Anesthesia depth monitoring method and system based on heart rate dynamic characteristics

Abstract

The invention provides a heart rate dynamic characteristic-based anesthesia depth monitoring method, which comprises the following steps of: carrying out multi-dimensional signal synchronous acquisition on an anesthetized patient to obtain a multi-source original signal, and then carrying out self-adaptive multi-mode signal purification treatment to obtain a purified RR interval sequence; performing HRV feature layered extraction and respiratory sinus arrhythmia decoupling processing on the RR interval sequence to obtain an anesthesia depth specific feature set; constructing an individual baseline model based on preoperative data of a patient, and performing drug effect quantitative compensation through anesthetic infusion data to obtain a calibrated feature set; and performing index operation on the anesthesia depth specific feature set and the calibrated feature set through the constructed lightweight fusion model to obtain an anesthesia depth index and a result confidence coefficient. According to the method, through signal acquisition and purification, feature extraction and decoupling and individual baseline calibration technologies, the anesthesia depth is quantified, the confidence coefficient is evaluated, and the accuracy, real-time performance and reliability of anesthesia depth monitoring are improved.

Description

Technical Field

[0001] This invention relates to the field of medical monitoring technology, and in particular to a method and system for monitoring the depth of anesthesia based on dynamic heart rate characteristics. Background Technology

[0002] Anesthesia depth monitoring is a crucial aspect of clinical anesthesia. Its core objective is to accurately assess the patient's anesthetic status in real time. This involves avoiding both insufficient anesthesia leading to intraoperative awareness and pain stress, and excessive anesthesia causing respiratory and circulatory depression and postoperative cognitive impairment, directly impacting surgical safety and patient prognosis. Dynamic heart rate characteristics (especially heart rate variability, HRV), as a core physiological indicator reflecting autonomic nervous system function, have become an important research direction in anesthesia depth monitoring due to their inherent advantages such as non-invasive data collection, continuous monitoring, and low cost, gradually evolving from an auxiliary reference indicator to a core assessment dimension.

[0003] With the synergistic development of anesthesiology and biosensor technology, anesthesia depth monitoring technology has gradually evolved from early clinical sign observation (such as blood pressure and respiratory rate) to a quantitative assessment model based on physiological signals. Among traditional monitoring methods, technologies based on electroencephalogram (EEG) signals, such as the bispectral index (BIS) and the consciousness index (CSI), are widely used, but they have limitations such as high equipment costs, susceptibility to electromyographic interference, and sensitivity to different types of anesthetic drugs. Monitoring methods that rely solely on heart rate or blood pressure, without considering the dynamic correlation of physiological signals, are insufficient to meet the needs of complex surgical scenarios. In recent years, multimodal signal fusion has become a technological trend, which improves monitoring robustness by integrating multiple sources of signals such as ECG, pulse oximetry, and respiration. However, existing fusion schemes mostly remain at the level of simple parameter superposition, failing to fully explore the specific correlation between dynamic heart rate characteristics and anesthesia depth, and also failing to address the adaptability issues caused by complex interference and individual differences in clinical scenarios.

[0004] Although progress has been made in anesthesia depth monitoring technology based on dynamic heart rate characteristics, existing technologies still have significant shortcomings: First, the signal anti-interference ability is weak; factors such as surgical electrocautery, patient body movement, and respiratory movements can easily lead to distortion of ECG and pulse oximetry (PPG) signals, resulting in HRV calculation errors. Second, respiratory sinus arrhythmia (RSA) and HRV changes caused by anesthesia depth are superimposed, and existing methods lack effective decoupling mechanisms, leading to feature confusion. Third, individual differences and drug effects are not accurately calibrated; preoperative HRV baselines vary significantly among different patients, and the specific effects of some anesthetic drugs on HRV characteristics are not quantified, leading to biased assessment results. Fourth, there is an imbalance between real-time performance and confidence level; traditional HRV analysis relies on fixed time windows, making it difficult to capture sudden changes in anesthesia depth, and lacks quantitative assessment of the reliability of monitoring results, easily leading to clinical misjudgment. Therefore, it is essential to design an anesthesia depth monitoring method and system based on dynamic heart rate characteristics. Summary of the Invention

[0005] The purpose of this invention is to provide a method and system for monitoring anesthesia depth based on dynamic heart rate characteristics. By simultaneously acquiring multi-dimensional signals, purifying and processing them, extracting and decoupling features, calibrating individual baselines, and using lightweight fusion modeling, the method quantifies anesthesia depth-related indicators and assesses the confidence level of the results, thereby improving the accuracy, real-time performance, and reliability of anesthesia depth monitoring.

[0006] To achieve the above objectives, the present invention provides the following solution: A method for monitoring the depth of anesthesia based on dynamic heart rate characteristics includes the following steps: Multi-dimensional signals were simultaneously acquired from anesthetized patients, including electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and integrated into multi-source raw signals. Adaptive multimodal signal purification processing is performed on the original multi-source signals to obtain the purified RR interval sequence; HRV feature stratification extraction and decoupling processing of respiratory sinus arrhythmia were performed on RR interval sequences to obtain an anesthesia depth-specific feature set; An individual baseline model is constructed based on the patient's preoperative data. The drug effect of the individual baseline model is quantitatively compensated by anesthetic drug infusion data to obtain a calibrated feature set. A lightweight fusion model was constructed, and the anesthesia depth-specific feature set and the calibrated feature set were used to perform index calculations to obtain the anesthesia depth index and the result confidence level.

