Predicting Thromboembolism in Hospitalized Patients with Ventricular Thrombus

Background: Thromboembolism is associated with mortality and morbidity in patients with ventricular thrombus. Early detection of thromboembolism is critical. This study aimed to identify potential predictors of patient characteristics and develop a prediction model that predicted the risk of thromboembolism in hospitalized patients with ventricular thrombus. Methods: We performed a retrospective cohort study from the National Center of Cardiovascular Diseases of China between November 2019 and December 2021. Hospitalized patients with an initial diagnosis of ventricular thrombus were included. The primary outcome was the rate of thromboembolism during the hospitalization. The Lasso regression algorithm was performed to select independent predictors and the multivariate logistic regression was further verified. The calibration curve was derived and a nomogram risk prediction model was built to predict the occurrence of thromboembolism. Results: A total of 338 eligible patients were included in this study, which was randomly split into a training set (n = 238) and a validation set (n = 100). By performing Lasso regression and multivariate logistic regression, the prediction model was established including seven factors and the area under the receiving operating characteristic was 0.930 in the training set and 0.839 in the validation set. Factors associated with a high risk of thromboembolism were protuberant thrombus (odds ratio (OR) 5.03, 95% confidential intervals (CI) 1.14–23.83, p = 0.033), and history of diabetes mellitus (OR 6.28, 95% CI 1.59–29.96, p = 0.012), while a high level of left ventricular ejection fraction along with no antiplatelet therapy indicated a low risk of thromboembolism (OR 0.95, 95% CI 0.89–1.01, p = 0.098; OR 0.26, 95% CI 0.05–1.07, p = 0.083, separately). Conclusions: A prediction model was established by selecting seven factors based on the Lasso algorithm, which gave hints about how to forecast the probability of thromboembolism in hospitalized ventricular thrombus patients. For the development and validation of models, more prospective clinical studies are required. Clinical Trial Registration: NCT 05006677.


Introduction
It has long been a topic of discussion in medical settings on how to prevent thromboembolism, particularly cardiac embolism.Researchers reported that patients with ventricular thrombus had a high risk of stroke or systemic embolism (SSE) more than 20% before being discharged despite anticoagulation [1][2][3], and studies indicated that the in-hospital mortality rate of patients with ventricular thrombus was higher compared to patients without ventricular thrombus [4,5].With the advanced technology in imaging tools, the incidence of ventricular thrombus has increased in recent years, with a range of 4%-10% [6,7].As thromboembolism is currently the most noteworthy severe outcome in patients with ventricular thrombus [8,9], it is of vital importance to identify which patients are at a higher risk of thromboembolism, tending to decrease mortality or mobility.Prediction models in the prevention of atrial fib-rillation (AF)-related stroke have been developed [10][11][12], up to date, there is no prediction model built on the theme of thromboembolism secondary to ventricular thrombus, especially focusing on hospitalized medical patients.In our study, we aimed to build a prediction model by analyzing potential predictors including clinical characteristics, laboratory data, or imaging measurements, to better help clinicians target early awareness in hospitalized patients with high-risk factors, as well as to provide provoking thoughts or evidence in the management of patients with ventricular thrombus.

Patient Population
This retrospective cohort study was conducted from November 2019 to December 2021 using electronic medi-cal records of Fuwai Hospital, National Center of Cardiovascular Diseases in China, which was registered in Clini-calTrials.gov:NCT 05006677.This prediction model study was reported in accordance with the TRIPOD checklist [13].The inclusion criteria were: (1) Age ≥18 years; (2) Patients admitted to the center with the initial diagnosis of ventricular thrombus or occurred ventricular thrombus during the hospitalization.Patients diagnosed with inherited or acquired thrombophilia (e.g., antiphospholipid syndrome) were excluded since the risk of thromboembolism in these patients was established on a unique pathophysiological mechanism.

