An epidural space anesthesia assisting method and system based on pressure curve characteristics

Abstract

This invention discloses an epidural anesthesia assistance method and system based on pressure curve characteristics, including a data acquisition and preprocessing module, an anatomical layer segmentation module, a multidimensional feature extraction module, a template matching and initial classification module, a Bayesian probabilistic inference and fusion decision-making module, a risk assessment and feedback output module, and an online adaptive update module. By utilizing real-time pressure monitoring, negative pressure recognition, and biometric frequency extraction technologies, it assists in determining the position of the epidural puncture needle, significantly reducing the rate of epidural puncture mispuncture and improving clinical safety. Through a Bayesian online learning update mechanism, the model parameters can be continuously iterated based on clinically labeled data, adapting to different operator habits and sensor drift, achieving long-term stable operation and individualized calibration. This invention is easy to integrate and has low hardware costs, facilitating large-scale clinical deployment.

Description

Technical Field

[0001] This invention relates to a biosignal processing technology, and more particularly to an epidural anesthesia support method and system based on pressure curve characteristics. Background Technology

[0002] Currently, clinical epidural puncture mainly relies on methods such as the resistance disappearance method or the hanging drop method to determine whether the needle tip has entered the epidural space. These methods suffer from problems such as dependence on the doctor's subjective sense, high false positive rate, and poor adaptability due to anatomical differences. Patent CN202511675262.2 discloses a spinal anesthesia puncture guidance system with optical visualization function. It uses an optical probe fixed to the needle to acquire images of the environment in front of the needle to help determine whether it has entered the epidural space. However, the space of the epidural needle is limited, making it difficult and costly to fix the optical probe on it. In addition, due to individual differences and different needle insertion positions, various factors may interfere with the imaging and affect the accuracy of the judgment. Summary of the Invention

[0003] Purpose of the invention: The purpose of this invention is to provide an epidural anesthesia support method based on pressure curve characteristics. Another purpose of this invention is to provide a system for automating the execution of this method.

[0004] Technical solution: The present invention provides an auxiliary method for epidural anesthesia based on pressure curve characteristics, comprising:

[0005] (1) Collect real-time pressure signals during the puncture process and construct a pressure time series dataset;

[0006] (2) Based on prior knowledge of anatomy, construct a puncture layer segmentation model to automatically identify the tissue layer where the puncture needle is located;

[0007] (3) Extract the multidimensional time-frequency features of the pressure curve and construct the puncture state feature vector;

[0008] (4) Based on the dynamic time warping algorithm, a multi-category pressure curve template library is constructed to perform similarity measurement and preliminary classification of hierarchical transformation patterns;

[0009] (5) Construct a Bayesian neural network probabilistic inference model to determine the location of the epidural space;

[0010] (6) Establish a multi-level risk assessment and real-time feedback output mechanism for clinical decision-making;

[0011] (7) Establish an online adaptive update mechanism to continuously learn and individually calibrate model parameters.

[0012] In step (1), the puncture process continuously collects the real-time contact pressure signal p(t) between the tip of the puncture needle and the tissue interface through a miniature high-precision pressure sensor connected to the tail end of the puncture needle. At the same time, an optical encoder collects the puncture needle advance displacement signal d(t). After all sensor data are synchronized with timestamps, they are converted from analog to digital and uploaded to the embedded signal processing unit to form a two-dimensional time-series observation pair. ; where t k For the k-th sampling time, K is the total number of sampling points in the current puncture operation.

[0013] The pretreatment process for the puncture includes:

[0014] (1) Baseline drift correction: An adaptive piecewise linear baseline estimation algorithm is used to subtract the estimated baseline from the original pressure signal to obtain the baseline-de-baseline pressure signal. ;

[0015] (2) Sliding median filtering: with window length W m right Apply sliding median filtering to suppress impulse noise;

[0016] (3) Low-pass Butterworth filtering: The median-filtered signal is smoothed using a zero-phase low-pass Butterworth filter to obtain the final preprocessed pressure sequence. .