[0007] Optionally, multi-dimensional signal synchronous acquisition is performed on anesthetized patients to obtain electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and these are integrated into multi-source raw signals, including: Data was collected from anesthetized patients using a three-lead ECG module, a fingertip dual-light source PPG module, a chest and abdominal piezoelectric respiratory sensor, and a drug infusion pump data interface, to obtain electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data. A synchronization signal is obtained by adding timestamps and performing spatiotemporal alignment to electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data using a high-precision clock chip; Preliminary R-wave detection was performed on the electrocardiogram signal, and pulse wave peak value was identified on the pulse oxygenation signal to obtain the raw RR interval and pulse interval data; The raw RR interval and pulse interval data are integrated with the synchronization signal to obtain a multi-source raw signal.

[0008] Optionally, adaptive multimodal signal purification processing is performed on the original multi-source signals to obtain the purified RR interval sequence, including: Wavelet packet decomposition and signal-to-noise ratio filtering were performed on the electrocardiogram (ECG) signal to obtain the reconstructed ECG signal; A state equation is constructed using the RR interval sequence of the reconstructed ECG signal as the state variable; An observation equation was constructed based on the RR interval of the reconstructed ECG signal and the pulse interval data from the multi-source original signal. The first denoised signal is obtained by suppressing the RR interval jump of the multi-source original signal through the state equation and the observation equation; Three-dimensional index determination was performed on the multi-source raw signals to obtain motion artifacts; the three-dimensional index determination included: ECG signal mutation feature determination, PPG perfusion fluctuation feature determination, and triaxial acceleration feature determination; motion artifacts included: surgical electrocautery artifacts, patient body motion artifacts, and respiratory motion artifacts. The motion artifacts are matched and differentially interpolated to obtain the second denoised signal; The first denoised signal and the second denoised signal are fitted together to obtain the purified RR interval sequence.

[0009] Optionally, HRV feature stratification extraction and respiratory sinus arrhythmia decoupling processing are performed on the RR interval sequence to obtain an anesthesia depth-specific feature set, including: The RR interval sequence is divided into windows by a sliding window, and the time domain features, frequency domain features and nonlinear features of the windows are extracted to obtain the basic HRV features; Calculate the feature difference rate of adjacent windows, and obtain the feature trend entropy by calculating the feature difference rate of consecutive windows; The respiratory signal was divided into inspiratory and expiratory phases, and the HF component and RSA contribution of the inspiratory and expiratory phases were calculated respectively. The RR interval sequence was decoupled based on the HF component and RSA contribution to obtain specific HF. The basic HRV features, feature trend entropy, and specific HF are integrated into an anesthesia depth-specific feature set.

[0010] Optionally, an individual baseline model is constructed based on the patient's preoperative data. The individual baseline model is then compensated for with anesthetic drug infusion data to quantify the drug effect, resulting in a calibrated feature set, including: Collect resting state data of anesthetized patients, construct a baseline prediction model based on the resting state data and preoperative patient characteristics, and obtain individual baseline reference ranges through the baseline prediction model; A drug impact database was constructed based on clinical data of patients with different anesthesia regimens, and drug impact values ​​were calculated based on the impact database and anesthetic drug infusion data. The anesthesia depth-specific feature set is calibrated based on the individual baseline reference range and drug effect value to obtain the calibrated feature set.

[0011] Optionally, the baseline prediction model includes: an input layer, two LSTM hidden layers, a fully connected layer, and an output layer; the input dimension of the input layer is 9; the first LSTM hidden layer has 48 hidden units, and the second LSTM hidden layer has 24 hidden units; the fully connected layer has 16 neurons, and the activation function of the fully connected layer is the ReLU function; the output layer is used to output the upper and lower limits of the reference range of the core baseline features of the anesthesia state.

[0012] Optionally, a lightweight fusion model is constructed, and the anesthesia depth-specific feature set and the calibrated feature set are used to perform index calculations to obtain the anesthesia depth index and result confidence, including: Construct a lightweight fusion model; the lightweight fusion model includes: an input layer, a feature attention layer, a shallow CNN layer, an LSTM layer, and a fully connected fusion layer; The core indicators are calculated based on the calibrated feature set; the core indicators include: basic drug correction characteristics, dynamic compensation characteristics, and individual baseline fit coefficient. Using core indicators and anesthesia depth-specific feature sets as inputs, a lightweight fusion model is used to output the anesthesia depth index and result confidence.

[0013] Optionally, the process of outputting the result confidence level includes: The first confidence level was calculated using the electrocardiogram signal quality index and the pulse oximetry signal quality index. The percentage of core indicators and anesthesia depth-specific features that conform to the individual baseline reference range is used as the second confidence level. The third confidence level is calculated by the variance of the hidden state vector output by the LSTM layer. The first confidence level, the second confidence level, and the third confidence level are weighted and summed to obtain the final confidence level.

[0014] An anesthesia depth monitoring system based on dynamic heart rate characteristics includes: The signal acquisition module is used to simultaneously acquire multi-dimensional signals from anesthetized patients, including electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and integrate them into multi-source raw signals. The signal purification module is used to perform adaptive multimodal signal purification processing on multi-source raw signals to obtain purified RR interval sequences; The feature extraction module is used to perform HRV feature hierarchical extraction and decoupling processing of respiratory sinus arrhythmia on RR interval sequences to obtain an anesthesia depth-specific feature set. The feature optimization module is used to build an individual baseline model based on the patient's preoperative data, and to perform drug effect quantification compensation on the individual baseline model through anesthetic drug infusion data to obtain a calibrated feature set. The indicator monitoring module is used to construct a lightweight fusion model, and to perform indicator calculations on the anesthesia depth-specific feature set and the calibrated feature set through the lightweight fusion model to obtain the anesthesia depth index and result confidence.

