Comparison of Clinical Trials and Administrative Claims to Identify Stroke Among Patients Undergoing Aortic Valve Replacement: Findings from the Extending Trial-Based Evaluations of Medical Therapies Using Novel Sources of Data (EXTEND) Study

Abstract

Background:

Cerebrovascular events (CVEs) are devastating complications after aortic valve replacement. We assessed whether billing claims accurately identify CVEs in place of clinical event adjudication in structural heart disease trials.

Methods:

Adult participants in the US CoreValve High Risk and SURTAVI trials were linked to Medicare inpatient claims from 1/1/2006 to 12/31/2016. Claims consistent with CVEs within 14 days of a similar trial-adjudicated CVE were considered a “match.” The sensitivity, specificity, and positive (PPV) and negative (NPV) predictive values of International Classification of Diseases, 9^th and 10^th Revisions, Clinical Modification (ICD-9-CM and ICD-10-CM) billing codes for cerebrovascular disease were determined against trial-defined CVEs as the criterion standard. Kaplan-Meier estimates of claims-defined vs. trial-defined CVEs were compared.

Results:

Among 4230 linked trial participants (linkage rate 79.8%), 550 (13.0%) sustained 630 adjudicated CVEs over a 5-year follow-up period. Linked and non-linked individuals were similar. An algorithm using four ICD-9-CM codes (434.91, 434.11, 433.11, and 997.02) had a sensitivity of 60.9%, specificity of 99.0%, PPV of 86.5%, and NPV of 95.8% for identifying trial-adjudicated ischemic stroke. An algorithm using three ICD-10-CM codes (I63.9, I63.40, I63.49) had a sensitivity of 66.7%, specificity of 99.4%, positive predictive value of 88.9%, and negative predictive value of 97.6%.

Conclusions:

In linked clinical trial and Medicare claims data, four ICD-9-CM and three ICD-10-CM billing codes identified half of trial-adjudicated CVEs during follow-up with high specificity and predictive value, but imperfect sensitivity. While low sensitivity may limit the use of claims to substitute for traditional trial outcomes to identify CVEs, high specificity suggests claims could be used to trigger evaluation of neurological events, potentially improving the efficiency of evaluation of techniques and devices designed to reduce such events.

INTRODUCTION

Stroke is an important cause of morbidity and mortality among patients undergoing aortic valve replacement (AVR), occurring in 4.9% of transcatheter (TAVR) and 6.2% of surgical AVR (SAVR) patients at 30 days.¹ While several devices for cerebral embolic protection have been tested, they have not yet demonstrated a reduction in clinical stroke.² Current studies are unable to assess differences in rates of stroke, in part due to the high cost of running an adequately powered trial.³ Centralized event adjudication is an important driver of high trial costs. As an alternative, use of hospital billing claims to identify adverse events could reduce trial costs while maintaining validity, enabling the conduct of larger, more highly powered trials. However, use of claims, collected for reimbursement, to substitute for clinical events committee (CEC)-adjudicated stroke in a pivotal cardiovascular device trial has not been tested, and concerns exist about miscoding or undercoding of events leading to inaccurate or biased event rates and effect estimates. ^{4, 5}

We report the performance of Medicare claims to identify CEC-adjudicated stroke and transient ischemic attack (TIA) in the US CoreValve Program (CVPT),^1,6–7 a collection of large trials and registries evaluating the efficacy of the self-expanding CoreValve TAVR for treatment of severe aortic stenosis. Our primary objectives were to determine the concordance between claims-based and trial-adjudicated stroke and TIA outcomes, and assess whether similar rates of these events were captured by the trial and claims.

METHODS

Data were used according to prespecified data use agreements and are not publicly accessible.

Study Population

We evaluated patients undergoing AVR included in the CVPT from 1/1/2006–12/31/2016 who could be linked to the Medicare Provider Analysis and Review (MedPAR) database (Figure 1). The program consists of 3 large trials (N = 3,049) comparing TAVR using the Medtronic self-expanding CoreValve bioprosthesis with SAVR: the SURTAVI study,⁷ High Risk study,¹ and the Extreme Risk Study.⁶ Additionally, participants in the Continued Access Study (N = 2,732), a single-arm cohort of individuals receiving TAVR, were included.⁸ The MedPAR database consists of 100% inpatient discharge claims (Part A claims) for Medicare Fee-for-Service beneficiaries.^9–12