[0017] After pretreatment, the displacement is advanced using a puncture needle. With pressure as the horizontal axis Construct a two-dimensional time series curve of pressure-displacement with the vertical axis as the ordinate. This serves as the foundational data for subsequent feature extraction and recognition.

[0018] In step (2), the puncture level segmentation model uses a hidden Markov model to constrain the hierarchical order, including:

[0019] (1) Based on anatomical priors, the state transition matrix of the hidden Markov model is ordered because the puncture needle can only move forward in one direction.

[0020] (2) The stress observation distribution at each organizational level is defined as a Gaussian mixture model;

[0021] (3) After obtaining the stress observation probabilities at different organizational levels, the Viterbi algorithm is used to perform maximum a posteriori decoding on the observation sequence to obtain the optimal hierarchical labeling sequence;

[0022] (4) Based on the optimal hierarchical labeling sequence results, the pressure-displacement curve is divided into sub-sequence segments corresponding to each anatomical level.

[0023] In step (3), the process of constructing the puncture state feature vector specifically includes:

[0024] (1) Extracting time-domain statistical features of pressure: For the segmented subsequence of the current level, calculate the time-domain statistical features;

[0025] (2) Extract pressure-displacement joint features: Define unit displacement pressure and pressure-displacement curve slope sequence, and further calculate the slope mean and slope standard deviation from the pressure-displacement curve slope sequence to obtain the number of slope sign changes that characterize the frequency of pressure fluctuations;

[0026] (3) Extracting features of resistance disappearance events: Define a local pressure drop detection function, and further extract the drop magnitude, drop rate and the current hidden Markov model annotation level for the detected resistance disappearance events;

[0027] (4) Extracting the frequency domain and time-frequency features of the pressure curve: Perform short-time Fourier transform on the segmented subsequence of the current level to extract the power spectral density features, characterize the frequency components of pressure oscillation, and distinguish between mechanical resistance features and physiological noise components.

[0028] (5) Based on all the extracted features, construct a complete puncture state feature vector.

[0029] In step (4), the process of performing similarity measurement and preliminary classification of hierarchical transformation patterns includes:

[0030] (1) Constructing a pressure curve template library: From the clinical annotation dataset, pressure-displacement sub-curve segments were extracted according to five event types to construct a template library; the five event types include the disappearance of true resistance at the normal ligamentum flavum-epidural interface, the disappearance of false resistance in the interspinous ligament space, the disappearance of false resistance in the ligamentum flavum space, the disappearance of false resistance in the paravertebral space, and the disappearance of false resistance in the muscle space; for each event type, the existing pressure-displacement curves accurately annotated by clinicians were used as the original templates, and the average template sequence of the event type was calculated by the dynamic time warping centroid averaging algorithm;

[0031] (2) Real-time similarity measurement based on DTW distance: For the currently collected real-time data points, take the current sampling point as the center, take a certain length of sliding window forward to obtain the pressure subsequence, and calculate the DTW distance between it and the average sequence of various templates;

[0032] (3) Preliminary classification decision based on DTW distance: Based on the calculated DTW distance vector, the preliminary category probability distribution is obtained by Softmax normalization. This distribution is used as an auxiliary feature and concatenated to the extracted puncture state feature vector to form an enhanced feature vector, which is then input into the subsequent Bayesian neural network for probability inference.

[0033] In step (5), determining the location of the epidural space includes constructing a BNN network structure, BNN probability prediction and inference, and constructing a multi-source evidence fusion mechanism for identifying the disappearance of false resistance.

[0034] In step (7), the process of establishing the online adaptive update mechanism includes:

[0035] (1) After the puncture is completed, the anesthesiologist will mark the results of the puncture. If the real resistance disappears, it will be marked directly. If the false resistance disappears, it will be marked according to experience. The enhanced feature vector sequence of this puncture will be stored as a new sample.

[0036] (2) With a fixed number of new samples as the interval, the variational parameters are fine-tuned online by updating the objective function by minimizing the lower bound of evidence in the newly added training sample set.