Definitions
The diagnosis of ventricular thrombus was confirmed by transesophageal or transthoracic echocardiography with or without contrast, computer tomography (CT), or cardiac magnetic resonance (CMR) imaging.When these imaging tools were not consistent, X.Q.(Ph.D., majoring in echocardiography) and other professors would review images and reach a conclusion.A ventricular thrombus was identified as a ventricular cavity with an aberrant echo mass or intensity, whose edge was different from the ventricular endocardium [14].The existence of the thrombus was confirmed by several sections, including parasternal short and long-axis views, as well as apical 2-, 3-, and 4-chamber images.When a thrombus was detected, its morphology was categorized as either mural (if its borders are generally continuous with the adjacent endocardium) or protuberant (if its borders are distinct from the adjacent endocardium and protrude into the ventricular cavity) [15].
Information on thromboembolism events during the hospitalization was obtained by searching our institutional database.Thromboembolism events were defined as the composite of ischemic stroke or transient ischemic attack, pulmonary embolism (PE), and systemic embolic events, with the exclusion of deep venous thrombosis [16].Ischemic stroke and transient ischemic attack were defined as the presence of acute focal neurological deficit with clinical symptoms or signs [17].PE and peripheral embolic events were documented by angiography or objective testing [18].

Model Development
Two colleagues (Q.Y. and X.Q.) extracted the data independently and compared the results to ensure coherence, and an additional scholar resolved the discrepancies.A total of 46 variables including patient demographics, laboratory results, and imaging measurements were collected in the initial model.
The data were randomly split into a training set (70% of the sample) and a validation set (30% of the sample).The training set was the terminology used in univariate regression as well as Lasso regression to find out clinical potential factors.Variables with a p value < 0.10 in univariate analysis were considered to be linked to the outcome and then performed stepwise predictor selection in three directions separately (forward, backward, and both), defined as Model 1 followed by multiple logistic regression.Odds ratio (OR) and 95% confidence interval (CI) were calculated using logistic regression models.We also conducted Lasso regression with L1-penalized least absolute shrinkage to select other potential factors and then formed Model 2 by performing multivariate analysis based on the Lasso method.The reliability of the predictive model was assessed concerning discrimination and calibration.The discrimination analysis and the mean area under the receiver operating characteristic curve (AUROC) obtained by repeated cross-validation (ten-fold), were used to select models.This procedure was repeated many times and the performance on the validation set was averaged to select the model with the greatest external validity.The reliability of the model was then evaluated using a concordance index (C-index) and a calibration plot via the bootstrap method which was tested with a Hosmer-Lemeshow goodness-of-fit test (R 2 ) [19].The regression model with the minimum Akaike's information criterion was used in the nomogram formulation.To quantitatively visualize the net benefit of clinical decisions, the decision curve analysis (DCA) was also conducted.

Statistics Analysis
Descriptive statistics were computed using the CBCgrps-Package in R [20].Continuous variables were presented as mean (standard deviation, SD) or median (interquartile range, IQR) and as frequency (percentage) for categorical variables [21].Analysis of variance was used to compare normally continuous variables and Pearson chi-squared test for categorical data.The Fisher exact test and Kruskal-Wallis H test were used as appropriate.Missing data for predictor variables were handled by using multiple imputations by chained equations with predictive mean matching (MICE-Package in R) creating 5 imputed data sets.Categorical variables were encoded by binary with the first category dropped.The car package in R was used to detect collinearity between variables, and a variance inflation factor <10 was tolerated.All analyses were scheduled for completion with R Studio and R, Version 3.5.1 (The R Project for Statistical Computing, Vienna, Austria).