Figure 1:

Schematic Representing Linkage Strategy and Included Studies

Linkage between the CVPT and MedPAR databases was performed as part of the Extending Trial-Based Evaluations of Medical Therapies Using Novel Sources of Data (EXTEND) Study as previously described (Figure 1).¹³ Patients < 65 years and those whose index AVR was at a Veterans Affairs or European hospital were excluded (N = 479). Claims for Medicare Advantage patients were not available for linkage. As Medicare Advantage patients represented 13–30% of Medicare enrollees during the time period evaluated,¹⁴ they likely represented the majority of non-matched individuals. The study was approved by the Institutional Review Board at Beth Israel Deaconess Medical Center.

Covariates and Outcomes

Clinical variables were determined for AVR recipients using baseline trial variables.^1,6–7 Codes classified as “Cerebrovascular Diseases” or “Cerebral Infarction” (codes 430.X-438.X; 997.02) based on the International Classification of Diseases, 9^th Revision, Clinical Modification (ICD-9-CM), or those classified as “Cerebrovascular Diseases” (codes I60.X-I69.X) or “Transient Cerebral Ischemic Attacks and Related Symptoms” (codes G45.X) based on the International Classification of Diseases, 10^th Revision, Clinical Modification (ICD-10-CM) were considered a priori to represent a possible cerebrovascular event (CVE) based on published literature and clinical plausibility (eTables 1 and 2). The ICD-10-CM classification was only applied to hospitalizations after its adoption on October 1, 2015. CVE claims in any position from hospitalizations with an admission date 14 days before or after a trial-adjudicated event were considered a “match.” If a claim for CVE occurred during a hospitalization with an admission date outside the 28-day window of a trial event, the claim was considered a “false positive.” Trial-defined CVEs were based on the Valve Academic Research Consortium (High Risk, Extreme Risk, and Continued Access Studies) and Valve Academic Research Consortium-2 (SURTAVI) definitions and had been adjudicated by independent CECs ^15,16 (eTable 3).

Statistical Analysis

Baseline characteristics of linked and non-linked individuals are presented using means and standard deviations (SDs) or frequencies and percentages, and compared using t-tests for continuous variables and the chi square or Fisher’s exact test for categorical variables.

A hierarchical algorithm was used to determine the most sensitive and specific codes for CVEs using trial-adjudicated CVE as the criterion standard. First, codes representing CVEs were matched with trial events and ranked based on frequency. If multiple codes represented a trial CVE, only the most frequent was considered. Second, for each claims classification and CVE type, we identified the sensitivity, specificity, positive (PPV) and negative (NPV) predictive values of the most frequent claim. To avoid low-frequency codes, those matching 5 or more CVEs were subsequently added to the algorithm beginning with the most frequent. To improve specificity, codes were removed if they commonly occurred in hospitalizations without matching trial events. ICD-9-CM codes 433.X were excluded for low specificity. However, as 433.11 (“occlusion and stenosis of carotid artery with cerebral infarction”) improved agreement with trial-defined CVEs, it was retained. As CVE codes during trial follow-up could represent historical events, individuals with at least one CVE claim in the year prior to AVR were excluded. The sensitivity, specificity, PPV and NPV of the final algorithm to detect trial-adjudicated events were calculated. Unweighted kappa statistics were determined as measures of agreement between claims and trials.

The cumulative incidence of CVEs based on claims versus trial-adjudicated endpoints was plotted using the Kaplan-Meier method using the trial date of censoring, and compared using the log-rank test. Only the first trial event was considered. Among individuals with a trial-adjudicated CVE without a matching claims code, we calculated the frequency and proportion of individuals with a hospitalization of any kind in the 14 days before or after the trial event. All analyses were conducted in SAS 9.4 (SAS Institute, Cary, NC) using a two-tailed p < 0.05.

Sensitivity Analysis

We performed recursive partitioning to identify the most parsimonious ICD-9-CM codes for ischemic stroke using a tree-based approach as a secondary mechanism. We grew the tree using an entropy reduction criterion until the node had only one observation or a tree depth of 10 had been reached. We pruned the tree using a cost-complexity strategy with the complexity parameter selected through cross-validation, limiting the maximum number of leaves to 4. We used 10-fold cross validation to evaluate the optimism of this approach, and determined Harrell’s concordance statistic adjusted for model optimism on cross validation. Recursive partitioning was not used for hemorrhagic stroke or TIA (ICD-9-CM) or ICD-10-CM codes due to the low number of events. Subgroup analyses were performed to assess relative performance of algorithms for TAVR vs. SAVR and within vs. after 30 days from the index procedure. Landmark analyses were performed for CVEs starting at 30 days post-procedure, censoring events or deaths within 30 days.

