Ethical approval for this study was provided by the Ethical Committee Institutional Helsinki committee of Galilee Medical Center, Nahariya, Israel (Ethical Committee No. 0137-13-NHR)
Originally, we tested various aspects of pathophysiology of pneumoperitoneum during laparoscopic cholecystectomy, and the participants gave informed consent to the current study, including the use of total intravenous anesthesia (TIVA). We delayed the publication of the results of our study until we succeeded in refining features that reliably represent cognitive parameters.
2.1 Patients
Our study included 17 patients who were electively scheduled for laparoscopic cholecystectomy due to symptomatic cholecystolithiasis. All patients provided written informed consent to participate in the experiment. All surgeries occurred at the Galilee Medical Center, Nahariya, Israel.
Mean age was 45.7 ± 14.63 years (range 23–73; only one patient was aged 73, and age distribution in both cases was the same in both groups); M/F ratio 6/11. All patients were in a good medical state, with ASA scores I-II. Pre-operative blood tests, including renal and liver function tests, were within normal ranges. All participants did not take any regular medication, and their cognitive statuses were normal. During surgery, patients were ventilated by conventional intermittent positive pressure (IPPV). The patients were randomly divided into two groups: those undergoing volatile anesthesia (9 patients) and those undergoing TIVA (8 patients). Randomization was achieved by choosing a sealed envelope containing the mode of anesthesia. Both groups were statistically matched in relation to their demographics and medical parameters.
2.2 Procedure
The main laparoscopic approach included abdominal CO2 insufflation (digitally controlled) to 14 mmHg and introduction of 4 trocars/cannula into the peritoneal cavity. Following abdominal exploration via a video camera, the cystic duct and vessels were identified, clipped, and divided, and the gallbladder dissected, resected, and evacuated from the abdomen by an Endo Bag. Surgery was terminated after peritoneal CO2 evacuation. Mean duration of surgery was 38.2 ± 8.46 minutes (range 25–55 minutes).
Anesthesia was administered by a senior anesthesiologist who is familiar with both TIVA and volatile anesthesia. Monitoring of the depth of anesthesia was done by clinical assessment according to standard pathophysiological criteria, in conjunction with the BIS (44–45), keeping BIS scores between 40–60. When considering older subjects, we kept the BIS index levels close to their higher thresholds, in addition to using physiologic parameters to assist in adjusting depths of anesthesia. Induction of anesthesia was the same in both anesthetized groups. Anesthesia was induced with the intravenous administration of 1–3 mg of dormicum (midazolam), 0.1 mg of fentanyl, 150–200 mg of propofol, and 0.5 mg/kg of esmeron. The volatile anesthesia group received sevoflurane with nitrous oxide (60%) for maintenance. The concentration of sevoflurane ranged between 1.5% and 2.5%, making the adjustment according to the minimal alveolar concentration (MAC), hemodynamic parameters, and BIS. For analgesia, the patients under volatile anesthesia received fentanyl or remifentanil. Patients in the TIVA group received 2–4 ng/mL remifentanil (46) and 2–4 µg/mL propofol (46, 47) for maintenance. The titration of both drugs was done according to TCI regulations, and the exact concentrations were affected by physiological parameters (clinical assessment) and BIS (47–51). For those drugs’ administration, we used the B Braun Perfusor Space Target Controlled Infusion Pump (B Braun Medical Ltd., UK). Regarding the pump, we used the Schnider algorithm pharmacological model for propofol and the Minto algorithm for remifentanil, adhering to TCI regulation, and in accordance with BIS and basic clinical assessments. Esmeron was used in both treatment groups as needed for muscle relaxation.
The anesthesiologist attached three electrodes on each patient’s forehead. Real-time activity of the EEG features (described below) was presented during the surgery. For the data analysis, all participants’ data was downloaded as second-by-second activity after all surgeries were done.
2.3 EEG device
The EEG signal acquisition system included a three-electrode patch attached to each subject’s forehead (Aurora by Neurosteer®, Herzliya, Israel). The medical-grade electrode patch included dry gel for optimal signal transduction. The electrodes were located at Fp1 and Fp2, with a reference electrode at Fpz. The EEG signal was amplified by a factor of 100 and sampled at 500 Hz. Signal processing was performed automatically by Neurosteer® in the cloud (see Section 2.1.4 below and Appendix A). We were therefore provided with a sample per second of activity of the brain oscillations (i.e., delta, theta, alpha, beta, and gamma) and three selected features (i.e., VC9, ST4, and A0); see signal processing below.
2.4 Signal processing
Full technical specifications regarding the signal analysis are provided in Appendix A. In brief, the signal processing algorithm interprets the EEG data using a time/frequency wavelet-packet analysis, instead of the commonly used spectral or wavelet analysis. The best basis algorithm (52) constructs an orthogonal decomposition of the single EEG channel and, together with the frequency bands, creates a presentation of 121 features, which we term brain activity features (BAFs). Unlike the standard-frequency band features, the time/frequency wavelet packets also include, in addition to the time/varying components, higher harmonics.