Patients Characteristics
A total of 498 patients were identified in the electronic records from November 2019 to December 2021, while 7 out of 498 patients were without ventricular thrombus.153 patients were excluded, of these, 136 patients were already diagnosed with ventricular thrombus before this hospitalization, 12 patients were aged <18 years, and 5 patients had a suspected diagnosis of thrombophilia (2 antiphospholipid syndrome) at discharge.Overall, we included 338 eligible patients in this study, which were ran-domly split into a training set (n = 238) and a validation set (n = 100) (Supplementary Fig. 1).Among 338 patients, 20 (5.9%) patients underwent thrombectomy therapy, either with or without ventricular aneurysm resection, and 9 (2.7%) patients had heart transplantation in the hospital.288 (85.2%) patients were male and 71 (21%) patients were overweight (defined as body mass index (BMI) ≥28%).Patients who were diagnosed with myocardial infarction (MI) at admission accounted for 62% (n = 208).At baseline, the median level of D-dimer was more than two-fold higher than the reference value (<0.5 g/L) while the level of fibrin degradation products (FDP) with a median range of 2.6 (IQR 2.5, 5.5) g/L was negatively normal (0-5 g/L).Most patients (79.9%) had a creatinine clearance (CrCl) of more than 50 mL/min while 54 (16%) patients had moderate renal dysfunction with the range of 30 mL/min to 49 mL/min and 14 (4.1%) patients had a CrCl of less than 30 mL/min.The median of N-Terminal pro-brain natriuretic peptide (NT-proBNP) was 2408.0 pg/mL, and 123 (36.4%) patients had a more than 10% decline in NT-proBNP at discharge (Supplementary Table 1).
In our study, 282 (83.4%) patients were diagnosed with ventricular thrombus confirmed by echocardiography and 13 (6.8%)patients depended on CMR to find ventricular thrombus while their echocardiograms were negative.Another 43 (12.7%)patients had a record of ventricular thrombus only with CT in our center.Patients had a median left ventricular ejection fraction (LVEF) of 35.0% and a left ventricular end-diastolic diameter of 60 mm.287 (85%) patients had a mural thrombus, and the remaining patients had a protuberant thrombus with or without a mobile free edge.In terms of anticoagulation therapy, 176 patients (52%) had heparin injections whereas 239 patients (71%) received oral anticoagulation during the period of hospitalization, of which 72% were on non-vitamin K antagonist oral anticoagulants (NOACs) and 28% on warfarin.Of the 173 patients who took NOACs, 165 (95.4%) received rivaroxaban (almost half of whom took 20 mg daily), and the remaining 8 (4.6%) were given dabigatran 110 mg twice daily.Given the high percentage of patients with coronary artery diseases, 164 (49%) patients got antiplatelet therapy, with 86 receiving mono antiplatelet therapy (20 on aspirin and 76 on clopidogrel) and 78 receiving dual antiplatelet therapy (66 on aspirin plus clopidogrel and 12 on aspirin plus ticagrelor).Above all, no significant differences were found comparing the training cohort and validation cohort in demography and clinic characteristics (Table 1).

Factors Selected by Univariate and Lasso Regression
We included 46 characteristics in our models.A total of 15 factors were selected from the univariate analysis (Table 2) and 5 factors remained after performing a multiple logistic regression model which formed Model 1 (Table 3).They were BMI, ventricular aneurysm, history of diabetes mellitus (DM), prior SSE, and therapy of antiplatelet.And with the Lasso regression, Lambda = 0.000010 was chosen (minimum criteria) according to ten-fold cross-validation of the Lasso coefficient profiles of the 46 features, and 11 factors were selected (Fig. 1 and Supplementary Fig. 2).A multiple logistic regression model was established using Lasso regression and the analysis results were shown in Table 3.The following four risk factors were not associated with the outcome (p < 0.05): history of heart failure (HF), therapy of heparin, site of thrombus, and FDP change.Finally, a total of 7 factors (BMI, diastolic blood pressure, LVEF, thrombus morphology, medical history of DM, prior SSE, and antiplatelet therapy) were extracted into Model 2. By comparing the AUROC, Model 2 showed a greater AU-ROC in the training set than Model 1 (Model 1: 0.904, 95% CI 0.850-0.958;Model 2: 0.930, 95% CI 0.883-0.977,p = 0.205), as well as Model 2 performed better in the validation set (Model 1: 0.805, 95% CI 0.609-1.000;Model 2: 0.839, 95% CI 0.669-1.000,p = 0.354) (Fig. 2, and Supplementary Figs.3,4).Positive agreements between ideal curves and calibration curves were also observed Supplementary Figs.5,6).The DCA curve revealed a range of cutoff probabilities shown by the nomogram (Supplementary Fig. 7).In summary, we chose Model 2 as the final model to make a prediction.The prediction result of Model 2 after incorporating the 7 factors into the model was presented in Fig. 2 with the AUROC being 0.930 in the training set and 0.839 in the validation set in Model 2. And by conducting the leave-one-out cross-validation, the accuracy of Model 2 was 0.937 while the Kappa value was 0.413.

Prediction Model in the Prediction of Thromboembolism
According to Model 2 (factors included prior SSE, medical history of DM, thrombus morphology, diastolic blood pressure, BMI, LVEF, and antiplatelet therapy), we established a nomogram risk prediction model containing independent risk factors (R 2 0.52, C index 0.93, 95% CI Variables are presented as n (%), mean ± SD, and median (IQR).0.87-0.99)(Fig. 3).The scores of the items displayed in the nomogram should be added up.For example, if a patient with ventricular mural thrombus, had a level of BMI of 28 kg/m 2 and diastolic blood pressure of 70 mmHg, had no medical history of DM or SSE, had a level of LVEF of 30%, and he/she was not on antiplatelet therapy during the one-week hospitalization, then the total score was approximately 106, indicating an estimated thromboembolism event of <10%.And considering the wide CI in the factors of the prior SSE, the results needed to be critically evaluated, which could be accounted for by the very small sample of patients who had a history of SSE.Other factors that were related to a high risk of thromboembolism were protuberant thrombus (OR 5.03, 95% CI 1.14-23.83,p = 0.033), a higher level of diastolic blood pressure (OR 1,07, 95% CI 1.01-1.14,p = 0.019), and history of DM (OR 6.28, 95% CI 1.59-29.96,p = 0.012), while a relatively high

