Artificial neural network based intracranial pressure mean forecast algorithm for medical decision support

Abstract

Although the future mean of intracranial pressure (ICP) is of critical concern of many clinicians for timely medical treatment, the problem of forecasting the future ICP mean has not been addressed yet. In this paper, we present a nonlinear autoregressive with exogenous input artificial neural network based mean forecast algorithm (ANN(NARX)-MFA) to predict the ICP mean of the future windows based on features extracted from past windows and segmented sub-windows. We compare its performance with nonlinear autoregressive artificial neural network algorithm (ANN(NAR)) without features extracted from window segmentation. Experimental results showed that, ANN(NARX)-MFA algorithm outperforms ANN(NAR) algorithm in prediction accuracy, because additional features extracted from finer segmented sub-windows help to catch the subtle changes of ICP trends. This verifies the effectiveness of decomposing the whole window into sub-windows to obtain features in making predictions on future windows. Based on the forecast of ICP mean, medical treatments can be planned in advance to control ICP elevation, in order to maximize recovery and optimize outcome.

Development of Traumatic Brain Injury Associated Intracranial Hypertension Prediction Algorithms: A Narrative Review

“…In 2011, Zhang and associates 49 described a non-linear autoregressive with exogenous input artificial neural network-based mean forecast algorithm (ANN NARX -MFA) to predict ICP. This algorithm, which extracted features from past windows and segmented sub-windows for processing, was shown to be superior (15-min coefficient of determination [R 2 ] 0.93 ± 0.05 and relative absolute error [RAE] 9% ± 3%) to an ANN NAR algorithm (15-min R 2 0.88 ± 0.07 and RAE 15% ± 5%), which did not incorporate feature extraction.…”

“…This algorithm, which extracted features from past windows and segmented sub-windows for processing, was shown to be superior (15-min coefficient of determination [R 2 ] 0.93 ± 0.05 and relative absolute error [RAE] 9% ± 3%) to an ANN NAR algorithm (15-min R 2 0.88 ± 0.07 and RAE 15% ± 5%), which did not incorporate feature extraction. 49 In a 2012 follow-up study by the same group, Feng and co-workers demonstrated that a forecasting model utilizing temporal information on historical ICP readings and related parameters (MAP, pressure reactivity index [PRx], and brain tissue oxygenation [P bt O 2 ]) improved forecasting performance by an average of 20% ( p -value <0.001) compared with algorithms that did not make use of such information. 50 …”

“…A variety of recurrent ML algorithms have also been used for both tIH prediction and ICP forecasting with varying degrees of success. 29 , 37 , 44 , 49 , 50 , 67 , 79 Recurrent algorithms, which include RNNs, LSTMs, and other deep neural networks, are a family of algorithms designed to make predictions from temporal dynamic evolution of the input features. 80 , 81 Temporal dynamics is the concept that driver–response relationships are not always constant through time, but rather are conditioned to varying degrees by both recent and remote past events.…”

“…Integration of waveform data, as opposed to averaged data, for both training and operation of the latest generation of memory-enabled recurrent algorithms is a logical next step. 49 , 50 , 58 , 83 …”

Forecasting intracranial hypertension using multi-scale waveform metrics

“…It uses simple features such as the last measured ICP value or the time to the last ICH crisis. Besides tackling the classification task directly, other models have been suggested that predict the future ICP mean value, for example by using nearest-neighbor regression (Bonds et al 2015b), neural networks (Shieh et al 2004, Zhang et al 2011 or ARIMA models (Zhang et al 2012).…”