RESULTS

Linkage Results and Baseline Characteristics

A total of 4,230 (79.8%) individuals in the CVPT were linked to Medicare. Linked and non-linked individuals were similar (Table 1), although mean age, Society of Thoracic Surgeons risk score, and the proportion with prior heart failure were higher in the linked group. Baseline patient characteristics were similar among those events identified via claims, via trial adjudication, and by both approaches, although patients in whom events were identified by claims had a higher rate of prior stroke (eTable 4).

Table 1:

Characteristics of Linked and Non-Linked Individuals Included in the EXTEND Study

Characteristic	Linked group (N = 4230)	Non-linked group (N = 1072)	p-value
Age — years ± SD	83.0 ± 6.7	82.4 ± 7.1	0.02
Female sex — no. (%)	1939 (45.8)	478 (44.6)	0.47
Body Mass Index – kg/m2 ± SD	28.2 ± 6.2	28.3 ± 6.4	0.14
New York Heart Association class — no. (%)			0.69
Class II	811 (19.2)	218 (20.3)
Class III	2781 (65.7)	696 (20.0)
Class IV	638 (15.1)	158 (19.9)
Society of Thoracic Surgeons Risk Score — % ± SD	7.9 ± 4.5	7.5 ± 4.4	0.02
Logistic EuroSCORE — % ± SD	19.5 ± 14.6	20.0 ± 15.7	0.0011
Diabetes mellitus — no. (%)
All	1585 (37.5)	400 (37.3)	0.94
Controlled by insulin	545 (12.9)	143 (13.3)	0.68
History of hypertension — no. (%)	3955 (93.5)	1000 (93.3)	0.80
Peripheral vascular disease — no./total no. (%)	1787 (42.4)	474 (44.2)	0.28
Prior stroke — no./total no. (%)	496 (11.7)	113 (10.6)	0.28
Prior transient ischemic attack — no./total no. (%)	425 (10.1)	93 (8.7)	0.19
Cardiac risk factors— no./total no. (%)
Coronary artery disease	3189 (75.4)	813 (75.8)	0.76
Prior coronary-artery bypass surgery	1315 (31.1)	320 (29.9)	0.46
Prior percutaneous coronary intervention	1468 (34.7)	364 (34.0)	0.65
Balloon Valvuloplasty	417 (9.9)	98 (9.1)	0.53
Pre-Existing Pacemaker of Implantable Cardioverter-Defibrillator	818 (19.3)	234 (21.8)	0.07
Prior myocardial infarction	1025 (24.2)	252 (23.5)	0.63
Congestive heart failure	4126 (97.5)	1033 (96.4)	0.04
Prior atrial fibrillation or atrial flutter	1711 (40.5)	427 (39.9)	0.73

SD = standard deviation. No. = number.

Of 4,230 individuals in the linked dataset, 550 (13.0%) were adjudicated as having at least one of 630 CVEs in the trial database (1.15 events per person; annualized incidence 2.6%). Of these individuals, 479 (87.1%) had one recorded CVE, 61 (11.1%) had two events, and 10 (1.8%) had three events.

Frequency of Claims-based vs. Trial-based Events

Of 80 codes, 53 (66%) had at least one matched trial-adjudicated CVE (eTables 5 and 6). Of the 630 trial-adjudicated CVEs (220 major ischemic strokes, 201 minor ischemic strokes, 50 hemorrhagic strokes, 140 TIAs, and 19 undetermined CVEs), a corresponding CVE claim within 14 days was identified in 408. Of these, there were 179 (43.9%) claims representing major ischemic stroke, 123 (30.1%) representing minor ischemic stroke, 66 (16.2%) representing TIA, 35 (8.6%) representing major hemorrhagic stroke, and 5 (1.2%) undetermined. Among those with trial-adjudicated CVEs, the five most frequent codes were 434.91, 434.11, 435.9, 433.10, and 431.X for ICD-9-CM, and I63.9, I63.40, G45.9, I69.391, and I62.9 for ICD-10-CM.

