A clinical and proteomics approach to predict the presence of obstructive peripheral arterial disease: From the Catheter Sampled Blood Archive in Cardiovascular Diseases (CASABLANCA) Study

Abstract

Background

Peripheral arterial disease (PAD) is a global health problem that is frequently underdiagnosed and undertreated. Noninvasive tools to predict the presence and severity of PAD have limitations including inaccuracy, cost, or need for intravenous contrast and ionizing radiation.

Hypothesis

A clinical/biomarker score may offer an attractive alternative diagnostic method for PAD.

Methods

In a prospective cohort of 354 patients referred for diagnostic peripheral and/or coronary angiography, predictors of ≥50% stenosis in ≥1 peripheral vessel (carotid/subclavian, renal, or lower extremity arteries) were identified from >50 clinical variables and 109 biomarkers. Machine learning identified variables predictive of obstructive PAD; a score derived from the final model was developed.

Results

The score consisted of 1 clinical variable (history of hypertension) and 6 biomarkers (midkine, kidney injury molecule‐1, interleukin‐23, follicle‐stimulating hormone, angiopoietin‐1, and eotaxin‐1). The model had an in‐sample area under the receiver operating characteristic curve of 0.85 for obstructive PAD and a cross‐validated area under the curve of 0.84; higher scores were associated with greater severity of angiographic stenosis. At optimal cutoff, the score had 65% sensitivity, 88% specificity, 76% positive predictive value (PPV), and 81% negative predictive value (NPV) for obstructive PAD and performed consistently across vascular territories. Partitioning the score into 5 levels resulted in a PPV of 86% and NPV of 98% in the highest and lowest levels, respectively. Elevated score was associated with shorter time to revascularization during 4.3 years of follow‐up.

Conclusions

A clinical/biomarker score demonstrates high accuracy for predicting the presence of PAD.

1. INTRODUCTION

Peripheral arterial disease (PAD) is a global health problem, with 202 million people living with the diagnosis.1 As symptoms of PAD are protean, its diagnosis is challenging until its advanced stages. Accordingly, PAD is often underdiagnosed and undertreated, with most patients not receiving optimal management that might delay progression of disease2 and prevent future ischemic events.3 Tools exist for detection of PAD. For example, the ankle‐brachial index (ABI) is the most common noninvasive tool utilized to diagnose lower‐extremity PAD; however, its diagnostic accuracy is limited in patients with stiff, calcified ankle arteries, and it is also subject to significant variability.4, 5 In addition, in some cases (such as in proximal disease location), sensitivity of ABI may be reduced without exercise. Imaging modalities are often utilized to delineate the anatomy of carotid, renal, and lower‐extremity arteries; but imaging is limited by cost, availability, and need to utilize intravenous contrast and/or ionizing radiation. Most important, an anatomic diagnosis of PAD does not necessarily predict functional severity and/or prognosis. For all these reasons, a need exists for alternative means for evaluating PAD.

One option might be the use of biomarkers to identify presence and prognosticate course of the diagnosis. We previously demonstrated the utility of combining clinical variables with machine learning supported proteomics for diagnosis6 and prognosis of coronary artery disease (CAD).7 Accordingly, we sought to identify clinical and biomarker predictors of clinically significant PAD in an at‐risk population of subjects enrolled in the Catheter Sampled Blood Archive in Cardiovascular Diseases (CASABLANCA) study undergoing peripheral angiography.8 We hypothesized that the combination of biomarkers and clinical risk factors might increase the accuracy of predicting clinically significant PAD.

2. METHODS

The CASABLANCA study was a prospective, single‐center, investigator‐initiated, observational cohort study performed at Massachusetts General Hospital in Boston; 1251 subjects undergoing coronary and peripheral angiography ±intervention between 2008 and 2011 were enrolled.8

For the purpose of this study, we included 354 patients who underwent peripheral angiography only (n = 140), peripheral and coronary angiography but without angiographically significant CAD (n = 11), and coronary angiography alone without angiographically significant CAD and no history of PAD (n = 203). Given their lack of peripheral angiography results, the latter group was included to increase cohort size and assumed to be negative for peripheral obstruction, given their medical history. The indications for peripheral angiography included claudication (n = 96), hypertension (HTN; n = 21), carotid artery stenosis with/without stroke (n = 11), and other PAD without claudication (n = 25).8 The locations of the peripheral angiographies included the lower extremity (n = 129), renal arteries (n = 59), and carotid/subclavian vessels (n = 18). All study procedures were approved by the Partners Healthcare Institutional Review Board and carried out in accordance with the Declaration of Helsinki.

