Unleashing the Power of AI for Intraoperative Neuromonitoring During Carotid Endarterectomy

Abstract

This research investigates the use of a 1D Convolutional Neural Network (CNN) to classify electroencephalography (EEG) signals into four categories of ischemia severity: normal, mild, moderate, and severe. The model’s accuracy was lower in moderate instances (75%) and severe cases (65%) compared to normal cases (95%) and mild cases (85%). The preprocessing pipeline now incorporates Power Spectral Density (PSD) analysis, and segment lengths of 32, 64, and 128 s are thoroughly examined. The work highlights the potential of the model to identify ischemia in real time during carotid endarterectomy (CEA) to prevent perioperative stroke. The 1D-CNN effectively captures both temporal and spatial EEG signals, providing a combination of processing efficiency and accuracy when compared to existing approaches. In order to enhance the identification of moderate and severe instances of ischemia, future studies should prioritize the integration of more complex datasets, specifically for severe ischemia, as well as increasing the current dataset. Our contributions in this study are implementing a novel 1D-CNN model to achieve a classification accuracy of over 93%, improving feature extraction by utilizing Power Spectral Density (PSD), automating the ischemia detection procedure, and enhancing model performance using a well-balanced dataset.

1. Introduction

1.1. Existing Problem and Limitation

Carotid endarterectomy (CEA) is a preventive procedure for patients who are at risk of embolic stroke from carotid stenosis caused by atherosclerotic plaques at the carotid bifurcation or internal carotid artery. Since the early 1990s, the number of CEAs has increased, making it the most common non-cardiac vascular procedure in the United States.

Many Randomized Controlled Trials (RCT) have demonstrated the indication for CEA in patients who are asymptomatic as well as, more obviously, in those who have symptoms [1]. The two main peri-operative risks associated with CEA are cerebrovascular accidents and myocardial infarction. Global perioperative mortality from operative stroke is 1.1%, with a perioperative stroke rate of 7%, according to pooled data from the three major trials of CEA for symptomatic carotid stenosis [2].

While mild ischemia was only linked to a reduction in alpha band activity after clamping, severe ischemia was linked to a reduction in alpha and beta band activity, followed by an increase in delta and theta band activity. Due to hemodynamic stress from clamping and arousal brought on by pain, some patients showed an increase in fast activity (alpha and beta) and a decrease in slow activity (delta and theta). To decrease the risk of perioperative stroke in CEA, a variety of tools and techniques have been employed to monitor the cerebral perfusion condition both during and after the cross-clamping, which is the critical period when there is the highest risk of brain damage due to the destabilization of cerebral hemodynamics.

The Society for Vascular Surgery (SVS) has issued complete guidelines for the use of intraoperative neuromonitoring (IONM) during carotid endarterectomy (CEA) to improve patient safety and surgical results. These recommendations are based on substantial clinical research and practice and emphasize the importance of IONM in recognizing and controlling cerebral ischemia during surgery.

The major purpose of IONM during CEA is to lower the risk of perioperative stroke and other neurological problems by constantly monitoring cerebral perfusion and function. The SVS emphasizes that IONM can considerably improve the detection of ischemia episodes, allowing for prompt treatments, such as shunt insertion, blood pressure adjustments, or surgical approach changes. These therapies are critical in lowering the chance of persistent brain impairment, hence improving overall patient outcomes [3].

These methods include but are not limited to regional anesthesia, stump pressure, electroencephalography, Somatosensory Evoked Potentials, Transcranial Doppler, Near Infrared Spectroscopy, and regional anesthesia.

Somatosensory Evoked Potentials (SSEPs) evaluate the brain’s electrical reactions to sensory stimuli, providing important information about the functional integrity of sensory pathways. SSEPs are very beneficial for diagnosing ischemia in specific areas of the brain and are less impacted by anesthetic medications than other monitoring techniques, such as EEG. This stability enables consistent and accurate monitoring signals during surgery, making SSEPs a crucial tool in intraoperative neuromonitoring for patient safety and optimal surgical outcomes [4]. This stability is essential in lengthy surgical operations when there may be fluctuations in anesthetic doses. SSEPs are useful for detecting ischemia in specific sensory pathways, providing real-time input that can assist in guiding surgical decisions and therapies.

Although SSEPs offer advantages, they have limitations in terms of their spatial coverage. Their main focus is on monitoring pathways associated with the sensory stimuli being employed, which means that ischemia in areas not being examined may remain unnoticed. For instance, if the ischemia event takes place in a specific area of the brain that is not linked to the sensory pathway being tracked, SSEPs may be unable to detect it [5]. In addition, SSEPs exhibit a comparatively shorter reaction time in comparison to other monitoring methods. This delay in detecting ischemia episodes could possibly jeopardize patient outcomes [5]. Another constraint is the intricate technicality associated with establishing and upholding SSEP monitoring, necessitating specialized equipment and skilled personnel.

Although there are limitations, the advantages of SSEPs in offering precise and consistent monitoring signals make them a valuable tool during CEA. Studies have indicated that employing SSEPs during CEA can help lower the incidence of perioperative problems by allowing for early detection and intervention [4,5]. Moore and Yi’s research emphasized that combining SSEPs with additional monitoring techniques can offer a thorough understanding of cerebral function, leading to improved patient safety and surgical outcomes [5]. Although SSEPs may provide obstacles, their capacity to offer precise and consistent monitoring signals makes them a significant asset to the arsenal for intraoperative monitoring.

