Anesthesia depth evaluation method based on multi-channel physiological signal fusion
By using multi-channel signal fusion and machine learning models, the inaccuracy of assessing anesthesia depth with a single signal is solved, enabling precise and stable monitoring of anesthesia depth and automated drug administration control, thus supporting postoperative quality management.
Patent Information
- Application Number
- CN202511056408.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-30
- Publication Date
- 2025-10-31
AI Technical Summary
Existing methods for assessing the depth of anesthesia rely on a single physiological signal, are susceptible to noise interference, and cannot fully reflect the central and autonomic nervous system responses, resulting in inaccurate and unstable assessments. Furthermore, they lack automated drug administration control and data recording.
Multi-channel sensors are used to simultaneously collect EEG, ECG, EMG, blood oxygen, and blood pressure signals. These signals are then fused using deep neural networks, convolutional neural networks, or long short-term memory networks. The signal weights are dynamically adjusted and linked with the anesthetic drug infusion device to form a closed-loop control.
It enables accurate assessment and stable monitoring of anesthesia depth, reduces misjudgments, provides automated dosing adjustments and data recording, and supports postoperative quality review.
Smart Images
Figure CN120859438A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical diagnostics and monitoring technology, specifically a method for assessing the depth of anesthesia by fusing multi-channel physiological signals. Background Technology
[0002] Accurate monitoring and assessment of the depth of anesthesia during anesthesia and surgery are crucial for ensuring patient safety and anesthetic efficacy. Traditional assessments of anesthesia depth primarily rely on the anesthesiologist's clinical experience and subjective judgment, as well as limited single-signal monitoring methods. For example, the widely used bispectral index (BPI) monitoring system relies solely on the patient's electroencephalogram (EEG) signals to infer anesthesia depth. Similarly, devices based on EEG entropy values or single-lead EEG pattern recognition also obtain anesthesia depth indicators from EEG signals. These single-signal monitoring methods have limitations: due to individual differences and surgical interference, single physiological parameters are susceptible to noise and artifacts, potentially leading to inaccurate anesthesia depth assessments. For instance, EEG signals are easily affected by electrosurgical procedures, electromagnetic interference, and electromyographic noise; when muscle relaxants are used or individual patient differences result in insignificant changes in EEG signals, relying solely on EEG indicators may not reliably reflect anesthesia depth. Furthermore, relying solely on changes in vital signs such as heart rate and blood pressure cannot comprehensively represent the patient's level of consciousness and analgesia status. Existing technologies lack comprehensive consideration of information on central nervous system activity, cardiovascular response, and muscle activity under anesthesia, making it difficult to meet clinical needs for real-time, objective, and precise monitoring of anesthesia depth.
[0003] Existing technologies still have the following drawbacks when in use:
[0004] Existing technologies typically rely on a single physiological signal (such as EEG or blood pressure fluctuations) to assess the depth of anesthesia. This approach is prone to misjudgment when the signal is interfered with by electrosurgical excision, high-frequency noise, or electromyography artifacts. At the same time, a single signal cannot simultaneously reflect both central nervous system inhibition and autonomic nervous system response, resulting in unstable and unreliable monitoring results under painful stimuli or deep anesthesia.
[0005] Traditional methods for determining the depth of anesthesia typically employ empirical thresholds or linear algorithms, treating all types of physiological characteristics equally. This approach fails to dynamically adjust signal weights based on patient age, disease type, and intraoperative stimuli. Consequently, significant errors occur when the signal feature distribution deviates from the reference range used during training, greatly diminishing the accuracy and generalization ability of the assessment.
[0006] Existing monitoring equipment is mostly passive, relying solely on manual adjustment of medication when the anesthesia depth index exceeds the safe range. This results in problems such as delayed physician response, operational errors, and the inability to control intraoperative instantaneous fluctuations in a timely manner. Furthermore, the lack of automatic recording and retrospective analysis of medication and anesthesia depth curves makes it difficult to provide complete data support for postoperative quality control and protocol optimization. Summary of the Invention
[0007] A method for assessing anesthesia depth by fusing multi-channel physiological signals, characterized by the following specific steps:
[0008] Step S1: Simultaneously acquire multiple physiological signals from the anesthetized patient using a multi-channel sensor. These physiological signals include at least electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, electromyogram (EMG) signals, blood oxygen saturation (SpO2) signals, and blood pressure signals.