After obtaining informed consent, detailed clinical and historical variables were recorded at the time of the procedure. For the purposes of this analysis, we characterized obstructive PAD as ≥50% luminal obstruction in ≥1 peripheral vessel. We obtained 15 mL of blood immediately before angiography through a centrally placed vascular‐access sheath. The samples were stored in a 4 °C refrigerator until centrifuging took place. After a single freeze–thaw cycle, 200 μL of plasma was analyzed for a panel of 109 biomarkers on the 100/200 xMAP technology platform (Luminex, Austin, TX). The 109 biomarkers were acquired in the form of a commercially available Multi‐Analyte Profiling (MAP) kit (Myriad RBM, Austin, TX). This incorporates biomarkers that reflect a wide variety of pathways including acute phase reactants, inflammation, and atherosclerosis. This technology utilizes multiplexed, microsphere‐based assays in a single reaction vessel. The assay‐specific capture antibody on each microsphere binds to the protein of interest. Similar to a flow cytometer, as each individual microsphere passes through a series of excitation beams, it is analyzed for size and encoded fluorescence signature, and the amount of fluorescence generated is proportionate to the protein concentration.

2.1. Statistical analysis

Baseline characteristics between patients with and without PAD were compared; dichotomous variables were compared using a 2‐sided Fisher exact test, and continuous variables were compared using a 2‐sided, 2‐sample t test. The biomarkers compared were tested with the Wilcoxon rank‐sum test, as their concentrations were not normally distributed. A complete case analysis was performed; 2 patients were missing ≥1 of the concentration readouts for the 6 proteins in the final panel, so they were excluded, leaving 352 samples available for analysis. For any biomarker result that was unmeasurable, we utilized a standard approach of imputing concentrations 50% below the limit of detection.

We utilized machine learning, a subset of artificial intelligence, to identify predictors of significant PAD. To facilitate the predictive analysis, the concentration values for all proteins underwent the following transformations: (1) they were log‐transformed to achieve a normal distribution; (2) outliers were clipped at the value of 3× the median absolute deviation; and (3) the values were rescaled to a distribution with zero mean and unit variance. The starting sets of variables consisted of all 109 proteins as well as clinical factors in the CASABLANCA dataset that were chosen for their possible clinical relevance. Candidate panels of proteins and clinical features were generated via least‐angle regression.9 In this method, factors were included in the model one at a time, with their coefficients determined by their correlation with the outcome. This was repeated until all factors were included in the model, and the step at which the performance plateaued resulted in our initial panel of interest. With this panel of interest, predictive analyses were run on the training set using least absolute shrinkage and selection operator (LASSO)10 with logistic regression, predicting the outcome of obstructive PAD using only the variables in the panel of interest. Mathematically, the final diagnostic model is a linear formula, in the following form: score = a + b1 × 1 + b2 × 2 + b3 × 3 + b4 × 4 + b5 × 5 + b6 × 6 + b7 × 7, where ×1 represents the presence or absence of the clinical variable (1 = yes, 0 = no), ×2 through ×7 represent the normalized and scaled protein concentrations, and b1 through b7 represent the coefficients of the model, which were determined by the model‐building process. This model‐development process was done via Monte Carlo cross‐validation, using 400 iterations with an 80:20 (training:test) split.

The final panel was used to create a final model with the entire sample, and this model was then evaluated to predict obstructive PAD. We used the predictions to generate a receiver operating characteristic (ROC) curve and to calculate sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV). The predictive score generated by the diagnostic model was rescaled to a range of 0 to 10 to facilitate interpretation, and the score was then partitioned into 5 different risk levels, corresponding to multiple levels of PAD risk. The PPVs and NPVs were calculated at each risk level. Finally, time to revascularization as a function of elevated PAD score was calculated from 7 days after index angiography over a 4.3‐year follow‐up period and displayed as a Kaplan–Meier survival curve.

