Pulse Wave Velocity and Prognosis in End-Stage Kidney Disease

The protocol was in conformity to the ethical guidelines of our institutions, and informed consent was obtained from each participant. Data are available on request from the Authors.

The etiologic and the prognostic value of PWV for all-cause and cardiovascular mortality was investigated in 2 dialysis populations. The Manhes cohort (test set) included a series of 287 hemodialysis patients (participation rate: 75%) with no exclusion criterion⁸ except diseases interfering with the pulse waves recording like absence of femoral pulse or arrhythmia. The Quebec cohort (validation set) was an incident-prevalent cohort.¹⁸ The exclusion criteria of this cohort were similar to those of the Manhes cohort, that is, any clinical condition hampering hemodynamic measurements (arrhythmia, like of arterial pulses, marked hypotension) or acute episode of illness (infection, acute heart failure, active bleeding).¹⁸ For the scope of the present analysis, we included all hemodialysis patients (n=246) of the Quebec cohort.¹⁸ We excluded the 65 patients on peritoneal dialysis in this cohort because the test cohort (Manhes cohort) did not include peritoneal dialysis patients. Hemodialysis patients of both cohorts were treated thrice weekly with standard bicarbonate dialysis and various noncellulosic dialysis filters. The enrollment period was from January 1996 to January 1998 for the Manhes cohort and from August 2006 to July 2012 for the Quebec cohort. Demographic and clinical details of patients included in the 2 cohorts are given in Table 1 and in previous publications.^8,18

Because the 2 ARO calculators were validated to predict the risk of all-cause and cardiovascular mortality over a 2-year period, for the purpose of this analysis, the follow-up was restricted to 2 years. In both cohorts, the causes of death and fatal cardiovascular events (myocardial infarction, ECG documented arrhythmia, heart failure, and stroke and other thrombotic events) were defined according to prespecified criteria and events were accurately recorded in the Manhes Hospital and Hotel de Dieu de Québec research databases.

In the Manhes Hospital, cohort aortic PWV was determined from transcutaneous Doppler flow recordings and the foot-to-foot method. As described in detail in a previous study,⁸ 2 simultaneous Doppler flow tracings were recorded, the first at the suprasternal notch level (aortic arch level) and the second at the groin (femoral artery level). These tracing were obtained using a handheld probe connected with a nondirectional Doppler unit (SEGA M842, 10 MHz) and recorded on a Gould 8188 recorder (Gould Electronique) at a speed of 100 to 200 mm/s. The time delay (t) was measured between the bases of the flow waves recorded at these different points and was averaged over 10 beats. The distance (D) traveled by the pulse wave was measured over the body surface as the distance between the 2 recording sites, and PWV was calculated as PWV5D/t. The mean±SD intraobserver repeatability of aortic PWV measurements was 5.8±1%.

As reported elsewhere,¹⁸ PWV in the Quebec cohort was measured by a well-validated instrument, the Complior SP (Artech Medical, Pantin, France).²¹ Pressure waveforms are digitized at 500 Hz (carotid-femoral) and 800 Hz (carotid-radial) PWV and stored in a recirculating memory buffer and shown in the same computer display in 2 separate tracings. The upper tracing shows the pressure waveform of the proximal recording site (carotid) and the lower that of the distal site (femoral). The delay between the 2 pulse waves is determined by performing a correlation between the data of the 2 waveforms by a correlation algorithm adopting appropriate time-shifts. The calculated pulse delay is then printed.²¹ The regression coefficient value for the repeatability of automatic PWV recordings by the Complior SP instrument is 0.935 and does not differ from that of manual recording (0.938).²¹ The intrasession and intersession coefficient of variation of PWV at Quebec Research center are, respectively, 2.9% and 8.9%.¹⁸

The procedure for the calculation of the ARO scores for both 2-year all-cause¹⁸ and cardiovascular mortality¹⁹ is summarized in Tables S1 and S2 in the online-only Data Supplement.

In both test (Manhes) and validation (Quebec) cohorts, data are summarized as mean±SD, median and interquartile range or as percent frequency, as appropriate. The association between PWV and risk scores (for both all-cause and cardiovascular mortality) was investigated by correlation analyses, and data were expressed as correlation coefficients and P values. The relationship between PWV and the incidence rate of all-cause and cardiovascular mortality in the 2 study cohorts was investigated by univariate and multiple (risk score adjusted) Cox regression analyses. Data were expressed as hazard ratio, 95% confidence interval (CI), and P value.

