Intraoperative nerve stimulation monitoring system, method, equipment and medium

Abstract

The invention discloses an intraoperative nerve stimulation monitoring system, method and device and a medium. The system comprises a signal acquisition module which is used for acquiring a target somatosensory evoked potential signal and target physiological parameter information in an operation process; the signal processing module is used for carrying out noise reduction and standardization processing on the target somatosensory evoked potential signal; the signal anomaly detection module is used for judging whether the target somatosensory evoked potential signal is abnormal or not according to the waveform interpretation model and the target physiological parameter information; wherein the waveform interpretation model is constructed based on a convolutional neural network, a long-short-term memory network and an attention mechanism, and is trained by adopting a double-branch knowledge distillation structure; and the alarm module is used for triggering real-time alarm if the target somatosensory evoked potential signal is abnormal. According to the technical scheme, the waveform interpretation model can be combined with the physiological parameter information to carry out efficient and accurate signal anomaly detection and alarm on the SEP signal in the operation, so that the safety and the accuracy of the operation are guaranteed.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and in particular to an intraoperative nerve stimulation monitoring system, method, device and medium. BACKGROUND

[0002] As an important auxiliary technology of modern surgery, intraoperative nerve stimulation monitoring has been widely used in neurosurgery, spine surgery, ear-nose-throat surgery and other fields. Its core goal is to monitor and evaluate the functional state of the nervous system in real time, identify potential nerve damage risks, and ensure the safety and accuracy of surgery.

[0003] Somatosensory Evoked Potential (SEP) is a transient neural response signal induced by external electrical stimulation, reflecting the conduction function between peripheral nerves, spinal cord and cerebral cortex. The existing intraoperative neurophysiological monitoring technology for SEP mainly relies on manual identification of waveform abnormalities, which has problems such as poor real-time performance, high false alarm rate and strong subjectivity; and is easily affected by non-neural damage factors, such as abnormal misjudgment caused by changes in anesthesia depth, with a high false positive rate. SUMMARY

[0004] The present application provides an intraoperative nerve stimulation monitoring system, method, device and medium, which can use a waveform interpretation model combined with physiological parameter information to perform efficient and accurate signal anomaly detection and alarm on intraoperative SEP signals, thereby ensuring the safety and accuracy of surgery.

[0005] According to an aspect of the present application, an intraoperative nerve stimulation monitoring system is provided, the system comprising a signal acquisition module, a signal processing module, a signal anomaly detection module and an alarm module; wherein:

[0006] The signal acquisition module is configured to acquire target somatosensory evoked potential signals and target physiological parameter information during surgery;

[0007] The signal processing module is configured to perform noise reduction and standardization processing on the target somatosensory evoked potential signals;

[0008] The signal anomaly detection module is configured to determine whether the target somatosensory evoked potential signals are abnormal according to a waveform interpretation model and the target physiological parameter information; wherein the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained using a double-branch knowledge distillation structure;

[0009] The alarm module is configured to trigger real-time alarm if the target somatosensory evoked potential signals are abnormal.

[0010] According to another aspect of the present application, there is provided an intraoperative nerve stimulation monitoring method, comprising:

[0011] acquiring target somatosensory evoked potential signals and target physiological parameter information during the surgery through a signal acquisition module;

[0012] performing noise reduction and standardization processing on the target somatosensory evoked potential signals through a signal processing module;

[0013] judging whether the target somatosensory evoked potential signals are abnormal according to a waveform interpretation model and the target physiological parameter information through a signal anomaly detection module; wherein the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained using a double-branch knowledge distillation structure;

[0014] if the target somatosensory evoked potential signals are abnormal, triggering real-time alarm through an alarm module.

[0015] According to another aspect of the present application, there is provided an electronic device, comprising:

[0016] at least one processor; and,

[0017] a memory in communication connection with the at least one processor; wherein,

[0018] the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to execute the intraoperative nerve stimulation monitoring method according to any one of the embodiments of the present application.

[0019] According to another aspect of the present application, there is provided a computer readable storage medium storing computer instructions for enabling a processor to implement the intraoperative nerve stimulation monitoring method according to any one of the embodiments of the present application when executed by the processor.

[0020] The technical scheme of the embodiment of the present application provides an intraoperative nerve stimulation monitoring system, which comprises a signal acquisition module, a signal processing module, a signal anomaly detection module and an alarm module; wherein: the signal acquisition module is used for acquiring target somatosensory evoked potential signal and target physiological parameter information in a surgical process; the signal processing module is used for carrying out noise reduction and standardization processing on the target somatosensory evoked potential signal; the signal anomaly detection module is used for judging whether the target somatosensory evoked potential signal is abnormal according to a waveform interpretation model and the target physiological parameter information; wherein, the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained by using a double-branch knowledge distillation structure; the alarm module is used for triggering real-time alarm if the target somatosensory evoked potential signal is abnormal. The technical scheme can use the waveform interpretation model combined with the physiological parameter information to carry out efficient and accurate signal anomaly detection and alarm on the intraoperative SEP signal, so as to guarantee the safety and accuracy of the operation.

[0021] It should be understood that the content described in this part is not intended to identify key or important features of the embodiments of the present application, nor is it used to limit the scope of the present application. Other features of the present application will become apparent from the following description. BRIEF DESCRIPTION OF DRAWINGS

[0022] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can also be obtained by those skilled in the art without creative labor.

