2.1 Patient’s Cohort
In the present investigation, randomly selected patients with established carotid atherosclerosis and subjects without apparent atherosclerotic manifestations were enrolled after informed consent. 136 patients, from the Vascular Surgery Clinic or the Cardiology Clinic of the University General Hospital “Attikon”, were enrolled in the study. The study protocol conformed to the ethical guidelines of the 1975 Helsinki Declaration and was approved by the Ethics Committee of “Attikon” University Hospital. Patients were subsequently examined and assigned into the following groups :
Symptomatic (high risk) group (stenosis degree ≥ 50%, which is proven to be responsible for ischemic stroke in the last six months): the number of patients in this group was 82.
Asymptomatic (low risk) group (stenosis degree ≤ 50%): it consisted of 54 individuals to identify differences from the patients with symptoms.
Moreover, it is well known that atherosclerosis can lead to stroke, also, this disease starts when the arteries endothelium becomes damaged; therefore, arterial stenosis, in a relatively high percentage, happens [1, 2].
2.2 Clinical Data Collection and Imaging Tests
Patients demographic data were recorded: including gender, age, smoking habits, body mass index (BMI, underweight < 18.5; normal weight 18.5–25; overweight 25–30; obese > 30), systolic blood pressure (Systolic BP, normal 100–140 mmHg), triglycerides (normal < 150 mg/dL; mildly high 150–199 mg/dL; high 200–499 mg/dL; very high > 500 mg/dL) white blood cells (WBC, average normal range 4500–10,000 counts/mm3). Blood pressure was measured twice with the patient seated for at least 15 min and with an intermediate time interval of 5 min between the measurements. The average of the measurements was evaluated and recorded, without any differences between the right and the left arm.
To define the stenosis degree, all participants underwent carotid ultrasound followed by image analysis at the study initiation. The ultra-sonographic test was performed by one experienced vascular surgeon to avoid inter-observer variability, using a linear transducer signal of 12 MHz (General Electric LogiqE, Riverside, USA). The patients were in a supine position with a slight stretch of the head, to be able to display the ipsilateral carotid in an appropriate suitable longitudinal and lateral projection. Both carotids, were tested with an angle between head and transducer < 60° and the patient being in apnea. To ensure the reproducibility of carotid measurements, a carotid ultrasound assessment protocol was developed, by the University of Athens team, based on co-evaluation of sonographic and angiographic findings. In that way, the longitudinal ultrasound imaging was taking place in standard view of the carotid, while all other parameters were stable (imaging modality, B-mode; dynamic range, 60 dB; persistence, low; frame rate, higher than 25 frames/s). To have comparability of the measured parameters between the groups, anatomical areas of interest were predetermined based on guide points that included 3 cm length of the proximal internal carotid artery, carotid bulb and 1 cm length of the common carotid artery. Atherosclerotic lesions were distributed in the 3 carotid regions as follows: (a) carotid division (50%), (b) internal carotid artery (30%), and (c) common carotid artery (20%). In addition, to avoid inter-observer variability, all patients were evaluated by a single and experienced vascular surgeon.
For all patients, blood sampling was performed in the morning (8:00 and 10:00 a.m.) after an overnight fast. Serum and plasma were isolated after centrifugation at 700 g. Samples were stored at − 80 °C. Levels of all proteins, except fibrinogen, were measured in serum using multiplex analysis. Using commercially available kits, inflammation related biomarkers were selected. This included measurement of plasma concentrations of fibrinogen, matrix metalloproteinase-1 (MMP-1), tissue inhibitor of metalloproteinase (TIMP-1), soluble intercellular adhesion molecule (SiCAM), soluble vascular cell adhesion molecule (SvCAM), adiponectin and insulin (EMD Millipore Corporation, Darmstadt, Germany). Bead assays were performed according to the manufacturer’s protocol (Millipore, Billerica, MA, USA) and were analyzed on a Luminex 3D platform (Luminex Corp, Austin, TX, USA).
Intra-assay precision for all proteins was generated from the mean of the %CV’s from eight reportable results across two different concentrations of analytes in a single assay; thus for an overnight protocol, the values were: < 15%CV for fibrinogen, ≤ 10%CV for TIMP-1, < 15%CV for SiCAM and SvCAM, 2%CV for adiponectin and < 10% for insulin; while for a 2-h protocol, the values were: 2.6%CV for MMP-1. Inter-assay precision was generated from the mean of the %CVs across two different concentrations of analytes and across 4 different assays. The values were: < 20% CV for fibrinogen, ≤ 10%CV for TIMP-1, < 20%CV for SiCAM and SvCAM, 10%CV for adiponectin, < 15% for insulin and 8.4%CV for MMP-1, respectively.
2.4 Statistical Analysis
The variation of the biomarkers’ levels between symptomatic and asymptomatic patients was performed by the independent-samples t test. Results were reported as mean value ± SD. The distributions were tested for normality by Kolmogorov–Smirnov analysis. A p value < 0.05 was considered to be statistically significant. SPSS statistics software (version 22.0; SPSS Inc., Chicago, IL, USA) was used both for statistical analysis and CART construction.
2.5 Classification and Regression Tree (CART)
Classification and regression tree  is a recursive, partitioning, machine learning technique that builds tree-like structures for predicting or discriminating continuous variables (hence regression) or categorical variables (classification). In this study, the produced CART tree-like algorithm, created a set of if–then logical/split rules, eventually allowing assignment of patients as symptomatic or asymptomatic. The probability of a case belonging to one of the two categories was also provided by the CART. The CART was composed of nodes (i.e., points where decisions are made), moreover nodes had a hierarchy of layers: the first layer had only one node (“root” node), nodes in subsequent layers were linked with nodes in two other layers (parent and children nodes, called “branches”), while this hierarchical structure ends with terminal nodes (having only parents but not children, called “leafs”).
The CART was built in a recursive manner on the basis of the supplied clinical characteristics and serum biomarkers. During each recursion, statistical measures for all supplied parameters were calculated, subsequently the parameter that allows better separation of the cases was identified and used to separate the case as symptomatic or asymptomatic (i.e., a new if–then rule was created), moreover two (or more) new children nodes were created. The patients were distributed to each one of the “children” nodes, which contained the number of patients being symptomatic and asymptomatic as well as the relevant probability. Subsequently, and if each child node contained both symptomatic and asymptomatic cases, the algorithm was repeated for each new child node that was produced. Once again the parameter allowed better separation of cases, was statistically identified and was used to create a new if–then rule and new “children” nodes.
The CART minimum parent size was set to 10, the minimum child size was set to five and the maximum allowed depth was set to eight, the QUEST growing algorithm was employed, while a tenfold cross-validation was performed. Note that these limits were not reached, the selected values are typical values and rather heuristic.