The prognostic value of PWV beyond and above that provided by the risk score was investigated by calibration (Hosmer–May test)¹⁴ and discrimination (Harrell C index)¹⁵ analysis, by explained variation (R², an index combining discrimination and calibration),¹⁶ and risk reclassification analyses.¹⁷ The analysis of reclassification was performed by calculating the Integrated Discrimination Improvement (IDI),²² this latter allowing to measure the risk reclassification provided by a risk factor by avoiding data categorization. Briefly, calibration measures how much the prognostic estimate of a predictive model including one or more biomarkers/indicators matches the “real” probability of the outcome (ie, the observed proportion of an event in a given time period). In calibration analysis, predicted and observed probabilities of the event are compared by the May–Hosmer test. A not significant May–Hosmer test indicates that predicted and observed probabilities of the event do not differ between them and then the calibration of the model is satisfactory. Discrimination measures how well a prognostic model distinguishes (discriminates) patients with and without the outcome of interest (all-cause or cardiovascular death in our case). Discrimination, as measured by the Harrell C-index, may take values ranging from 0.5 (no discrimination) to 1.0 (perfect discrimination). The higher the Harrell C-index, the higher the accuracy of the model for predicting the event of interest. The analysis of risk reclassification by the IDI allows to measure the risk reclassification by a risk factor without data categorization. IDI quantifies whether a new biomarker (PWV in our instance) provides a clinically relevant improvement in prediction beyond and above that provided by model including the ARO risk score (the score for either all-cause mortality or cardiovascular mortality). The basic idea behind IDI is that a valuable new biomarker will increase the predicted risk for patients with the event of interest and will decrease predicted risk for patients without such events.

All calculations were done by a standard statistical package (STATA/IC 13.1 for Windows, December 9, 2015, College Station, TX).

During a 2-year follow-up period, 77 patients died in the Mahnes cohort (16.5 deaths/100 person-year; 95% CI, 13.0–20.7 deaths/100 person-year) and a similar figure (15.7 deaths/100 person-years; 95% CI, 12.0–20.1 deaths/100 person-year; n=62 deaths) was observed in the Quebec cohort (P=0.76). Likewise, the incidence rate of cardiovascular death in the Manhes cohort (11.8 cases/100 person-year; 95% CI, 8.9–15.4 cases/100 person-year; n=55 cardiovascular death) did not differ (P=0.27) from that observed in the Quebec cohort (9.4 cases/100 person-year; 95% CI, 6.6–12.9 cases/100 person-year; n=37 cardiovascular death).

On univariate Cox regression analyses (Table 2), the ARO risk scores and PWV significantly predicted the incidence rate of all-cause and cardiovascular mortality in both cohorts. After adjustment for PWV (Table 2), the ARO risk scores predicted both outcomes in the Manhes cohort and the same was true for PWV adjusted for the ARO risk scores. In the Quebec cohort, the ARO risk score for all-cause mortality (adjusted for PWV) but not for the risk score for cardiovascular mortality significantly predicted the corresponding outcome. In Quebec cohort, PWV adjusted for the ARO risk scores largely failed to predict both study outcomes (Table 2).

The discriminatory power (Harrell C-index) of the ARO risk scores was consistently higher than that provided by the PWV in the 2 cohorts for both all-cause and cardiovascular mortality (Figure 2). The inclusion of PWV into Cox analyses including the ARO risk scores provided a modest increase in the discriminatory power of the same models in both the Manhes (all-cause death: Δ Harrell C-index, +1.8%; P=0.02 and cardiovascular mortality: Δ Harrell C-index, +3.3%; P<0.001) and Quebec cohort (all-cause death: Δ Harrell C-index, +0.8%; P=0.32 and cardiovascular mortality, Δ Harrell C-index, +0.5%; P=0.16). The additional prognostic value of PWV as assessed by the explained variation in mortality (R²) confirmed a very modest increase in prognostic accuracy by this biomarker beyond and above the ARO risk scores (Manhes cohort: all-cause death, Δ: +2.8%; cardiovascular mortality, Δ: +7.4% and Quebec cohort: all-cause death, Δ: +1%; cardiovascular mortality, Δ:+3.1%; Figure 2), and this was also true when the data were analyzed by the IDI index (Manhes cohort: all-cause death, IDI=+2.7%; P=0.02; cardiovascular death, IDI=+5.1%; P=0.02 and Quebec cohort: all-cause death, IDI=+0.9%; P=0.16; cardiovascular death, IDI=+1.2%; P=0.18). Of note, the introduction of PWV into Cox analyses containing the risk score determined a worsening in models calibration because the models including this biomarker produced prognostic estimates that differed from the observed probabilities of study outcomes (Manhes cohort: all-cause death, Hosmer–May test: χ²=11.20; P=0.004; cardiovascular death, Hosmer-May test: χ²=6.00; P=0.049 and Quebec cohort: all-cause death, Hosmer-May test: χ²=6.61; P=0.09; cardiovascular death, Hosmer-May test: χ²=7.75; P=0.051).

