B-mode ultrasound measurements of carotid intima media thickness and the assessment of global cardiovascular risk

Carotid intima-media thickness (C-IMT) has been shown to be related to vascular risk factors (VRFs), prevalent cardiovascular disease (CVD), and atherosclerosis in coronary and peripheral arteries. Despite these relationships only a few studies have evaluated the potentiality of C-IMT to identify patient at high risk of CVD. In these, C-IMT included in a risk function for the assessment of global risk does not increases its predictivity. This can be due to a real lack of prediction capacity of C-IMT but also to the use of statistical tools unable to disentangle the non linear relationships connecting IMT to the global risk. Artificial neural networks (ANNs) are highly sophisticated computer algorithms able to recognise even the more hidden non linear relationships relating different variables, and to absolve complex classification tasks. In the present study the potentiality of C-IMT, alone or added to established VRFs, to identify patients at high risk of vascular disease (VD) was investigated by using ANNs and the classical statistical approach based on discriminant analyses. Patients were arbitrarily assigned to the high risk group when suffering from overt cardio- (myocardial infarction or angina), cerebro- (transient ischemic attack or stroke) or peripheral-VD. Arterial near and far wall of left and right carotid arteries were measured from 578 patients (464 at low and 114 at high risk for VD) by using B-Mode ultrasound. The results show that ANNs can be trained to identify low and high risk subjects with a greater accuracy than discriminant analyses. In addition, with optimal settings, a prediction accuracy of about 87% was achieved using conventional VRFs as input variables in the ANN classification task. When only ultrasonic variables were used, a prediction accuracy of about 77% was observed. The addition to this set of variables obtained without any additional cost (gender, age, weight, height and body mass index) led the accuracy of prediction to 86%. Pooling data of all ultrasonic variables and all VRFs did not significantly improve the performance of ANNs in the classification task (prediction accuracy = 83%). Finally, when ANNs were allowed to choose automatically the relevant input data (I.S. system-Semeion), 31 variables were selected and, among these, 6 were ultrasonic variables. By using this set of variables as input data the performance of ANNs in the classification task increased, reaching a prediction accuracy close to 92%, with 100% of correct classification of high risk patients. In conclusion, with the ANN technology C-IMT may increase the discriminant capacity of vascular risk factors in the classification of patients into low or high risk classes.