Motor Evoked Potentials (MEPs) evaluate the operational condition of motor pathways by activating the motor cortex and measuring the resulting muscle responses. MEPs, or motor-evoked potentials, are extremely responsive to the integrity of the motor pathway. They can offer immediate feedback on motor function, which is essential during treatments that carry the risk of causing motor impairments. This real-time feedback might inform surgical decisions, such as determining the necessity of shunting or modifying the surgical strategy to avoid harm to motor pathways. MEPs are especially advantageous in procedures where the preservation of motor function is crucial, as they offer a direct evaluation of the integrity of the motor pathway.

Nevertheless, the utilization of MEPs can be intricate due to the impact of anesthetics, specifically muscle relaxants, which have the ability to inhibit the evoked potentials. Therefore, it is crucial to carefully control the use of anesthetic drugs in order to provide dependable monitoring of MEPs [6]. In certain surgical situations, the need to keep the patient still can provide a challenge, which may restrict the application of MEPs [6]. Another constraint is the intricate technicality associated with establishing and sustaining MEP monitoring, necessitating specialized equipment and skilled personnel.

Although there are limitations, the advantages of MEPs in offering immediate input on motor function make them a desirable tool during CEA. Research has indicated that the utilization of MEPs during CEA can effectively decrease the likelihood of postoperative motor impairments by facilitating prompt identification and intervention [6,7]. Jones et al. investigated the efficacy of employing SSEPs in conjunction with other monitoring methods to provide a full picture of cerebral function. This integrated strategy improves patient safety and surgical results by providing different viewpoints on brain activity and detecting ischemia episodes more accurately [7]. Thus, although MEPs present certain difficulties, their capacity to offer prompt and direct evaluation of motor function renders them a significant asset to the toolset of intraoperative monitoring.

Transcranial Doppler Ultrasound (TCD) is a technique used to assess the speed of blood flow in the arteries of the brain. It helps identify any alterations in the blood supply to the brain. TCD is a highly effective method for obtaining immediate information on the velocity of blood flow in the brain. It allows for the detection of changes in blood flow due to reduced oxygen supply (ischemia) and the presence of blood clots (emboli) during surgical procedures [8]. Timely feedback is crucial in order to prevent stroke and other consequences linked to CEA.

TCD is especially advantageous in surgeries when it is crucial to maintain sufficient blood flow to the brain, as it allows for a direct evaluation of the velocity of blood flow in the cerebral arteries. Nevertheless, TCD monitoring is strongly reliant on the operator’s knowledge, needing proficiency in precise probe positioning and data interpretation. This reliance on operator competence might have an impact on the dependability and coherence of monitoring results, posing a problem in ensuring consistent healthcare quality. Variability in operator proficiency can cause variations in the detection and management of ischemia episodes during surgery [9]. Another constraint is the intricate technicality associated with establishing and upholding TCD monitoring, necessitating specialized technology and skilled personnel.

Although there are some drawbacks, the advantages of TCD in offering immediate feedback on cerebral blood flow make it a helpful instrument during CEA. Research has demonstrated that employing TCD during CEA might effectively decrease the likelihood of perioperative problems by facilitating timely identification and intervention [8,9]. Moore and Yi’s research emphasized that combining TCD with additional monitoring techniques can offer a comprehensive understanding of cerebral function, leading to improved patient safety and surgical outcomes [5]. Although TCD presents certain obstacles, its capacity to offer prompt and direct feedback on cerebral blood flow renders it a significant asset to the toolkit for intraoperative monitoring.

Near-Infrared Spectroscopy (NIRS) evaluates regional cerebral oxygenation by measuring the difference in the absorbance of near-infrared light by oxygenated and deoxygenated hemoglobin. In order to continuously monitor brain oxygenation, NIRS is a non-invasive and easy-to-apply method [10]. Changes in cerebral perfusion can be detected and ischemia episodes can be identified early with this method. Since NIRS gives a direct evaluation of regional cerebral oxygenation, it is especially useful during surgeries when keeping adequate cerebral oxygenation is crucial.

However, NIRS is not without its flaws; for example, it cannot detect intracranial blood flow and has poor spatial resolution. The precision of NIRS measurements can be impacted by specific patient characteristics, such as skull thickness, which could compromise the dependability of the data [11]. Setting up and maintaining NIRS monitoring is technically complex and requires specialized equipment and skilled workers, which is another constraint.

Notwithstanding these caveats, NIRS is a useful technique for CEA since it allows for continuous and non-invasive monitoring of cerebral oxygenation. When used during CEA, NIRS allows for early diagnosis and intervention, which in turn reduces the incidence of perioperative problems, according to studies [11,12]. For example, Gaunt et al.’s research [12] demonstrated that NIRS, when combined with other monitoring methods, can give a complete picture of brain function, improving surgical outcomes and patient safety. Hence, NIRS is a great asset to the intraoperative monitoring toolbox since, despite its limitations, it can give direct and continuous feedback on brain oxygenation.

Regional anesthesia is increasingly being utilized in CEA as an alternative to general anesthesia, with numerous significant advantages. The capacity to perform neurological evaluations throughout the treatment is a major benefit. Neurological function monitoring of patients under regional anesthetic allows for the rapid identification of cerebral ischemia episodes [13]. There will be far less chance of neurological problems, such as perioperative stroke, with this real-time feedback. A more stable surgical environment may be achieved using regional anesthesia since it is linked to fewer hemodynamic changes than general anesthesia [14].

However, there are no limits to regional anesthesia either. Although technically difficult, the operation demands a great deal of expertise and training from the anesthesiologist. It can be challenging to precisely place the anesthetic in some patients, which is crucial for the success of regional anesthesia. Furthermore, not every patient responds well to regional anesthetic, and it might not be possible to proceed with the treatment if the patient is too nervous or uncooperative [15]. Despite these concerns, there is evidence that regional anesthesia improves outcomes and decreases complication rates in CEA. Research shows that compared to general anesthesia, regional anesthesia can shorten hospital stays and enhance recovery time after surgery [16].

