Artifact removal from neurophysiological signals: impact on intracranial and arterial pressure monitoring in traumatic brain injury

Restricted access

OBJECTIVE

Monitoring intracranial and arterial blood pressure (ICP and ABP, respectively) provides crucial information regarding the neurological status of patients with traumatic brain injury (TBI). However, these signals are often heavily affected by artifacts, which may significantly reduce the reliability of the clinical determinations derived from the signals. The goal of this work was to eliminate signal artifacts from continuous ICP and ABP monitoring via deep learning techniques and to assess the changes in the prognostic capacities of clinical parameters after artifact elimination.

METHODS

The first 24 hours of monitoring ICP and ABP in a total of 309 patients with TBI was retrospectively analyzed. An artifact elimination model for ICP and ABP was constructed via a stacked convolutional autoencoder (SCAE) and convolutional neural network (CNN) with 10-fold cross-validation tests. The prevalence and prognostic capacity of ICP- and ABP-related clinical events were compared before and after artifact elimination.

RESULTS

The proposed SCAE-CNN model exhibited reliable accuracy in eliminating ABP and ICP artifacts (net prediction rates of 97% and 94%, respectively). The prevalence of ICP- and ABP-related clinical events (i.e., systemic hypotension, intracranial hypertension, cerebral hypoperfusion, and poor cerebrovascular reactivity) all decreased significantly after artifact removal.

CONCLUSIONS

The SCAE-CNN model can be reliably used to eliminate artifacts, which significantly improves the reliability and efficacy of ICP- and ABP-derived clinical parameters for prognostic determinations after TBI.

ABBREVIATIONS ABP = arterial blood pressure; CNN = convolutional neural network; CPP = cerebral perfusion pressure; GCS = Glasgow Coma Scale; ICP = intracranial pressure; IQR = interquartile range; KSVM = kernel support vector machine; LSVM = linear support vector machine; PRx = pressure reactivity index; ReLu = rectified linear unit; SCAE = stacked convolutional autoencoder; TBI = traumatic brain injury.

Article Information

Correspondence Dong-Joo Kim: Korea University, Seoul, South Korea. dongjookim@korea.ac.kr.

INCLUDE WHEN CITING Published online May 10, 2019; DOI: 10.3171/2019.2.JNS182260.

Disclosures ICM+, a signal processing software used in this study, is licensed through Cambridge Enterprise Ltd., Cambridge, UK. M. Czosnyka has a financial interest in 10% of the licensing fee, and P. Smielewski receives a portion of the licensing fee.

© AANS, except where prohibited by US copyright law.

Headings

Figures

  • View in gallery

    A graphic overview of signal processing: onset detection (A), normalization (B), generation of representative image (C), artifact classification (D), and artifact removal (E). After onset detection, segmented signals were converted into a 2D image (64 × 64). The SCAE was used to generate a representative image of the ICP pulse. This allows morphological feature extraction, including the general characteristics of the ICP signal. A CNN suitable for morphology classification was used to classify normal and artifactual pulses. The signals classified as artifacts were removed based on the detected pulse onset. All the convolutional layers in the SCAE and CNN used a ReLu activator. The classification layer in the CNN is the softmax dense layer. Conv = convolution. Figure is available in color online only.

  • View in gallery

    Representative cases of ICP and ABP artifact detection by the CNN. Examples of high-frequency artifacts usually generated by patient motion are shown in A and B. Artifacts generated by mechanical problems are shown in C and D. The signal corrupted by these artifacts may heavily distort clinical parameters. Temporal biological artifacts are shown in E and F. This type of artifact could contaminate mean pressure values. The shaded area is the section that is classified as an artifact by the proposed method.

  • View in gallery

    Radial plot for a comparison of the performance of 4 methods in detecting artifacts of ABP (A) and ICP (B). The minimum value of the axis is 50%, and the maximum value of the axis is 100%. Acc = accuracy; F = F-score; G = G-score; Net = net prediction; Sen = sensitivity; Spe = specificity. Figure is available in color online only.

  • View in gallery

    Bar graphs showing the incidence of systemic hypotension (systolic blood pressure < 90 mm Hg), poor cerebrovascular pressure reactivity (PRx > 0.3), cerebral hypoperfusion (CPP < 60 mm Hg), intracranial hypertension I (ICP > 20 mm Hg), and intracranial hypertension II (ICP > 22 mm Hg), measured as the counts (A), duration (B), and proportion (C) in the entire recordings. Dark gray bars show the mean prevalence detected from the artifact-removed signals, whereas light gray bars show the mean prevalence detected from the raw signals. The asterisks above the bars indicate statistical significance (*p < 0.05, **p < 0.01) according to the Mann-Whitney U-test.

  • View in gallery

    Bar graph showing the number of events when ICP was > 20 mm Hg and CPP was < 60 mm Hg, when ICP was > 22 mm Hg and CPP was < 60 mm Hg, and when ICP was > 25 mm Hg and CPP was < 60 mm Hg. The light gray bars show the mean number of counts detected from the raw signals, whereas the dark gray bars show the mean prevalence detected from the signals after artifact removal. The asterisks above the bars indicate statistical significance (**p < 0.01).

References

TrendMD

Metrics

Metrics

All Time Past Year Past 30 Days
Abstract Views 134 134 134
Full Text Views 34 34 34
PDF Downloads 23 23 23
EPUB Downloads 0 0 0

PubMed

Google Scholar