Applying a Random Forest Approach in Predicting Health Status in Patients with Carotid Artery Stenosis 30 Days Post Stenting

Abstract

BACKGROUND

Approximately 20% of ischemic strokes in the United States result from carotid artery stenosis. Carotid artery stenting (CAS) can reduce stroke risk, but variability in poststenting health outcomes and their predictors is poorly understood. We examined the 30‐day post‐CAS health status and derived its most important predictors.

METHODS

The Stenting and Angioplasty with Protection of Patients with High Risk for Endarterectomy (SAPPHIRE) Worldwide Registry measured preprocedural and 30‐day health status for patients undergoing transfemoral CAS using the 36‐item Short Form Survey Mental Component Summary and Physical Component Summary, EuroQoL 5‐Dimension Index Value, and Visual Analogue Scale. Random Forest models ranked 66 preprocedural candidate variables by relative importance (RI) in predicting 30‐day post‐CAS health status, stratified by patient symptomatic status. Variables with the highest relative importance were identified and used to develop predictive multivariable linear regression models, which were evaluated using R‐square (coefficient of determination) and root mean square error.

RESULTS

Health status was assessed using the 36‐item Short Form Survey in 3017 patients and EuroQoL 5‐Dimension in 3930 patients. Random forest models identified 9 key predictors of post‐CAS health status: preprocedural health status (RI 100%), Modified Rankin Scale score (RI 26.2%–76.5%), National Institutes of Health Stroke Scale score (RI 12.1%–28.0%), history of stroke (RI 9.2%–19.8%), congestive heart failure (RI 12.3%–19.7%), spinal immobility (RI 6.7%–31.0%), diabetes mellitus (RI 8.1%–32.9%), severe pulmonary disease/chronic obstructive pulmonary disease (RI 13.8%–45.6%), and non‐Hispanic/Latino ethnicity (RI 8.4%–32.4%). Multivariable linear regression models explained ∼36%–61% of the health status variance, with 36‐item Short Form Survey models (R‐square = 36%–61%) outperforming EuroQoL 5‐Dimension models (R‐square 37%–44%) with regard to R‐square and visual fit of observed versus predicted values.

CONCLUSIONS

We derived multivariable linear regression‐based prediction models that partially explained 30‐day post‐CAS health status outcomes. Preprocedural health status scores, stroke scale scores, and medical comorbidities may have utility in appropriately risk‐stratifying patients under consideration for CAS and should be considered when discussing health status benefits in pre‐CAS treatment shared decision‐making discussions.

Clinical Trial Registration Information

This study analyzed data from the SAPPHIRE Worldwide: Stenting and Angioplasty with Protection of Patients with High Risk for Endarterectomy trial.

ClinicalTrials.gov ID

NCT00403078.URL: https://clinicaltrials.gov/study/NCT00403078

In the United States, around 20% of ischemic strokes present secondary to carotid artery stenosis. ¹ Medical management and carotid revascularization can decrease ischemic stroke risk for carotid artery stenosis through carotid artery stenting (CAS) and carotid endarterectomy (CEA). ¹ , ² , ³ CAS has been shown to be noninferior to CEA in patients with severe carotid artery stenosis. ³ However, revascularization benefits must be weighed against nonnegligible risks such as periprocedural stroke and mortality, which can occur in up to 5.6% and 1.4% of patients undergoing urgent revascularization, respectively. ² , ⁴ Nonetheless, the American College of Cardiology recommends that CAS is reasonable to consider in symptomatic patients with ≥50% stenosis or asymptomatic patients with ≥70% stenosis who possess anatomical or physiological risk factors that make them poor candidates for alternative forms of revascularization. ⁵

Besides stroke risk, CAS has implications for individuals’ health status. Although short‐term health status changes have been noted for CAS over CEA at 2 weeks post intervention, it is unknown what the main predictors of poststenting health status outcomes are. ⁶ As current CAS trials actively compare medical management, CAS, and CEA head to head, it is important to compare health status outcomes from clinical trials against real‐world observational cohorts who have undergone revascularization procedures, such as CAS, to inform expected benchmarks and interpret effect sizes for meaningful health status changes associated with treatment. ⁷ Additionally, understanding which clinical patient profiles are associated with health status improvement post CAS is critical in informing shared decision‐making in carotid artery stenosis.