[0015] According to specific embodiments provided by the present invention, the following technical effects are disclosed: The anesthesia depth monitoring method based on dynamic heart rate characteristics provided by the present invention includes: synchronously acquiring multi-dimensional signals from anesthetized patients to obtain electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and integrating them into multi-source raw signals; performing adaptive multimodal signal purification processing on the multi-source raw signals to obtain purified RR interval sequences; performing HRV feature hierarchical extraction and decoupling processing of respiratory sinus arrhythmia on the RR interval sequences to obtain an anesthesia depth-specific feature set; constructing an individual baseline model based on the patient's preoperative data, and performing drug effect quantification compensation on the individual baseline model through anesthetic drug infusion data to obtain a calibrated feature set; constructing a lightweight fusion model, and performing index calculations on the anesthesia depth-specific feature set and the calibrated feature set through the lightweight fusion model to obtain an anesthesia depth index and result confidence. This method, through synchronous acquisition of multi-dimensional signals, purification processing, feature extraction and decoupling, individual baseline calibration, and lightweight fusion modeling, quantifies anesthesia depth-related indicators and evaluates result confidence, thereby improving the accuracy, real-time performance, and reliability of anesthesia depth monitoring. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the anesthesia depth monitoring method based on dynamic heart rate characteristics according to the present invention. Figure 2 This is a schematic diagram of the anesthesia depth monitoring system based on dynamic heart rate characteristics according to the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] like Figure 1 As shown, this invention provides a method for monitoring the depth of anesthesia based on dynamic heart rate characteristics, comprising the following steps: Step 100: Perform multi-dimensional synchronous signal acquisition on the anesthetized patient, obtain electrocardiogram signals, pulse oximetry signals, respiratory signals and anesthetic drug infusion data, and integrate them into multi-source raw signals; Step 200: Perform adaptive multimodal signal purification on the original multi-source signals to obtain the purified RR interval sequence; Step 300: Perform HRV feature layer extraction and respiratory sinus arrhythmia decoupling processing on the RR interval sequence to obtain an anesthesia depth-specific feature set; Step 400: Construct an individual baseline model based on the patient's preoperative data, and perform drug effect quantification compensation on the individual baseline model using anesthetic drug infusion data to obtain a calibrated feature set; Step 500: Construct a lightweight fusion model, and use the lightweight fusion model to perform index calculations on the anesthesia depth-specific feature set and the calibrated feature set to obtain the anesthesia depth index and the result confidence level.

[0021] Preferably, multi-dimensional signal synchronous acquisition is performed on the anesthetized patient to obtain electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and these are integrated into multi-source raw signals, including: Data was collected from anesthetized patients using a three-lead ECG module, a fingertip dual-light source PPG module, a chest and abdominal piezoelectric respiratory sensor, and a drug infusion pump data interface, to obtain electrocardiogram signals, pulse oximetry (PPG) signals, respiratory signals, and anesthetic drug infusion data. A synchronization signal is obtained by adding timestamps and performing spatiotemporal alignment to electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data using a high-precision clock chip; Preliminary R-wave detection was performed on the electrocardiogram signal, and pulse wave peak value was identified on the pulse oxygenation signal to obtain the raw RR interval and pulse interval data; The raw RR interval and pulse interval data are integrated with the synchronization signal to obtain a multi-source raw signal.

[0022] In the specific implementation process, step 100 uses a disposable silver-silver chloride (Ag / AgCl) electrode, and a modified II-lead three-lead ECG module is formed. An integrated 50Hz notch filter suppresses power frequency interference, while a 0.5-30Hz bandpass filter is superimposed to retain the R-wave characteristic frequency band, and the raw ECG signal is output in real time. A dual-source light-emitting diode with 660nm (red light) and 940nm (infrared light) is used, along with a silicon photodiode, as a fingertip dual-source PPG module to receive pulse oximetry signals. The output voltage change of a piezoelectric thin-film sensor reflects the amplitude of chest and abdominal fluctuations, and a 1-5Hz bandpass filter is used to remove heartbeat interference and environmental vibration noise, while marking the rising edge (initiation of inspiration) and falling edge (initiation of expiration) of the respiratory signal. Communication is established with the anesthetic drug infusion pump through an RS485 serial interface to obtain the drug type, instantaneous infusion rate, cumulative infusion dose, and infusion start or stop timestamp in real time.

[0023] A DS3231 high-precision real-time clock (RTC) chip is used to provide a unified clock reference for each signal. Each signal sampling point is embedded with a 13-bit timestamp (formatted as "year-month-day-hour-minute-second-millisecond", accurate to 1ms). For ECG and PPG signals, one timestamp is embedded for each sampling point; for respiratory and drug infusion signals, one timestamp is embedded for every 10 sampling points (respiratory) and 1 sampling point (drug), ensuring fine-grained coverage of the timestamps. Using the ECG signal with the highest sampling rate as the time reference, linear interpolation is used to map respiratory signal sampling points to the ECG time axis. A zero-order hold method is used to assign drug parameter values ​​within one second of the drug infusion data to the time frames corresponding to all ECG sampling points within that time period, thereby generating synchronization signal frames. Each frame contains one ECG sampling point, one PPG sampling point, one interpolated respiratory signal value, and one drug parameter value. The frame time interval is consistent with the ECG sampling period.

[0024] Then, the difference between the maximum and minimum values ​​of the ECG signal within a 1-second window is calculated and used as the R-wave amplitude feature. The first derivative of the ECG signal is calculated, and the maximum value of the derivative is used as the slope feature. The duration of ECG signal amplitude greater than 0.5 times the R-wave amplitude feature is calculated and used as the waveform width feature. A feature threshold library is established based on the valid R-waves detected in the previous 5 minutes, where the R-wave amplitude feature threshold is 0.7 times the mean of the R-wave amplitude features in the previous 5 minutes, the slope feature threshold is 0.6 times the mean of the slope features in the previous 5 minutes, and the waveform width feature threshold is 0.3 times the mean of the waveform width features in the previous 5 minutes. The ECG signal is traversed, and when a sampling point simultaneously satisfies all three features being greater than their respective thresholds, it is marked as an R-wave candidate point. The time interval between the candidate point and the previous R-wave is further calculated. ,like <300ms or If the interval is greater than 2000ms, it is considered a false detection and the candidate point is removed. The final retained candidate points are the R-wave positions. The timestamps and corresponding ECG signal amplitudes are recorded to form the original RR interval sequence.