[0009] Step S2: Preprocess the acquired physiological signals, including filtering and denoising, baseline correction, and artifact removal, and extract features from the preprocessed signals, specifically including:
[0010] Frequency domain analysis was performed on the EEG signal to extract the power spectrum intensity, EEG entropy value, and complexity features of the δ, θ, α, and β frequency bands;
[0011] Calculate the time-domain and frequency-domain indices of heart rate and heart rate variability (HRV) from electrocardiogram signals;
[0012] Extracting amplitude or spectral energy characteristics of electromyographic signals;
[0013] Extracting saturation values and trends from blood oxygen signals;
[0014] Extract mean arterial pressure, systolic / diastolic blood pressure and fluctuation characteristics from blood pressure signals;
[0015] Step S3: Input the extracted multi-channel physiological feature parameters into the pre-trained anesthesia depth assessment model, and calculate and output the anesthesia depth index or anesthesia depth level through a multi-source signal fusion algorithm. The index or level reflects the comprehensive anesthesia state of the patient's central nervous system inhibition and autonomic nervous system response.
[0016] Furthermore, the anesthesia depth assessment model in step S3 is a machine learning-based neural network model, including any of the following structures:
[0017] A deep neural network (DNN) consists of an input layer, at least one hidden layer, and an output layer. It converts multidimensional feature vectors into an anesthesia depth index through nonlinear mapping.
[0018] Convolutional Neural Networks (CNNs) are used to extract the spatiotemporal features of physiological signals.
[0019] Long Short-Term Memory (LSTM) networks are used to capture the temporal dependence of physiological signals; the model integrates an attention mechanism to dynamically allocate the weights of physiological features in each channel.
[0020] Furthermore, in the preprocessing step:
[0021] EEG signals were filtered using a 0.5-45Hz bandpass filter and independent component analysis (ICA) to remove eye movement and electromyography artifacts.
[0022] The ECG signal was filtered using a 0.05-100Hz bandpass filter to eliminate baseline drift, and a cardiac interval sequence was generated based on R-peak detection.
[0023] High-pass filtering and envelope extraction were performed on the electromyographic signals at a frequency of 20Hz or higher.
[0024] Sliding window smoothing filtering or interpolation processing is used for blood oxygen and blood pressure signals.
[0025] Furthermore, the anesthesia depth index is expressed on a scale of 0-100, wherein:
[0026] A score below 20 indicates a state of deep anesthesia;
[0027] 40-60 is the appropriate range for anesthesia in surgical procedures;
[0028] A score of 80 or higher indicates insufficient anesthesia or near-consciousness.
[0029] When the index continuously exceeds the preset threshold, an audible and visual alarm is triggered to prompt anesthesia adjustment.
[0030] Furthermore, the following steps are included:
[0031] The anesthesia depth index is linked to the anesthetic drug infusion device in real time, and the infusion rate of sedatives or analgesics is dynamically adjusted according to the index value to form a closed-loop feedback control.
[0032] Compared with the prior art, the present invention has the following advantages:
[0033] This invention acquires physiological signals from different channels by simultaneously deploying multiple sensors on the patient's body surface, including electroencephalogram (EEG), electrocardiogram (ECG), electromyography (EMG), blood oxygen saturation (SpO2), and blood pressure. After preprocessing the signals through bandpass filtering, baseline correction, and artifact removal, frequency domain, time domain, and temporal features are extracted. These features include EEG power and entropy values at different frequency bands, heart rate variability, EMG envelope amplitude, blood oxygen fluctuation trends, and average and amplitude blood pressure. By aggregating these multi-dimensional and complementary physiological features, the limitations of single signals under interference, individual differences, or artifacts can be overcome. This allows the assessment of anesthesia depth to reflect both the degree of central nervous system inhibition and the autonomic nervous system's response to pain or stimuli, significantly improving the stability and reliability of the assessment results and providing a more comprehensive and objective basis for monitoring anesthesia status in clinical practice.
[0034] This method employs machine learning models such as deep neural networks (DNN), convolutional neural networks (CNN), or long short-term memory networks (LSTM), combined with pre-labeled large-scale clinical data for supervised training. This enables the model to efficiently learn and map the complex nonlinear relationship between physiological characteristics of each channel and the depth of anesthesia. Simultaneously, an attention mechanism is embedded in the network, allowing the model to automatically allocate weights based on the dynamic changes in input features at different stages of anesthesia—for example, prioritizing autonomic nerve signals such as heart rate and blood pressure when the patient experiences pain or shows initial signs of awakening, while focusing more on EEG slow waves and entropy changes during deep anesthesia. This adaptive modeling approach not only significantly reduces assessment errors caused by differences in disease type, age, and medication regimen, but also continuously optimizes model parameters through online learning or incremental fine-tuning, ensuring consistently high accuracy in assessments of new patients.