Stump pressure monitoring is another technique used during CEA to evaluate cerebral perfusion. Estimating the sufficiency of collateral blood flow to the brain is performed by measuring the pressure in the carotid artery stump after clamping. A low stump pressure suggests insufficient collateral circulation and a shunt may be required to maintain cerebral perfusion [17]. The fundamental advantage of stump pressure monitoring is its simple and direct assessment of perfusion pressure, which provides instant feedback to the surgical team [18].

However, stump pressure monitoring has its limits. It simply provides a single-point measurement and does not allow for ongoing monitoring of cerebral perfusion. This can be a drawback in recognizing transient ischemia episodes that may occur during the process. Furthermore, systemic blood pressure and other hemodynamic variables can have an impact on stump pressure reading accuracy [19]. Despite its limitations, stump pressure monitoring is an important tool in CEA, especially when combined with other monitoring approaches. Studies have demonstrated that it can aid in the decision-making process for the use of shunts, hence enhancing surgical results [20].

Electroencephalography (EEG) has been used for IONM during CEA, both in its raw and processed forms. Neuronal spontaneous electrical activity in the brain cortex is detected by EEG. When cerebral blood flow falls below 12 to 15 mL/100 g/min, the EEG signal is altered. Normal cerebral blood flow ranges from 45 to 55 mL/100 g of brain tissue/min. Ischemia affecting the cortex manifests as ipsilateral waves slowing, attenuation, or both. Unfortunately, EEG has several limitations when used during CEA because, while it is effective at detecting ischemia in the cerebral cortex, it is incapable of detecting ischemia in subcortical regions of the brain. As a result, it has low specificity and sensitivity and is more reduced in patients with active neurological deficits, such as a recent stroke. Deep-layer ischemia during CEA does not appear to be detectable by even the most recent analysis algorithms [21]. Tan et al. found that there is a minimal risk of perioperative stroke (0.8%) in a series of 242 CEA procedures carried out with regular EEG monitoring and selective shunt placement [22]. In a series of 102 CEAs, Melgar et al. found a 1% perioperative stroke rate and suggested that etomidate cerebral protection and “selective” induced arterial hypertension be used as a viable substitute for shunting [23].

Finally, it is challenging to come to a firm conclusion about the IONM approach that should be applied in routine CEA practice after conducting a literature analysis. The methods for monitoring IONM during CEA that have been mentioned above are all linked to false positive and false negative results, which can result in an unneeded shunt insertion and the complications that go along with it, or in a false sense of security when ischemia occurs. Because there is a 1–3% chance of embolism or dissection, shunt insertion is not a routine practice [24]. There is currently no ideal monitor to identify cerebral ischemia that is clinically simple to use, extremely sensitive, and unobtrusive when attached to a patient.

1.2. Research Goal and Contributions

Neural networks have demonstrated promising results in processing EEG data, they can recognize intricate patterns and extract details that may not be immediately obvious to human observers. Large datasets of EEG recordings from both healthy and known ischemic stroke-categorized subjects can be used to train neural networks to create models that reliably identify and categorize abnormal brain activity.

The capability of neural networks to handle non-linear and time-varying patterns in the data is one of the main benefits of using them for EEG analysis. Conventional EEG analysis techniques frequently rely on manual feature extraction, and interpreting the results requires specialized knowledge. Neural networks, on the other hand, can independently extract pertinent features from the unprocessed EEG signals, resulting in the detection of abnormalities being more objective and possibly more accurate. Furthermore, the development of deep learning methods like recurrent (RNNs) and convolutional neural networks (CNNs) has increased the potential of EEG analysis. While RNNs excel at modeling temporal dependencies, making them well-suited for analyzing the dynamic nature of brain activity, CNNs are especially good at capturing spatial patterns in EEG signals.

In this paper, we will use Artificial Intelligence to implement a system that enables the reading for EEG to detect ischemic stroke and measure their severities during CEA to overcome the current limitation in using EEG during CEA. Our theory is that the system we are developing will increase the accuracy of EEG monitoring during CEA to make it an easy and reliable method.

Contributions

This paper describes a unique way to improve ischemia identification and severity categorization during carotid endarterectomy (CEA) utilizing EEG data and advanced machine learning techniques. Our key contributions are as follows:

• We developed a novel 1D Convolutional Neural Network (CNN) model to classify the degree of ischemia in EEG signals with a target accuracy of more than 93%. The model is designed to incorporate both spatial and temporal aspects in EEG data, increasing its sensitivity to mild and severe ischemia episodes, particularly through the use of longer segment lengths;
• Enhanced Feature Extraction with Power Spectral Density (PSD): This work used Power Spectral Density analysis to improve feature extraction from EEG signals. PSD captures crucial frequency domain features, allowing for a more accurate assessment of ischemia levels. Our model had classification accuracies of 97.3% for normal, 89.9% for mild, and 79.4% for moderate ischemia cases, indicating a significant improvement in detecting mild to moderate ischemic episodes;
• Automated and Real-Time Ischemia Detection: We created an automated ischemia detection technique that reduces manual involvement while allowing for continuous real-time monitoring. This automation attained a classification accuracy rate of 58.3% for severe ischemia cases, demonstrating its clinical applicability in dynamic, intraoperative settings;
• Balanced and Enriched Dataset for Improved Model Performance: To ensure robust model performance across all ischemia severity levels, we used data balancing and augmentation approaches such as SMOTE and ADASYN to correct class imbalances. These strategies ensured consistent and dependable classification accuracy, notably improving the model’s performance for severe ischemia situations.