[0025] Next, the first and second derivatives of the PPG signal are calculated. When the first derivative changes from positive to negative and the second derivative is less than -0.05 mV / ms... 2 At any given time, the point is marked as a candidate point for the pulse wave peak, and its timestamp and amplitude are recorded. For each peak candidate point's timestamp, the timestamp of its preceding ECG-R wave is searched. If the difference between the two timestamps is ∈ [100ms, 400ms], it is determined to be a valid peak. If it exceeds this range, the difference between the peak candidate point's timestamp and the subsequent ECG-R wave's timestamp is further calculated. If the difference is ∈ [100ms, 400ms], the candidate point is adjusted to be the pulse wave peak corresponding to the next R wave; otherwise, it is determined to be a false detection. Simultaneously, 0.7 times the average amplitude of the first 5 valid peaks is used as the dynamic amplitude threshold. If the amplitude of a candidate point is less than the dynamic amplitude threshold, it is determined to be noise and discarded, even if the phase condition is met. The valid pulse wave peaks after the elimination are sorted by time, and the time interval between adjacent peaks is calculated to obtain the pulse interval sequence.

[0026] It should be noted that the former uses multi-feature fusion adaptive R-wave detection through three-dimensional features of R-wave amplitude, slope and width, and combines the physiological conduction time relationship of 100-400ms after the R-wave to identify the ECG-R-wave phase-correlated pulse wave peak, which reduces the false detection rate of R-wave and pulse wave peak and greatly improves the accuracy of key physiological feature identification.

[0027] Preferably, the original multi-source signals are subjected to adaptive multimodal signal purification processing to obtain the purified RR interval sequence, including: Wavelet packet decomposition and signal-to-noise ratio filtering were performed on the electrocardiogram (ECG) signal to obtain the reconstructed ECG signal; A state equation is constructed using the RR interval sequence of the reconstructed ECG signal as the state variable; An observation equation was constructed based on the RR interval of the reconstructed ECG signal and the pulse interval data from the multi-source original signal. The first denoised signal is obtained by suppressing the RR interval jump of the multi-source original signal through the state equation and the observation equation; Three-dimensional index determination was performed on the multi-source raw signals to obtain motion artifacts; the three-dimensional index determination included: ECG signal mutation feature determination, PPG perfusion fluctuation feature determination, and triaxial acceleration feature determination; motion artifacts included: surgical electrocautery artifacts, patient body motion artifacts, and respiratory motion artifacts. The motion artifacts are matched and differentially interpolated to obtain the second denoised signal; The first denoised signal and the second denoised signal are fitted together to obtain the purified RR interval sequence.

[0028] In the specific implementation process, step 200 first inputs the ECG signal from the multi-source original signal into the wavelet packet decomposition module, selects the db4 wavelet basis, and sets the decomposition level to 4 levels. The first level decomposition covers the 0-125Hz frequency band, the second level covers 0-62.5Hz, the third level covers 0-31.25Hz, and the fourth level covers 0-15.625Hz, completely covering the core frequency bands of the R-wave, T-wave, and P-wave in the ECG signal. After decomposition, 16 wavelet packet nodes are obtained, and the signal-to-noise ratio (SNR) of each node is calculated. The SNR threshold is set to 15dB, and nodes with SNR ≥ 15dB are selected and reconstructed using wavelet packets to obtain the reconstructed ECG signal.

[0029] Then, using the reconstructed RR interval sequence of the ECG signal as the core, the RR interval change rate is introduced as an auxiliary state variable to construct a linear discrete state equation, the expression of which is: ; ; ; ; in, Let k be the state vector at time k. To reconstruct the RR interval of the ECG at time k, Let be the rate of change of the RR interval at time k. It is a 2×2 state transition matrix. This is the process noise vector. This is random fluctuation noise between RR intervals. This is noise caused by fluctuations in the rate of change. The expression for the observation equation is: ; ; ; ; in, Let k be the observation vector at time k. To reconstruct the RR interval of the ECG at time k, The original pulse interval at time k. It is a 2×2 observation matrix. To observe the noise vector, This is for ECG-RR interval observation noise. The noise from the PPG pulse interval observation is then addressed. Based on the constructed state equation and observation equation, an iterative optimization algorithm using Kalman filtering is employed to suppress RR interval jumps, generating the first denoised signal.

[0030] Specifically, the criteria for determining ECG signal mutation characteristics are: a single ECG signal mutation amplitude > 500 μV, duration < 100 ms, and the simultaneous electrosurgical device being turned on; the criteria for determining PPG perfusion fluctuation characteristics are: a PPG perfusion index < 0.3, instantaneous acceleration detected by a triaxial accelerometer > 0.5 g, and respiratory signal amplitude mutation > 50%; the criteria for determining triaxial acceleration characteristics are: a correlation between the RR interval fluctuation period and the respiratory cycle > 0.8, and an ECG-PPG signal quality index > 0.6. Signals meeting each criterion are respectively classified as surgical electrosurgical artifacts, patient body motion artifacts, and respiratory motion artifacts.