[0035] Building upon accurate assessment, this invention further links the real-time calculated anesthesia depth index with the anesthetic drug infusion device, forming an automated closed-loop drug delivery control system. When the index is too high (too shallow anesthesia) or too low (too deep anesthesia) exceeding a preset threshold, the system can adjust the administration rate of sedatives or analgesics within milliseconds, simultaneously triggering an audible and visual alarm to alert the anesthesiologist for verification. This reduces the operational burden on clinical staff and quickly maintains the patient's anesthesia level within a safe and effective target range. Furthermore, this system can record complete drug delivery and anesthesia depth monitoring curves, facilitating postoperative quality review and protocol optimization, and providing comprehensive data support for surgical safety management. Attached Figure Description
[0036] Figure 1 This is a flowchart of the present invention; Detailed Implementation
[0037] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed merely to enable those skilled in the art to better understand and implement the subject matter described herein, and are not intended to limit the scope, applicability, or examples set forth in the claims. The function and arrangement of the elements discussed may be changed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the various examples. For example, the described methods may be performed in a different order than described, and steps may be added, omitted, or combined. Furthermore, features described in some examples may be combined in other examples.
[0038] As used herein, the term "comprising" and its variations are open terms meaning "including but not limited to". The term "based on" means "at least partially based on". The terms "one embodiment" and "an embodiment" mean "at least one embodiment". The term "another embodiment" means "at least one other embodiment". The terms "first", "second", etc., may refer to different or the same objects. Other definitions, whether explicit or implicit, may be included below. Unless explicitly indicated by the context, the definition of a term shall remain consistent throughout the specification.
[0039] Example
[0040] A method for assessing anesthesia depth by fusing multi-channel physiological signals, characterized by the following specific steps:
[0041] Step S1: Simultaneously acquire multiple physiological signals from the anesthetized patient using a multi-channel sensor. These physiological signals include at least electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, electromyogram (EMG) signals, blood oxygen saturation (SpO2) signals, and blood pressure signals.
[0042] Step S2: Preprocess the acquired physiological signals, including filtering and denoising, baseline correction, and artifact removal, and extract features from the preprocessed signals, specifically including:
[0043] Frequency domain analysis was performed on the EEG signal to extract the power spectrum intensity, EEG entropy value, and complexity features of the δ, θ, α, and β frequency bands;
[0044] Calculate the time-domain and frequency-domain indices of heart rate and heart rate variability (HRV) from electrocardiogram signals;
[0045] Extracting amplitude or spectral energy characteristics of electromyographic signals;
[0046] Extracting saturation values and trends from blood oxygen signals;
[0047] Extract mean arterial pressure, systolic / diastolic blood pressure and fluctuation characteristics from blood pressure signals;
[0048] Step S3: Input the extracted multi-channel physiological feature parameters into the pre-trained anesthesia depth assessment model, and calculate and output the anesthesia depth index or anesthesia depth level through a multi-source signal fusion algorithm. The index or level reflects the comprehensive anesthesia state of the patient's central nervous system inhibition and autonomic nervous system response.
[0049] This embodiment: Hardware platform and sensor layout: EEG front-end acquisition:
[0050] Electrode placement: Using the international 10–20 system, the four active electrodes are placed at Fp1, Fp2, Cz and the reference site (mastoid), respectively, and Ag / AgCl gel electrodes are used to improve contact quality.
[0051] Analog-to-digital converter: The TIADS1299 multi-channel physiological signal acquisition chip is used, featuring 24-bit ADC resolution and independently programmable gain (1–12×), with a sampling rate set to 500Hz. The ADS1299 has a built-in programmable gain amplifier and low-pass filter, which directly amplifies and anti-aliasing the EEG signal.
[0052] ECG and EMG front-end data acquisition:
[0053] ECG: ADS1292R dual-channel ADC was used, with electrodes placed in the chest lead II configuration, sampling rate of 500Hz, and gain set to 6×.
[0054] EMG: Using the same ADS1292R channel or TIOPA2376 dual op-amp plus external ADC, a 20–450Hz bandpass filter (3rd order Butterworth) and a sampling rate of 1kHz are selected to capture electromyographic details.
[0055] SpO2 and blood pressure collection:
[0056] SpO2: A commercially available finger-clip photoplethysmography (PPG) module (such as MAX30102) with a sampling rate of 100Hz is selected. The module outputs a pulse intensity signal, which is sent to the main controller via a built-in ADC.