We aimed to identify the most important patient characteristics in predicting health status 30 days post stenting in patients with carotid artery stenosis using a random forest‐based (RF) machine‐learning approach in the Stenting and Angioplasty with Protection of Patients with High Risk for Endarterectomy (SAPPHIRE) Worldwide Registry.

METHODS

The data that support the findings of this study can be requested from the principal investigators of the SAPPHIRE Worldwide study. ⁸ , ⁹

Study Design and Patient Cohorts

The SAPPHIRE Worldwide Registry is an international (United States, Canada), prospective, observational registry containing data from 390 clinical sites encompassing patients enrolled between 2006 and 2014 who underwent either transfemoral‐CAS or CEA for asymptomatic or symptomatic carotid artery stenosis. ⁸ , ⁹ The 36‐item Short Form Survey (SF‐36) was administered to patients enrolled between 2010–2012, and the EuroQoL 5‐Dimension (EQ‐5D) to patients enrolled between 2012 and 2014. ¹⁰ , ¹¹ The National Institutes of Health Stroke Scale and Modified Rankin Scale were administered to patients enrolled between 2010 and 2014. ¹² , ¹³ Patient information was obtained from medical records and interviews conducted by clinicians and clinical volunteers.

Nonstandard Abbreviations and Acronyms

CAS

carotid artery stenting

CEA

carotid endarterectomy

EQ‐5D

EuroQoL 5‐Dimension

MCS

Mental Component Summary

OOB

out‐of‐bag

PCS

Physical Component Summary

RF

random forest

RI

relative importance

SAPPHIRE

Stenting and Angioplasty with Protection of Patients with High Risk for Endarterectomy

SF‐36

36‐item Short Form Survey

VAS

Visual Analogue Scale

CLINICAL PERSPECTIVE

What Is New?

• This study applied a random Forest machine‐learning approach to identify 9 preprocedural patient characteristics for asymptomatic and symptomatic carotid artery stenosis patients, which can significantly predict health status outcomes (as measured by the EuroQoL‐5D or 36‐item Short Form Survey) at 30 days following transfemoral carotid artery stenting. These patient characteristics include baseline health status scores, stroke scale scores (Modified Rankin and National Institutes of Health Stroke Scale scores), key medical comorbidities (diabetes, congestive heart failure, chronic obstructive pulmonary disease, spinal immobility), history of prior stroke, and non‐Hispanic ethnicity.

• Our 9‐variable, multivariable regression‐based predictive models explained ∼37% to61% of variability in 30‐day poststenting health status outcomes, performing best in predicting post‐stenting Short Form‐36 Physical Component Summary scores.

What Are the Clinical Implications?

• These initial insights could serve to inform shared decision‐making frameworks for patients with carotid artery stenosis considering stenting, supporting both clinicians and patients in their discussions about the expected benefits and risks of carotid revascularization procedures, particularly given the variability in poststenting individual health status trajectories and the nonnegligible clinical risks associated with revascularization.

From SAPPHIRE, we derived 2 patient cohorts: SF‐36 and EQ‐5D. Inclusion criteria were (1) stenting performed between 2006 and 2014, (2) stenting performed during 2010–2012 (SF‐36 administered) or 2012–2014 (EQ‐5D administered), (3) ≥18 years old, and (4) having ≥1 high surgical risk factor. ⁸ , ⁹ Exclusion criteria were (1) Instructions for Use for the stent (Cordis PRECISE Nitinol Stent System) used were not followed during stenting or information regarding whether Instructions for Use were followed is missing, (2) no stent was deployed at the intended location, (3) health status (ie, SF‐36 or EQ‐5D) was not assessed at baseline, (4) baseline health status was not assessed <28 days prior to procedure, (5) baseline health status assessment score was missing, (6) 30‐day poststenting health status was not assessed, (7) 30‐day poststenting health status was assessed ±7 days out of the 30‐day poststenting window, and (8) 30‐day poststenting health status assessment score was missing.