Among 3,827 trial participants without a trial-adjudicated CVE, 558 (14.6%) had at least one claim for a possible CVE in Medicare data (eTables 7 and 8). These codes appeared 886 times in 729 hospitalizations in those without a matched CVE in the trial. Among those without trial-adjudicated CVEs, the five most common codes were 433.10, 433.30, 438.20, 438.89, and 437.0 for ICD-9-CM, and I65.23, I65.29, I65.22, I69.354, and I69.398 for ICD-10-CM.

Performance Characteristics of Optimal Coding Algorithms

Table 2 shows the performance of the optimal coding algorithm to identify trial-adjudicated CVEs. The combination of four codes (434.11, 434.91, 433.11, 997.02) had an 86.5% PPV for identifying ischemic stroke. While frequent, 433.10 (“occlusion and stenosis of the carotid artery without mention of cerebral infarction”) had poor sensitivity and specificity for trial CVEs (sensitivity 4.3%, specificity 43.6%, PPV 3.9%, NPV 46.5%). For hemorrhagic stroke, four codes (430.X, 431.X, 432.1, 432.9) had moderate sensitivity (61.2%) and high specificity (99.5%), with limited PPV (61.2%). While ICD-9-CM code 435.9 was specific (98.5%) for TIA, both sensitivity (31.5%) and PPV (40.2%) were lower.

Table 2:

Sensitivity, Specificity, Positive and Negative Predictive Values of the Optimal Codes Used to Represent CVEs in the US CoreValve Pivotal Trials

ICD-9-CM Classification
Ischemic Stroke – codes 434.11, 434.91, 433.11, 997.02
Present in Claims (N)	Present in Trial (N)			Sensitivity (%)	Specificity (%)	PPV (%)	NPV (%)	Kappa Statistic (95% CI)
	Yes	No	Total
Yes	249	39	288	60.9	99.0	86.5	95.8	0.69 (0.65–0.73)
No	160	3667	3827
Total	409	3706	4115
Hemorrhagic Stroke – codes 430.X, 431.X, 432.1, 432.9
Yes	30	19	49	61.2	99.5	61.2	99.5	0.61 (0.49–0.72)
No	19	3997	4016
Total	49	4016	4065
Transient Ischemic Attack – code 435.9
Yes	39	58	97	31.5	98.5	40.2	97.9	0.34 (0.25–0.42)
No	85	3902	3987
Total	124	3960	4084
ICD-10-CM Classification
Ischemic Stroke – codes I63.9, I63.40, I63.49
Yes	8	1	9	66.7	99.4	88.9	97.6	0.75 (0.55–0.96)
No	4	164	168
Total	12	165	177
Transient Ischemic Attack– codes G45.9, I69.391
Yes	3	1	4	18.8	99.4	75	92.6	0.27 (0.02–0.53)
No	13	162	175
Total	16	163	179

N = number of events. As there was only one hemorrhagic stroke in SURTAVI, identified through ICD-10-CM code I62.9, hemorrhagic stroke is not listed above under the ICD-10-CM classification.

Under ICD-10-CM, the combination of three codes (I63.9, I63.40, I63.49) had a sensitivity of 66.7%, specificity of 99.4%, PPV of 88.9%, and NPV of 97.6% for identifying ischemic stroke. This algorithm was moderately sensitive (66.7%) but very specific (99.4%) for ischemic stroke. The combination of two codes (G45.9, I69.391) identified TIA with 99.4% specificity with lower sensitivity (18.8%) and PPV (75.0%). Only one patient in the linked SURTAVI dataset had a hemorrhagic stroke (identified by I62.9).

Time-to-Event Outcomes

At 5 years (Figure 2), using all possible CVE claims, the cumulative event rate was 39.5% (claims) vs. 21.2% (trials; p <0.001). Excluding 433.X codes (except 433.11), the event rate was 31.2% (claims) vs. 21.2% (trials; p < 0.001). Further excluding individuals with historical CVEs, the event rate was 29.3% (claims) vs. 20.8% (trials; p = 0.02). After applying the optimized coding algorithm and excluding those with historical CVEs, estimates were similar (claims vs. trial event rates: 20.4% and 20.8%; p = 0.003). Comparisons of claims to trial CVEs by subtype are presented in eFigures 1–3.

Figure 2: Kaplan-Meier Comparison of Trial and Non-trial Surrogate CVEs.