Discussion
Our study first conducted a prediction model established on Lasso regression to predict the risk of thromboembolism in hospitalized patients with ventricular thrombus.And we concluded that patients were more likely to experience thromboembolism in hospital, who had a medical history of SSE and DM, a lower BMI and LVEF but a higher diastolic blood pressure at baseline, along with protuberant thrombus and without antiplatelet therapy during hospitalization.
It is well established that DM and prior SSE have been widely used to stratify the risk of stroke, which were proved to be predictors of thromboembolism events in the study.Patients with DM had a higher risk of thrombotic events due to the pathophysiological underpinnings of endothelial dysfunction and vascular inflammation.Recurrent thromboembolism was more common among patients who had previously experienced it, and its incidence was seven times greater than that of newly discovered cases.Patients with a first PE had more than a two-fold risk of developing a second PE [22].In the model built on the ROCKET-AF trial, prior thromboembolism was the strongest independent pre-dictor of thromboembolism [10], which was similar to our results.Along with a history of DM and stroke, we observed a strong relationship between the history of HF and the occurrence of thromboembolism in univariate analysis, whereas HF has been identified as a risk factor for thromboembolic events in previous research [23,24].Patients who experienced HF or cardiac dysfunction (e.g., a high NT-proBNP, a low LVEF, or a large left ventricular enddiastolic volume) at baseline, faced a higher rate of thromboembolism, and it could be attributed to complex pathophysiological mechanisms such as neurohormonal activation or decreased myocardial contractility, resulting in an increased vulnerability to thromboses [25].And the abnormal blood flow as well as other requirements of Virchow's triad including hypercoagulability, and endothelial injury was satisfied in patients with HF [26,27].In a population-based 30-year cohort study, patients with HF had an increased risk of stroke compared with the general population group [28].And by pooling 2 trials related to HF, researchers reported stroke occurrence in 4.7% of patients with AF and 3.4% of patients without AF [29].A large prospective study reported that HF hospitalization increased the risk of MI or stroke [30], which provided the clear message that HF should no longer be considered a minor risk factor for thromboembolism.
In summary of studies that predicted the embolism events, factors including the level of D-dimer indicated a higher additional risk besides the major persistent risk factors [22,23].D-dimer and FDP levels at admission were significantly related to a high risk of embolism, otherwise, neither D-dimer nor FDP with more than a one-fold increase at discharge had a significant relationship with events in the study.Without a doubt, patients who had a high D-dimer had a higher risk of any embolism events since D-dimer was inherently an indicator of thrombus formation.Interestingly, another laboratory indicator also showed an opposite relationship with thromboembolism.The lower the level of lymphocyte count was, the risk of thromboembolism increased.Whether the level of lymphocyte count could indicate thromboembolism remained unknown, and more evidence mechanism is needed to explore.It reported that in COVID-19 patients the lymphocyte count (p = 0.004) showed a lower value in the patients with PE compared with those without PE [31].And previous studies have concluded that the increased inflammation increased the risk of thromboembolism as well, which mostly happened to patients who had inspiratory diseases [25,32].Moreover, researchers found that in 60 patients who developed left ventricular thrombus in COVID-19, 21.5% and 16.9% of patients had stroke events and PE separately, while 12.3% of patients had peripheral arterial embolism [33].
When assessing the effect of the amount or location of thrombus on the risk of thromboembolism, as most patients were diagnosed by echocardiographic assessments, it remained to explore a more accurate embolism rate in CMR or CT or contrast echo since CMR has been regarded as gold criteria could find small and more ventricular thrombus [9].And patients who had biventricular thrombus were more likely to occur thromboembolism, and one of the reasons might be accounted that they had severe cardiac dysfunction as well as a complex inner condition at admission.In terms of thrombus morphology, protuberant or mobile thrombi were related to a higher risk of embolism compared with mural thrombi, though data on the subtype of thrombus were limited.Researchers demonstrated that transthoracic echocardiography implemented with pulsed wave tissue doppler imaging could provide a more precise definition of mass mobility over visual assessment, and concluded that a ≥10 cm/s mass peak Va was considered the most significant predictor of embolic risk in hospitalized patients [34].In the 2022 statement for left ventricular thrombus [15], researchers suggest that for a protuberant thrombus as well as a newly diagnosed mural thrombus, it would be prudent to give anticoagulation therapy.And a shared decisionmaking approach is recommended for organized or calcified thrombi.