[0031] For surgical electrosurgical artifacts, three normal RR intervals before and after the artifact segment were used as samples. Weights were assigned according to an exponential decay rule, with greater weight given to intervals closer to the artifact segment. Outliers were removed from the samples using robust linear regression. For patient motion artifacts, the artifact segment was divided into 2-3 segments, each ≤500ms in length. A piecewise Hermite polynomial was constructed using the normal RR intervals at the beginning and end of each segment and their first derivative as constraints. Piecewise cubic Hermite interpolation was then performed to eliminate outliers. For respiratory motion artifacts, the respiratory phase corresponding to each missing time point within the artifact segment was first marked. Then, matching samples were selected from the normal RR intervals of the same respiratory phase within 10s before and after the artifact segment. The selection criteria were: the deviation between the respiratory depth corresponding to the sample and the current respiratory depth was <20%. The RR interval at the corresponding time point was obtained through linear interpolation. The three signals after differential interpolation were integrated, and the interpolated RR interval sequence was used as the second denoising signal. Finally, the first and second denoised signals were weighted and fused with weights of 0.6 and 0.4 respectively to obtain the purified RR interval sequence.

[0032] Preferably, the RR interval sequence is subjected to HRV feature stratification extraction and decoupling processing for respiratory sinus arrhythmia to obtain an anesthesia depth-specific feature set, including: The RR interval sequence is divided into windows by a sliding window, and the time domain features, frequency domain features and nonlinear features of the windows are extracted to obtain the basic HRV features; Calculate the feature difference rate of adjacent windows, and obtain the feature trend entropy by calculating the feature difference rate of consecutive windows; The respiratory signal was divided into inspiratory and expiratory phases, and the HF component and RSA contribution of the inspiratory and expiratory phases were calculated respectively. The RR interval sequence was decoupled based on the HF component and RSA contribution to obtain specific HF. The basic HRV features, feature trend entropy, and specific HF are integrated into an anesthesia depth-specific feature set.

[0033] In the specific implementation process, step 300 first uses a dynamic sliding window to process the purified RR interval sequence. When the RR interval variation coefficient is >10%, the window length is set to 5 seconds; when the RR interval variation coefficient is ≤10%, the window length is set to 3 seconds, and the sampling step size is fixed at 1 second. Then, HRV basic features are extracted hierarchically. Time-domain features include: the standard deviation of all sinus RR intervals (SDNN), the number of adjacent RR interval differences >50ms (NN50), and the percentage of NN50 to the total number of heartbeats (pNN50). Frequency-domain features include: low-frequency band (LF), high-frequency band (HF), and the logarithmic index of the frequency band power ratio (ln(LF / HF)). Nonlinear features include: Poincare plot index, sample entropy, and permutation entropy. All features are set as the basic HRV feature set. Next, the feature difference rate is calculated using the following formula: ; in, Let f(i-1) be a certain HRV feature value of the i-th window, and f(i-1) be the same feature value of the previous window. Then, the feature trend entropy is calculated. The feature difference rate of 5 consecutive windows is selected and divided into 5 intervals: 0-5%, 5%-10%, 10%-15%, 15%-20%, and >20%. The difference rate percentage of each interval is counted, and the feature trend entropy is calculated using the weighted Shannon entropy formula.

[0034] Next, decoupling processing for respiratory sinus arrhythmia was performed. First, the synchronously acquired respiratory signals were divided into inspiratory and expiratory phases. The inspiratory phase is the stage from the trough to the peak of the respiratory signal, during which the signal slope > 0 and the amplitude continuously increases. The expiratory phase is the stage from the peak to the trough, during which the slope < 0 and the amplitude continuously decreases. Simultaneously, apnea segments were removed. Then, frequency domain analysis was performed on the RR interval subsequences corresponding to the inspiratory and expiratory phases respectively to extract their respective HF components, using the formula... The RSA contribution was calculated, where, For the HF power of the getter phase, For expiratory phase HF power, This is the respiratory rate correction factor, when the respiratory rate f resp When the frequency is ≤15 times / minute, K=1, and 15<f resp When the frequency is ≤20 times / minute, K=0.9, f resp When the frequency is >20 times / min, K=0.8. Subsequently, based on the RSA contribution, RR interval sequence decoupling is performed, and an adaptive filter is constructed to... Using the original RR interval sequence as the input signal and the reference signal as the input signal, the filter coefficients are iteratively updated using the least mean square error algorithm to separate the RSA-related components and anesthesia depth-related components in the RR interval sequence, and then the specific HF is calculated. The calculation formula is as follows: ; in, The total HF power is calculated. Finally, the basic HRV feature set, feature trend entropy, and specific HF are integrated according to feature importance to form an anesthesia depth-specific feature set.

[0035] Preferably, an individual baseline model is constructed based on the patient's preoperative data, and the individual baseline model is compensated for by drug effect quantification using anesthetic drug infusion data to obtain a calibrated feature set, including: Collect resting state data of anesthetized patients, construct a baseline prediction model based on the resting state data and preoperative patient characteristics, and obtain individual baseline reference ranges through the baseline prediction model; A drug impact database was constructed based on clinical data of patients with different anesthesia regimens, and drug impact values ​​were calculated based on the impact database and anesthetic drug infusion data. The anesthesia depth-specific feature set is calibrated based on the individual baseline reference range and drug effect value to obtain the calibrated feature set.

[0036] Preferably, the baseline prediction model includes: an input layer, two LSTM hidden layers, a fully connected layer, and an output layer; the input dimension of the input layer is 9; the first LSTM hidden layer has 48 hidden units, and the second LSTM hidden layer has 24 hidden units; the fully connected layer has 16 neurons, and the activation function of the fully connected layer is the ReLU function; the output layer is used to output the upper and lower limits of the reference range of the core baseline features of the anesthesia state.