Symptomatic patients experience symptoms secondary to their carotid artery stenosis or associated pathology (ie, stroke, transient ischemic attack) whereas asymptomatic patients do not. Symptomatic patients face markedly higher stroke risk and are clinically managed differently from asymptomatic patients. ¹ , ² , ⁴ , ⁵ , ¹⁴ We therefore stratified SF‐36 and EQ‐5D cohorts by symptomatic status to contrast findings across symptomatic status.

Health Status Measures

The main outcome was patient health status scores at 30±7 days post transfemoral CAS on the following health status measures: SF‐36 Physical Component Summary (PCS), SF‐36 Mental Component Summary (MCS), EQ‐5D Index Value, or EQ‐5D Visual Analogue Scale (VAS).

The SF‐36 and EQ‐5D are standardized, generic health status instruments that ask patients to rate their physical and mental health status across various domains. Both have been extensively described. ¹⁰ , ¹¹ Scores on both instruments can be converted to summary scores (ie, PCS and MCS for the SF‐36, Index Value and VAS for EQ‐5D) using country‐specific weights. Higher scores indicate better health status. Both instruments see widespread use in calculating quality‐adjusted life years, health economic analyses, and evaluation of patients with cardiovascular disease, as lower health status scores correlate with the level of anatomic disease, disease progression, and risk of poor clinical outcomes. ¹⁰ , ¹¹ , ¹⁵ , ¹⁶ , ¹⁷

Candidate Variables

We considered 66 preprocedural candidate variables, which have been described in detail in previous work ¹⁸ and are listed in Tables S1–S4. Candidate variables included preprocedural health status scores. We combined variables with prevalence ≤1% among patients in any cohort within a “Rare Characteristics” variable in our analysis. Candidate variables were categorized as follows: demographic, high‐risk criteria, medical history, neurological history, procedural and lesion characteristics, and baseline stroke scale or health status scores. Information regarding health inequalities between socioeconomic groups in our patient cohorts, other than health insurance status (ie, public insurance, private insurance, or no insurance), was not included in our data set. High‐risk criteria are the most common risk factors among SAPPHIRE‐eligible patients who underwent CEA.

The preprocedural National Institutes of Health Stroke Scale and Modified Rankin Scale scores, which measure patient disability in stroke, were included as candidate variables. ¹² , ¹³ Both correlate strongly with postintervention health status for patients experiencing a stroke, can be reliably reproduced, and predict inpatient length of stay after experiencing a stroke. ¹⁹ , ²⁰ , ²¹

Random Forest Model Development

We constructed 8 RF models, 1 for each health status score (SF‐36 PCS and MCS, EQ‐5D Index, and VAS) stratified by patient symptomatic status (asymptomatic or symptomatic). Each RF ranked 66 candidate variables based on their RI in predicting patient health status 30 days post ‐stenting.

RF was selected for its superior performance in handling variables with multicollinearity and nonlinear effects on the outcome. ²² To reduce overfitting and multicollinearity, RFs train each tree using a subset of randomly selected variables using bootstrap sampling (constructed using two thirds of the data). The remaining one third, known as out‐of‐bag (OOB), was used to assess RF error. ²²

First, the SF‐36 and EQ‐5D subcohorts were randomly divided into training samples (80% of data) to build RFs and testing samples (20%) to validate RFs. We assessed candidate variable similarity between these samples using standard differences (d). Continuous variables were summarized as means with SDs and compared using Student t‐tests. Categorical variables were summarized as frequencies and were compared using chi‐square or Fisher's exact tests. We assessed effect sizes between cohorts using standard differences (d). An absolute value |d| ≤0.1, ≤0.2, ≤0.5, or ≤0.8 was considered a negligible, small, moderate, or large effect size, respectively. ²³

Second, we tuned RF hyperparameters, the number of decision trees, and the number of randomly selected variables, using a grid search method in the training sample. ²² To minimize OOB error, RFs were trained using 500 decision trees, with each tree considering a random subset of 30 out of 66 candidate variables.

Then, using testing sample data, 66 candidate variables were ranked based on their RI (Breiman–Cutler importance) in predicting 30‐day poststenting health status scores. ²² For each RF, we determined the most important variables via graphical relative importance (RI) plots. On graphical plots of RI, a noticeable and discontinuous inflection point in RI served as the selection threshold for the most important variables.