[0023] Figure 1 is a structural schematic diagram of an intraoperative nerve stimulation monitoring system provided by the embodiment of the present application;

[0024] Figure 2 is a training process schematic diagram of a waveform interpretation model provided by the embodiment of the present application;

[0025] Figure 3 is a process schematic diagram of intraoperative nerve stimulation monitoring provided by the embodiment of the present application;

[0026] Figure 4 is a schematic diagram of self-supervised model training provided by the embodiment of the present application;

[0027] Figure 5 is a structural schematic diagram of another intraoperative nerve stimulation monitoring system provided by the embodiment of the present application;

[0028] Figure 6 is a flowchart of an intraoperative nerve stimulation monitoring method provided by the embodiment of the present application;

[0029] Figure 7 is a structural schematic diagram of an electronic device for implementing an intraoperative nerve stimulation monitoring method according to an embodiment of the present application. DETAILED DESCRIPTION

[0030] In order to make the personnel in the art better understand the present application, the technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work should belong to the scope of protection of the present application.

[0031] It should be noted that the terms "first", "second", "target" and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily have to be used to describe a specific order or sequence. It should be understood that the data thus used can be interchanged under appropriate circumstances, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, for example, a process, method, system, product or device including a series of steps or units does not have to be limited to those steps or units clearly listed, but can include other steps or units not clearly listed or inherent to these processes, methods, products or devices.

[0032] Embodiment one

[0033] Figure 1 A structural schematic diagram of an intraoperative nerve stimulation monitoring system according to embodiment one of the present application is provided, and the present embodiment can be applicable to real-time monitoring and alarm of intraoperative SEP signals. The intraoperative nerve stimulation monitoring system can be implemented in the form of hardware and / or software, and the intraoperative nerve stimulation monitoring system can be configured in an electronic device with data processing capability.

[0034] As Figure 1As shown, the system comprises a signal acquisition module, a signal processing module, a signal anomaly detection module and an alarm module; wherein: the signal acquisition module is used to acquire the target somatosensory evoked potential signal and the target physiological parameter information in the surgical process; the signal processing module is used to perform noise reduction and standardization processing on the target somatosensory evoked potential signal; the signal anomaly detection module is used to determine whether the target somatosensory evoked potential signal is abnormal according to the waveform interpretation model and the target physiological parameter information; wherein, the waveform interpretation model is constructed based on convolutional neural network, long short-term memory network and attention mechanism, and trained using a double-branch knowledge distillation structure; the alarm module is used to trigger real-time alarm if the target somatosensory evoked potential signal is abnormal.

[0035] Wherein, the target somatosensory evoked potential signal can refer to the intraoperative real-time SEP signal that needs to be detected for abnormalities. It should be noted that, unlike the spontaneous periodic electrical signals such as electrocardiogram and electroencephalogram, SEP signal is a transient neural response signal induced by external electrical stimulation, which has the following characteristics: (1) event locking: each signal corresponds to an external stimulation event and must be accurately aligned; (2) low signal-to-noise ratio: the signal amplitude is only microvolt level, which is easily disturbed by electromyogram and stimulation artifacts; (3) special abnormal characteristics: mainly manifested as wave peak latency delay and amplitude drop, rather than rhythm change. The target physiological parameter information can refer to the physiological parameter information of the object to which the target somatosensory evoked potential signal belongs. For example, the physiological parameter information can include heart rate, body temperature, non-invasive systolic pressure, blood oxygen saturation and arterial diastolic pressure, etc.

[0036] In this embodiment, the target somatosensory evoked potential signal and the target physiological parameter information of the target object (such as a patient) can be obtained from the intraoperative monitoring device through the signal acquisition module, and then the target somatosensory evoked potential signal is preprocessed through the signal processing module. The artifacts (such as metal artifacts and motion artifacts) can be removed by band-pass filtering and noise reduction. Because the amplitude and distribution of signal data often have large differences under different conditions of different objects, data standardization processing is also needed to adjust the mean value of all data to 0 and the variance to 1, so that the data distribution is unified. For details, see formula . Wherein, is the data before standardization, is the data after standardization, is the mean value, is the standard deviation.

[0037] In this embodiment, a waveform interpretation model needs to be obtained in advance through model training, which can be used for abnormal detection of intraoperative SEP signals in combination with physiological parameter information. The waveform interpretation model is constructed based on a convolutional neural network (CNN), a long short-term memory network (LSTM), and an attention mechanism. In addition, in order to improve the robustness and generalization ability of the waveform interpretation model in a noisy environment, a double-branch knowledge distillation structure is used for model training, thereby improving the model performance without significantly increasing the inference overhead. It should be noted that the waveform interpretation model can be pre-deployed in the signal anomaly detection module, or can be deployed outside the signal anomaly detection module and obtained by calling when needed.

[0038] In this embodiment, optionally, the system further includes a model training module, which is configured to: construct a teacher branch based on a spatiotemporal attention denoising network and an attention mechanism, and construct a student branch based on a convolutional neural network, a long short-term memory network, and an attention mechanism; wherein the spatiotemporal attention denoising network is constructed based on multi-scale convolution and temporal attention; supervise training of the teacher branch and the student branch according to candidate physiological parameter information, candidate somatosensory evoked potential signals, and corresponding label information; in the process of supervising training of the teacher branch and the student branch, migrate the representation knowledge and the discrimination knowledge of the teacher branch to the student branch; determine a target loss of the student branch based on the supervising training process and the knowledge migration process, and determine whether the target loss meets a training completion condition; if yes, end the model training process, and determine the student branch at the end of the training as the waveform interpretation model.