[0037] In the specific implementation process, step 400 first collects the resting state data of the anesthetized patient three times within 24 hours before the operation, each time for 10 minutes, and removes the periods of movement and emotional fluctuation, and extracts 9-dimensional input features in the resting state, including 3 HRV time domain features: resting SDNN, resting NN50, and resting pNN50, 3 frequency domain features: resting LF power, resting HF power, and resting LF / HF, and 3 patient baseline features: age, weight, and ASA classification, and constructs a baseline prediction model based on these features. The model employs a deep learning architecture consisting of an input layer, two LSTM hidden layers, a fully connected layer, and an output layer. The input layer matches 9-dimensional features. The first LSTM hidden layer has 48 hidden units with tanh activation and a forget gate bias initialized to 1.0. The second LSTM hidden layer has 24 hidden units with tanh activation and overfitting is suppressed by a dropout layer with a dropout rate of 0.2. The fully connected layer has 16 neurons with ReLU activation and the output is normalized by a BatchNorm layer. The output layer uses linear activation and outputs the upper and lower limits of the reference range for five core baseline features of the anesthesia state: SDNN power, HF power, LF / HF, SD1, and sample entropy. The Adam optimizer was used for model training with an initial learning rate of 0.001, which decayed by 10% every 50 rounds. The loss function was root mean square error. The model was trained based on preoperative resting data and postoperative retrospective labeled data from 1,000 patients with different ASA grades. Five-fold cross-validation was used to ensure that the model's generalization error was less than 5%. After training, the 9-dimensional preoperative features of the current patient were input to obtain the individual baseline reference range.

[0038] Next, based on clinical data from 500 patients with different anesthesia regimens, a matrix of influence coefficients for each drug was established using multiple linear regression. The influence coefficients in the matrix reflect the degree of change in a specific HRV characteristic per unit dose. Subsequently, combined with the current patients' anesthetic drug infusion data, the following formula was used to... Calculate the drug effect value for a specific HRV characteristic f, where, The drug effect value of characteristic f, and The coefficients representing the influence of the two currently used anesthetic drugs, propofol and remifentanil, on characteristic f are respectively. and This corresponds to the instantaneous infusion rate of the drug. and This represents the cumulative infusion time of the corresponding drug. For the patient's weight, This is a drug interaction correction factor; if it is a combination of propofol and remifentanil anesthesia, =0.95, if it is a single-drug anesthesia, =1.

[0039] Finally, for each feature f in the anesthesia depth-specific feature set, a two-factor calibration formula is used. Feature calibration is performed, among which... These are the calibrated feature values. This is a specific characteristic value of the original anesthesia depth. The mean of the individual baseline reference range for this feature is used to compensate for the inhibitory or enhancing effect of the drug on the feature. Furthermore, the individual physiological benchmark is anchored by superimposing the correction term of the baseline mean, thus avoiding the problem of ignoring individual differences by simply compensating with drugs.

[0040] Preferably, a lightweight fusion model is constructed, and the anesthesia depth-specific feature set and the calibrated feature set are used to perform index calculations to obtain the anesthesia depth index and the result confidence level, including: Construct a lightweight fusion model; the lightweight fusion model includes: an input layer, a feature attention layer, a shallow CNN layer, an LSTM layer, and a fully connected fusion layer; The core indicators are calculated based on the calibrated feature set; the core indicators include: basic drug correction characteristics, dynamic compensation characteristics, and individual baseline fit coefficient. Using core indicators and anesthesia depth-specific feature sets as inputs, a lightweight fusion model is used to output the anesthesia depth index and result confidence.

[0041] Preferably, the process of outputting the confidence level of the result includes: The first confidence level was calculated using the electrocardiogram signal quality index and the pulse oximetry signal quality index. The percentage of core indicators and anesthesia depth-specific features that conform to the individual baseline reference range is used as the second confidence level. The third confidence level is calculated by the variance of the hidden state vector output by the LSTM layer. The first confidence level, the second confidence level, and the third confidence level are weighted and summed to obtain the final confidence level.

[0042] In the specific implementation process, step 500 first constructs a lightweight fusion model. This model adopts a compact architecture of input layer - feature attention layer - shallow CNN layer - LSTM layer - fully connected fusion layer to balance the real-time performance and accuracy of monitoring. The dimension of the input layer matches the total dimension of the anesthesia depth-specific feature set and the calibrated feature set. Before input, the features are Min-Max standardized to ensure the uniformity of feature dimensions. The feature attention layer adopts a dual attention mechanism of channel and space. First, the importance weight of each feature channel is calculated through the channel attention module, and then the temporal correlation weight between features is captured through the spatial attention module. Finally, the weighted feature matrix is ​​output. The shallow CNN layer has 32 3×1 volumes. The kernel, along with a batch normalization layer and ReLU activation function, is used to extract local correlation information from the feature matrix. The LSTM layer has 16 hidden units, and layer normalization and a 0.15 Dropout layer are introduced to suppress overfitting. At the same time, a simplified gating structure is adopted, removing the forget gate bias initialization of the traditional LSTM, in order to capture the long-term temporal dynamics of the features (in some embodiments, the gradual change of HRV features as the anesthesia depth increases). The fully connected fusion layer has 2 hidden sublayers and 1 output layer. The first layer has 8 neurons and the second layer has 4 neurons, both activated by ReLU. The output layer has 2 neurons, which output the anesthesia depth index and the result confidence, respectively. The model was trained using the AdamW optimizer with an initial learning rate of 0.0005, which decreased by 0.8 every 20 rounds. The loss function was a weighted sum of the anesthesia depth index (MSE) loss and the confidence cross-entropy loss with a weight ratio of 0.7:0.3. The training data came from anesthesia monitoring data of 500 different surgical types. Five-fold cross-validation was used to ensure that the model inference latency was <100ms and the index prediction error was <3%.

[0043] Specifically, the drug-corrected baseline features are the weighted averages of SDNN, specific HF, and sample entropy in the calibrated feature set, with weights of 0.4, 0.3, and 0.3, respectively. Dynamic compensation features are obtained by calculating the 5-second sliding average of the feature trend entropy. The individual baseline fit coefficient is the cosine similarity between the calibrated feature set and the mean of the individual baseline reference range.