Panels A-D represent four different comparisons of claims (blue dashed line) with adjudicated trial data (red solid line) to identify CVEs. Panel A includes all claims matching to at least one CVE in the CVPT. Panel B = Panel A minus carotid vascular disease claims (e.g. all 433.X codes except 433.11). Panel C = Panel B minus individuals with ≥1 CVE within a year prior to AVR. Panel D = optimized coding algorithm (ICD-9-CM: 434.91, 434.11, 433.11, 997.02, 430.X, 431.X, 432.1, 432.9, 435.9; ICD-10-CM: I63.9, I63.4, I63.49, G45.9, I69.391, I62.9), excluding those with CVEs within a year prior to AVR.

Among trial CVEs without a corresponding claims event (N = 236), 104 (44.1%) had a hospitalization within 14 days of the trial event (eTables 9 and 10). The five most common ICD-9-CM discharge codes in these hospitalizations were 272.4 (N = 58), 424.1 (N = 54), 428.0 (N = 54), 401.9 (N = 46), and V70.7 (N = 35). The five most common ICD-10-CM discharge codes were I35.0 (N = 5), Z0.06 (N = 4), I10.X (N = 3), I25.10 (N = 3), and E78.5 (N = 2).

Sensitivity Analysis

Recursive partitioning identified three ICD-9-CM codes (434.91, 434.11, 433.11) with a sensitivity and specificity for CEC-adjudicated ischemic stroke events of 60.1% and 99.2%, respectively, on cross-validation, with a model c-statistic of 0.80 (eFigure 4). As these codes were nearly identical to those selected through the hierarchical algorithm with similar performance characteristics after cross-validation, the coding algorithm is likely robust to optimism or significant biases in algorithm selection.

The relative performance of the parsimonious coding algorithms was similar in the SAVR and TAVR arms (eTables 11 and 12), although the sensitivity to detect ischemic stroke was modestly higher and sensitivity to detect hemorrhagic stroke modestly lower using ICD-9-CM in SAVR vs. TAVR. Likewise, the performance of the parsimonious coding algorithms was similar for CVEs occurring within or after 30 days from AVR, although the sensitivity to detect ischemic stroke was modestly higher within 30 days (eTables 13 and 14). In the landmark analyses (eFigures 5–7), ischemic stroke ascertainment after 30 days was not significantly different between claims and trial data (log rank p-value = 0.66).

DISCUSSION

We found that a parsimonious algorithm using three to four ICD-9-CM or ICD-10-CM billing codes identified over half of trial-adjudicated CVEs with high PPV and specificity, but moderate sensitivity. Excluding claims of individuals with prior CVEs within one year pre-procedure and excluding codes for carotid artery stenosis or occlusion from the diagnostic algorithms improved agreement. These findings demonstrate relatively high agreement between appropriately selected claims for CVEs and centrally adjudicated CVEs, and suggest that while claims may not be interchangeable with trial-adjudicated events, they could feasibly augment information on trial participants during follow-up.

Few studies have attempted to validate the use of administrative codes for stroke, and none have done so in the structural heart disease setting.^{5, 17–19} This is important as the accuracy of claims to identify stroke differs based on the setting of evaluation, stroke rates, and the rigor of ascertainment.^3,18–19

In this study, an algorithm using four ICD-9-CM claims had a cumulative sensitivity of 60.9%, specificity of 99.0%, PPV of 86.5%, and NPV of 95.8% for identification of ischemic stroke compared with trial adjudication. Similarly, an algorithm using three ICD-10-CM claims had a cumulative sensitivity of 66.7%, specificity of 99.4%, PPV of 88.9%, and NPV of 97.6%. While claims for ischemic stroke and TIA had >95% specificity under both ICD-9-CM and ICD-10-CM schemes, sensitivity was significantly lower, similar to prior reports,^18–19 reflecting the challenges of accurately coding CVEs. Four ICD-9-CM codes (430.X, 431.X, 432.1, 432.9) had moderate sensitivity (61.2%) and excellent specificity (99.5%) for identifying hemorrhagic stroke, though the PPV was considerably lower (61.2%) reflecting the lower percentage (7.9%) of hemorrhagic vs. non-hemorrhagic events.

As patients in the CVPT underwent rigorous pre- and post-procedural neurologic assessment, it is possible that documentation of events in the trial was better than in routine practice where such neurologic testing is not protocolized. In sensitivity analyses, the sensitivity of claims to detect CVEs was higher in the first 30 days than afterward, which may reflect differential ascertainment of events early in the trial and accrual of CVEs early post-AVR.