[0039] Figure 2 A training process diagram of a waveform interpretation model provided in this embodiment of the present application is shown in FIG. 1. In the diagram, ce loss is a cross-entropy loss, kl loss is a relative entropy loss, and feature loss is used to represent the feature difference between the CNN_LSTM network output and the spatiotemporal attention denoising network output. Specifically, as shown in FIG. 1, the model training process includes the following steps: Figure 2As shown, firstly, a spatio-temporal attention denoising network is constructed based on multi-scale convolution and temporal attention, and a teacher branch is constructed according to the spatio-temporal attention denoising network and an attention mechanism, while a student branch is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism. Then, the teacher branch and the student branch are supervised and trained according to candidate physiological parameter information and candidate SEP signals with labeled information, and the representation knowledge and the discriminative knowledge of the teacher branch are transferred to the student branch in the process of supervised training. The candidate physiological parameter information and the candidate SEP signals can be physiological parameter information and SEP signals participating in model training, respectively, and both of them are from the same object. The labeled information is obtained by manual annotation, and can be used to indicate whether the candidate SEP signal is a normal signal or an abnormal signal. The representation knowledge can be used to reflect the intermediate features of the branch, and the discriminative knowledge can be used to reflect the classification features (i.e. output features) of the branch.

[0040] In the process of supervised training and knowledge transfer, distillation training is combined with the anomaly detection task to determine the target loss of the student branch, and the student branch is supervised and trained based on the target loss. At the end of each training round, it is necessary to determine whether the target loss meets the training completion condition. The training completion condition can be set as the target loss being less than a preset loss value or reaching a preset training number. If the target loss meets the training completion condition, the model training process is ended, and the student branch at the end of training is determined as the waveform interpretation model; if the target loss does not meet the training completion condition, the next round of training is continued.

[0041] In this embodiment, optionally, the model training module is further configured to: perform denoising and standardization processing on the candidate SEP signal to obtain a training signal; extract features of the training signal through the convolutional neural network and the long short-term memory network in the student branch to obtain first feature information; encode the candidate physiological parameter information into a feature vector to obtain second feature information; fuse the first feature information and the second feature information through the attention mechanism in the student branch to obtain first fusion feature information corresponding to the student branch; and determine a first classification result corresponding to the student branch according to the first fusion feature information.

[0042] Specifically, in order to ensure that the training data is accurate and effective, the candidate SEP signal needs to be preprocessed first, the pseudo signal is removed by band-pass filtering and denoising, and the formula The data standardization processing is performed to obtain the training signal. Then, the training signal is input into the student branch, and the convolutional neural network and the long short-term memory network in the student branch are used to extract the features of the training signal to obtain the first feature information. The CNN is used to extract the local waveform features, especially the key features related to the waveform form, such as the wave peak, wave trough, rising edge and falling edge. Through the multi-layer convolution operation and the nonlinear activation function, the change mode of the local region is captured, including the amplitude change, the local signal sharpness, etc. The features extracted by the convolution layer retain the local spatial information of the signal, and the dimensionality is reduced through the pooling layer, which effectively reduces the noise interference and the computational complexity. Through the multi-layer convolution operation, the fine-grained understanding of the SEP signal waveform can be realized, and the amplitudes, intervals of the wave peaks and wave troughs, and the occurrence of abnormal spikes can be accurately identified. It should be noted that the local receptive field characteristic of the CNN makes it highly sensitive to local details in the signal, providing high-quality input features for the subsequent network. By stacking multiple convolution layers, the model can gradually capture higher-level features, such as the trend of changes in specific frequency components or the features of abnormal waveforms. In addition, the CNN not only performs well in extracting local features, but also greatly reduces the computational overhead of the model through shared convolution kernel weights, making it suitable for real-time intraoperative monitoring scenarios. At the same time, the long short-term memory network in the student branch captures the temporal dependence relationship, such as the dynamic characteristics of the SEP signal in the time dimension, especially the trend of the signal change over time. The memory gating mechanism of the LSTM enables it to simultaneously learn long-term dependencies and short-term change features, capturing dynamic information such as latency delay and waveform frequency change in electrophysiological signals.

[0043] Since the SEP signal will be disturbed by the changes in physiological indicators, after obtaining the signal feature encoding (i.e., the first feature information), the candidate physiological parameter information needs to be encoded into a feature vector as the second feature information, and the second feature information is input into the student branch. The attention mechanism in the student branch is used to fuse the first feature information and the second feature information to form a comprehensive feature expression as the first fused feature information. It should be noted that the changes in physiological parameters have no correlation with nerve damage, but may cause abnormal changes in neural signals, so the physiological parameters are introduced as key features for the model to reduce false positives.

[0044] where the Attention mechanism focuses on the relationship between physiological parameters and signal features, assigning weights to the features extracted by CNN and LSTM, and focusing on key areas in the signal, such as the vertical distance between wave peaks and troughs, time latency, and other important features. By calculating feature weights, the model dynamically focuses on features that contribute most to the prediction task and weakens features that are irrelevant to the task, thereby improving the model's ability to identify abnormal signals. The Attention mechanism can adaptively select features and combine clinical key information (such as latency or amplitude threshold) to improve the accuracy of model prediction. The specific calculation steps are as follows:

[0045] 1. Feature transformation and activation: Transform the input features through two learnable parameter matrices and , and extract the interaction between features through a nonlinear activation function tanh, as shown in the formula . Where and represent the first and second feature information, respectively.