All statistics were performed by using R software, version 3.4 (R Foundation for Statistical Computing, Vienna, Austria). P values are 2‐sided, with a value <0.05 considered significant.

3. RESULTS

Of the 354 patients included in this study, 132 had obstructive PAD. The vascular territories involved are detailed in the Supporting Information, Table 1, in the online version of this article.

Baseline characteristics of study subjects, dichotomized as a function of presence or absence of significant PAD, are detailed in Table 1. Patients with obstructive PAD were older, more likely to be male, and had a higher prevalence of HTN, CAD, diabetes, and prior coronary revascularization (Table 1). Notably, of the biomarkers measured, those with severe PAD had lower concentrations of follicle‐stimulating hormone (FSH) and angiopoietin‐1 and higher concentrations of kidney injury molecule‐1 (KIM‐1), midkine, interleukin‐23 (IL‐23), and eotaxin‐1 (see Supporting Information, Table 2, in the online version of this article).

Table 1.

Baseline characteristics in patients with or without obstructive PAD

Characteristics	Subjects With Obstructive PAD, n = 132	Subjects Without Obstructive PAD, n = 222	P Value
Demographics
Mean age, y	68 ± 11.3	63 ± 11.6	<0.001
Male sex	72.0	55.4	0.002
Caucasian race	93.9	91.4	0.54
Medical history
Smoker	17.7	14.1	0.36
AF/flutter	15.9	26.1	0.03
HTN	93.2	64.9	<0.001
CAD	63.6	20.3	<0.001
Prior MI	23.5	8.1	<0.001
HF	18.9	24.8	0.24
COPD	21.2	20.3	0.89
DM (type 1 or 2)	43.2	16.7	<0.001
CVA/TIA	18.9	8.1	0.004
CKD	22.0	3.6	<0.001
RRT	2.3	0.9	0.37
Prior angioplasty	33.3	3.6	<0.001
Prior CABG	35.6	2.3	<0.001
Prior PCI	54.5	10.8	<0.001
Medications
ACEI/ARB	62.9	47.7	0.008
β‐Blocker	67.4	57.9	0.09
Aldosterone antagonist	6.1	2.3	0.08
Loop diuretic	19.7	21.3	0.79
Nitrate	18.2	9.1	0.02
CCB	32.8	21.7	0.02
Statin	85.6	55.7	<0.001
ASA	87.9	60.6	<0.001
Warfarin	13.7	22.6	0.05
Clopidogrel	35.6	7.7	<0.001
Echocardiographic results
LVEF, %	60.4 ± 12.9	56.5 ± 16.1	0.15
RVSP, mm Hg	44.2 ± 13.6	40.6 ± 12.5	0.32
Laboratory testing
Sodium, mEq/L	139.7 ± 3.1	139.8 ± 3.2	0.74
BUN, mg/dL	21 (16–29)	17 (14–21)	<0.001
Cr, mg/dL	1.2 (1.0–1.5)	1.0 (0.9–1.2)	<0.001
eGFR, mL/min/1.73 m2	89.6 (64–104.5)	102.5 (84.4–111.9)	<0.001
TC, mg/dL	142.6 ± 35.7	168.5 ± 45.8	<0.001
LDL‐C, mg/dL	74.1 ± 27.4	95.6 ± 32.7	<0.001
HbA1c, %	6.8 (5.9–9.2)	5.8 (5.5–6.5)	0.02
Glucose, mg/dL	106 (93.8–124.8)	101 (89.0–116.0)	0.02
Hb, g/dL	12.8 ± 1.7	13.4 ± 1.7	0.003
Biomarkers
ANG‐1, ng/mL	5.9 (4.5–7.6)	7.4 (5.2–11.0)	<0.001
Eotaxin‐1, pg/mL	111 (42.5–163.5)	91 (42.5–135.0)	0.001
FSH, mIU/mL	7.95 (3.9–27.3)	10 (4.5–41.0)	0.05
IL‐23, ng/mL	2.7 (2.3–3.5)	2.5 (1.9–3.1)	<0.001
KIM‐1, ng/mL	0.05 (0.03–0.08)	0.01 (0.01–0.05)	<0.001
Midkine, ng/mL, median	20 (13–28)	12 (9.6–18.0)	<0.001

