Factors Influencing Physician Adherence to Venous Thromboembolism Risk Assessment Model Recommendations: A Best–Worst Scaling Authors: Nicola B Campoamor, Benjamin G Mittman, Stephanie Moss, Michael B Rothberg, John F P Bridges Published: 2026 ABSTRACT AbstractPurposeRisk assessment models (RAM) use predictive algorithms to support physicians in personalizing care. For example, a RAM to determine the risk of venous thromboembolism (VTE-RAM) in hospitalized patients offered treatment recommendations that physicians adhered to at varying rates. We sought to investigate which factors influenced physician adherence to prophylaxis recommendations from a VTE-RAM implemented in an inpatient setting.MethodsWe assessed adherence to treatment recommendations for a VTE-RAM in the inpatient setting using both survey and real-world data sources. Inclusion criteria included being a hospitalist or resident in the inpatient setting, using a VTE-RAM at least 50 times, and having complete data from both sources available. Variation in treatment adherence from the two data sources was compared, and physicians were inductively segmented into high- and low-adherence groups. Seven possible factors influencing adherence were identified using prior research and prioritized using best–worst scaling (BWS). Results were stratified by adherence group, and differences in priorities were tested using Wald tests.ResultsWe obtained data for 188 physicians, most of whom (56%) were unable to accurately state their rate of adherence. Two groups were identified: those with relatively high adherence (≥ 70%; n = 58) and those with relatively low adherence (< 70%; n = 130). High-adherence physicians prioritized that the VTE-RAM increases decision-making confidence (p < 0.001), reduces overtreatment of low-risk patients (p < 0.001), and facilitates documentation of low-risk patients (p = 0.029). Low-adherence physicians were more likely to prioritize that important risk factors are missing from the algorithm (p < 0.001), that the VTE-RAM disrupts workflow (p < 0.001), and that it is used owing to institutional policy requirements (p = 0.021).ConclusionsObserved physician adherence to VTE-RAM recommendations often differs from self-reported behavior. We found that several factors associated with high- and low-adherence physicians. High adherence was linked to confidence, reduced overtreatment, and documentation support, while low adherence reflected concerns about missing factors, workflow disruption, and institutional requirements. Aligning RAM design with physician priorities may improve adherence and clinical utility.Supplementary InformationThe online version contains supplementary material available at 10.1007/s40256-026-00799-4. Key Points This study investigated factors influencing physician adherence to treatment recommendations from a venous thromboembolism risk assessment model (VTE-RAM) in the inpatient setting.Physicians with high adherence emphasized the ability of the VTE-RAM to increase decision-making confidence, reduce overtreatment, and facilitate documentation, while low adherence was associated with concerns about missing risk factors, workflow disruption, and institutional mandates.These findings underscore the importance of aligning RAM design and implementation with physician priorities to enhance clinical utility and uptake. Introduction Risk is ubiquitous in medicine and impacts the ability to predict future outcomes [1]. Physicians seek to both assess and manage risk by understanding what factors influence the likelihood of certain adverse events occurring [2]. Risk assessment and management can help protect patients during a hospitalization from infections, falls, and blood clots [3, 4]. Venous thromboembolism (VTE), wherein a blood clot forms in a vein, is a well-known risk among hospitalized patients. VTE affects at least 600,000 people annually, causing around 100,000 deaths [5, 6]. The risk of VTE is highest after surgery, injury, and chronic inflammation [7]. Several randomized placebo-controlled trials have demonstrated that rates of symptomatic VTE can be reduced with prophylaxis measures, such as chemical agents, compression devices, and activity-based interventions [8–10]. Physicians may have difficulty in identifying all relevant risk factors or weighing them correctly, resulting in inconsistent estimations of risk for individual patients [11, 12]. Risk assessment models (RAM) can be used by physicians to aid in assessing the likelihood of an event [13–16]. These models utilize a combination of patient data, medical history, and other relevant information to predict specific diseases or outcomes and improve clinical decision-making [17]. RAM have been developed for various conditions, including risk of falls for elderly patients [18], risk of breast cancer [19, 20], and risks related to general oncology [21]. RAM can be standalone (i.e., accessed through an independent website or application) [22] or integrated into other information systems, such as electronic health record (EHR) systems [23–25]. EHR integration allows for seamless access to patient data and enhances the efficiency of clinical workflows [26]. RAM can be assisted