[0037] The present invention discloses an epidural anesthesia support system based on pressure curve characteristics, comprising:

[0038] Data acquisition and preprocessing module: Receives real-time contact pressure signals acquired by miniature pressure sensors and displacement signals acquired by optical encoders, and uses an adaptive piecewise linear baseline estimation algorithm, a sliding median filter, and a low-pass Butterworth filter to perform baseline removal and smoothing and noise reduction on the original signals, outputting a pressure-displacement two-dimensional time series curve with a high signal-to-noise ratio.

[0039] Anatomical layer segmentation module: Input the two-dimensional time series curve into the pre-trained hidden Markov-Gaussian mixture model, calculate the observation probability of each anatomical tissue layer where the needle tip is located, and perform maximum a posteriori decoding through the Viterbi algorithm to automatically segment the continuous long time series signal into sub-sequence segments of anatomical structures such as ligaments and epidural space.

[0040] Multidimensional feature extraction module: For the segmented subsequence segments, extract time-domain statistical features, pressure-displacement joint features, sudden drop features of resistance disappearance event, and frequency-domain features extracted based on short-time Fourier transform. Concatenate the multidimensional features to form an enhanced feature vector to characterize the current tissue resistance state.

[0041] Template matching and initial classification module: Extract the pressure subsequence within the current sliding window, use the dynamic time warping algorithm to calculate the elastic distance between it and the built-in multi-class real or false resistance disappearance curve template, use the Softmax function to transform the distance into a preliminary classification probability distribution, and add it as an auxiliary feature to the enhanced feature vector;

[0042] Bayesian Probability Inference and Fusion Decision Module: The enhanced feature vector with additional auxiliary features is input into the Bayesian neural network. The mean posterior prediction probability of each category is calculated through Monte Carlo sampling to quantify uncertainty. The Dempster-Shafer evidence theory is used to fuse the posterior prediction probability of the Bayesian neural network with the preliminary classification probability through multi-source evidence fusion, eliminate judgment conflicts, and output the final true resistance disappearance fusion probability.

[0043] Risk assessment and feedback output module: Dynamically compares the final fusion probability with the system's preset high and low thresholds for epidural confirmation, and triggers the human-computer interaction interface based on the comparison results to issue corresponding three-level clinical decision instructions: confirmation has been reached, suspected case requires manual confirmation, or non-epidural area with high risk of mis-puncture.

[0044] Online adaptive update module: Receives the actual results labeled by clinicians after puncture and stores them in a local buffer. When the accumulated samples reach a preset number, the objective function is updated by variational minimization of the lower bound of evidence and combined with KL divergence regularization to perform online fine-tuning and parameter calibration of the Bayesian neural network.

[0045] Beneficial effects: Compared with the prior art, the present invention has the following significant advantages: (1) Significantly reduces the rate of accidental puncture of epidural puncture and improves clinical safety: Through pressure time series acquisition, anatomical layer segmentation, multidimensional feature extraction, DTW The synergistic fusion of template matching and Bayesian neural network probability inference can accurately distinguish between the real resistance disappearance and false resistance disappearance signals such as interspinous ligament, ligamentum flavum, paravertebral ligament, and intermuscular space, transforming the puncture positioning from relying on the doctor's subjective feeling to objective quantitative judgment, thereby reducing complications such as vascular, nerve, and spinal cord injury from the root and significantly improving the safety of epidural anesthesia operation; (2) Get rid of subjective experience dependence and lower the operation threshold: Compared with the traditional resistance disappearance method and the hanging drop method, which are highly dependent on the operator's feeling and experience, this invention realizes full-process automated identification and real-time judgment, which can complete accurate positioning without the operator having long-term clinical accumulation, effectively shortening the training cycle of young anesthesiologists and improving the success rate of operation for grassroots and novice; (3) Strong anti-interference ability and adaptable to complex clinical scenarios: Through the preprocessing process of baseline correction, median filtering, and Butterworth low-pass filtering, physiological interferences such as breathing, heartbeat, and hand tremors can be suppressed; combined with the hidden Markov model to constrain the puncture layer sequence, it can adapt to patients of different ages, body types, and spinal anatomy variations, solving the problem of poor individual adaptability and susceptibility to environmental interference in the existing technology. The problems; (4) High positioning accuracy, and the judgment results are quantifiable and traceable: Based on the time-frequency features, pressure-displacement joint features, resistance disappearance event features, and DTW similarity features, a multi-dimensional feature vector is constructed. Combined with the Bayesian neural network, uncertainty quantification and DS evidence fusion decision are realized, and a three-level risk assessment result is output. The positioning result is objective, quantifiable, and the process is traceable, which is superior to the spatial limitation, high cost, and easy interference defects of optical visualization and other technologies; (5) It has online adaptive capability and continuously optimizes performance: The system has a Bayesian online learning and update mechanism, which can continuously iterate the model parameters according to the clinical annotation data, adapt to different operator habits and sensor drift, realize long-term stable operation and individualized calibration, and solve the problem that traditional fixed models cannot adapt to clinical dynamic changes; (6) Low hardware cost, easy integration, and strong clinical promotion: Only a miniature high-precision pressure sensor and an optical encoder are needed to realize data acquisition. There is no need to integrate complex components such as optical probes into the needle body. The hardware structure is simple, the processing cost is low, and it is easy to be compatible with existing epidural puncture needles, which is convenient for large-scale clinical popularization. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention;