[0046] 2. Weight calculation (Softmax normalization): Use a learnable weight vector to calculate the importance score of each feature, as shown in the formula and . Where and represent the importance scores of the first and second feature information, respectively. Then normalize the weights through the Softmax function, as shown in the formula and . Where .

[0047] 3. Weighted output features: According to the calculated attention weights, weight the sum of the original features to obtain a new feature vector, as shown in the formula . Where represents the first fused feature information. After obtaining the first fused feature information, the classification head in the student branch can classify the first fused feature information to obtain the first classification result corresponding to the student branch.

[0048] It should be noted that the attention mechanism can improve the sensitivity of the model to abnormal signal features, and can more quickly and accurately identify potential abnormalities in the SEP signal, such as sudden amplitude reduction or significant latency. And through the learning of physiological parameter features, it can reduce the interference of irrelevant information and improve the overall discrimination performance of the model, providing more reliable signal determination for real-time monitoring. The dynamic weight distribution characteristics of the attention mechanism make it highly adaptable in complex signal processing scenarios, and it is particularly suitable for identifying abnormal signals and risk warning. In the collaborative analysis of multi-dimensional features, the attention mechanism can significantly improve the feature selection ability of the model, thereby more efficiently supporting intraoperative decision-making.

[0049] In this embodiment, optionally, the model training module is further configured to: perform feature extraction on the to-be-trained signal by the spatio-temporal attention denoising network in the teacher branch to obtain third feature information; fuse the first feature information, the second feature information and the third feature information by the attention mechanism in the teacher branch to obtain second fusion feature information corresponding to the teacher branch; determine a second classification result corresponding to the teacher branch according to the second fusion feature information; and determine a first loss according to the second classification result and label information corresponding to the candidate somatosensory evoked potential signal, and supervise training of the teacher branch based on the first loss.

[0050] Specifically, after obtaining the to-be-trained signal, it is input into the spatio-temporal attention denoising network in the teacher branch. Different time scale features are captured by multi-scale convolution in the network, and time attention is introduced to automatically enhance SEP neural response components and suppress stimulation artifacts and electromyographic interference, while standardized processing is performed to obtain third feature information, realizing adaptive enhancement of key periods and automatic purification of neural response components. After obtaining the third feature information, the first feature information, the second feature information and the third feature information can be fused by the attention mechanism in the teacher branch. The specific fusion process can be referred to the Attention calculation steps in the student branch, the difference being that the number of input features is increased to 3, thereby obtaining second fusion feature information corresponding to the teacher branch. Further, the classification head in the teacher branch classifies the second fusion feature information to obtain the second classification result corresponding to the teacher branch. Then, the second classification result and the label information corresponding to the candidate somatosensory evoked potential signal are used to calculate the first loss, and the loss is used as the training loss of the teacher branch, so as to supervise the training of the teacher branch based on the first loss. Exemplarily, the first loss can adopt cross-entropy loss.

[0051] In the embodiment, optionally, the model training module is further configured to: determine a second loss according to the first classification result and label information corresponding to the candidate somatosensory evoked potential signal; determine a third loss according to the first feature information and the third feature information, and perform representation alignment on the teacher branch and the student branch based on the third loss; determine a fourth loss according to the first classification result and the second classification result, and perform discriminative alignment on the teacher branch and the student branch based on the fourth loss; determine a target loss of the student branch according to the second loss, the third loss and the fourth loss, and perform supervised training on the student branch based on the target loss.

[0052] Specifically, after obtaining the first classification result, the second loss can be calculated using the first classification result and the label information corresponding to the candidate somatosensory evoked potential signal. Exemplarily, the second loss can adopt cross-entropy loss. Then, the third loss is calculated using the intermediate feature of the student branch (i.e., the first feature information) and the intermediate feature of the teacher branch (i.e., the third feature information), which can be used to reflect the difference between the intermediate features of the teacher branch and the student branch, so as to realize representation alignment (i.e., intermediate feature distillation) between the teacher branch and the student branch based on the third loss. If the dimensions of the two are inconsistent, a projection operator can be set to complete dimension matching. The projection operator can adopt a linear projection operator or a convolution projection operator. Exemplarily, the third loss can be measured by mean square error, as shown in the formula . Alternatively, the third loss can be measured by cosine similarity, as shown in the formula . Wherein, is the third loss, is the first feature information, is the third feature information, is a weight coefficient, preferably in the range of [0.1, 1.0]. For variable-length sequences, an effective bit mask can be introduced to calculate the third loss only at non-padding positions.

[0053] Further, the fourth loss is calculated after temperature smoothing of the first classification result and the second classification result respectively, which can be used to reflect the difference between the class posterior distribution of the teacher branch and the student branch, so as to perform discriminative alignment (i.e., output distillation) on the teacher branch and the student branch based on the fourth loss. Exemplarily, the fourth loss can adopt relative entropy loss, as shown in the formula . Wherein, is the fourth loss, is a temperature parameter, preferably in the range of [2, 6], which is used to restore the gradient magnitude and prevent loss imbalance caused by temperature scaling. Then, a total loss is determined according to the second loss, the third loss and the fourth loss as a target loss of the student branch, as shown in the formula , so as to perform supervised training on the student branch based on the target loss. Wherein, a target loss, a second loss, , and are weight values of the second loss, the third loss and the fourth loss, respectively.