Finally, using testing sample data, we evaluated RF accuracy using OOB accuracy (ie, 1–OOB error), which estimates RF generalization to unseen data, and the root mean square error. ²³

Multivariable Linear Regression Model Development

Because RFs do not assess the directionality of a variable's effect, we used linear regression models to create more easily implementable and interpretable clinical models. We constructed 8 multivariable linear regression models, 1 for each health status instrument stratified by patient symptomatic status. Each regression model included variables previously identified as the “most important” in any RF, with care center incorporated as a random effect variable, to derive coefficients (β) with a 95% CI. Model performance was evaluated based on the R‐square, along with plots of observed versus predicted health status scores.

This study was conducted and reported in accordance with the Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis guidelines for prognostic model development and validation. ²⁴ A formal study protocol was prepared and shared with the institutional review board and sponsor of the database.

All statistical tests were 2 tailed with an alpha of 0.05. Statistical analyses were performed using Stata Statistical Software: Release 17 (StataCorp LLC; College Station, TX).

Missing Data

Missing data in RFs were handled during the RF's growth process, as RFs consider a subset of randomly selected variables during each tree's construction. Missing data in linear regression models were handled using multiple imputation by chained equations, with 10 imputed data sets. Each imputed data set was then analyzed separately, and the parameter estimates (including SEs) were subsequently combined using Rubin's Rules. ²³ This pooling procedure accounts for both within‐imputation and between‐imputation variance, yielding valid overall estimates and inference in the presence of missing data.

RESULTS

Overview of SF‐36 and EQ‐5D Cohorts and Rare Characteristics

The SF‐36 cohort consisted of 3017 participants (14.36% of the registry), and the EQ‐5D cohort consisted of 3390 participants (18.71% of the registry) (Figure 1). After stratifying cohorts by symptomatic status, SF‐36 subcohorts included 2098 asymptomatic patients and 919 symptomatic patients. EQ‐5D sub‐cohorts included 2840 asymptomatic patients and 1090 symptomatic patients.

Figure 1.

Patient flow diagrams showing the derivation of patient cohorts from the SAPPHIRE registry for (A) the SF‐36 cohort and (B) the EQ‐5D cohort. The diagram shows inclusion and exclusion criteria applied to derive patient cohorts, the number of patients included or excluded at each step, and the number of patients in each cohort after applying inclusion and exclusion criteria. EQ‐5D indicates EuroQoL 5‐Dimension; SAPPHIRE, Stenting and Angioplasty with Protection of Patients with High Risk for Endarterectomy; and SF‐36, 36‐item Short Form Survey.

We categorized 11 variables as “Rare Characteristics” in RFs: having contralateral laryngeal palsy, being on an organ transplant list, having a tracheostomy stoma, requiring cardiac valve surgery, recent myocardial infarction, no health insurance, previous mitral valve replacement, previous extracranial‐intracranial bypass, significant aortic arch disease, no medical history available, and history of open‐heart surgery.

The rate of missing data ranged from 0% to 4.3% (target lesion length variable) in the SF‐36 cohort and from 0% to 4.5% (lesion reference diameter variable) in the EQ‐5D cohort. ¹⁸

Patient Characteristics for EQ‐5D and SF‐36 Cohorts

A detailed description and comparison of all 66 patient characteristics and candidate variables considered for the EQ‐5D and SF‐36 cohorts, as well as asymptomatic versus symptomatic subcohorts, has been previously described by prior work. ¹⁸ In summary, baseline characteristics across the SF‐36 versus EQ‐5D cohorts were not different (all |d|<0.20), with a mean age of 72.00±9.30 years versus 72.43±9.26 years, 92.8% versus 92.5% of patients being White, and 61.5% versus 63.2% of patients being male, respectively. In the SF‐36 cohort, 69.5% of patients were asymptomatic compared with 72.3% in the EQ‐5D cohort (d = 0.06). Patient characteristics for SF‐36 and EQ‐5D asymptomatic and symptomatic subcohorts have also been described and compared in prior work. ¹⁸ Additionally, sociodemographic characteristics were similar between asymptomatic and symptomatic patients in both SF‐36 and EQ‐5D cohorts (|d|<0.20). However, symptomatic patients in the SF‐36 subcohort had higher rates of bilateral artery stenosis, prior coronary artery disease, neurological history, lower diameter stenosis percentage, and ulcerated lesions (|d|≥0.20), whereas those in the EQ‐5D subcohort exhibited a higher prevalence of neurological history, lower diameter stenosis percentage, and ulcerated lesions (|d|≥0.20). Patient characteristics for training and testing samples for RFs are described and compared in Tables S1–S4, with no significant differences (all |d|<0.20) in patient characteristics between testing and training samples across RF models.