Abbreviations: ACEI, angiotensin‐converting enzyme inhibitor; AF, atrial fibrillation; ANG‐1, angiopoietin‐1; ARB, angiotensin II receptor blocker; ASA, acetylsalicylic acid (aspirin); BUN, blood urea nitrogen; CABG, coronary artery bypass grafting; CAD, coronary artery disease; CCB, calcium channel blocker; CKD, chronic kidney disease; COPD, chronic obstructive pulmonary disease; Cr, creatinine; CVA, cerebrovascular accident; DM, diabetes mellitus; eGFR, estimated glomerular filtration rate; FSH, follicle‐stimulating hormone; Hb, hemoglobin; HbA1c, glycated hemoglobin; HF, heart failure; HTN, hypertension; IL‐23, interleukin 23; IQR, interquartile range; KIM‐1, kidney injury molecule‐1; LDL‐C, low‐density lipoprotein cholesterol; LVEF, left ventricular ejection fraction; MI, myocardial infarction; PCI, percutaneous coronary intervention; RRT, renal replacement therapy; RVSP, right ventricular systolic pressure; TC, total cholesterol; TIA, transient ischemic attack

Data are presented as %, mean ±SD, or median (IQR).

After the machine‐learning model‐building process, the final panel consisted of systemic HTN as the sole clinical variable and 6 biomarkers: FSH, angiopoietin‐1, KIM‐1, midkine, IL‐23, and eotaxin‐1. With the data represented in the 0 to 10 scale, the optimal cutoff for the score was determined to be 5.546 using the optimal Youden index, which identifies the optimal balance of sensitivity and specificity. In ROC testing, the score generated an in‐sample area under the curve (AUC) of 0.85 for obstructive PAD and a cross‐validated AUC of 0.84 (Figure 1).

Figure 1.

ROC curve for the PAD score to predict obstructive PAD. The score had a very robust AUC. Abbreviations: AUC, area under the curve; PAD, peripheral arterial disease; ROC, receiver operating characteristic

At its optimal cutoff to diagnose PAD, for the group as a whole, the score had an in‐sample sensitivity of 65%, specificity of 88%, NPV of 81%, and PPV of 76%. We found a higher prevalence of obstructive PAD in those with higher scores and lower prevalence among those with lower scores (see Supporting Information in the online version of this article). Partitioning the score into 5 categories yielded a PPV of 86% and NPV of 98% in the highest and lowest scores, respectively (Table 2). When the score was divided into 5 categories of predicted risk, an increasing score correlated with an increasing degree of mean PAD stenosis (Figure 2). The score had a similar in‐sample performance in each individual vascular territory using the optimal cutoff. For example, in the diagnosis of obstructive (≥50%) renal artery disease, the score had a sensitivity of 63%, specificity of 73%, PPV of 91%, and NPV of 31%. The score had a sensitivity of 67%, specificity of 76%, PPV of 95%, and NPV of 26% for obstructive lower‐extremity arterial disease. For the renal arteries, the corresponding PPV and NPV were 100% and 100%; and for lower‐extremity obstruction, the PPV and NPV were also 100% and 100%.

Table 2.

Diagnostic characteristics of the panel for obstructive PAD when partitioned into 5 categories

Score	No. of Patients	PPV	NPV
5	35	0.86	—
4	118	0.58	—
3	72	0.33	0.67
2	84	—	0.93
1	44	—	0.98

Abbreviations: NPV, negative predictive value; PAD, peripheral arterial disease; PPV, positive predictive value

Figure 2.

Correlation between PAD score and mean degree of arterial stenosis. Abbreviations: PAD, peripheral arterial disease

In adjusted Cox proportional hazards models, from 7 days after index angiography to the end of follow‐up, those with a dichotomously elevated score had higher risk for revascularization compared with patients with a lower PAD score (hazard ratio: 4.18; 95% confidence interval: 2.43–7.2, P < 0.001); those with higher scores also had shorter time to first revascularization event (Figure 3).

Figure 3.