by artificial intelligence and machine learning [27]. RAM are recommended by clinical guidelines for assessing and managing risk [28, 29]. The adoption of RAM may be hindered by time constraints, lack of familiarity, and perceived complexity [17, 30]. Not all physicians agree with treatment recommendations from RAM. Physicians may also be skeptical about the accuracy of these models, often preferring to rely on their own clinical judgment and maintain autonomy [31]. Efforts made to address limited adoption of RAM need to identify specific barriers and facilitators and to identify constructs that influence the processes and outcomes surrounding the use of the RAM [32–34]. Improvements are observed when RAM are tailored to the context of clinical practice and by enhancing physician trust in the accuracy and value of using the RAM and considering its output [35, 36]. The Cleveland Clinic developed and implemented a RAM focused on VTE prophylaxis in 2017. This VTE-RAM gathers information from the EHR and presents it to the physician, who then verifies it and enters the information back into the RAM, which then returns an estimated risk of VTE in the next 14 days together with a recommendation as to whether prophylaxis is required. VTE was the focus because it is a common inpatient problem, the risk factors are readily available in the EHR, and physicians relying solely on their judgment tend to overestimate risk, resulting in overuse of prophylaxis among patients at low risk for VTE. This puts them at risk of unnecessary bleeding, which is a side effect of prophylaxis. We sought to investigate factors influencing physician adherence to VTE prophylaxis recommendations from RAM in an inpatient setting. Substantively, our study provides important insights into the real-world experiences of physicians using RAM that may inform future implementation of RAM and other decision support tools. Methodologically, we demonstrate novel approaches to explore stratification on the basis of rates of adherence and illustrate how theory-driven prioritization methods can identify end-user perspectives with regards to predictive algorithms. Methods Our methodology can be described as a learning health systems approach that combines a brief survey about clinical practice with real-world data [37, 38]. This approach emphasizes iterative cycles of data generation, analysis, and feedback to practice, enabling continuous refinement of clinical tools and processes [39]. Our study integrates routinely collected EHR data with targeted survey methods to generate actionable evidence that can inform subsequent refinements to VTE-RAM implementation. The goal is to increase uptake of the VTE-RAM in a real-world setting and ultimately to ensure that prophylaxis decisions are concordant with patient risk. These combined data were used to compare physicians’ self-reported rates of adherence to the VTE-RAM with their observed rates in clinical practice. We then segmented physicians into groups with relatively high and low adherence with VTE-RAM treatment recommendations in their prescribing of prophylaxis. We defined adherence as whether the physician’s decision to prescribe or withhold prophylaxis matched the recommendation made by the VTE-RAM. To assess the possible drivers of adherence, we developed and implemented a brief survey that incorporated a best–worst scaling (BWS) instrument to assess physician experience with the VTE-RAM. This BWS allowed us to first quantify the importance of these factors in aggregate and then test for differences between physicians with high and low adherence. Further details for each of these components are developed in the subsections that follow. Identifying Physicians with High and Low Adherence Inpatient physicians, including hospitalists and internal medicine residents, from the Department of Hospital Medicine at the Cleveland Clinic were contacted via email to participate in a research study. The invitation came from the study team. Three reminder emails were sent to nonresponding physicians 1–2 weeks following the first email. Physicians were offered a US$10 Starbucks gift card as compensation for their participation. To ensure stable estimates of physician adherence to RAM treatment recommendations, only physicians who had used the VTE-RAM at least 50 times were included in the initial sample for analysis. The survey was administered online from 22 June to 8 November 2023. We evaluated the relationship between stated and observed adherence by plotting the distribution of observed adherence to VTE-RAM treatment recommendations across four self-reported adherence groups: “less than 25%,” “25–50%,” “50–75%,” and “greater than 75%.” We assessed accuracy by calculating the proportion of respondents in each group whose observed adherence matched their self-reported estimates. We then tested the normality of the observed adherence distribution using the Shapiro–Wilk test and the skewness and kurtosis test. Both tests indicated significant deviation from normality (Shapiro–Wilk: p = 0.009; skewness and kurtosis: p = 0.010). Quantifying the Drivers of Adherence BWS was used to measure possible physician beliefs toward the VTE-RAM. BWS