[0047] Figure 2 This is a pressure-displacement curve of epidural puncture according to the present invention. Detailed Implementation

[0048] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0049] Example 1

[0050] This embodiment provides a method for determining the location of an epidural puncture needle using real-time pressure monitoring, negative pressure identification, and biometric frequency extraction technology. The flowchart of this method is shown below. Figure 1 As shown, the specific process includes the following steps:

[0051] Step S10: Acquire real-time pressure signals during the puncture process and construct a pressure time-series dataset.

[0052] In a standard epidural puncture procedure, the puncture needle sequentially traverses the skin and subcutaneous tissue, the supraspinous ligament, the interspinous ligament, and the ligamentum flavum, finally penetrating the ligamentum flavum to enter the epidural space. Throughout the entire puncture process, a miniature high-precision pressure sensor connected to the tail of the puncture needle continuously collects the real-time contact pressure signal p(t) between the tip of the puncture needle and the tissue interface. Simultaneously, an optical encoder collects the puncture needle advancement displacement signal d(t). All sensor data are synchronized with timestamps, then converted from analog to digital and uploaded to the embedded signal processing unit to form a two-dimensional time-series observation pair. , where t k For the k-th sampling time, K is the total number of sampling points in the current puncture operation.

[0053] To suppress high-frequency noise introduced by physiological disturbances such as hand tremors, respiratory movements, and heartbeats during the puncture procedure, the following preprocessing steps are applied sequentially to the original pressure signal:

[0054] (1) Baseline drift correction: An adaptive piecewise linear baseline estimation algorithm is used to subtract the estimated baseline from the original pressure signal to obtain the baseline-de-baseline pressure signal.

[0055] (2) Sliding median filtering: with window length W m right Apply a sliding median filter to suppress impulse noise. m Typically, 7-11 sampling points are set.

[0056] (3) Low-pass Butterworth filter: with cutoff frequency f c =20Hz, order N f A zero-phase low-pass Butterworth filter with a phase ratio of 4 is used to smooth the median-filtered signal, resulting in the final preprocessed pressure sequence. .

[0057] After pretreatment, the displacement is advanced using a puncture needle. With pressure as the horizontal axis Construct a two-dimensional time series curve of pressure-displacement with the vertical axis as the ordinate. This serves as the foundational data for subsequent feature extraction and recognition.

[0058] Step S20: Based on prior anatomical knowledge, construct a puncture layer segmentation model to automatically identify the tissue layer where the puncture needle is located.