Kaplan–Meier survival curves depicting time to revascularization as a function of PAD score. Patients in the positive group had a score > or = the optimal cutoff for the score, which was determined to be 5.546 using the optimal Youden index (with the model's output rescaled to the range of 0–10). Patients in the negative group had a score < 5.546. Abbreviations: PAD, peripheral arterial disease

4. DISCUSSION

Using an approach leveraging clinical information plus proteomic screening, we describe a novel method to predict angiographically significant PAD. Based on the statistical model developed, we derived a score for clinical use that combines 1 clinical variable (HTN) with 6 biomarkers. The performance of the diagnostic model was consistent across vascular territories, including renal and lower‐extremity arteries.

The biomarkers in this model all have plausible biologic links to atherosclerosis and/or vascular calcification; several may also be associated with angiogenesis, a process strongly associated with PAD, particularly of the lower extremities. Midkine is a basic heparin‐binding growth factor associated with vascular and inflammatory cell migration, proliferation, atherogenesis, angiogenesis, and PAD.11, 12, 13 Using similar unbiased proteomics methods, we previously demonstrated association between midkine concentrations and presence/severity of CAD.6 Eotaxin‐1, also known as CCL11, is a chemotactic hormone responsible for, among other things, attraction of eosinophils to mediate allergic and inflammatory states. Eosinophils may play an important role in atherogenesis,14 and previous data have found higher concentrations in patients with CAD.15 Additionally, much like midkine, eotaxin‐1 may be angiogenic. IL‐23, another inflammatory cytokine, has previously been associated with PAD and disease progression in patients with carotid atherosclerosis16, 17; concentrations of IL‐23 also appear to be associated with risk for plaque rupture, and may be modulated by statin therapy.18 KIM‐1 is a marker of ischemic or nephrotoxic insults to proximal tubular cells, but it also has been associated with CAD presence and severity6; KIM‐1 may additionally promote neovascularization in ischemic tissues. We found lower concentrations of FSH in patients with PAD. Indeed, this is in keeping with a prior study describing FSH being inversely related to carotid intima‐media thickness19 and increased risk of atherosclerotic cardiovascular disease.20 Finally, angiopoietin‐1 promotes endothelial‐cell survival, stabilizes endothelial interactions with supporting cells, and limits the permeability‐inducing effects of vascular endothelial growth factor.21 Further, depletion of angiopoietin‐1 results in defects in endocardial and myocardial tissue, which supports its role in vascular health and pathology.22 Thus, it may not be surprising that lower concentrations of angiopoietin‐1 were associated with PAD.

Combined, these biomarkers provide a distinctive pathophysiological mix of processes associated with atherosclerosis and angiogenesis. From a clinical perspective, this model could be calculated by a hardware platform in the laboratory, or in a distributed system hosted in a cloud. The concentration of the biomarkers and the presence or absence of systemic HTN could be electronically submitted to the model's supporting software system, which would perform the calculation and return the risk score to the clinician almost immediately.

Our study reflects the importance of unbiased proteomics to inform relative importance of various proteins to diagnose cardiovascular processes such as PAD. We have argued for the value of this approach using biomarkers providing “orthogonal” information, showing incremental value, for example, in diagnosis of CAD or prognostication of major adverse cardiovascular events in this cohort. For example, much as in the present study, we previously found the composite of 2 clinical variables (male sex and previous percutaneous coronary intervention) and 4 biomarkers (midkine, adiponectin, apolipoprotein C‐I, and KIM‐1) could predict obstructive CAD with high accuracy (odds ratio: 9.74, P < 0.001),6 whereas a multiple‐marker strategy combining N‐terminal pro B‐type natriuretic peptide, KIM‐1, osteopontin, and tissue inhibitor of metalloproteinase‐1 was prognostic for the composite of cardiovascular death/myocardial infarction/stroke during follow‐up.7

The combination of biomarkers and clinical variable might be useful to clinicians. Given a range of score values that provides both strong PPV and NPV, one could theoretically argue that the use of a clinical/biomarker tool such as this could act as a gatekeeper to imaging or invasive testing, thereby reducing costs and exposures to intravenous contrast and/or ionizing radiation by avoiding expensive imaging modalities when unwarranted. Notably, the score performed similarly well in specific vascular regions (renal and lower‐extremity arteries). Thus, the scoring system could potentially be used in the diagnosis of renal‐artery stenosis or lower‐extremity arterial disease. It goes without saying that pretest probability determines post‐test accuracy of any diagnostic modality, and our results are no exception. In this setting, the score should serve as an adjunct to a thorough history and physical examination. Our score outperforms the only existing clinical score to diagnose lower‐extremity PAD, which used clinical variables alone (age, sex, smoking, diabetes mellitus, body mass index, HTN, and history of heart failure, CAD, and cerebrovascular disease) to yield a c‐statistic of 0.64.23 However, if there are clinical indications of significant PAD, angiography will be needed even if proteomics are not performed.