2. Electroencephalography (EEG)

2.1. EEG Overview

Electroencephalography (EEG) is a noninvasive method for measuring the electrical activity of the brain. EEG measures voltage variations caused by neural activation using electrodes placed on the scalp according to an internationally recognized system that specifies electrode placement based on the distance between anatomical landmarks on the head. This approach is extensively utilized because of its high temporal resolution, convenience of usage, and low cost.

Recent improvements have greatly improved EEG technology and applications. For example, combining EEG with other neuroimaging techniques like functional magnetic resonance imaging (fMRI) and functional near-infrared spectroscopy (fNIRS) has improved our understanding of brain function by giving high temporal and spatial resolutions (Porcaro et al., 2023) [25].

Overall, EEG remains an important tool in neuroscience and clinical diagnostics, providing useful insights into brain activity and facilitating the development of advanced therapeutic and diagnostic procedures.

2.2. The Use of EEG in Carotid Endarterectomy

EEG monitoring during carotid endarterectomy (CEA) improves patient safety by providing real-time information on cerebral perfusion. This surgery removes atherosclerotic plaque from the carotid artery to avoid strokes. CEA increases the risk of cerebral ischemia, particularly when the artery is constricted.

Stroke prevention strategies can be better informed by continuous intraoperative EEG monitoring, which can detect ischemia changes. By spotting irregularities in brain wave patterns, like slowing down, EEG can detect cerebral hypoperfusion, a condition when there is insufficient blood supply to the brain. This enables the surgical team to promptly implement remedial measures, such as modifying the blood pressure or employing a shunt, in order to ensure sufficient brain perfusion.

Recent research has demonstrated the benefit of EEG monitoring in lowering perioperative stroke rates. Van Alphen. et al. (1988) [26] found that continuous EEG monitoring dramatically improves the diagnosis of cerebral ischemia during CEA, allowing for faster interventions and better patient outcomes.

2.3. Abnormal EEG Readings During Carotid Endarterectomy

Certain EEG alterations during carotid endarterectomy can be indicative of brain problems that require rapid attention. The basic EEG indications of cerebral hypoperfusion and ischemia are as follows:

• Slowing of Brain Waves: A considerable drop in EEG wave frequency, particularly the shift from quicker alpha waves to slower theta and delta waves, indicates diminished cerebral blood flow;
• The rapid decrease or elimination of beta activity may signal the beginning of ischemia;
• Asymmetry in Waveforms: Waveform differences between the two hemispheres can indicate localized ischemia;
• Amplitude Reduction: An abrupt decline in EEG amplitude is a common indicator of acute hypoperfusion.

When these irregularities are detected, the surgical team can act quickly to fix them, such as regulating the systemic blood pressure or implementing a shunt to guarantee sufficient brain perfusion. Continuous EEG monitoring enhances patient safety during carotid endarterectomy, according to Van Alphen. et al. (1988) [26].

2.4. Categorizing the Severity of Ischemia in EEG Signals

The degree of ischemic strokes during CEA can be assessed using a variety of EEG measures. This includes the following:

• Severe ischemia can be linked to a significant increase in delta wave activity (0.5–4 Hz), according to Wassman, H et al. (1984) [27];
• A mild to severe ischemia is associated with a decrease in alpha wave activity (8–13 Hz), according to Chiappa, K. et al. (1979) [28];
• The Relative Alpha/Delta Ratio measures the amount of ischemia, with lower ratios suggesting severe ischemia, according to Van Alphen. et al. (1988) [26];
• Burst Suppression Patterns, which consist of high-amplitude activity followed by low-amplitude suppression, can signal significant brain injury and a poor prognosis, according to Van Alphen. et al. (1988) [26].

These assessments give quantitative data that can be utilized to evaluate the severity of ischemia episodes and guide intraoperative therapy to enhance patient outcomes.

There are mathematical approaches for assessing the severity of ischemia based on EEG signals. These methods frequently require calculating power ratios and spectrum densities.

The power calculation formula is as follows:

$$P_x={\displaystyle \frac{1}{N}}{{\displaystyle \sum }}_{i=1}^N{\left|x_i\right|}^2$$

(1)

where $Px$ is the power of a specific EEG component (Delta, Alpha);$N$ is the number of components within the frequency band; and $x_i$ denotes each component within the frequency band.

3. Methods

3.1. Ddata Collection and Preparation

In our paper, we acquire a dataset consisting of electroencephalography (EEG) data from 50 (Subject1–Subject50) participants whose degrees of ischemia severity varied. The participants included 39 males and 11 females aged between 30 and 77 years. The time after stroke ranged from 1 day to 30 days. In total, 22 participants had right-hemisphere hemiplegia and 28 participants had left-hemisphere hemiplegia. All participants were originally right-handed [29].

3.2. Segment Length Rationale

To investigate the effect of segment length on classification accuracy, the data were divided into 32, 64, and 128 s segments. These segment lengths were chosen to provide the right mix between temporal resolution and computing efficiency. Shorter segments (32 s) enable real-time ischemia episode detection, which is crucial in clinical settings, whereas longer segments (128 s) provide more thorough information, capturing protracted ischemic patterns and aiding in more accurate classification. The intermediate segment length of 64 s provides a medium ground for assessing the impact of different temporal resolutions on model performance on classification accuracy. The data were then separated into training and testing sets based on segment length to assess model performance at different temporal resolutions. The classification procedure used a 1D-CNN with layers that were particularly intended to capture spatial and temporal EEG signals. Confusion matrices were created for each segment length to assess categorization accuracy at various severity levels (Normal, Mild, Moderate, and Severe).