is a type of choice experiment grounded in random utility theory [40], using an experimental design to generate choice tasks, each composed of a subset of objects (i.e., factors driving the use of the VTE-RAM) [40]. This method of prioritization has been growing in popularity [41, 42], and its application can be guided by current best practices [43, 44]. Multiple sources can guide the identification of factors to include in a BWS [40, 42, 45]. We used a previously published qualitative study conducted by members of our study team to identify relevant factors [23]. This study interviewed hospitalists and advanced practice providers about the barriers to and facilitators of using VTE-RAM. This study identified that VTE-RAM was easy to use and provided useful information to guide prophylaxis decisions, though took issue when the model did not align with their clinical judgement. Ten themes were identified from this study and our team refined these using an abductive approach to establish the seven factors used in the BWS (Table 1). Table 1Factors included in the best–worst scaling experimentFactorDescriptionReferencesDisrupts workflowThe model disrupts my workflow[65]Reduces overtreatmentThe model reduces overtreatment for low-risk patients[66, 67]Increases confidenceThe model increases my decision-making confidence[68]Factors missingImportant risk factors are missing from the model[64]Institutional policiesI use the model because I need to follow institutional policies[69]TrustworthyThe model is trustworthy[68, 70, 71]Facilitates documentationThe model facilitates documentation of low-risk patients[69] Factors included in the best–worst scaling experiment Survey development followed established good practices for preference elicitation instruments, including the ISPOR checklist for patient preference studies and reporting guidance for formative qualitative work that supports instrument design [46, 47]. Procedures aligned with ISPOR recommendations intended to increase the usefulness and impact of patient preference studies for decision-makers [43]. Validity assessment incorporated established conventions in stated-preference research, including evidence for content and respondent understanding and checks of internal coherence and choice consistency [45, 48]. A copy of the BWS instrument and other survey questions are included in the Supplementary Materials. One of the key strengths of BWS is the experimental design, which allows for subsets of the factors being explored to be considered across multiple tasks. The is generally done with a balanced incomplete block design (BIBD) that ensures that each of the subsets is of equal size, that each factor is presented the same number of times, and that each factor is presented with every other factor an equal number of times. We used a Youden experimental design, which is a special type of BIBD that also ensures that each factor is presented in each possible position of the choice task and an equal number of times [49–52]. We specifically used a Youden design for seven factors that used seven tasks containing a subset of four factors. Each factor appeared four times in the instrument, each time in a different position (i.e., listed first, second, third, or fourth). This design is one of the most commonly used designs in the literature [41]. Physicians were asked to select which factors were most and least reflective of their beliefs regarding the VTE-RAM (Fig. 1). After completing all the choice tasks, we asked four evaluation questions to determine whether the choice tasks were easy to understand and relevant to them, and whether their answers showed their real preferences. Study data were collected and managed using REDCap electronic data capture tools hosted at the Cleveland Clinic [53]. Example best–worst scaling task Analysis We report the number of times each factor was chosen as best and worst. A simple score of importance was calculated by subtracting the number of worst votes from the number of best votes and dividing by the number of times it was shown. The resulting best-minus-worst (B–W) score falls between 1 (always chosen as best) and −1 (always chosen as worst). We then estimated a conditional logistic regression using effects coding for the independent variables to analyze the BWS assuming sequential best and worst choices [54]. The regression was stratified by physicians’ adherence level. A Wald test was conducted to assess differences between the individual regression coefficients of the two physician groups. Importance scores of the regression coefficients were calculated using a probability rescaling procedure to convert scores to a range from 0 to 100 [55]. Statistical analyses were conducted using Stata SE version 17 (StataCorp, College Station, TX, USA). All study procedures were reviewed and approved by the Cleveland Clinic Institutional Review Board (identifier no. 20-321). Results A total of 434 physicians were invited to take the survey (Fig. 2). Approximately half (n = 213; 49%) did not take the survey, and 30 physicians only partially completed it. This resulted in the survey dataset having a total sample of 191 physicians. Between 16 December 2020, and 