[0054] Figure 3 is a flowchart of an intraoperative nerve stimulation monitoring provided by an embodiment of the present application. As shown in Figure 3 After obtaining the waveform interpretation model, the preprocessed target SEP signal and the feature vector encoded target physiological parameter information signal can be input into the waveform interpretation model to determine whether the waveform is abnormal. Specifically, the preprocessed target SEP signal is first extracted by the CNN_LSTM network, and then the feature extracted target SEP signal and the feature vector encoded target physiological parameter information are adaptively fused by the Attention mechanism. Then, the classification head determines the classification result of the target SEP signal according to the fused features. If the classification result is greater than a preset threshold, it can be determined that the target SEP signal has waveform abnormalities; otherwise, it is determined that the target SEP signal has no waveform abnormalities, thereby realizing real-time abnormality detection of the intraoperative SEP signal.

[0055] Further, if the target SEP signal has waveform abnormalities, real-time alarm can be triggered to remind medical personnel to check and handle in time; if the target SEP signal has no waveform abnormalities, no alarm is given and the detection continues. Exemplarily, the alarm mode can adopt a pop-up reminder, a voice reminder and / or an audible and visual alarm, etc. The alarm content can include basic information of the abnormal object (such as the patient), abnormal waveband, abnormal type and timestamp, etc. In addition, a postoperative report can also be generated according to the alarm content to record the complete intraoperative abnormality for postoperative tracing.

[0056] The technical scheme of the embodiment of the present application proposes an intraoperative nerve stimulation monitoring system, which comprises a signal acquisition module, a signal processing module, a signal anomaly detection module and an alarm module; wherein: the signal acquisition module is used for acquiring target somatosensory evoked potential signal and target physiological parameter information in the surgical process; the signal processing module is used for denoising and standardizing the target somatosensory evoked potential signal; the signal anomaly detection module is used for judging whether the target somatosensory evoked potential signal is abnormal according to a waveform interpretation model and the target physiological parameter information; wherein, the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained by using a double-branch knowledge distillation structure; the alarm module is used for triggering real-time alarm if the target somatosensory evoked potential signal is abnormal. The technical scheme can use the waveform interpretation model combined with the physiological parameter information to perform efficient and accurate signal anomaly detection and alarm on the intraoperative SEP signal, thereby ensuring the safety and accuracy of the surgery. Specifically, the present scheme uses a waveform interpretation model with a CNN-LSTM-Attention hybrid architecture to detect and alarm the intraoperative SEP signal. The model uses a double-branch knowledge distillation architecture for model training, which retains the accuracy and real-time performance of anomaly detection during inference, can accurately capture the peak, trough and latency changes of the SEP signal, and thus realizes real-time and accurate discrimination of abnormal signals. The model can exclude the influence of environmental factors on the signal by combining physiological parameter information, improve the calculation difference caused by small changes in the baseline under different conditions, and thus reduce the false alarm rate. The model shows higher sensitivity to abnormal signals by combining the attention mechanism, and can provide timely intraoperative warning at critical moments to assist medical staff to make rapid intervention.

[0057] In the present embodiment, optionally, the model training module is further configured to: determine a target self-supervised task before constructing a teacher branch based on the spatio-temporal attention denoising network and the attention mechanism, and constructing a student branch based on the convolutional neural network, the long short-term memory network and the attention mechanism; and perform self-supervised feature learning on the spatio-temporal attention denoising network, the convolutional neural network and the long short-term memory network based on the target self-supervised task and the unlabeled somatosensory evoked potential signal.

[0058] It should be noted that the supervised training method requires a large amount of labeled data and needs to match hardware conditions with high computing power. The labeled data requires a lot of time and manpower, and the model trained by the labeled data has high dependence on data and insufficient generalization ability, and requires high hardware resources. In view of the above problems, the present application adds a model pre-training stage, which trains the basic network by a self-supervised model to still have high-precision analysis and real-time anomaly warning ability under the conditions of small sample, weak labeling and insufficient hardware conditions, so as to improve the safety and reliability of the clinic.

[0059] Figure 4 A schematic diagram of self-supervised model training is provided for an embodiment of the present application. Wherein, represents an input signal, represents the feature information learned by the encoder, represents an output signal. Specifically, as shown in Figure 4 , first, a target self-supervised task (such as signal masking recovery) is constructed, and then the spatio-temporal attention denoising network, the convolutional neural network and the long short-term memory network are used as encoders respectively, while adding a decoder, and based on the similarity loss, self-supervised representation learning is performed on the unlabelled somatosensory evoked potential signal, so as to obtain a general feature representation, and the trained encoder is used as a basic network, so as to construct a double-branch knowledge distillation network based on the basic network after self-supervised training (i.e. the network in Figure 2 ).

[0060] Through such a setting, by adding a self-supervised model training link, the data dependence and hardware resource requirements of the model can be effectively reduced, while the training efficiency and generalization ability of the model are improved, so that the model still has high-precision analysis and real-time abnormality early warning ability under the conditions of small sample, weak annotation and insufficient hardware conditions.

[0061] Embodiment Two

[0062] Figure 5 A structural schematic diagram of an intraoperative neural stimulation monitoring system is provided for the second embodiment of the present application, which is optimized based on the above-mentioned embodiment.

[0063] As shown in Figure 5 , the system further comprises a security management module and a remote collaboration module; wherein the security management module is used for data encryption storage and access control of the system, and the remote collaboration module is used for remote control of the system.