[0059] For step S20, different anatomical layers have different mechanical resistance characteristics during epidural puncture, such as... Figure 2 The typical epidural puncture pressure-displacement curves shown exhibit statistically regular segmented morphological characteristics. Specifically: the skin and subcutaneous tissue layers correspond to the initial rapid pressure rise segment; the supraspinous ligament layer corresponds to the high pressure plateau segment; the interspinous ligament layer corresponds to the moderate pressure fluctuation segment; the ligamentum flavum layer corresponds to the highest pressure plateau segment; a characteristic sudden drop in negative pressure response occurs upon entering the epidural space. To accurately distinguish these layers, this invention establishes a sequential segmentation method for puncture layers based on a hidden Markov model, dividing the two-dimensional pressure-displacement time-series curves... It is divided into sub-sequence segments corresponding to each anatomical level.

[0060] Define the organizational hierarchical state sequence as These correspond to five tissue layers: skin and subcutaneous tissue, supraspinous ligament, interspinous ligament, ligamentum flavum, and epidural space. A Hidden Markov Model (HMM) is used to constrain the hierarchical order. First, the state transition matrix A of the HMM is constrained based on anatomical priors, namely, that the puncture needle can only move in one direction and the hierarchical state transitions have strict order, resulting in:

[0061]

[0062] The distribution of stress observations at each organizational level is defined as a Gaussian mixture model, where the probability of stress observations at the i-th organizational level is:

[0063]

[0064] In the formula: G is the number of mixed components. It is a Gaussian function. , , These are the mixture weights, mean, and variance of the g-th Gaussian component at the i-th tissue level, respectively. These parameters were obtained offline by training on a historical dataset containing labeled puncture pressure data using the expectation-maximization algorithm.

[0065] After obtaining the observed probabilities of stress at different organizational levels, the Viterbi algorithm is used to analyze the observed sequences. Perform maximum a posteriori decoding to obtain the optimal hierarchical labeled sequence:

[0066]

[0067] Finally, based on the optimal hierarchical annotation sequence results, the pressure-displacement curve is divided into sub-sequence segments corresponding to each anatomical level, providing structured input for subsequent refined analysis of the ligamentum flavum-epidural interface features.

[0068] Step S30: Extract multi-dimensional time-frequency features from the pressure curve and construct a puncture state feature vector.

[0069] For step S30, using the optimal hierarchical annotation sequence obtained in step S20, the pressure-displacement two-dimensional time-series curve is divided into 5 sub-sequences, corresponding sequentially to the five processes of epidural puncture: skin and subcutaneous tissue, supraspinous ligament, interspinous ligament, ligamentum flavum, and epidural space. For each sub-sequence, the following multidimensional features are extracted to construct a feature vector f. k .

[0070] Step S301: Extract time-domain statistical features of pressure;

[0071] For the segmented subsequence with a current hierarchical segmentation window length of W Calculate the following time-domain statistical characteristics:

[0072]

[0073]

[0074]

[0075]

[0076]

[0077] In the formula: , These represent the mean and standard deviation of the pressure within the window, respectively. This represents the net change in window pressure. , These are the maximum first-order and second-order differences, respectively. The sampling interval is denoted as .

[0078] Step S302: Extract pressure-displacement joint features;

[0079] Define the pressure per unit displacement as:

[0080]

[0081] In the formula: This represents the displacement increment between adjacent sampling times.

[0082] Define the slope sequence of the pressure-displacement curve:

[0083]

[0084] In the formula: To prevent the use of tiny constants for division by zero, it is typically 1 × 10⁻⁶. -6 .

[0085] Depend on The sequence is further calculated to calculate the mean slope. Standard deviation of slope Based on this, the number of times the slope sign changes can be further determined. It is used to characterize the frequency of pressure fluctuations.

[0086] Step S303: Extract the characteristics of the resistance disappearance event;

[0087] Define a local pressure drop detection function:

[0088]

[0089] In the formula: For the length of the lookback window, The threshold for sudden pressure drop. The absolute value threshold of the pressure. This is an indicator function.

[0090] For the detected resistance disappearance event, the following descriptive features are further extracted: magnitude of drop. , rate of sudden drop and the hierarchy of current Hidden Markov Model annotations. .

[0091] Step S304: Extract the frequency domain and time-frequency characteristics of the pressure curve;

[0092] For the segmented subsequence with a current hierarchical segmentation window length of W Perform short-time Fourier transform to extract power spectral density features To characterize the frequency components of pressure oscillations and distinguish between mechanical resistance characteristics and physiological noise components.