[0057] Non-invasive blood pressure (NIBP): Uses a cuff-type blood pressure module (such as SunTech OSC+) to automatically measure systolic / diastolic blood pressure every 3 minutes and send the measurement results via UART; if continuous monitoring is required, it can be replaced with an arterial catheter + pressure sensor (such as a medical pressure sensor with 5mV / V output) with a sampling rate of 200Hz.
[0058] Main control and communication:
[0059] Main control chip: STM32H743 (Cortex-M7, 480MHz)
[0060] 512KB RAM + 2MB Flash, running FreeRTOS.
[0061] Data bus: Each ADC / sensor connects via SPI or I / O. 2 C connects to the main controller; the NIBP module connects via UART; SpO2 connects via I... 2 C.
[0062] Host computer / infusion pump interface: CAN2.0B or RS-232 is used to upload the anesthesia depth index in real time and receive the infusion pump status to build a closed-loop control.
[0063] Signal preprocessing and feature extraction (corresponding to step S2):
[0064] Bandpass filtering and notch filtering:
[0065] Implementing this using the CMSIS-DSP library on an STM32:
[0066] EEG: 0.5–45Hz 4th order IIR bandpass; 50Hz power frequency notch (2nd order notch).
[0067] ECG: 0.05–100Hz 3rd order bandpass.
[0068] EMG: 20–450Hz 3rd order high-pass + low-pass filter (450Hz).
[0069] Baseline correction and artifact removal:
[0070] Independent component analysis (ICA) was performed on the EEG signal window (2s) to separate and remove eye movement artifacts before reconstructing the signal.
[0071] Baseline drift was removed by using a two-way difference method for ECG; drift was removed by using a sliding window (500ms) mean for PPG / blood pressure.
[0072] Feature extraction (per 1-second window):
[0073] EEG spectrum: CMSIS-DSPFFT is used to calculate the 512-point spectrum, and the power proportions of δ (0.5–4Hz), θ (4–8Hz), α (8–13Hz), and β (13–30Hz) are statistically analyzed, and the sample entropy is calculated.
[0074] ECGHRV: Real-time R-peak detection (threshold + differentiation method), calculation of SDNN and RMSSD on the RR interval sequence; estimation of the low-frequency (LF) / high-frequency (HF) ratio using the Welch method.
[0075] EMG envelope: After rectifying and low-passing the filtered signal (10Hz), calculate the mean and peak values of the envelope amplitude within the window.
[0076] SpO2: Takes the minimum, average, and standard deviation of the original saturation value of the finger clip module.
[0077] Blood pressure: Directly read the systolic / diastolic blood pressure data from the NIBP module. If it is a continuous waveform, calculate the mean arterial pressure (MAP) and its fluctuation amplitude within 1 second.
[0078] Feature vector construction:
[0079] The 10–12 features from each of the above channels are concatenated into a 1×N floating-point vector, where N≈50, and this vector is then fed into the evaluation model as input.
[0080] Multi-source signal fusion and online evaluation (corresponding to step S3):
[0081] Model Deployment:
[0082] Train DNN / CNN / LSTM models offline on a PC using TensorFlow / Keras, and then quantize them into TFLiteMicro format.
[0083] The model is burned into the STM32 Flash memory and inference is performed on the Cortex-M7 using the TFLiteMicro library, which is called once per second, with an inference latency of <20ms.
[0084] Attention mechanism:
[0085] A self-attention layer is inserted into the model structure to dynamically calculate channel attention weights based on the current feature vector, thereby strengthening key channel signals (such as heart rate and blood pressure when subjected to painful stimuli).
[0086] Exponential output:
[0087] The inference outputs an anesthesia depth index of 0–100, which is mapped to three alert ranges: 20, 40–60, and 80. An 8-byte data packet is then sent via CAN.
[0088] Closed-loop drug delivery control and alarm:
[0089] Automatic adjustment: The infusion pump is equipped with a control module with a CAN interface. After receiving the anesthesia depth index, if the index is >80 (too shallow), the pump speed increases by ΔV mL / h; if it is <20 (too deep), the pump speed decreases by ΔV mL / h. ΔV and the adjustment frequency are preset by the physician.
[0090] Audible and visual alarm: The STM32 is equipped with an LED indicator and a buzzer. When the index exceeds the limit and fails to return to the safe zone for 3 consecutive times (3 seconds), the red light will light up and the buzzer will sound.
[0091] Data recording and backtracking: All physiological characteristics and indices, as well as drug administration commands, are stored on an SD card, with one line of CSV written every 1 second; after surgery, the data can be imported into a PC and a complete curve report can be generated, supporting quality auditing and model retraining.