[0044] Then, the core indicators and the weighted anesthesia depth-specific feature matrix are input into the lightweight fusion model. After the feature attention layer outputs the weighted features, the local features are extracted by the shallow CNN layer, and then the temporal dynamics are captured by the LSTM layer. Finally, the fully connected fusion layer outputs the original anesthesia depth index through nonlinear mapping, which ranges from 0 to 100, where 0 represents deep anesthesia and 100 represents a conscious state.

[0045] More specifically, the formula for calculating the first confidence level is: ; in, ECG signal quality level This refers to the PPG signal quality level.

[0046] The third confidence score is calculated based on the model's predictive stability, using the following formula: ; in, Let be the variance of the hidden state vector output by the LSTM layer. This represents the maximum variance of the vector in the training set. The final confidence score is obtained by weighted summation, with weights of 0.4, 0.3, and 0.3 for the first, second, and third confidence scores, respectively.

[0047] like Figure 2 As shown, the present invention also provides an anesthesia depth monitoring system based on dynamic heart rate characteristics, comprising: The signal acquisition module is used to simultaneously acquire multi-dimensional signals from anesthetized patients, including electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and integrate them into multi-source raw signals. The signal purification module is used to perform adaptive multimodal signal purification processing on multi-source raw signals to obtain purified RR interval sequences; The feature extraction module is used to perform HRV feature hierarchical extraction and decoupling processing of respiratory sinus arrhythmia on RR interval sequences to obtain an anesthesia depth-specific feature set. The feature optimization module is used to build an individual baseline model based on the patient's preoperative data, and to perform drug effect quantification compensation on the individual baseline model through anesthetic drug infusion data to obtain a calibrated feature set. The indicator monitoring module is used to construct a lightweight fusion model, and to perform indicator calculations on the anesthesia depth-specific feature set and the calibrated feature set through the lightweight fusion model to obtain the anesthesia depth index and result confidence.

[0048] The beneficial effects of this invention are as follows: 1) By suppressing RR interval jumps through wavelet packet decomposition and Kalman filtering, and combining three-dimensional indicators such as ECG mutations, PPG perfusion fluctuations, and triaxial acceleration to determine motion artifacts, and then performing matched differential interpolation for different artifact types to cancel interference from surgical electrocautery, body movement, respiration, etc., the signal anti-interference ability is improved and the data purity is guaranteed. 2) By dividing the respiratory inspiratory / expiratory phases, calculating the contribution of HF components and RSA, and combining adaptive filters to separate RSA-related components and anesthesia depth-related components in the RR interval sequence, specific HF was extracted, avoiding feature confusion caused by RSA interference, achieving decoupling of RSA and anesthesia depth-related HRV features, and improving feature specificity. 3) An LSTM baseline prediction model was constructed based on preoperative resting data and patient baseline characteristics to generate individual baseline reference ranges. The specific effects of anesthetic drugs on HRV characteristics were quantified through a drug effect database, avoiding assessment bias caused by differences in preoperative HRV baselines among different patients and drug interference. This improved individual suitability while reducing assessment bias. 4) The use of a dynamic sliding window balances timeliness and feature stability, while the lightweight fusion model meets the real-time monitoring requirements. At the same time, the confidence of the results is calculated from multiple dimensions such as signal quality, baseline adaptation and model prediction stability, which quantifies the reliability of the monitoring results and avoids clinical misjudgment. 5) Based on non-invasive acquisition signals such as ECG and pulse oximetry, it does not rely on expensive EEG monitoring equipment, reducing application costs. Multimodal signal fusion and hierarchical feature extraction strategies enable the system to work stably under different surgical types and different anesthesia protocols, adapting to complex clinical scenarios.

[0049] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0050] Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of this invention. Furthermore, those skilled in the art will recognize that, based on the ideas of this invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A method for monitoring the depth of anesthesia based on dynamic heart rate characteristics, characterized in that, Includes the following steps: Multi-dimensional signals were simultaneously acquired from anesthetized patients, including electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and integrated into multi-source raw signals. The original multi-source signals are subjected to adaptive multimodal signal purification processing to obtain the purified RR interval sequence; The RR interval sequence was subjected to HRV feature layer extraction and decoupling processing for respiratory sinus arrhythmia to obtain an anesthesia depth-specific feature set. An individual baseline model is constructed based on the patient's preoperative data. The drug effect quantification compensation of the individual baseline model is performed using the anesthetic drug infusion data to obtain a calibrated feature set. A lightweight fusion model is constructed, and the anesthesia depth-specific feature set and the calibrated feature set are used to perform index calculations to obtain the anesthesia depth index and the result confidence level.

2. The method for monitoring anesthesia depth based on dynamic heart rate characteristics according to claim 1, characterized in that, Multi-dimensional signal synchronous acquisition was performed on anesthetized patients, obtaining electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, which were then integrated into multi-source raw signals, including: Data is collected from the anesthetized patient using a three-lead ECG module, a fingertip dual-light source PPG module, a chest and abdominal piezoelectric respiratory sensor, and a drug infusion pump data interface to obtain the electrocardiogram signal, the pulse oxygenation signal, the respiratory signal, and the anesthetic drug infusion data. A synchronization signal is obtained by adding timestamps to the electrocardiogram signal, the pulse oximetry signal, the respiration signal, and the anesthetic drug infusion data using a high-precision clock chip and performing spatiotemporal alignment. The ECG signal is subjected to preliminary R-wave detection, and the pulse wave peak value is identified from the pulse oxygenation signal to obtain the original RR interval and pulse interval data. The original RR interval and pulse interval data are integrated with the synchronization signal to obtain the multi-source original signal.

