A Deep Learning Framework for Deriving Noninvasive Intracranial Pressure Waveforms from Transcranial Doppler

Abstract

Increased intracranial pressure (ICP) causes disability and mortality in the neurointensive care population. Current methods for monitoring ICP are invasive. We designed a deep learning framework using a domain adversarial neural network to estimate noninvasive ICP, from blood pressure, electrocardiogram, and cerebral blood flow velocity. Our model had a mean of median absolute error of 3.88 ± 3.26 mmHg for the domain adversarial neural network, and 3.94 ± 1.71 mmHg for the domain adversarial transformers. Compared with nonlinear approaches, such as support vector regression, this was 26.7% and 25.7% lower. Our proposed framework provides more accurate noninvasive ICP estimates than currently available. ANN NEUROL 2023;94:196-202.

Figure 1:

The proposed Domain Adversarial Residual Network framework, the input $x_i^0$ to the framework $(AB{P}_i,EC{G}_i,FV)$. It consists of three layers: a) Feature extractor, b) ICP Estimator, and c) Domain classifier.

Figure 2:. Illustrates the performance of different models for deriving non-invasive intracranial pressure (nICP).

The proposed frameworks DANN and DAT performed best when compared to the current state of the art methods for nICP estimation. Green dot represents the mean. SV nu: Support Vector, SVR: Support Vector Regression, CNN: Convolutional Neural Network, LSTM: Long short-term memory, DACNN: Domain adversarial convolutional neural network, DANN: Domain adversarial neural network, DARN: Domain adversarial residual network, MAE: Median absolute error, DAT: Domain adversarial transformers.

Figure 3:. Bland-Altman Plots for comparing true intracranial pressure (ICP) vs non-invasive ICP.

Panels A, B, and C display the reference ICP (mmHg) versus the ICP error (mmHg) for three different methods: SV Nu, DANN, and DAT. Corresponding scatter plots of true ICP versus predicted ICP are shown in panels A1, B1, and C1. Panel D shows the percentage of data with an error of less than 5mmHg for each method. The standard error for the proposed DARN and DANN network is between +/− 10 mmHg, when compared to the current state of the art technique using SV nu. The percentage of ICP error less than 5 mmHg is 74.45% for the proposed framework (DAT), which is better when compared with 61.8% (SV Nu) and 39.02 % (SVR) for current non-linear approaches. SV nu: Support Vector, SVR: Support Vector Regression, CNN: Convolutional Neural Network, LSTM: Long short-term memory, DACNN: Domain adversarial convolutional neural network, DANN: Domain adversarial neural network, DARN: Domain adversarial residual network, DAT: Domain adversarial transformers.