[0092] It should be noted that not all steps and units in the above process and system structure diagrams are mandatory; some steps or units can be omitted according to actual needs. The execution order of each step is not fixed and can be determined as needed. The device structure described in the above embodiments can be a physical structure or a logical structure; that is, some units may be implemented by the same physical entity, or some units may be implemented by multiple physical entities, or they may be jointly implemented by certain components in multiple independent devices.
[0093] The specific embodiments described above are exemplary embodiments, but do not represent all embodiments that can be implemented or fall within the scope of the claims. The term "exemplary" as used throughout this specification means "serving as an example, instance, or illustration" and does not imply that it is "preferred" or "advantageous" compared to other embodiments. Specific details are included to provide an understanding of the described techniques. However, these techniques can be practiced without these specific details. In some instances, well-known structures and apparatuses are shown in block diagram form to avoid obscuring the concepts of the described embodiments.
[0094] The foregoing description of this disclosure is provided to enable any person skilled in the art to implement or use this disclosure. Various modifications to this disclosure will be apparent to those skilled in the art, and the general principles defined herein can be applied to other variations without departing from the scope of this disclosure. Therefore, this disclosure is not limited to the examples and designs described herein, but is consistent with the widest scope of the principles and novel features disclosed herein.
Claims
1. A method for assessing anesthesia depth by fusing multi-channel physiological signals, characterized in that: The specific steps are as follows: Step S1: Simultaneously acquire multiple physiological signals from the anesthetized patient using a multi-channel sensor. These physiological signals include at least electroencephalogram (EEG) signals, electrocardiogram (ECG) signals, electromyogram (EMG) signals, blood oxygen saturation (SpO2) signals, and blood pressure signals. Step S2: Preprocess the acquired physiological signals, including filtering and denoising, baseline correction, and artifact removal, and extract features from the preprocessed signals, specifically including: Frequency domain analysis was performed on the EEG signal to extract the power spectrum intensity, EEG entropy value, and complexity features of the δ, θ, α, and β frequency bands; Calculate the time-domain and frequency-domain indices of heart rate and heart rate variability (HRV) from electrocardiogram signals; Extracting amplitude or spectral energy characteristics of electromyographic signals; Extracting saturation values and trends from blood oxygen signals; Extract mean arterial pressure, systolic / diastolic blood pressure and fluctuation characteristics from blood pressure signals; Step S3: Input the extracted multi-channel physiological feature parameters into the pre-trained anesthesia depth assessment model, and calculate and output the anesthesia depth index or anesthesia depth level through a multi-source signal fusion algorithm. The index or level reflects the comprehensive anesthesia state of the patient's central nervous system inhibition and autonomic nervous system response.
2. The method for assessing anesthesia depth by multi-channel physiological signal fusion according to claim 1, characterized in that: The anesthesia depth assessment model in step S3 is a machine learning-based neural network model, including any of the following structures: A deep neural network (DNN) consists of an input layer, at least one hidden layer, and an output layer. It converts multidimensional feature vectors into an anesthesia depth index through nonlinear mapping. Convolutional Neural Networks (CNNs) are used to extract the spatiotemporal features of physiological signals. Long Short-Term Memory (LSTM) networks are used to capture the temporal dependence of physiological signals; the model integrates an attention mechanism to dynamically allocate the weights of physiological features in each channel.
3. The method for assessing anesthesia depth by multi-channel physiological signal fusion according to claim 1, characterized in that: In the preprocessing step: EEG signals were filtered using a 0.5-45Hz bandpass filter and independent component analysis (ICA) to remove eye movement and electromyography artifacts. The ECG signal was filtered using a 0.05-100Hz bandpass filter to eliminate baseline drift, and a cardiac interval sequence was generated based on R-peak detection. High-pass filtering and envelope extraction were performed on the electromyographic signals at a frequency of 20Hz or higher. Sliding window smoothing filtering or interpolation processing is used for blood oxygen and blood pressure signals.
4. The method for assessing anesthesia depth by multi-channel physiological signal fusion according to claim 1, characterized in that: The anesthesia depth index is expressed on a scale of 0-100, wherein: A score below 20 indicates a state of deep anesthesia; 40-60 is the appropriate range for anesthesia in surgical procedures; A score of 80 or higher indicates insufficient anesthesia or near-consciousness. When the index continuously exceeds the preset threshold, an audible and visual alarm is triggered to prompt anesthesia adjustment.
5. The method for assessing anesthesia depth by multi-channel physiological signal fusion according to claim 1, characterized in that: Further steps include: The anesthesia depth index is linked to the anesthetic drug infusion device in real time, and the infusion rate of sedatives or analgesics is dynamically adjusted according to the index value to form a closed-loop feedback control.