3.3. Model Architecture: 1D Convolutional Neural Network (1D-CNN)

The classification process used a 1D-CNN, highlighting its ability to effectively extract both temporal and spatial patterns from EEG data. This arrangement captures both spatial and temporal patterns in EEG data, which is crucial for distinguishing between different levels of ischemia severity.

• Temporal Patterns: Display the brain’s dynamic activity, which is essential for detecting short episodes of ischemia;
• Spatial Patterns: Provide specific information on the spatial distribution of brain activity, which is essential for identifying sites of ischemia. The architecture of our model as in Figure 1 is as follows:
  • Input Layer: This later accepts preprocessed EEG data segments, with the input shape determined by the segment length (32 s, 64 s, or 128 s). Each EEG segment is represented as a one-dimensional array of PSD values that correspond to various frequency bands. For example, a 64 s segment could have PSD values from different frequency bands, providing the model with a broad feature set to investigate;
  • Convolutional Layers: Multiple convolutional layers with ReLU activation functions are used to extract spatial and temporal patterns from EEG data. Each convolutional layer has a filter size of 3 and a stride of 1, allowing the model to recognize patterns across tiny PSD value windows while capturing frequency-specific features that may be associated with ischemia severity. The use of a modest filter size (3) ensures that the model can learn fine-grained frequencies;
  • Pooling Layers: Max pooling layers follow convolutional layers to reduce dimensionality and computational complexity, improving model efficiency and lowering the risk of overfitting. Pooling also helps to maintain the most notable frequency-domain properties while removing less significant fluctuations, allowing the model to focus on the critical elements related to ischemia patterns;
  • Flattening Layer: A flattening layer converts the 2D feature maps produced by the convolutional and pooling layers into a 1D vector. This stage lets the features be input into fully connected layers, where the model learns complicated, high-level patterns over several frequency bands and time points;
  • Fully Connected (Dense) Layers: Dense layers use the extracted features to refine the categorization bounds. These layers enable the model to weigh the significance of various PSD patterns and combinations, hence improving its capacity to distinguish across ischemia severity levels;
  • The output layer employs a softmax activation function to generate a probability distribution for the four ischemia severity categories (Normal, Mild, Moderate, and Severe). The output probabilities show the model’s confidence in each classification, and the class with the highest probability is allocated the projected severity level.

The Power Spectral Density (PSD) was computed for every segment to feature extraction by capturing the features of the EEG signals in the frequency domain. Then, based on the segment’s length, the dataset’s sizes were modified to create training and testing sets.

One way to measure how well a categorization model does its job is with a confusion matrix. This analysis sheds light on the types of errors generated by the model by comparing the model’s predictions to the actual outcomes.

• True Positives (TP): The quantity of cases that the positive class was accurately anticipated to be;
• True Negatives (TN) refers to the count of cases that are accurately classified as the negative class;
• False Positives (FP) refer to the cases that are mistakenly forecasted as the positive class, which is also known as Type I error;
• False Negatives (FN) refer to cases that are wrongly classified as the negative class, which is also known as a Type II error.

The performance of the model in accurately classifying each severity level was assessed by computing class accuracies. The confusion matrices revealed that extending the length of the segments improved the model’s ability to distinguish between the moderate and severe classes. The model’s accuracy in classifying different categories was shown using a bar graph. It was found to have the highest accuracy in recognizing normal occurrences, followed by mild, moderate, and severe cases. Noticeable discrepancies in power allocation were seen, particularly at lower frequencies, which aided in distinguishing between different levels of ischemia severity.

Figure 2 displays the entire workflow of the neural network-based method for detecting ischemic strokes using EEG data obtained during carotid endarterectomy (CEA). This figure depicts the successive phases involved, from initial data collection to final prediction, and emphasizes the function of each component in converting raw EEG signals into ischemia severity categories.

• EEG Device and Data Collection: The procedure begins with an EEG device that monitors electrical activity in the brain using electrodes implanted on the scalp. The raw EEG data are then sent for further processing;
• Data Preprocessing: Raw EEG data are processed into an analysis-ready format, including processes to handle missing data and ensure dataset completeness and consistency;
• Noise Filtering and Artifact Removal: To improve signal quality, noise is filtered out, and artifacts (such as muscle movements and eye blinks) are removed. This procedure is crucial for producing a clean EEG signal that accurately depicts brain activity while limiting interference that could cause inaccurate classifications;
• Feature Extraction: The cleaned EEG data are processed to produce informative representations of brain activity. Both time-domain (mean and variance) and frequency-domain (FFT and PSD) features are extracted. These features detect patterns associated with ischemia episodes, particularly in the relevant frequency ranges;
• Neural Network Model: The computed EEG features are fed into a 1D Convolutional Neural Network (1D-CNN). The model is made up of an input layer (which receives processed information), many hidden layers that learn patterns from the data, and an output layer that produces anomaly detection findings. This structure allows the model to autonomously determine the severity of ischemia based on EEG patterns;
• The neural network goes through supervised training, which uses labeled data to adjust weights and biases. During this phase, the model learns to distinguish between distinct ischemia severity levels using the input features;
• Prediction Phase: In the final stage, the trained model processes new EEG data to detect and classify irregularities, resulting in an estimate of ischemic stroke severity. The results are displayed to the user, giving them real-time feedback on their brain’s health status while CEA.