RFs: Relative Importance in Predicting 30‐Day Poststenting Health Status and Performances

The RI of the 66 candidate variables in predicting 30‐day poststenting SF‐36 (MCS and PCS) and EQ‐5D (Index Value and VAS) scores stratified by symptom status are presented in Tables S5 and S6. RF plots of variables with the highest RI for each health status outcome stratified by symptomatic status are presented in Figures S1 and S2.

We identified 9 variables as “most important” in predicting the 30‐day health status scores. Baseline health status score emerged as the most important variable, with an RI of 100% across all health status outcomes. The other “most important” variables were Modified Rankin Scale score (RI 26.2%–76.5%), National Institutes of Health Stroke Scale score (RI 12.1%–28.0%), history of stroke (RI 9.2%–19.8%), congestive heart failure (RI 12.3%–19.7%), spinal immobility (RI 6.7%–31.0%), diabetes mellitus (RI 8.1%–32.9%), severe pulmonary disease/chronic obstructive pulmonary disease (RI 13.8%–45.6%), and not having Hispanic/Latino ethnicity (RI 8.4%–32.4%).

The RFs showed good performance, with OOB accuracy ranging from 86.70% to 95.16% (Table S7). RF Root mean square errors are presented in Table S7.

Multivariable Linear Regression

When evaluating 30‐day SF‐36 PCS and MCS, higher preprocedural scores were associated with better 30‐day health status regardless of patient symptomatic status (P‐values <0.001) (Figure 2). In asymptomatic patients, having severe pulmonary disease/chronic obstructive pulmonary disease was associated with decreased 30‐day SF‐36 PCS only (β = −1.15, 95% CI −1.88 to −0.41, P value = 0.002), whereas a higher preprocedural Modified Rankin Scale score was associated with decreased 30‐day SF‐36 PCS and MCS (SF‐36 PCS: β = −0.92, 95% CI −1.39 to −0.44, P value <0.001; SF‐36 MCS: β = −0.79, 95% CI −1.27 to −0.31, P value = 0.001) (Figure 2a). In symptomatic patients, having diabetes mellitus was associated with a decreased SF‐36 PCS score (β = −1.23, 95% CI −2.30 to ‐0.16, P value = 0.024) (Figure 2b).

Figure 2.

Forest plots of variable coefficients from multivariable linear regression models used to predict 30‐day post‐stenting SF‐36 PCS and MCS scores. A, variable coefficients and 95% CIs for asymptomatic patients’ scores 30 days post‐stenting. B, variable coefficients and 95% CIs for symptomatic patients’ scores 30 days post‐stenting. Coefficients represent the estimated change in 30‐day post‐stenting SF‐36 Physical Component Summary or Mental Component Summary score associated with each variable. Variables were chosen using random forest modeling. * indicates coefficients with P value <0.05. ** indicates coefficients with P value <0.01. COPD indicates chronic obstructive pulmonary disease; MCS, Mental Component Summary; NIH, National Institutes of Health; PCS, Physical Component Summary; and SF‐36, 36‐item Short Form Survey.

Similarly, when evaluating 30‐day EQ‐5D Index Value and VAS, higher preprocedural scores were associated with an increase in 30‐day health status regardless of patient symptomatic status (P‐values <0.001). In asymptomatic patients, congestive heart failure and severe pulmonary disease/chronic obstructive pulmonary disease were most strongly associated with a decreased 30‐day EQ‐5D Index Value and VAS Score (Figure S3a), whereas in symptomatic patients, a higher preprocedural Modified Rankin Scale score was most strongly associated with decreased 30‐day EQ‐5D Index Value and VAS (Figure S3b).

