Intracranial Pressure Monitoring Signals After Traumatic Brain Injury: A Narrative Overview and Conceptual Data Science Framework

Figures

• 

  Figure 1. ICP pulse morphology: (A) Normal ICP Waveform with dominant P1, (B) Abnormal ICP waveform with dominant P2. ICP waveforms during normal intracranial physiologic conditions (A) differ from waveforms recorded when brain compliance is reduced (B). The waveform (A) represents the dominant P1 with amplitude P1 > P2 > P3 while waveform (B) represent the dominant P2 with amplitude P2 > P1, P3.

• 

  Figure 2. The frequency components in a typical ICP signal. The ICP signal can be decomposed into frequency domain components based on contributing sources. The bandpass filter technique has been applied to raw ICP signal (A) to extract cardiac cycle (B), respiratory waves (C), Mayer's waves (D), and slow waves (E). The passband frequency ranges (Hz) for bandpass filter is set to [1, 1.3], [0.26, 0.3], [0.1, 0.101], and [0.0055, 0.5], respectively.

• 

  Figure 3. Framework of data mining approaches for ICP signals in patients with TBI. In this framework, the raw ICP signal is fed into a data processing layer to remove high frequency noise and artifacts. Then, features are extracted in the information layer, across two major categories: morphology and segment signal characteristics, including time and frequency domain features. In order to extract the morphology of the ICP waveform, ECG or ABP serve as reference signal to help identify ICP pulse waves and peak detection can validate the accuracy of extracted features. Final, an application layer provides a link between the ICP waveform and clinically relevant events.

• 

  Figure 4. High- frequency Noise Removal for ICP signal. (A) PSD estimation of ICP signal. (B) Comparison of before and after signal filtering. The high frequency noise can be removed by low-pass filtering. The periodogram is obtained using fast Fourier transform (FFT). In order to conserve the total power, both the positive and negative frequencies multiply by a factor of 2. The zero frequency (DC) equation is psdx= 1Fs×N×|fft(x)1:N2+1|2, where psdx is the estimated power spectral density, Fs is the ICP signal sample rate, N is the ICP total sample size and fft is the FFT function. The PSD results indicate the cut-off frequency rate can be set to 10 Hz in other cases because most of the energy associated with the signal is confined under the threshold of 8 Hz. (B) shows the effects of removing high-frequency noise from the ICP signal via a low-pass filter (Fs = 128, Fpass = 10 Hz, Fstop = 15 Hz, Apass = 1, Astop = 50, minimum order).

• 

  Figure 5. (A–D) Artifacts in the time series of the continuous ICP signal. Several low-frequency artifacts are shown.

• 

  Figure 6. Overview of intracranial hypertension detection and prediction modeling. Intracranial hypertension detection uses intracranial hypertension episodes and control episodes while intracranial prediction uses pre-hypertension episodes and control episodes. The statistical analysis is usually evaluated by p-value which is concerned with the significant difference of individual features between pre-intracranial hypertension/intracranial hypertension and control groups. Machine learning models are often evaluated by accuracy, AUC values etc. to demonstrate the predictive power to distinguish pre-intracranial hypertension/intracranial hypertension groups from other control groups. ACC, Accuracy; AUC, Area Under receiver operating characteristic Curve; CNN, Convolutional Neural Networks; MLR, Multivariate linear regression; SEN, Sensitivity; SR-DA, Spectral Regression- Discriminant Analysis; SVM, Support Vector Machine; SPE, Specific; SR-KDA, Spectral Regression-Kernel Spectral Regression; QD, Quadratic Classifier.

• 

  Figure 7. Strengths and limitations of modeling techniques by prediction accuracy and explainable degree.