To demonstrate this process, let g and h be a set of biorthogonal quadrature filters created from the filters G and H, respectively. These are convolution-decimation operators, where in a simple Haar wavelet, g is a set of averages and h is a set of differences.
Let \({\psi }_{1}\) be the mother wavelet associated with the filters \(s\in H\), and \(d\in G\). Then, the collection of wavelet packets \({\psi }_{n}\), is given by:
(1)
$${\psi }_{2n}=H{\psi }_{n};{ \psi }_{2n}\left(t\right)=\sqrt{2}{\sum }_{j\in Z}\text{s}\left(j\right){\psi }_{n}\left(2t-j\right),$$
(2)
\(\) \({\psi }_{2n+1}=G{\psi }_{n};{ \psi }_{2n+1}\left(t\right)=\sqrt{2}{\sum }_{j\in Z}\text{d}\left(j\right){\psi }_{n}\left(2t-j\right).\) The recursive form provides a natural arrangement in the form of a binary tree. The functions \({\psi }_{n}\) have a fixed scale. A library of wavelet packets of any scale s, frequency f, and position p is given by
(3)
$${ \psi }_{sfp}\left(t\right)={2}^{-s/2}{\psi }_{f}\left({2}^{-s}t-p\right).$$
The wavelet packets \({\{\psi }_{sfp}: p\in Z\}\) include a large collection of potential orthonormal bases. An optimal basis can be chosen by the best basis algorithm (52). Furthermore, an optimal mother wavelet can be chosen by (53). Following robust statistics methods to prune some of the basis functions, one gets 121 basis functions, which we term brain activity features (BAFs). Based on a given labeled-BAFs dataset, various models can be created for different discriminations of these labels. In the linear case, these models are of the form:
(4)
$${V}_{k}\left(w,x\right)={\Psi }\left({\sum }_{i}{w}_{i}{x}_{i}\right),$$
where w is a vector of weights and \(\varPsi\) is a transfer function that can either be linear, e.g., \(\varPsi \left(y\right)=y,\) or sigmoidal for logistic regression \(\varPsi \left(y\right)=1/\left(1+{e}^{-y}\right)\).
2.5 Dependent variables
For details on creating the higher-level features VC9, A0, and ST4, see Appendix A and (39–41). Validation of the cognitive properties of these features on healthy individuals appears in (39, 42), and usage in different contexts appears in (40, 41, 43).
We provide here a short description of the creation of the high-level features.
The 121 BAFs were created in an unsupervised way, namely that no labels of brain status were used. The higher-level features were created using linear and nonlinear machine-learning techniques from previously collected labeled datasets. These datasets included EEG data collected from participants undergoing different cognitive and emotional challenges. Specifically, VC9 was found with linear discriminant analysis technique (LDA) on the 121 BAFs to be the best separator between an auditory detection task (higher cognitive load) and an auditory classification task (lower cognitive load, 43); A0 was calculated to discriminate between resting state task with music and auditory detection task using PCA; and ST4 was calculated to discriminate between high-load and low-load conditions in auditory n-back task, using unsupervised PCA.
On separate datasets, these features were found to correlate with different tasks/levels/states. For example, VC9 correlated with cognitive load using the n-back task, auditory detection task, and interruptions paradigm (39–43). VC9 also correlated with cognitive load and individual performance of medical interns who were undergoing a surgery simulation task (40). A0 and ST4 activity correlated with cognitive decline scores (i.e., mini-mental score, 54) and separated between groups with different levels of cognitive impairment (i.e., mild, moderate ,and severe) (39, 43). ST4 correlated with individual performance for healthy participants undergoing the n-back task (39, 41) and also showed some usefulness in differentiating between cognitive load levels (39, 43). Neurosteer® provided the real-time activity of the BAFs and a dataset with a sample-per-second activity of the EEG bands (i.e., delta (0.5-4 Hz), theta (4–7 Hz), alpha (8–15 Hz), and beta (16–31 Hz)), and features (i.e., VC9, A0, and ST4).
2.6 Statistical analysis
Due to non-normal distribution of the data, we examined the differences (of median activities) between the two anesthetics using the Mann-Whitney test for continuous variables. For each comparison, d-prime was calculated to estimate effect size. Additionally, we compared the effect sizes of all dependent variables from smallest to highest to explore which of these effects was significantly larger than the others, using the following:
If d is the observed Cohen's d-value, then the sampling variance of d is approximately equal to:
$$v=\frac{1}{n1}+\frac{1}{n2}+\frac{{d}^{2}}{2\left(n1+n2\right)},$$
So, to test H0 : δ1 = δ2 (where δ1 and δ2 denote the true d-values of the two studies), we computed:
$$z=\frac{d1- d2}{\sqrt{v1+v2} },$$
which follows an approximately standard normal distribution under H0. So, if |z| ≥ 1.96, we rejected H0 at α = 0.05 (two-sided).
The data was analyzed using the These analyses were performed using Python Statsmodels (55) and SciPy (56).