Performances of multivariable linear regression models are summarized in Figures 3 and 4. R‐squares ranged from 37% for 30‐day EQ‐5D VAS scores in symptomatic patients to 61% for SF‐36 PCS scores in asymptomatic patients. Root mean square errors of regression models was comparable to those of RFs, suggesting that including additional variables to predict the health status at 30 days poststenting would not improve the models’ predictive performance.

Figure 3.

Scatter plots of observed versus predicted 30‐day post‐stenting EQ‐5D Index Value scores for (A) asymptomatic and (B) symptomatic patients, as well as EQ‐5D VAS scores in (C) asymptomatic and (D) symptomatic patients. Predicted scores were obtained from by multivariable linear regression models. A red dashed line is plotted to represent the y=x line for visual reference. R‐square (coefficient of determination) and RMSE are listed for each scatter plot. EQ‐5D indicates EuroQoL 5‐Dimension; RMSE, root‐mean‐square‐error; VAS, Visual Analogue Scale.

Figure 4.

Scatter plots of observed versus predicted 30‐day post‐stenting SF‐36 PCS scores for (A) asymptomatic and (B) symptomatic patients, as well as SF‐36 MCS scores in (C) asymptomatic and (D) symptomatic patients. Predicted scores were obtained from multivariable linear regression models. A red dashed line is plotted to represent the y=x line for visual reference. R‐square (coefficient of determination) and RMSE are listed for each scatter plot. MCS indicates Mental Component Summary; PCS, Physical Component Summary; RMSE, root‐mean‐square error; and SF‐36 indicates 36‐item Short Form Survey.

DISCUSSION

Our study used machine learning to develop 30‐day health status prediction models for patients undergoing transfemoral CAS that captured patients’ preprocedural health status, stroke scale scores, and medical comorbidities. Our models explained ∼37% to61% of the variance in health status 30 days post transfemoral CAS, with SF‐36 PCS models performing best.

Health status evaluation in patients with carotid artery stenosis is still emergent. Currently, no disease‐specific instruments or prediction models effectively and prospectively capture multidimensional demographic, socioeconomic, clinical, and psychosocial determinants of patients’ health status outcomes post revascularization to understand which patients will benefit from revascularization based on their characteristics. Previous work shows that not all patients with carotid artery stenosis experience meaningful improvements in health status post CAS, further highlighting the importance of evaluating risks versus benefits of CAS for patients. ¹⁸ With clinical trials for carotid artery stenosis treatment underway, such as the ongoing Carotid Revascularization Endarterectomy versus Stenting Trial, and with Centers for Medicare & Medicaid Services expanding reimbursement for CAS, there is an urgent need to consider shared decision‐making approaches in deciding which treatment modality fits best with one's expected risk profile, and health status outcomes become invaluable in weighing those decisions. ⁷ , ²⁵ Our findings will help interventionalists and patients better weigh risks and benefits of CAS to make more tailored treatment decisions by providing insight into whether a patient can expect to see health status improvements post stenting based on their characteristics and comorbidities in light of the nonnegligible risk of periprocedural adverse events. ² , ⁴

The 9 variables we identified from the SAPPHIRE registry can reliably predict the health status changes of patients with carotid artery stenosis, explaining 61% and 54% of changes in asymptomatic and symptomatic patients’ SF‐36 PCS scores post stenting, respectively. These variables, obtainable through patient charts or clinical evaluation, include demographics, comorbidities, preprocedural stroke scale scores, and preprocedural health status. While our study used RF models, their computational complexity likely limits near‐term use. As a more feasible alternative, our 9‐variable multivariable regression models may be incorporated into clinical practice to better predict post‐CAS physical and mental health status outcomes. However, the SAPPHIRE registry lacks detailed information regarding patients’ psychosocial determinants of health, socioeconomic status, and cognitive status. Incorporating more variables that capture multidimensional aspects of health may have improved our performance in predicting poststenting health status.

