Neural network modelling of dynamic cerebral autoregulation: assessment and comparison with established methods

A time lagged recurrent neural network (TLRN) was implemented to model the dynamic relationship between arterial blood pressure (ABP) and cerebral blood flow velocity (CBFV) and its performance was compared to classical linear model such as transfer function analysis, Aaslid's dynamic autoregulation model, and the Wiener-Laguerre moving average filter. A simple linear regression was also tested as a naive estimator. In 16 normal subjects, CBFV was continuously recorded with Doppler ultrasound and ABP with the Finapres device during six repeated thigh cuff manoeuvres. Using mean beat-to-beat values of ABP as input and CBFV as output, the performance of each method was assessed by the model's predicted velocity correlation coefficient and normalized mean square error (MSE). Cross-validation was performed using three thigh cuff manoeuvres for the training data set and the other three for the validation set. The four methods studied performed significantly better than the zero-order naive estimator. The TLRN performed better than transfer function analysis, but was not significantly different from the time-domain techniques, despite showing the minimum predictive MSE. CBFV step responses could be extracted from the TLRN showing the presence of non-linear behaviour both in terms of amplitude and directionality.