Generally, patients with ventricular thrombus ought to be governed by anticoagulation in the absence of contraindications.More than 70% of patients received oral anticoagulation and nearly 50% were on heparin in hospital.Patients who had no history of AF were less likely to be pretreated with anticoagulants, which increased the risk of thromboembolism without long-term anticoagulation [8].In a pooled meta-analysis of studies of ventricular thrombus after MI, the use of anticoagulants (either warfarin or heparin) reduced the risk of stroke by 81% [35].On the other hand, the results of studies that compared the use of NOACs to vitamin K antagonists in the prevention of embolism risk were controversial [36][37][38], requiring more randomized clinical trials (RCTs) to provide robust evidence.Antiplatelet therapy and anticoagulation therapy, which have different targets, both have an effect on reducing the risk of thromboembolism [39,40].Upon the topic of antiplatelet therapy secondary to anticoagulation treatment in the field of prevention of thromboembolism, studies have demonstrated that antiplatelet therapy was effective for the primary prevention of embolism events [41][42][43].Other large RCTs have demonstrated a significant reduction ranging from 20% to 69% in recurrent thromboembolism with aspirin versus placebo after anticoagulants were discontinued in patients with a history of embolic events [44,45].But the treatment of triple antithrombotic therapy which was associated with a higher rate of bleeding remained unknown for patients with ventricular thrombus [46].Personalized management for the prevention and treatment of ventricular thrombus should be developed to take into account of patient characteristics.
Concerning other predictors in the final model of this study, we outlined the findings as follows.A high risk of thromboembolism was linked to higher diastolic blood pressure.In the RE-LY trial's subgroup analysis, patients with high diastolic blood pressure (≥90 mmHg) had a high risk of developing SSE [47].The elevated diastolic blood pressure was found to be significantly associated with an increased risk of stroke in another RCT with 22,672 patients, with a 1.5-fold risk for diastolic blood pressure of 80-89 mmHg and a 4-fold risk for 90 mmHg or more [48].Likewise, a remarkable correlation was observed between the BMI and the outcome of the study.Previous results from three RCT trials (ARISTOTLE [49], ROCKET-AF [50], and ENGAGE AF-TIMI 48 [51]) showed that a higher BMI was independently related to a decreased risk of SSE.The reason for the apparent protective effect of obesity is unclear, and we hypothesized that patients in the higher BMI categories are offered earlier and more intensive treatments to manage the risk of stroke events.Several limitations were as followed.First, the validation set was based on the same dataset with a small sample, which restricted the power and the practical utility of our model.Second, limited to patient resources, the result of the study could not greatly expand to a large population.Third, even if ventricular thrombus mobility is a major prognostic determinant of increased thromboembolism [34], this retrospective analysis did not include a detailed assessment of thrombotic mass mobility.Additionally, it was also undetermined whether or when to implement a strategy to prevent embolism, since this study focused on developing a novel prediction model to identify patients who were at high risk of embolism.

Conclusions
This study conducted a prediction model by selecting seven factors based on the Lasso algorithm, aiming to identify the risk prediction of thromboembolism in hospitalized patients with ventricular thrombus.Patients who had a medical history of SSE and DM, a lower level of BMI and LVEF but a higher diastolic blood pressure at baseline, along with protuberant thrombus and without antiplatelet therapy during hospitalization, were more likely to experience thromboembolism in hospital.More prospective clinical trials are required to develop and validate models, and individualized discussion and shared decision-making are of critical importance in managing patients with ventricular thrombus.

Fig. 1 .
Fig. 1.Tuning parameter (Lambda) selection in the Lasso Model used ten-fold cross-validation based on the minimum criteria (left dotted vertical line) or the 1 standard error criteria (right dotted vertical line).

Fig. 2 .
Fig. 2. ROC curves of Model 2 for predicting the risk of thromboembolism.(A) Training set.(B) Validation set.ROC, receiver operating characteristic; AUC, area under the ROC curve.