The etiologic role of arterial stiffness in cardiovascular disease is supported by a variety of studies including studies in animal models,²⁵ a large series of observational studies in various conditions⁹ and by a clinical trial showing that, at identical brachial pressure, a drug regimen that produces a greater fall in central blood pressure associates with a lower incident risk for cardiovascular events as compared with another, standard regimen.²⁶ ESKD is considered as a condition paradigmatic of very severe arterial stiffness and inflammation and the ensuing propensity to calcification in large part explain the almost unique severity of arterial damage in this disease.²⁷ In longitudinal studies in patients with ESKD, a decline in PWV in response to antihypertensive treatment predicts better cardiovascular outcomes²⁸ and even though there is still no specific trial targeting this alteration in dialysis patients, it is felt that arterial stiffness most likely represents a modifiable risk factor in patients with CKD.⁷ Thus, the etiologic role of arterial stiffness seems to be beyond question in the dialysis population.

As previously mentioned, studies in various populations seem to support a prognostic role of PWV. In the meta-analysis by Ben-Shlomo et al⁹ PWV improved model fit in analyses adjusting for classical risk factors. It has to be noted that, in this meta-analysis, gains in discrimination power (area under the receiver operating characteristic curve) and in IDI—that is, the centerpieces of modern prognostic testing—by PWV were very modest and statistically trivial. However, in individuals at intermediate risk, PWV improved the 5- and 10-year risk reclassification for death and various cardiovascular outcomes by the 6.1% and 27.2%,⁹ which are improvements of mild to moderate relevance. The interpretation of risk reclassification indices and their relevance for clinical practice is an intriguing but still much debated area in biostatistical research because methods pertaining to these indices are not yet well developed.²⁹ Uncertainties and problematic aspects of net reclassification indices have been recently emphasized by Kerr at al.³⁰ Scientific debate apart, a fundamental tenet of prognostic research is that risk models need to be developed and validated in the specific diseases, for example, hypertension or CKD, and in the populations where they are planned to be applied in clinical practice.³¹ In this regard, it should be emphasized that, to date, no study in ESKD specifically investigated, by the state-of-the-art statistical methods, the additional prognostic value of this biomarker over and above simple prediction instruments based on easily available clinical information. The issue is of importance because prognostic biomarkers need to be accurate, easy, and cheap to measure, should add prognostic value to easily available prognostic factors, and be validated in independent cohorts of patients.³²

Beyond the classical Cox’s regression analysis, prognostic tests evaluate various characteristics that are considered central for the validation of prognostic biomarkers. Discrimination quantitatively assesses how much a prognostic model discriminates patients with from those without a given disease outcome,¹⁵ whereas calibration is the ability of a prognostic model to correctly match the probability of the event of interest across the whole range of prognostic estimates.¹⁴ Reclassification reflects how much a candidate prognostic biomarker increases the proportion of individuals correctly reclassified as having or not a given outcome as compared with a previous classification based on an existing prognostic model.¹⁷ Finally, explained variation is a test combining discrimination and calibration, which compares the predictive accuracy of prognostic models with and without a given key predictor (PWV in our case).¹⁶ In our study, the discriminatory power of PWV in the 2 cohorts was consistently lower than that provided by the ARO risk scores for both all-cause and cardiovascular mortality and the addition of PWV into Cox analyses including the same risk scores provided just a modest increase in the discriminatory power of the same models in both the test (Manhes) and the validation (Quebec) cohorts (Figure 2). The modest gain in prognostic accuracy by PWV was confirmed also when the performance of this biomarker was investigated by the explained variation in mortality (gain ranging from 1% to 7.4%) as well as by IDI (gain ranging from 0.9% to 5.1%). Of note, the gain in prognostic power by PWV for all-cause and cardiovascular mortality was lower than that achieved by a commonly measured, easily available, and cheap biomarker of inflammation such as IL-6 (interleukin-6) for the same outcomes (all-cause death, +9.9%; cardiovascular mortality, +7.7%).³³ Furthermore, the introduction of PWV into Cox models based on the ARO risk scores determined a worsening in calibration because the models including this biomarker produced prognostic estimates that differed from the observed probabilities of study outcomes. Overall, these analyses coherently show that PWV had a limited prognostic value over and above the ARO risk scores in the dialysis population. Although the relevance of the measurement of PWV in pathophysiology studies and in clinical research remains beyond question,^7,9 our data show that this measurement only for risk stratification is unwarranted in patients with ESKD.

Although PWV remains a valuable instrument to investigate the pathophysiology of arterial disease, to monitor treatment efficacy and a potential treatment target, this technique is inferior to well-validated clinical risk scores for the prediction of mortality and fatal cardiovascular events in this condition. When combined to the clinical risk scores, PWV produces a fairly modest gain in prediction power. Thus, the measurement of PWV solely for risk stratification seems to be unwarranted in patients with ESKD.

The results of this study imply that PWV remains a treatment target in patients with ESRD and a tool to monitor the efficacy of specific therapeutic strategies, but it is not useful as prognostic biomarker in the dialysis population.