Assessment of degree of internal carotid artery stenosis based on duplex velocity measurements using an artificial neural network

Background: Duplex imaging is increasingly used as a stand‐alone investigation before carotid endarterectomy. There is concern regarding the wide variation in diagnostic methods of grading internal carotid artery (ICA) stenosis using duplex velocity criteria. The aim of the study was to design an artificial neural network (ANN) that is able to provide a measure of the degree of ICA stenosis based on the four most commonly used duplex velocity profiles. Methods: One hundred and four consecutive patients who were admitted for assessment of carotid occlusive disease were included. All patients underwent arch injection digital subtraction angiography and carotid duplex ultrasonography. Peak systolic velocity (PSV) and end‐diastolic velocity (EDV) in the ICA and common carotid artery (CCA) were measured. The degree of angiographic stenosis was measured by two blinded investigators. A three‐layered perceptron ANN was constructed using the NeuroSolutions TM version 1 software program (NeuroDimension, Gainesville, FL, USA). The accuracy of the model in predicting the degree of ICA stenosis was expressed as the mean squared error (MSE) of the actual degree of angiographic stenosis. The performance of the neural network model was compared with that of the linear regression model at predicting the degree of ICA stenosis and classifying a degree of ICA stenosis greater than 50 per cent. Results: One hundred and sixty‐eight carotid bifurcations were available for analysis. The MSE of the ANN at predicting the degree of ICA stenosis was 0·096 (overall accuracy of 84 per cent), which was superior to a linear regression model (MSE 0·132; overall accuracy 73 per cent) ( P < 0·001). Both models were equally accurate at identifying a degree of ICA stenosis greater than 50 per cent (ANN accuracy 0·91 versus regression model accuracy 0·89; P = 0·09). Conclusion: Neural network algorithms are able to predict the degree of ICA stenosis with reasonable accuracy and outperform linear regression analysis on the basis of a limited set of duplex velocity measurements. With further refinement, ANN models could replace duplex criteria in assessment of the degree of ICA stenosis. © 2001 British Journal of Surgery Society Ltd