[0064] In this embodiment, in order to improve the intelligent level of intraoperative neural stimulation monitoring, the system further sets a security management module and a remote collaboration module, which integrates the functions of real-time monitoring, intelligent analysis, data management and remote control of SEP signals, aiming to realize centralized data management and intelligent decision support across regions and multiple centers, so as to provide an efficient and safe solution for intraoperative neural monitoring. Specifically, the monitoring data can be uploaded to a cloud collaboration platform after encryption, supporting real-time viewing and intervention by remote experts, so as to realize data sharing and expert consultation across hospitals and operating rooms, and improve resource sharing and emergency response capabilities between hospitals through multi-center remote collaboration.

[0065] Embodiment Three

[0066] Figure 6A flowchart of an intraoperative nerve stimulation monitoring method provided for the third embodiment of the present application, this embodiment can be applicable to the case of real-time monitoring and alarming of intraoperative SEP signals, and the method can be executed by an intraoperative nerve stimulation monitoring system, with the corresponding beneficial effects of the system.

[0067] As shown in the method comprises: Figure 6

[0068] S110, acquiring target somatosensory evoked potential signals and target physiological parameter information during the operation by a signal acquisition module.

[0069] S120, performing noise reduction and standardization processing on the target somatosensory evoked potential signals by a signal processing module.

[0070] S130, determining whether the target somatosensory evoked potential signals are abnormal according to a waveform interpretation model and the target physiological parameter information by a signal anomaly detection module; wherein the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained using a double-branch knowledge distillation structure.

[0071] S140, if the target somatosensory evoked potential signals are abnormal, triggering real-time alarm by an alarm module.

[0072] Optionally, the method further comprises:

[0073] constructing a teacher branch based on a spatiotemporal attention denoising network and an attention mechanism, and constructing a student branch based on a convolutional neural network, a long short-term memory network and an attention mechanism; wherein the spatiotemporal attention denoising network is constructed based on multi-scale convolution and temporal attention;

[0074] supervising and training the teacher branch and the student branch according to candidate physiological parameter information, candidate somatosensory evoked potential signals and corresponding label information respectively;

[0075] migrating representation knowledge and discrimination knowledge of the teacher branch to the student branch during the supervising and training of the teacher branch and the student branch;

[0076] determining a target loss of the student branch based on the supervising and training process and the knowledge migration process, and determining whether the target loss meets a training completion condition;

[0077] if yes, ending the model training process, and determining the student branch at the end of the training as the waveform interpretation model.

[0078] Optionally, supervising and training the student branch according to candidate physiological parameter information, candidate somatosensory evoked potential signals and corresponding label information, comprises:

[0079] ​The candidate somatosensory evoked potential signal is denoised and standardized to obtain a to-be-trained signal;

[0080] The to-be-trained signal is feature-extracted by a convolutional neural network and a long short-term memory network in the student branch to obtain first feature information;

[0081] The candidate physiological parameter information is feature vector encoded to obtain second feature information;

[0082] The first feature information and the second feature information are fused by an attention mechanism in the student branch to obtain first fusion feature information corresponding to the student branch;

[0083] A first classification result corresponding to the student branch is determined according to the first fusion feature information.

[0084] Optionally, the teacher branch is supervised trained according to candidate physiological parameter information, candidate somatosensory evoked potential signals and corresponding label information, including:

[0085] The to-be-trained signal is feature-extracted by a spatio-temporal attention denoising network in the teacher branch to obtain third feature information;

[0086] The first feature information, the second feature information and the third feature information are fused by an attention mechanism in the teacher branch to obtain second fusion feature information corresponding to the teacher branch;

[0087] A second classification result corresponding to the teacher branch is determined according to the second fusion feature information.

[0088] A first loss is determined according to the second classification result and label information corresponding to the candidate somatosensory evoked potential signal, and the teacher branch is supervised trained based on the first loss.

[0089] Optionally, representation knowledge and discriminative knowledge of the teacher branch are transferred to the student branch, including:

[0090] A third loss is determined according to the first feature information and the third feature information, and representation alignment of the teacher branch and the student branch is performed based on the third loss;

[0091] A fourth loss is determined according to the first classification result and the second classification result, and discriminative alignment of the teacher branch and the student branch is performed based on the fourth loss;

[0092] A target loss of the student branch is determined based on a supervised training process and a knowledge transfer process, including:

[0093] determine a second loss according to the first classification result and label information corresponding to the candidate somatosensory evoked potential signal;

[0094] determine a target loss of the student branch according to the second loss, a third loss and a fourth loss;

[0095] Correspondingly, the student branch is supervised and trained according to candidate physiological parameter information, a candidate somatosensory evoked potential signal and corresponding label information, and the method further comprises:

[0096] The student branch is supervised and trained based on the target loss.

[0097] Optionally, the method further comprises:

[0098] Before constructing the teacher branch based on the spatio-temporal attention denoising network and the attention mechanism and constructing the student branch based on the convolutional neural network, the long short-term memory network and the attention mechanism, a target self-supervised task is determined;

[0099] Based on the target self-supervised task and the un-labeled somatosensory evoked potential signal, self-supervised representation learning is respectively performed on the spatio-temporal attention denoising network, the convolutional neural network and the long short-term memory network.

[0100] Optionally, the method further comprises:

[0101] The system is subjected to data encryption storage and access control through the security management module.

[0102] Optionally, the method further comprises:

[0103] The system is remotely controlled through the remote collaboration module.

