different areas. Second, numerous studies focused solely on developing predictive models without considering external validation or practical application of the model. The absence of external validation may affect the credibility of our pooled estimates. Third, since the traditional methods used to assess publication bias (such as funnel plots and Egger or Begg tests) are not applicable to AUC data, no quantitative assessment of publication bias was conducted for the included studies. However, the risk of publication bias remains: “positive” studies (eg, high AUC values) are more likely to be published than “negative” or inconclusive ones. Therefore, the studies included in this review may overestimate the true performance of these models. Finally, the included studies exhibited a high risk of bias and insufficient transparency. These methodological shortcomings could affect the accuracy of model predictions, therefore reducing the dependability and application of the conclusions.

The study identified 9 studies with a total of 34 ML models for predicting VTE after joint replacement. Most studies have modest to good discriminative ability from their AUC values. However, among these 9 studies, only 2 of them underwent effective external validation. The widespread neglect of external validation has left the generalization capabilities of these models, which performed well on the original dataset, completely unknown. In addition, these studies also show a high risk of bias and significant heterogeneity. Most of the studies had significant methodological shortcomings, including a lack of rigorous study design and absence of calibration measurements. These deficiencies will affect the reliability of the model in clinical applications and reduce the promotion value of the model. To improve the prediction capabilities of ML models, future research must closely follow the PROBAST criteria, which place a strong emphasis on meticulous study design and quality control. External validation is also necessary to improve the generalization and applicability of ML models in clinical settings.