31 December 2023, observed rates of adherence with the VTE-RAM were collected from 238 physicians, of whom 234 had complete data (i.e., real-world dataset). Respondent IDs from the survey dataset were matched with IDs from the real-world dataset. Data completeness was assessed, and three respondents were removed from the survey dataset owing to incomplete or missing real-world data. This resulted in a final analytic sample of 188 physicians. Flowchart of physician sample selection and inclusion To inform our comparison between self-reported and observed adherence, we first examined the distribution of observed adherence rates. We identified a bimodal pattern, with two distinct peaks suggesting natural clustering into low- and high-adherence groups. A local minimum between 65 and 70% served as a logical cutoff for stratification. On the basis this local minimum, we defined low-adherence physicians as those who were adherent with the model less than 70% of the time and high-adherence physicians as those who were adherent 70% of the time or more. Figure 3 illustrates this distribution and supports the decision to dichotomize the sample. Normality tests conducted on the stratified groups indicated that both distributions approximated normality (Shapiro–Wilk: p = 0.056; skewness and kurtosis: p = 0.155), further validating the use of this threshold for subsequent analyses. Bimodal distribution of observed adherence rates with a 70% cutoff for stratification Using the 70% cutoff, observed adherence measured continuously using EHR data showed that 130 physicians (69%) had low adherence (< 70%) and 58 (31%) had high adherence (≥ 70%). Self-reported adherence grouped physicians into four categories: 15 (8%) reported adherence less than 25% of the time, 45 (24%) reported 25–50%, 73 (39%) reported 50–75%, and 55 (29%) reported greater than 75% (Table 2). Comparison of self-reported and observed adherence showed that only 82 physicians (44%) accurately estimated their adherence. Table 2Comparison of self-reported and observed RAM treatment recommendation adherenceSelf-reported adherenceObserved adherenceAccuracyAggregateLow adherenceaHigh adherencebCorrect self-reportCounts (%)Counts (%)Counts (%)Counts (%)< 25% of the time15 (8%)15 (100%)0 (0%)2 (13%)25–50% of the time45 (24%)37 (82%)8 (18%)20 (44%)50–75% of the time73 (39%)56 (77%)17 (23%)33 (45%)> 75% of the time55 (29%)22 (40%)33 (60%)27 (49%)Total1881305882aLow adherence: observed adherence < 70%bHigh adherence: observed adherence ≥ 70% Comparison of self-reported and observed RAM treatment recommendation adherence aLow adherence: observed adherence < 70% bHigh adherence: observed adherence ≥ 70% Factors Driving Adherence Physicians prioritized several factors as most reflective of their experience using the VTE-RAM. The top three factors were that the VTE-RAM increases decision-making confidence (importance score 18.55), facilitates documentation of low-risk patients (17.61), and reduces overtreatment of low-risk patients (16.70). In contrast, the least prioritized factors were that the VTE-RAM disrupts workflow (importance score 8.96), is trustworthy (10.08), and is used owing to institutional policy requirements (11.97). These results suggest that physicians value the VTE-RAM most when it supports clinical efficiency and judgment, and least when it is perceived as burdensome or externally imposed (Table 3). Table 3Aggregate importance of factors influencing physician adherence to RAM treatment recommendationsAggregate (n = 188)BestWorstB–W scoreaCoefficientImportance scorebIncreases confidence2731290.1910.40118.55Facilitates documentation201980.1370.32317.61Reduces overtreatment2391420.1170.24716.70Factors missing2371600.1020.19816.13Institutional policies143215−0.096−0.20011.97Trustworthy68241−0. 230−0.41410.08Disrupts workflow164331−0. 222−0.5568.96aB−W score was calculated by subtracting the number of times a factor was selected as “worst” from the number of times it was selected as “best,” then dividing by the number of times the factor appearedbImportance scores were derived from a conditional logistic regression model and rescaled to a 0–100 range using a probability rescaling procedure Aggregate importance of factors influencing physician adherence to RAM treatment recommendations aB−W score was calculated by subtracting the number of times a factor was selected as “worst” from the number of times it was selected as “best,” then dividing by the number of times the factor appeared bImportance scores were derived from a conditional logistic regression model and rescaled to a 0–100 range using a probability rescaling procedure High-adherence physicians (n = 58) were significantly more likely to prioritize that the VTE-RAM improves decision-making confidence (importance score 22.98; p < 0.001), reduces overtreatment of low-risk patients (22.47; p < 0.001), facilitates documentation of low-risk patients (18.62; p = 0.029), and is trustworthy (11.67; p < 0.001) (Table 4). In contrast, low-adherence physicians (n = 130) were more likely to emphasize that important risk factors are missing from the model (18.76; p < 0.001), that they use the VTE-RAM owing to institutional policy