[0104] The technical scheme of the embodiment of the application first collects target somatosensory evoked potential signals and target physiological parameter information in a surgical process through a signal acquisition module; then performs denoising and standardization processing on the target somatosensory evoked potential signals through a signal processing module; further, judges whether the target somatosensory evoked potential signals are abnormal according to a waveform interpretation model and the target physiological parameter information through a signal anomaly detection module; wherein the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained using a double-branch knowledge distillation structure; if the target somatosensory evoked potential signals are abnormal, real-time alarm is triggered through an alarm module. The technical scheme can efficiently and accurately detect signal anomalies and alarm SEP signals in surgery by using a waveform interpretation model combined with physiological parameter information, thereby ensuring the safety and accuracy of surgery.

[0105] Embodiment four

[0106] Figure 7A structural diagram of an electronic device 10 that can be used to implement embodiments of the present application is shown. The electronic device is intended to represent various forms of digital computers, such as laptops, desktops, tablets, personal digital assistants, servers, blade servers, mainframes, and other appropriate computers. The electronic device can also represent various forms of mobile devices, such as personal digital assistants, cellular telephones, smartphones, wearable devices (e.g., headsets, glasses, watches, etc.), and other similar computing devices. The components shown here, their connections and relationships, and their functions, are meant to be examples only, and are not intended to limit the implementations of the present application described and / or claimed in this document.

[0107] As shown, the electronic device 10 includes at least one processor 11, and memory, such as read-only memory (ROM) 12, random access memory (RAM) 13, etc., communicatively connected to the at least one processor 11, where the memory stores computer programs executable by the at least one processor. The processor 11 can perform various appropriate actions and processes according to the computer programs stored in the read-only memory (ROM) 12 or loaded from the storage unit 18 into the random access memory (RAM) 13. In the RAM 13, various programs and data required for the operation of the electronic device 10 can also be stored. The processor 11, the ROM 12, and the RAM 13 are connected to each other through a bus 14. An input / output (I / O) interface 15 is also connected to the bus 14. Figure 7

[0108] Various components in the electronic device 10 are connected to the I / O interface 15, including an input unit 16, such as a keyboard, a mouse, etc., an output unit 17, such as various types of displays, speakers, etc., a storage unit 18, such as a magnetic disk, an optical disk, etc., and a communication unit 19, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 19 allows the electronic device 10 to exchange information / data with other devices through a computer network, such as the Internet, and / or various telecommunication networks.

[0109] The processor 11 can be various general and / or special purpose processing components with processing and computing capabilities. Some examples of the processor 11 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various specialized artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, a digital signal processor (DSP), and any appropriate processor, controller, microcontroller, etc. The processor 11 performs various methods and processes described above, such as the intraoperative neuromonitoring method.

[0110] ​In some embodiments, the intraoperative neurostimulation monitoring method can be implemented as a computer program tangibly embodied in a computer readable storage medium, e.g., storage unit 18. In some embodiments, portions of or all of the computer program can be loaded onto the electronic device 10 via, e.g., ROM 12 and / or communication unit 19. When a computer program is loaded onto the RAM 13 and executed by the processor 11, one or more of the steps of the intraoperative neurostimulation monitoring method described above can be performed. Alternatively, in other embodiments, the processor 11 can be configured to perform the intraoperative neurostimulation monitoring method by other means, e.g., with the aid of firmware.

[0111] Various implementations of the systems and techniques described above can be realized in digital electronic circuitry, integrated circuitry, a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), a system on a chip (SOC), a programmable logic device (PLD), a computer hardware, firmware, software, and / or combinations thereof. These various implementations can include implementation in one or more computer programs that are executable and / or interpretable on a programmable system including at least one programmable processor, which can be special or general purpose, coupled to receive data and instructions from, and to transmit data and instructions to, a storage system, at least one input device, and at least one output device.

[0112] Computer programs used to implement the methods of the present application can be written in any combination of one or more programming languages. These computer programs can be provided to a processor of a general purpose computer, special purpose computer, or other programmable data processing apparatus to produce a machine, such that the computer program, when executed by the processor, implements the functions / acts specified in the flowcharts and / or block diagrams. The computer program can be executed entirely on a machine, partially on a machine, partially on a machine as part of a standalone software package, or entirely on a remote machine or server.

[0113] In the context of the present application, a computer-readable storage medium can be a tangible medium that can contain or store a computer program for use by or in connection with an instruction execution system, apparatus, or device. A computer-readable storage medium can include, but is not limited to, an electronic, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium can be a machine-readable signal medium. More specific examples of a machine-readable storage medium will include one or more lines of a program of instructions in a transitory signal, a portable computer diskette, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or Flash memory), an optical fiber, a portable compact disc read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0114] To provide for interaction with a user, the systems and techniques described here can be implemented on an electronic device having a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user and a keyboard and a pointing device (e.g., a mouse or a trackball) by which the user can provide input to the electronic device. Other kinds of devices can be used to provide for interaction with a user as well; for example, feedback provided to the user can be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user can be received in any form, including acoustic, speech, or tactile input.

[0115] The systems and techniques described here can be implemented in a computing system that includes a back end component (e.g., as a data server), or that includes a middleware component (e.g., an application server), or that includes a front end component (e.g., a user computer having a graphical user interface or a Web browser through which a user can interact with an implementation of the systems and techniques described here), or any combination of such back end, middleware, or front end components. The components of the system can be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include a local area network (LAN), a wide area network (WAN), blockchain network, and the Internet.