3. The method for monitoring anesthesia depth based on dynamic heart rate characteristics according to claim 1, characterized in that, The original multi-source signals are subjected to adaptive multimodal signal purification processing to obtain the purified RR interval sequence, including: The ECG signal is subjected to wavelet packet decomposition and signal-to-noise ratio filtering to obtain the reconstructed ECG signal; The state equation is constructed using the RR interval sequence of the reconstructed ECG signal as the state variable; An observation equation is constructed based on the RR interval of the reconstructed ECG signal and the pulse interval data in the multi-source original signal. The first denoised signal is obtained by suppressing the RR interval jump of the multi-source original signal through the state equation and the observation equation. The multi-source raw signals are subjected to three-dimensional index determination to obtain motion artifacts; the three-dimensional index determination includes: ECG signal abrupt change feature determination, PPG perfusion fluctuation feature determination and triaxial acceleration feature determination; the motion artifacts include: surgical electrocautery artifacts, patient body motion artifacts and respiratory motion artifacts. The motion artifacts are subjected to matching differential interpolation to obtain a second denoised signal; The first denoised signal and the second denoised signal are fitted together to obtain the purified RR interval sequence.

4. The method for monitoring anesthesia depth based on dynamic heart rate characteristics according to claim 1, characterized in that, The RR interval sequence was subjected to HRV feature stratification extraction and decoupling processing for respiratory sinus arrhythmia to obtain an anesthesia depth-specific feature set, including: The RR interval sequence is divided into windows by a sliding window, and the time domain features, frequency domain features and nonlinear features of the windows are extracted to obtain the basic HRV features; Calculate the feature difference rate of adjacent windows, and obtain the feature trend entropy through the feature difference rate of consecutive windows; The respiratory signal is divided into an inspiratory phase and an expiratory phase, and the HF component and RSA contribution of the inspiratory phase and the expiratory phase are calculated respectively. The RR interval sequence is decoupled according to the HF component and the RSA contribution to obtain specific HF. The basic HRV features, the feature trend entropy, and the specific HF are integrated into the anesthesia depth specific feature set.

5. The method for monitoring anesthesia depth based on dynamic heart rate characteristics according to claim 1, characterized in that, An individual baseline model is constructed based on the patient's preoperative data. The drug effect is then quantified and compensated for using the anesthetic drug infusion data to obtain a calibrated feature set, including: The resting state data of the anesthetized patients are collected, a baseline prediction model is constructed based on the resting state data and the preoperative characteristics of the patients, and the individual baseline reference range is obtained through the baseline prediction model; A drug impact database was constructed based on clinical data of patients with different anesthesia regimens, and drug impact values ​​were calculated based on the impact database and the anesthetic drug infusion data. The anesthesia depth-specific feature set is calibrated based on the individual baseline reference range and the drug effect value to obtain the calibrated feature set.

6. The method for monitoring anesthesia depth based on dynamic heart rate characteristics according to claim 5, characterized in that, The baseline prediction model includes an input layer, two LSTM hidden layers, a fully connected layer, and an output layer. The input layer has an input dimension of 9. The first LSTM hidden layer has 48 hidden units, and the second LSTM hidden layer has 24 hidden units. The fully connected layer has 16 neurons, and the activation function of the fully connected layer is the ReLU function. The output layer is used to output the upper and lower limits of the core baseline feature reference range for the anesthesia state.

7. The method for monitoring anesthesia depth based on dynamic heart rate characteristics according to claim 1, characterized in that, A lightweight fusion model is constructed, and the anesthesia depth-specific feature set and the calibrated feature set are used to perform index calculations through the lightweight fusion model to obtain the anesthesia depth index and result confidence level, including: Construct the lightweight fusion model; the lightweight fusion model includes: an input layer, a feature attention layer, a shallow CNN layer, an LSTM layer, and a fully connected fusion layer; Core indicators are calculated based on the calibrated feature set; the core indicators include: basic drug correction features, dynamic compensation features, and individual baseline fit coefficient. Using the core indicators and the anesthesia depth-specific feature set as inputs, the anesthesia depth index and the result confidence level are obtained through the lightweight fusion model.

8. The method for monitoring anesthesia depth based on dynamic heart rate characteristics according to claim 7, characterized in that, The process of outputting the confidence level of the result includes: The first confidence level was calculated using the electrocardiogram signal quality index and the pulse oximetry signal quality index. The percentage of the core indicators and the anesthesia depth-specific features that conform to the individual baseline reference range is used as the second confidence level. The third confidence level is calculated by the variance of the hidden state vector output of the LSTM layer. The first confidence level, the second confidence level, and the third confidence level are weighted and summed to obtain the result confidence level.

9. An anesthesia depth monitoring system based on dynamic heart rate characteristics, characterized in that, include: The signal acquisition module is used to simultaneously acquire multi-dimensional signals from anesthetized patients, including electrocardiogram signals, pulse oximetry signals, respiratory signals, and anesthetic drug infusion data, and integrate them into multi-source raw signals. The signal purification module is used to perform adaptive multimodal signal purification processing on the multi-source original signal to obtain the purified RR interval sequence; The feature extraction module is used to perform HRV feature hierarchical extraction and respiratory sinus arrhythmia decoupling processing on the RR interval sequence to obtain an anesthesia depth-specific feature set. The feature optimization module is used to construct an individual baseline model based on the patient's preoperative data, and to perform drug effect quantification compensation on the individual baseline model using the anesthetic drug infusion data to obtain a calibrated feature set. The indicator monitoring module is used to construct a lightweight fusion model, and to perform indicator calculations on the anesthesia depth-specific feature set and the calibrated feature set through the lightweight fusion model to obtain the anesthesia depth index and result confidence.

Images