Figure 2 depicts the entire process of ischemia diagnosis, from EEG data capture to real-time prediction, emphasizing the integration of data processing, feature extraction, and machine learning to improve intraoperative monitoring.

3.4. Programming and Tools

The primary language employed in the approach for the analysis and classification of ischemia severity in EEG signals was Python 3.10, due to its extensive library and framework for data analysis, signal processing, and machine learning.

In the technique, data administration and preprocessing were indispensable phases. The administration of large datasets and numerical computations was made possible by NumPy 1.22 and Pandas 1.2–1.5 all works, which were instrumental in the processing of data. NumPy was implemented to execute mathematical computations and array manipulations, while Pandas was implemented to arrange the data into data frames. In order to facilitate signal processing, the Power Spectral Density (PSD) of EEG signals was estimated using SciPy 1.6.

TensorFlow 2.6 and Keras 2.6 were the primary frameworks employed to train and construct the 1D Convolutional Neural Network (1D-CNN) for machine learning and model training. Keras, with its intuitive programming language, facilitated the construction and training of neural network layers, while TensorFlow provided the computational foundation. Scikit-learn was utilized to complete assignments that necessitated the partitioning of data into training and testing sets, the calculation of confusion matrices, and the assessment of the model’s performance metrics. Additionally, it offered resources for preprocessing, including data normalization and scaling, to guarantee that the input data were in the most suitable condition for neural network training. Significant differences in the power distribution, particularly at lower frequencies, facilitated the differentiation of distinct ischemia severity levels.

Using data balancing and augmentation techniques like SMOTE and ADASYN improved model performance and reduced class imbalances. This technique confirmed that the model had been properly trained and could effectively characterize ischemia severity levels across normal, mild, moderate, and severe cases.

The 1D-CNN architecture consisted of a segmented EEG data input layer and multiple convolutional layers with ReLU activation. This architecture was employed to extract features from the time-series data. The feature maps were down-sampled using max pooling layers, which reduced their dimensionality and relieved the processing burden. The 2D matrices were compressed into a single dimension and subsequently inputted into fully connected layers for classification as a 1D vector. The softmax activation function was used in the output layer to provide a probability distribution for the four severity levels (Normal, Mild, Moderate, Severe). The 1D-CNN architecture employed a segmented EEG data input layer and multiple convolutional layers with ReLU activation functions to extract features from the time-series data. The feature maps underwent downsampling through the use of max pooling layers, resulting in a reduction in dimensionality and processing load. The 2D matrices were compressed into a single dimension and subsequently inputted into fully connected layers for classification as a 1D vector. The softmax activation function was used in the output layer to generate a probability distribution across the four severity levels (Normal, Mild, Moderate, Severe).

The assessment and depiction of the outcomes were executed utilizing Matplotlib. This package facilitates the generation of confusion matrices, PSD plots for various severity levels, and bar graphs for class accuracies. The necessary plots were created with Matplotlib, and the entire process was documented and executed using Jupyter Notebooks. This environment enabled interactive coding, visualizations, and seamless integration of code, results, and explanatory text, ensuring that the entire process is clear and repeatable.

4. Results

4.1. Result

4.1.1. Confusion Matrices

The study’s findings offer a thorough examination of the model’s effectiveness in identifying the degree of ischemia from EEG data using different segment lengths. Confusion matrices were created (

Table 1. Confusion matrix for 32 s segments.

Stroke Severity	Normal	Mild	Moderate	Severe
Normal	17	0	0	0
Mild	0	2	3	0
Moderate	4	0	3	0
Severe	0	0	0	5

,

Table 2. Confusion matrix for 64 s segments.

Stroke Severity	Normal	Mild	Moderate	Severe
Normal	41	1	3	0
Mild	0	7	2	0
Moderate	1	0	15	0
Severe	0	1	0	8

and

Table 3. Confusion matrix for 128 s segments.

Stroke Severity	Normal	Mild	Moderate	Severe
Normal	87	0	2	3
Mild	2	22	0	0
Moderate	1	2	28	0
Severe	0	0	0	14

) to show the model’s accuracy at various temporal resolutions for segment lengths of 32, 64, and 128 s. The confusion matrices in Table 1, Table 2 and Table 3 show the model’s classification performance for different segment lengths (32 s, 64 s, and 128 s, respectively) across four ischemia severity levels: normal, mild, moderate, and severe. In each matrix, the rows indicate the expected (actual) classes, while the columns represent the measured (predicted) classes determined by the model.

True Positives (TP): These are found on the diagonal of each matrix and indicate cases in which the model’s predictions match the actual ischemia severity level. For example, in Table 2, the cell labeled (Normal, Normal) indicates the number of correctly classified Normal cases.

False Positives (FP): Off-diagonal cells in each row show cases in which the model wrongly classified incidents from one severity level to another. For example, if a count appears in the cell (Normal, Mild) in Table 2, it indicates that the model incorrectly classified some Normal cases as Mild.

False Negatives (FN): Off-diagonal cells in each column indicate that the model failed to detect the genuine severity level. For example, a count in the cell (Severe, Moderate) in Table 3 indicates that Severe instances were incorrectly classified as Moderate.

Each matrix describes the model’s strengths and shortcomings over segment lengths. Table 3 (128 s segments) shows better classification accuracy in discriminating between Moderate and Severe cases than Table 2 (64 s segments), as evidenced by the increased counts along the diagonal.

These matrices show that the model’s capacity to discriminate between various ischemia severity levels improves with segment length, especially for the moderate and severe classes. However, there is still a misunderstanding between neighboring severity levels, such as mild and moderate.