requirements (12.75; p < 0.001), and that the model disrupts their workflow (11.51; p < 0.001) (Fig. 4). Table 4Stratified importance of factors by physician adherence levelLow adherence (n = 130)High adherence (n = 58)p-valueCoefficient (SE)Importance scoreCoefficient (SE)Importance scoreIncreases confidence0.232 (0.07)16.620.851 (0.10)22.98< 0.001Facilitates documentation0.258 (0.06)16.920.502 (0.09)18.620.029Reduces overtreatment0.033 (0.06)14.390.812 (0.11)22.47< 0.001Factors missing0.409 (0.07)18.76−0.300 (0.10)10.38< 0.001Institutional policies−0.126 (0.06)12.75−0.398 (0.10)9.580.021Trustworthy−0.551 (0.06)9.05−0.152 (0.09)11.67< 0.001Disrupts workflow−0.256 (0.07)11.51−1.315 (0.12)4.30< 0.001SE standard error Stratified importance of factors by physician adherence level SE standard error Factors influencing high and low adherence to VTE-RAM recommendations Discussion This study quantified physicians’ adherence to treatment recommendations provided by a VTE-RAM and investigated factors driving high and low adherence. We found that adherence rates were low overall, although some physicians consistently followed the treatment recommendations from the VTE-RAM. We also found that high adherence was associated with valuing the ability of the VTE-RAM to support decision-making, reduce overtreatment, and streamline documentation. Low adherence reflected concerns about missing risk factors, workflow disruption, and institutional requirements. These results highlight the importance of aligning RAM design with physician needs to improve adoption and clinical effectiveness. This study leveraged a unique opportunity to compare the validity of self-reported rates of adherence with VTE-RAM treatment recommendations with observed rates collected from EHR. Physicians with self-reported adherence rates of less than 50% generally underestimated their behavior, while those with self-reported rates of greater than 50% typically overestimated it. The discrepancy between observed and self-reported adherence rates became more pronounced as self-reported adherence rates increased. This validity check highlighted the higher accuracy and reliability of using the observed rates for the analysis, ensuring more valid and trustworthy results. Recent estimates suggest that approximately 20–30% of hospitalized patients are considered high-risk according to the VTE-RAM, depending on the threshold applied [56]. This means that the VTE-RAM would recommend against prophylaxis for at least 70% of patients. Physicians often treat most hospitalized patients as high-risk for VTE [57, 58]. Agreement with the low-risk recommendations from the VTE-RAM appears to be a key factor in adherence. High-adherence physicians likely found that the VTE-RAM supported their confidence in withholding prophylaxis from low-risk patients. Low-adherence physicians tended to view the VTE-RAM as disruptive and lacking clinical value, citing missing risk factors they use to identify more patients as high-risk. Physicians reliably identify individual risk factors but may struggle to integrate multiple factors into an accurate risk prediction [59]. Many physicians mistakenly treat a single risk factor as indicative of high risk, overlooking that individual factors differ in significance and that combining several modest factors can elevate a patient’s true risk [60]. This challenge reinforces the value of decision aids such as RAM that systematically integrate relevant risk factors to generate more precise risk estimates. Although trust in the VTE-RAM differed between high- and low-adherence groups, it was a relatively minor factor overall. Trust may be higher when the VTE-RAM aligns with physicians’ assessments, boosting their decision-making confidence and reducing overtreatment for low-risk patients. Conversely, trust may diminish when the VTE-RAM labels a patient as low risk but the physician believes the patient to be high risk. The key difference between the two groups may lie in how frequently they perceive patients as high risk, leading to more frequent disagreements with the VTE-RAM’s treatment recommendations. Barriers to adopting tools such as RAM often also include workflow disruptions from poor design and perceived limits on autonomy [61]. This loss of autonomy can lead to reactance, a negative motivational state that may also arise when a decision support tool contradicts a physician’s judgment [62]. Physicians may also experience alert fatigue whereby frequent noncritical alerts cause them to ignore all alerts [63]. The VTE-RAM used in this study minimized reactance and alert fatigue by presenting recommendations before decisions were made and by allowing physicians to choose whether to follow them. Despite these design features, low adherence was still observed in most physicians. While physicians generally did not find the VTE-RAM to be disruptive, low-adherence physicians were more likely to find it disruptive. This suggests that perceived value, rather than technical design, may be a more important factor in determining use. Proposed solutions may only address part of the problem if they focus solely on workflow integration without improving the tool’s perceived