While trial adjudication was used as the criterion standard for validation in this study, it is possible that neither clinical trials nor billing claims represent a true “gold standard”, with both providing complementary and distinct information. For example, the higher rate of events in claims may be due to underreporting of events in the clinical trial. In the quest to find the ideal pragmatic clinical trial design that includes relevant outcomes while minimizing costs, these data indicate that claims may be a useful adjunct to augment existing trial event ascertainment. Given the high specificity of the claims-based algorithm, a claim for a CVE could prospectively trigger evaluation and adjudication of an event that has not been reported in a trial. This may reduce workload and staffing needed to contact individual sites and patients about potential events, and may bring attention to true events not reported to sites or the trial CEC. Such triggers might be particularly valuable when an event results in admission to a hospital not participating in the trial. In addition, misclassification of events by claims would not be expected to create bias in a randomized trial. Whether or not events identified by claims or trial-adjudication have differing ability to identify long-term adverse outcomes such as mortality remains a question for future research.

This study has several limitations. First, 20.2% of trial participants could not be linked to claims. However, linked and non-linked individuals were similar across a range of clinical characteristics. Furthermore, the study did not use direct patient identifiers and could not examine outcomes of non-linked patients. Second, these results may not apply to individuals < 65 years old or with other insurance. Nevertheless, as age is a risk factor for stroke²⁰ and the mean age in the linked CVPT was 83, the majority of individuals with CVEs in the study were likely Medicare-enrolled. Third, the strict trial inclusion criteria may limit generalizability. Fourth, a limited number of CVEs occurred under the ICD-10-CM framework. Fifth, as claims were removed from the CVE-algorithm if they were nonspecific for trial-adjudicated CVEs, it is possible that multiple claims could have been chosen for removal. However, given the concordance between the claims chosen by recursive partitioning and the identified claims-based algorithm, it is likely that the selected codes are best suited for identifying stroke parsimoniously. Finally, the hierarchical iterative approach may be prone to overfitting. However, after adjusting for model optimism using 10-fold cross-validation in the recursive partitioning model, the model c-statistic remained high, suggesting that the algorithm could perform well in a different setting. External, prospective validation of this algorithm is warranted to test this hypothesis.

CONCLUSIONS

An algorithm using three to four ICD-9-CM or ICD-10-CM billing codes identified trial-adjudicated stroke events in the CVPT with limited sensitivity but excellent specificity and predictive values. These findings suggest that claims are imperfect substitutes for trial adjudicated events, but that they could potentially be incorporated into routine trial follow-up of CVEs to trigger event evaluations, improving the efficiency and cost of evaluating cardiovascular devices or therapeutics in the structural heart disease setting.

WHAT IS KNOWN:

• Cerebrovascular events (CVEs) are common after aortic valve replacement.

• Studies evaluating devices to reduce or prevent CVEs may be unable to meaningfully assess differences in CVE rates in part due to the high cost of conducting adequately powered trials.

• Whether administrative claims data could be used to substitute for more expensive trial adjudication of CVEs is unknown.

WHAT THIS STUDY ADDS:

• In a dataset linking the US CoreValve Pivotal Trials with Medicare claims, four billing codes identified half of trial-adjudicated CVEs during follow-up with high specificity but imperfect sensitivity.

• Linkage of trials with claims could be useful to support endpoint ascertainment, particularly for triggering a more thorough evaluation of a suspected adverse event, potentially improving the efficiency of future trials for which CVEs are an outcome.

NON-STANDARD ABBREVIATIONS AND ACRONYMS:

EXTEND

Extending Trial-Based Evaluations of Medical Therapies Using Novel Sources of Data Study

CVE

Cerebrovascular Event

ICD-9-CM

International Classification of Diseases, 9^th Revision, Clinical Modification

ICD-10-CM

International Classification of Diseases, 10^th Revision, Clinical Modification

AVR

aortic valve replacement

TAVR

transcatheter aortic valve replacement

SAVR

surgical aortic valve replacement

CEC

clinical events committee

TIA

Transient Ischemic Attack

SURTAVI

Surgical or Transcatheter Aortic-Valve Replacement in Intermediate Risk Patients Trial

CVPT

US CoreValve Pivotal Trials Program

SD

standard deviation

PPV

Positive predictive value

NPV

Negative predictive value