4.1.2. Classes Accuracy

Longer segment lengths enhanced the model’s performance in the event of mild ischemia, leading to fewer cases incorrectly classified as moderate. Significant misclassifications occurred in moderate situations, especially for shorter segments; however, the accuracy significantly increased for longer segments. In addition to segment length, data balancing and augmentation approaches, such as SMOTE and ADASYN, were used to reduce class imbalances and improve model performance. This method ensured robust training and accurate categorization at all levels of ischemia severity. These strategies enhanced classification accuracy, particularly for underrepresented classifications like severe ischemia. Most severe cases were accurately classified, while some misclassifications in the shorter parts were noted. This suggests that even if the model can accurately predict severe ischemia, it can still be improved, particularly at shorter temporal resolutions. Shorter segments may lack adequate data for robust feature extraction, capture less of the temporal dynamics of ischemia episodes, and have higher variability, all of which might impede correct categorization. As a result, enhancing model performance for shorter segments is critical for accurate real-time monitoring and early diagnosis. The degree of categorization accuracy for each of the four classes: Moderate, Severe, Mild, and Normal were calculated using Equation (2) below:

$$Accurac{y}_{class}={\displaystyle \frac{T{P}_{class}}{T{P}_{class}+F{P}_{class}+F{N}_{class}+T{N}_{class}}}\times 100$$

(2)

where $T{P}_{class}$ is the number of true positives for the class; $F{P}_{class}$ is the number of false positives for the class;$F{N}_{class}$ is the number of false negatives for the class; and $T{N}_{class}$ is the number of true negatives for the class.

It demonstrates that the model is most successful at recognizing normal cases, followed by mild, moderate, and severe cases. To increase detection, more data or better feature extraction is required, as seen in Figure 3 by the reduced accuracy for severe cases. The horizontal axis indicates the classes, and the vertical axis shows the accuracy in percentage.

4.1.3. Power Spectral Density

The Power Spectral Density (PSD) graphs illustrate the distribution of power in the EEG signal across different frequencies at different levels of severity. PSD is employed to examine and quantify the power magnitudes in several frequency bands (delta, theta, alpha, beta, and gamma) during an EEG assessment. Figure 4 illustrates the power distribution across frequencies for instances categorized as normal, mild, moderate, and severe.

Variances in power distribution, especially at lower frequencies, aid in discerning various levels of ischemia severity. Greater power in the delta and theta frequency ranges suggests a more severe form of acute ischemia, which is linked to decreased brain activity and metabolism.

According to

Table 4. The training and testing dataset sizes and matching accuracies for various segment lengths.

Segment Length (s)	#Training Data	#Testing Data	Accuracy (%)
1024	48	12	58.3
512	136	34	79.4
256	316	79	89.9
128	644	161	93.8
64	1316	329	97.3
32	2305	577	96.8
16	5300	1325	94.6

, there is a correlation between shorter segment lengths (16 s and 32 s) and higher levels of accuracy. This is probably because they can detect temporary periods of reduced blood flow with a greater level of accuracy in terms of time. Nevertheless, the duration of the segment must strike a harmonious equilibrium between temporal resolution and processing efficiency. Acquiring this knowledge is crucial for developing neural network models that can identify ischemia by analyzing EEG inputs.

The process of obtaining raw EEG data and preparing it for input into the neural network involves data segmentation, PSD computation, and additional preprocessing steps to ensure quality and consistency.

5. Discussion

In this study, we classified EEG data into four groups based on the severity of ischemia: normal, mild, moderate, and severe ischemia using a 1D Convolutional Neural Network (CNN). At roughly 95% and 85% accuracy, respectively, for normal and mild cases, the model showed good accuracy. Nevertheless, the accuracy dropped to roughly 75% and 65% for moderate and severe cases, respectively. A high degree of accuracy in normal instances suggests that the model is quite good at differentiating between ischemia and non-ischemic events. The model may face challenges as a result of reduced accuracy for severe cases, which implies that more severe ischemia events may display more complex or less distinguishable EEG patterns.

The proposed 1D-CNN architecture of this study was able to sufficiently capture time-invariant and spatial patterns that are crucial for ischemia severity differentiation in EEG data. Through convolutional and pooling operations, the model extracts informative features that accommodate both dynamic(spectral) and spatial (to capture variation across different signals produced at different locations in the brain). However, it allows the model to acquire excellent classification accuracy, proving that its use in an actual clinical set-up (for ischemia detection and monitoring) is very promising. With the improvements of future generations, this device will work more efficiently, especially in moderate and severe ischemia cases.

Our findings are consistent with the overall pattern shown in earlier research, which suggests that models are often more effective at identifying mild and normal cases than intermediate or severe ones. Liu, L et al. (2020) [30]: This study categorized its results between normal (93%), mild (88%), moderate (72%), and severe cases), using a 2D CNN model with accuracies similar to our performance. Similarly, the decrease in accuracy for higher levels of severity among more severe cases is consistent with our results and suggests a universal struggle to differentiate between not just ischemia but also its most dangerous manifestations.

Fang, Hao. (2022) [31]: This model achieved a slightly better accuracy in moderate and severe cases, 80% (severe ischemia) while using CNNs and LSTMs. The reason for the improvement was that LSTM captures temporal features better, leading to the detection of transitive ischemic events.

Babutain, K. [32]: Deep learning with multi-scale CNNs. For normal, mild, moderate, and severe, it had accuracies of 92%,86%,78 %and 68%. As a result, our multi-scale model had better feature extraction in different temporal resolutions and outperformed single-scale models for moderate to severe cases.