Although generic health status instruments may not have captured elements important to the experiences of patients with carotid artery stenosis, our RF and linear regression models performed better in predicting poststenting SF‐36 PCS and MCS scores than EQ‐5D Index Value and VAS scores. Therefore, the SF‐36 may correlate more strongly with commonly captured clinical characteristics among patients with carotid artery stenosis than the EQ‐5D. Alternatively, EQ‐5D scores may be more difficult to predict or model. ²⁶ Regardless, existing health status instruments would benefit from complementary carotid artery stenosis‐specific health status instruments that measure all relevant aspects of health status for patients with carotid artery stenosis, including not only patients’ clinical characteristics but also their psychosocial and socioeconomic determinants of health. Such instruments would allow clinicians to more accurately discern which patient characteristics are most predictive of positive outcomes post stenting, develop better health status prediction models, and enhance shared decision‐making discussions with patients surrounding whether or not to undergo CAS. Until disease‐specific health status instruments are developed, the SF‐36 may be a reasonable choice of a generic health status instrument for health status prediction for patients with carotid artery stenosis.

Our study should be validated and expanded upon. The development of more comprehensive registries of patients with carotid artery stenosis, which include more detailed information regarding patients’ clinical characteristics, psychosocial determinants of health, treatment preferences, and socioeconomic status, is important, as discussed previously. Future work should examine health status changes for patients undergoing alternative revascularization procedures, such as CEA or transcarotid artery revascularization, and examine whether the patient variables identified in this study meaningfully predict health status outcomes in patients with carotid artery stenosis who undergo CEA or transcarotid artery revascularization. Although we chose an RF‐based approach, future studies should validate whether alternative statistical models yield improved predictive power in predicting poststenting health status. Previous work has examined the rate of major adverse events (ie, stroke, myocardial infarction, death) in carotid artery stenosis patients who undergo revascularization with CAS ²⁷ ; future work should also examine whether the variables identified in our study meaningfully predict the risk of major adverse events in such patients.

Our study has limitations. First, our analysis may not be generalizable to patients undergoing revascularization procedures other than transfemoral‐CAS or to patients outside of the SAPPHIRE Worldwide registry. Second, although we stratified patients based on symptomatic status (symptomatic versus asymptomatic), we were unable to further stratify symptomatic patients by specific presenting symptoms (such as stroke, transient ischemic attack, or amaurosis fugax) or by presentation type (emergent versus nonemergent) due to the lack of more detailed patient information beyond the scope of the SAPPHIRE registry. ⁸ , ⁹ However, such information is clinically meaningful. Symptomatic patients with carotid artery stenosis can experience different clinical trajectories and risks of adverse events depending on which of their initial symptoms (eg, stroke, TIA, amaurosis fugax) led to their categorization as a symptomatic patient. ²⁸ Future work should explore whether further stratifying symptomatic patients in this manner affects health status outcome prediction in CAS patients. Third, we evaluated health changes based on health status scores 30 days post stenting, which may not be reflective of changes to patients’ health status >30 days post stenting. Future work should evaluate patient health status changes at further time intervals post stenting. Fourth, the SAPPHIRE Worldwide Registry enrolled patients who underwent revascularization for carotid artery stenosis between 2006–2014 and had a limited scope since it examined patients who underwent revascularization with only 1 type of stent (Cordis PRECISE Nitinol Stent) and 1 type of embolic filter (Cordis ANGIOGUARD XP Emboli Capture Guidewire). ⁶ , ⁸ , ⁹ Given the advancements in carotid artery revascularization techniques and devices since 2014, ²⁹ our findings may not fully reflect current clinical practices or effects on the health status of patients who are revascularized with more recently established techniques or devices. At the same time, we are not aware of any more up‐to‐date, large, real‐world cohorts that focus on patient‐centric experiences, which are outcomes data that are overdue for consideration in the clinical decision‐making process. Finally, our health status prediction model may have limited clinical utility when evaluating patients for whom pertinent data are either of poor quality or not available.

Using a machine‐learning‐based approach, we developed the first predictive linear regression models for health status variation 30 days post transfemoral CAS for patients with carotid artery stenosis. Our model encapsulated information regarding patients’ preprocedural health status, stroke scale scores, and medical comorbidities and may help guide patients’ and interventionalists’ decision‐making when considering transfemoral CAS for asymptomatic and symptomatic carotid artery stenosis.

