Summary¶ It is well known that intracranial pressure (ICP) is influenced by an array of predictable and unpredictable factors, which gives rise to a signal heavily loaded with stochastic, i.e. random components. Hence, statistical modelling of this signal has proved to be of limited utility, in spite of the very sophisticated mathematical methods applied. In recent years, neural network algorithms (ANN), which are an alternative to statistical methods, have proved their effectiveness in the prediction of trends, as applied in a variety of medical and non-medical tasks. We therefore attempted to test the efficiency of neural models in the on-line prediction of ICP values, compare their effectiveness to statistically oriented algorithms and combine ANN methods with some newer signal processing algorithms, like wavelet decomposition. Prediction horizons of up to 5 minutes have been tested with various architectures of the neural predictor. For a 3 minute prediction horizon, a satisfactory accuracy of forecasting has been achieved with “plain” ANN, as expressed by the “average relative variance coefficient”. This was measured by the ratio of the prediction error obtained, in relation to the error which would occur if a current value were taken as the forecasted one. The prediction quality with statistical autoregressive models has proved unsatisfactory, whilst the result obtained using the ANN model with the wavelet transform incorporated, performed significantly better than the ANN models alone. The prediction quality obtained with the ANN methodology seems to be satisfactory over a short time horizon, though no conclusion can be derived at this stage of the study, as to the clinical utility of this method. In particular, even with this methodology, it is not possible to forecast any sudden dehiscencies of the ICP signal with any practical reliability. From the point of view of modelling theory, such sharp deviations of the signal may be regarded as a “catastrophe”. This implies the necessity for a different approach to the ICP signal analysis with the artificial intelligence methodology; one, that is more oriented towards the global properties of the signal.
[1]
M. Daley,et al.
Detection of loss of cerebral vascular tone by correlation of arterial and intracranial pressure signals
,
1995,
IEEE Transactions on Biomedical Engineering.
[2]
P J Benkeser,et al.
A computer-based statistical pattern recognition for Doppler spectral waveforms of intracranial blood flow
,
1996,
Comput. Biol. Medicine.
[3]
Duc Truong Pham,et al.
Neural Networks for Identification, Prediction and Control
,
1995
.
[4]
Gwilym M. Jenkins,et al.
Time series analysis, forecasting and control
,
1971
.
[5]
M. Ursino,et al.
Intracranial pressure dynamics in patients with acute brain damage: a critical analysis with the aid of a mathematical model
,
1995,
IEEE Transactions on Biomedical Engineering.
[6]
Simon Haykin,et al.
Neural Networks: A Comprehensive Foundation
,
1998
.
[7]
Guanrong Chen,et al.
Kalman Filtering with Real-time Applications
,
1987
.
[8]
Truong Q. Nguyen,et al.
Wavelets and filter banks
,
1996
.
[9]
Kishan G. Mehrotra,et al.
Forecasting the behavior of multivariate time series using neural networks
,
1992,
Neural Networks.