[0116] The computing system can include clients and servers. A client and server are generally remote from each other and typically interact through a communication network. The relationship of client and server arises by virtue of computer programs running on the respective computers and having a client-server relationship to each other. The server can be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system, to solve the defects of large management difficulty and weak business scalability in traditional physical host and VPS service.

[0117] It should be understood that the various forms of flow shown above can be used to reorder, add or delete steps. For example, each step described in the present application can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solutions of the present application can be achieved, which is not limited herein.

[0118] The above detailed description does not constitute a limitation on the scope of protection of the present application. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present application shall be included in the scope of protection of the present application.

Claims

1. An intraoperative neuromonitoring system, comprising: The system comprises a signal acquisition module, a signal processing module, a signal anomaly detection module and an alarm module; wherein: The signal acquisition module is configured to acquire target somatosensory evoked potential signals and target physiological parameter information during a surgery; The signal processing module is configured to perform noise reduction and standardization processing on the target somatosensory evoked potential signals; The signal anomaly detection module is configured to determine whether the target somatosensory evoked potential signals are abnormal based on a waveform interpretation model and the target physiological parameter information; wherein the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained using a double-branch knowledge distillation structure; The alarm module is configured to trigger real-time alarms if the target somatosensory evoked potential signals are abnormal.

2. The system of claim 1, wherein, The system further comprises a model training module, which is configured to: construct a teacher branch based on a spatiotemporal attention denoising network and an attention mechanism, and construct a student branch based on a convolutional neural network, a long short-term memory network and an attention mechanism; wherein the spatiotemporal attention denoising network is constructed based on multi-scale convolution and temporal attention; perform supervised training on the teacher branch and the student branch based on candidate physiological parameter information, candidate somatosensory evoked potential signals and corresponding label information; migrate representation knowledge and discrimination knowledge of the teacher branch to the student branch during the supervised training of the teacher branch and the student branch; determine a target loss of the student branch based on the supervised training process and the knowledge migration process, and determine whether the target loss meets a training completion condition; if yes, end the model training process, and determine the student branch at the end of the training as the waveform interpretation model.

3. The system of claim 2, wherein, The model training module is further configured to: perform noise reduction and standardization processing on the candidate somatosensory evoked potential signals to obtain a training signal; extract features from the training signal through the convolutional neural network and the long short-term memory network in the student branch to obtain first feature information; perform feature vector encoding on the candidate physiological parameter information to obtain second feature information; fuse the first feature information and the second feature information through the attention mechanism in the student branch to obtain first fusion feature information corresponding to the student branch; determine a first classification result corresponding to the student branch based on the first fusion feature information.

4. The system of claim 3, wherein, The model training module is further configured to: extract features from the training signal through the spatiotemporal attention denoising network in the teacher branch to obtain third feature information; fuse the first feature information, the second feature information and the third feature information through the attention mechanism in the teacher branch to obtain second fusion feature information corresponding to the teacher branch; determine a second classification result corresponding to the teacher branch based on the second fusion feature information; determine a first loss based on the second classification result and label information corresponding to the candidate somatosensory evoked potential signals, and perform supervised training on the teacher branch based on the first loss.

5. The system of claim 4, wherein, The model training module is further configured to: determine a second loss according to the first classification result and label information corresponding to the candidate somatosensory evoked potential signal; determine a third loss according to the first feature information and the third feature information, and perform representation alignment on the teacher branch and the student branch based on the third loss; determine a fourth loss according to the first classification result and the second classification result, and perform discriminative alignment on the teacher branch and the student branch based on the fourth loss; determine a target loss of the student branch according to the second loss, the third loss and the fourth loss, and perform supervised training on the student branch based on the target loss.

6. The system of any one of claims 1-5, wherein, The model training module is further configured to: determine a target self-supervised task before constructing the teacher branch based on the spatiotemporal attention denoising network and the attention mechanism, and constructing the student branch based on the convolutional neural network, the long short-term memory network and the attention mechanism; perform self-supervised representation learning on the spatiotemporal attention denoising network, the convolutional neural network and the long short-term memory network based on the target self-supervised task and the unlabelled somatosensory evoked potential signal.

7. The system of claim 1, wherein, The system further comprises a security management module configured to perform data encryption storage and access control on the system.

8. The system of claim 1 or 7, wherein, The system further comprises a remote collaboration module configured to perform remote control on the system.

9. An intraoperative neurostimulation monitoring method, comprising: The method comprises: acquiring target somatosensory evoked potential signals and target physiological parameter information during a surgical procedure by a signal acquisition module; performing noise reduction and standardization processing on the target somatosensory evoked potential signals by a signal processing module; determining whether the target somatosensory evoked potential signals are abnormal according to a waveform interpretation model and the target physiological parameter information by a signal anomaly detection module, wherein the waveform interpretation model is constructed based on a convolutional neural network, a long short-term memory network and an attention mechanism, and is trained using a double-branch knowledge distillation structure; if the target somatosensory evoked potential signals are abnormal, triggering real-time alarm by an alarm module.

10. An electronic device, comprising: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor to enable the at least one processor to perform the intraoperative neural stimulation monitoring method of claim 9.

11. A computer readable storage medium, characterized in that, The computer readable storage medium stores computer instructions for enabling the processor to perform the intraoperative neural stimulation monitoring method of claim 9 when executed. The computer readable storage medium stores computer instructions for enabling the processor to perform the intraoperative neural stimulation monitoring method of claim 9 when executed.

Images