Furthermore, we found our score to have prognostic utility, with a shorter time to revascularization in those with an elevated score. As such, the scoring system also may be used to evaluate patients at risk for vascular complications. This information could be utilized to guide therapeutic intervention. Though aspirin and statins are standard therapies in PAD, novel oral anticoagulants and proprotein convertase subtilisin/kexin type 9 (PCSK9) inhibitors have recently demonstrate promise in this cohort.24, 25 Cost and side‐effect profile may limit their utility; thus, appropriate risk‐stratification will be important to guide therapy. A clinical/biomarker score may fit this need. Finally, it is plausible that our score may play a role in clinical trials to enrich for PAD‐related events or to identify patients at risk for adverse effects of drug therapies.

4.1. Study limitations

Despite being unique, our study has limitations. First, the number of patients from which we derived our findings was relatively small and included patients who only underwent coronary angiography as negative controls; though such patients were at low risk for PAD, unsuspected disease might have been present. As such, the results of our study should serve as preliminary evidence that requires confirmation in larger and ad hoc cohorts. The study population was predominantly Caucasian, which limits to the external validity of our score to African American patients. Biomarkers were measured at a single point in time, which may not reflect levels at future time periods. Furthermore, the diagnostic score was not compared with other noninvasive modalities such as ABI or ultrasonography, a comparison that will necessitate investigation in future studies. Our results need further validation and should not be extrapolated to the general population without suspected or known PAD, as these patients were not included in our study. Indeed, we reiterate the importance of clinical context when applying any diagnostic modality; the patients in our study were referred with clinical suspicion for significant PAD. Nonetheless, our clinical/proteomics model was able to deliver both excellent PPV and NPV when scaled according to high vs low score in this higher‐risk cohort. More studies will help to understand the value of our approach in different risk populations. Finally, utilizing the clinical/biomarker score alone will not be sufficient to differentiate the exact territory of PAD, and thus clinical correlation with history and physical examination will be a fundamental component of assessment.

5. CONCLUSION

We have developed a clinical and proteomics multiple‐marker scoring strategy to reliably diagnose obstructive PAD, also lending potential prognostic information regarding need for revascularization.

Conflicts of interest

Dr. Januzzi has received grant support from Roche Diagnostics, Siemens, Cleveland Heart Labs, and Prevencio; has received consulting income from Roche Diagnostics, Critical Diagnostics, Philips, and Novartis; and participates in clinical endpoint committees/data safety monitoring boards for AbbVie, Bayer, Pfizer, Novartis, Amgen, Janssen, and Boehringer Ingelheim. Mr. Magaret is a consultant to Prevencio, Inc. Dr. Gaggin has received grant support from Roche and Portola; has received consulting income from Roche Diagnostics, American Regent, Amgen, Boston Heart Diagnostics, and Critical Diagnostics; and has received research payments for clinical endpoint committees for EchoSense. Dr. Garasic has received consulting income from Siemens, Applied Clinical Intelligence, Bayer, Merck, Boehringer Ingelheim, and AbbVie. Ms. Rhyne and Dr. Barnes are employees of Prevencio, Inc. The authors declare no other potential conflicts of interest.

Supporting information

Supplemental Table 1 Locations of the peripheral artery lesions.

Supplemental Table 2: Baseline biomarker concentrations in patients with/without obstructive PAD.

Figure: Histogram showing distribution of the PAD Score to predict obstructive peripheral arterial disease

McCarthy CP, Ibrahim NE, van Kimmenade RRJ, et al. A clinical and proteomics approach to predict the presence of obstructive peripheral arterial disease: From the Catheter Sampled Blood Archive in Cardiovascular Diseases (CASABLANCA) Study. Clin Cardiol. 2018;41:903–909. 10.1002/clc.22939