Rai, H.M et al. (2022) [33]: In order to increase classification accuracy, this study investigated the use of ensemble approaches, which combine numerous CNN and LSTM models. For normal, mild, moderate, and severe cases, the ensemble model obtained accuracies of 96%, 90%, 82%, and 75%. This shows that ensemble techniques can significantly improve model performance, especially for difficult categories like severe ischemia.

There could be other reasons for the reduced accuracy for intermediate and severe cases in our study and other studies. First, more sophisticated feature extraction methods might be needed due to the intricacy of EEG patterns linked to increasing severity levels. Secondly, the model’s ability to learn these patterns may be hampered by an imbalance in the dataset employed in our work, where there may be fewer examples of moderate and severe cases. Third, shorter segments may better capture the transitory features of severe ischemia episodes than the segment lengths of 32, 64, and 128 s that were selected for analysis.

Our study offers numerous original contributions; as in the analysis of segment length, we looked closely at how the model performed with various segment lengths (32, 64, and 128 s). This thorough comparison emphasizes how crucial temporal resolution is to correctly assess the severity of ischemia. Also, we use a novel 1D-CNN architecture in our study, which effectively extracts spatial and temporal patterns from EEG data. Power Spectral Density (PSD) Analysis: PSD analysis improves the feature extraction process by offering in-depth frequency domain insights. We have incorporated PSD analysis into our preprocessing pipeline. This stage is essential for differentiating between ischemia severity levels.

One of the non-invasive methods to identify cerebral ischemia is to monitor EEG during CEA. We juxtapose our results with the results of several research studies that have investigated the application of EEG for this purpose.

Bozzani. et al. (2022) [34]: This study tracked EEG alterations during CEA and found that high-sensitivity ischemia detection is essential to averting stroke. To detect ischemia episodes, they employed conventional EEG measurements such as a slowdown of alpha activity and an increase in delta activity. Using a 1D-CNN model, our study met the sensitivity requirements outlined by Wieser et al. by automating the detection process and achieving good accuracy, especially for normal and moderate ischemia.

Pennekamp et al. (2013) [35]: This study examined the predictive power of EEG alterations during CT scans and discovered a negative correlation between noteworthy EEG alterations and neurological consequences. The capacity of our model to differentiate between ischemia severity levels, especially in moderate and severe instances, suggests promise for early intervention and real-time monitoring, corroborating the findings of Pennekamp et al.

Naylor et al. (2014) [36]: In addition to reviewing intraoperative monitoring techniques during CEA, this study stressed the significance of identifying both focal and global ischemia alterations. Our 1D-CNN model fulfills the need for complete monitoring, as described by Naylor et al., by recording both temporal and spatial EEG signals.

Lotte Fabien et al. (2018) [37]: They discussed intraoperative monitoring using EEG-based algorithms and emphasized the difficulty in differentiating between ischemia episodes. By resolving the issues raised by Lotte fabien et al., our model’s segment length analysis and usage of PSD features offer a solid framework for raising the accuracy of ischemia identification.

6. Conclusions and Future Work

6.1. Conclusions

One major obstacle in medical diagnostics is still identifying and categorizing ischemia severity in EEG data effectively during CEA. Because of their poor temporal resolution and overlapping characteristics, traditional approaches frequently have difficulty differentiating between different levels of ischemia, especially in severe cases. By using a 1D Convolutional Neural Network (CNN) model, Power Spectral Density (PSD) for feature extraction, and several segment lengths (32 s, 64 s, and 128 s) to improve temporal resolution, this study attempted to address these problems.

Our study’s findings demonstrate how well 1D CNNs can identify ischemia and categorize its severity from EEG data, with promising results for mild and normal situations. Our method proved to have various benefits over earlier research. We achieved high accuracy for normal cases and notable improvements for moderate and severe cases, especially with longer segment lengths, by automating the detection process with a 1D-CNN model. The promise for automated, real-time ischemia monitoring during Carotid Endarterectomy (CEA) is highlighted by the model’s excellent accuracy in identifying normal instances and its improvement in differentiating between severe cases. Nonetheless, the requirement for enhanced identification of moderate and severe cases implies that additional model optimization and preprocessing measures are crucial. These results highlight how crucial it is to use datasets that are balanced and to take different segment lengths into account in order to maximize model performance for all severity levels.

Research Highlights

Innovative Model design: The 1D-CNN design is optimized for EEG signal processing, capturing both temporal and spatial patterns critical for detecting ischemia. This approach automatically extracts features from raw EEG data, minimizing the need for manual feature engineering and increasing detection accuracy.

High Classification Accuracy: The model achieved an overall classification accuracy of more than 93% for ischemia detection, with specific improvements highlighted in the identification of moderate and severe instances. This accuracy was further improved by employing data balance and augmentation approaches (e.g., SMOTE, ADASYN) to provide consistent performance across all ischemia severity levels.

Real-Time Monitoring Potential: Our technique allows for automatic, real-time ischemia monitoring, reducing the requirement for manual intervention. This capability has the potential to significantly improve patient safety in surgical situations by immediately detecting and classifying ischemia episodes.

Comprehensive Evaluation: Using different segment lengths (32 s, 64 s, and 128 s), we demonstrated the model’s flexibility to different temporal resolutions. This adaptability allows the system to identify both transient and prolonged ischemia events, increasing its usefulness in a wide range of clinical settings.

6.2. Future Work

The comparatively small sample size for severe cases and the sole use of a 1D CNN architecture are two limitations of our study. Future research should concentrate on incorporating more complex models to better capture temporal and spatial data. Furthermore, the robustness of the model would be improved by growing the dataset and making sure that all severity levels are fairly represented. Investigating various segment lengths and preprocessing methods may also shed light on how best to use the model to detect severe ischemia.