Sources of Funding

This study was supported by an unrestricted research grant received by Cordis.

Disclosures

Dr Smolderen reports unrestricted research grants from Abbott, Merck & Co., Inc., Shockwave Medical Inc., and Janssen Pharmaceutical Companies of Johnson & Johnson. She is a consultant for Terumo, Cook Medical, Inc., NovoNordisk, and Dario Health; she is owner of BoboDream LLC. Dr Mena‐Hurtado reports grant funding from Shockwave Medical, Inc., Merck & Co., Inc. and is a consultant for Cook Medical, Inc., and Terumo. All other authors report no disclosures or conflicts of interest. This study was supported by an unrestricted research grant received by Yale University.

Ethical approval and Informed consent

This research was approved as a secondary analysis protocol exempt from annual institutional review board review under section 45CFR46.104(4). Patients participating in SAPPHIRE were previously consented per the original protocol of the registry. ⁸ , ⁹ There was no patient or public involvement in the design, conduct, reporting, interpretation, or dissemination of the study.

Code sharing statement

The code executed within Stata Statistical Software: Release 17 used to conduct this study can be shared upon reasonable request to the corresponding author.

Supporting information

Table S1. Baseline characteristics of asymptomatic patients within the SF‐36 cohort, stratified by the testing (20% of total patients) and training (80% of total patients) samples used to train the SF‐36 asymptomatic random forests. Values are presented as N (%) unless otherwise specified. SF‐36 indicates 36‐item Short Form Survey.

Table S2. Baseline characteristics of symptomatic patients within the SF‐36 cohort, stratified by the testing (20% of total patients) and training (80% of total patients) samples used to train the SF‐36 symptomatic random forests. Values are presented as N (%) unless otherwise specified. SF‐36 indicates 36‐item Short Form Survey.

Table S3. Baseline characteristics of asymptomatic patients within the EQ‐5D cohort, stratified by the testing (20% of total patients) and training (80% of total patients) samples used to train the EQ‐5D asymptomatic random forests. Values are presented as N (%) unless otherwise specified. EQ‐5D indicates EuroQoL 5‐Dimension.

Table S4. Baseline characteristics of symptomatic patients within the EQ‐5D cohort, stratified by the testing (20% of total patients) and training (80% of total patients) samples used to train the EQ‐5D symptomatic random forests. Values are presented as N (%) unless otherwise specified. EQ‐5D indicates EuroQoL 5‐Dimension.

Table S5. Relative importance and ranking of the 66 preprocedural candidate variables in predicting the SF‐36 Physical Component Summary (PCS) and Mental Component Summary (MCS) scores 30 days post stenting in asymptomatic and symptomatic patients.

Table S6. Relative importance and rank of the 66 preprocedural candidate variables in predicting the EQ‐5D Index Value and Visual Analogue (VAS) scores at 30 days post stenting in asymptomatic and symptomatic patients. EQ‐5D indicates EuroQoL 5‐Dimension.

Table S7. Random forest and multivariable linear regression model performance for each health status measure stratified by symptom status.

Figure S1 Top 10 candidate variables by relative importance for predicting 30‐day poststenting health status scores for EQ‐5D Index Value for asymptomatic (A) and symptomatic (B) patients as well as EQ‐5D Visual Analogue Scale for asymptomatic (C) and symptomatic (D) patients, as determined by random forest models.* EQ‐5D indicates EuroQoL 5‐Dimension.

Figure S2 Top 10 candidate variables by relative importance for predicting 30‐day poststenting health status scores for SF‐36 PCS for asymptomatic (A) and symptomatic (B) patients as well as SF‐36 MCS for asymptomatic (C) and symptomatic (D) patients, as determined by random forest models.* MCS indicates Mental Component Summary; PCS, Physical Component Summary; and SF‐36, 36‐item Short Form Survey.

Figure S3 Coefficients and 95% CIs derived from multivariable linear regression models for predicting 30‐day poststenting EQ‐5D Index Value and Visual Analogue Scale (VAS) scores for (A) asymptomatic patients and (B) symptomatic patients. EQ‐5D indicates EuroQoL 5‐Dimension.