A novel machine learning derived proteomics-leveraged panel to predict presence and prognosis of obstructive peripheral artery disease: the CASABLANCA Study
J Januzzi, C Magaret, C Hoti, R RhyneAbstract
Background
Peripheral artery disease (PAD) is associated with significant morbidity and mortality; it is frequently underdiagnosed until advanced stages, when complications and adverse outcomes become more likely. Although noninvasive approaches are available to assess presence and severity of PAD, these tools have important limitations, including variable accuracy, limited accessibility, high cost, and, in some cases, intravenous contrast and ionizing radiation. Availability of an easily interpretable biomarker-leveraged panel for evaluation of PAD would be of value. Such a tool would also optimally deliver prognostic information.
Purpose
To develop a commercially feasible clinical and multiprotein panel predicting the presence and prognosis of lower-extremity obstructive PAD.
Methods
Study participants prospectively underwent angiography of the iliac and infra-inguinal vessels, and a blood sample was obtained prior to the procedure. From more than 150 biomarkers and clinical variables, machine learning methods determined a final panel to identify peripheral artery stenosis ≥50%; it consisted of one clinical variable (history of hypertension) and concentrations of three biomarkers (midkine, kidney injury molecule-1, and angiopoietin-1). During development, performance of this diagnostic panel was assessed in patients using Monte Carlo cross validation, and a final model was evaluated with in-sample validation.
Results
In the study sample of 352 participants, the PAD panel had excellent performance for prediction of peripheral stenosis > 50% with an in-sample C-statistic of 0.81 (95% C.I.: 0.77, 0.86; P <0.001) for obstructive PAD; higher scores were associated with greater severity of angiographic stenosis (Figure 1). Using a cutoff derived from the optimal Youden’s Index, the model performed with a sensitivity of 80%, specificity of 75%, positive predictive value (PPV) of 61% and negative predictive value (NPV) of 85%. Translating the diagnostic score into a 3-level result, a score of 1 (low risk) had an NPV or 90% and a score of 3 (high risk) had a PPV of 71% for the detection of PAD. In Cox proportional hazards analysis adjusted for age, sex, smoking status, history of myocardial infarction, diabetes, and kidney function, over an average 3 year follow up an increasing score was associated with a shorter time to revascularization (hazard ratio = 1.75, 95% C.I.: 1.33, 2.31; p < 0.001; Figure 2).
Conclusions
Using commercially available biomarkers, we describe a clinical/ biomarker panel with high accuracy for predicting the presence, severity, and prognosis of patients with PAD.Fig 1:Panel result and mean vessel stenosisFor image description, please refer to the figure legend and surrounding text.Fig 2:Panel result and revascularizationFor image description, please refer to the figure legend and surrounding text.