clinical relevance. Physicians are more likely to adopt a tool that comprehensively incorporates relevant risk factors and patient data even if they have doubts about its clinical utility [64]. This suggests that the VTE-RAM does not need to be fully trusted for physicians to adhere to treatment recommendations. Further research should explore how end users interact with RAM and identify which features support or hinder workflow. This includes understanding what makes a tool valuable to physicians. Communicating with physicians across hospitals is difficult. In this case, information was shared through email, staff meetings, and local champions. The VTE-RAM also included a link to the published paper describing its derivation and validation. Most physicians likely did not click the link, read the emails, or attend the meetings. Implementing these tools effectively in large clinical systems presents an ongoing challenge. A potential limitation of this study is that the drivers were not introduced or explained before participants completed the choice tasks. Without a prior clarification of each factor’s meaning, participants may have interpreted them on the basis of personal assumptions, pervious experience, or contextual cues within the task itself. This lack of standardization could have led to variation in responses, affecting the reliability and comparability of the results. An additional consideration is that information on years of clinical experience was not collected, which may influence how physicians interpret and act on VTE-RAM recommendations. More experienced physicians may rely more heavily on prior practice patterns, while less experienced physicians may rely more on the VTE-RAM, potentially contributing to unexplained heterogeneity in responses. However, since an eligibility requirement for the survey was having used the VTE-RAM at least 50 times, all participants were likely familiar with the drivers presented, as they relate to various components of the VTE-RAM itself. We defined low and high adherence using statistical thresholds based on 3 years of clinical data, rather than grounding them in externally validated clinical standards. In real-world settings, clinicians often consider a 50% agreement rate with decision support tools to be a strong outcome, since most alerts are dismissed. Because we did not anchor our thresholds in such benchmarks, we may have overstated the clinical significance of adherence. We did not adjust analyses for differences in physicians’ VTE-RAM usage frequency, and our inclusion threshold of greater than or equal to 50 VTE-RAM uses per physician may therefore introduce selection bias. We selected this threshold to ensure stable adherence estimates, as very small denominators (e.g., 5–10 uses) yield imprecise and potentially misleading agreement rates. Nonetheless, because the VTE-RAM was embedded in routine inpatient workflow, this criterion likely excluded few physicians, limiting the practical impact of this bias on applicability. Conclusions By combining survey and real-world data, we identified key drivers of adherence that distinguish high- from low-adherence physicians. High adherence was associated with valuing the VTE-RAM’s ability to increase decision-making confidence, reduce overtreatment of low-risk patients, and support documentation. In contrast, low adherence was linked to concerns about missing risk factors, institutional mandates, and workflow disruption. These findings highlight the importance of aligning RAM design and implementation with physician priorities to improve adherence and maximize clinical utility. Future efforts should prioritize physician-centered design and communication strategies to enhance the impact of RAM in clinical practice. Supplementary Information Below is the link to the electronic supplementary material. Supplementary file1 (PDF 206 KB) Funding No funding was received to assist with the preparation of this manuscript. John F. P. Bridges holds an Innovation in Regulatory Science Award from the Burroughs Wellcome Fund. Benjamin G. Mittman was supported in part by the National Institutes of Health (NIH) under award T32GM152319. Declarations Conflict of interest Nicola B. Campoamor, Benjamin G. Mittman, Stephanie Moss, Michael B. Rothberg, and John F.P. Bridges declare that they have no potential conflicts of interest that might be relevant to the contents of this manuscript. Ethics approval All study procedures were reviewed and approved by the Cleveland Clinic Institutional Review Board (identifier no. 20-321). Consent to participate Not applicable. Consent for publication Not applicable. Data availability Request for data and materials from this study can be sent to the Cleveland Clinic via Michael Rothberg. Code availability Not applicable. Author contributions Study design: MBR, JFPB; study conduct: BGM, SM, MBR, JFPB; data analysis: NBC, JFPB; manuscript drafting: all authors; critical manuscript revision: all authors. References Associated Data Supplementary Materials Supplementary file1 (PDF 206 KB) Data Availability Statement Request for data and materials from this study can be sent to the Cleveland Clinic via Michael Rothberg.