[0093] Combining all the features extracted in steps S301 to S304 above, a complete puncture state feature vector is constructed:

[0094]

[0095] Step S40: Based on the dynamic time warping algorithm, construct a multi-category pressure curve template library to realize the similarity measurement and preliminary classification of hierarchical transformation patterns;

[0096] For step S40, under different operator, patient body type, and spinal anatomy variations, the epidural puncture pressure curve exhibits significant nonlinear stretching and deformation on the time axis, making it difficult for traditional Euclidean distance measurements to effectively capture its morphological similarity. Therefore, this invention introduces a Dynamic Time Warping (DTW) algorithm to establish a pressure curve similarity measurement framework with elastic time axis alignment.

[0097] Step S401: Construct a pressure curve template library;

[0098] From the clinical annotation dataset, pressure-displacement sub-curve segments were extracted according to the following five event types to construct a template library. :

[0099] Event Categories Event Type c=1 The true resistance at the normal ligamentum flavum-epidural interface disappears. c=2 The false resistance in the interspinous ligament space disappears c=3 The false resistance in the ligamentum flavum gap disappears c=4 The pseudo-resistance of the paravertebral space disappears c=5 Disappearance of pseudo-resistance between muscles

[0100] For each type of event c, N is defined by existing N values ​​accurately labeled by clinicians. c Using the pressure-displacement curve as the original template, the average template sequence for this type of event is calculated using a dynamic time warping centroid averaging algorithm. ,Right now:

[0101]

[0102] In the formula, Let be the average template sequence to be solved. This is the nth template curve for the c-th type of event. denoted as DTW distance between the two sequences.

[0103] Step S402: Real-time similarity measurement based on DTW distance;

[0104] For the currently collected real-time data points, take the length forward from the current sampling point as the center. Using a sliding window, obtain the pressure subsequence. Calculate the DTW distance between it and the average sequence of various templates:

[0105]

[0106] In the formula: To find the optimal regular path that satisfies the boundary conditions and the constraints of continuity and monotonicity, dynamic programming is used as a recursive solution:

[0107]

[0108] Step S403: Preliminary classification decision based on DTW distance;

[0109] Based on the calculated DTW distance vector The preliminary class probability distribution is obtained through Softmax normalization:

[0110]

[0111] Will As an auxiliary feature, it is concatenated into the feature vector extracted in step S30. , forming enhanced feature vectors The input is then used to perform probability inference in a subsequent Bayesian neural network.

[0112]

[0113] Step S50: Construct a Bayesian neural network probabilistic inference model to determine the location of the epidural space;

[0114] In step S50, the deterministic deep neural network only outputs point estimation predictions, which cannot quantify the cognitive uncertainty objectively existing during the puncture localization process due to insufficient training data or limited model expressive power, as well as the accidental uncertainty caused by sensor noise and individual patient differences. This is detrimental to risk assessment for clinical decision-making. Therefore, this invention constructs a Bayesian Neural Network (BNN), which, by applying a prior distribution to the network weights and estimating the posterior distribution based on variational inference, achieves complete quantification of the uncertainty of classification probabilities.

[0115] Step S501: Construct the BNN network structure;

[0116] The BNN network backbone adopts the following structure: the input layer receives enhanced feature vectors. The hidden layers are fully connected layers of type H, with N neurons per layer. h The activation function is ReLU, and the output layer is a 5-dimensional Softmax classification layer, outputting the posterior probabilities of the five event types of epidural puncture described in S401. The set of network weights is denoted as... .

[0117] Apply an isotropic Gaussian prior to the network weights w:

[0118]

[0119] in, It is a Gaussian function. It is the prior variance. Let be the covariance matrix.

[0120] Corresponding to the given training dataset obtained from real clinical practice ,in To label event types in clinical settings, the goal of Bayesian inference is to calculate the posterior distribution. Since the posterior distribution is difficult to solve analytically, this invention employs mean-field variational inference to parameterize the variational family. Approximate the true posterior by optimizing the variational parameters through minimizing the lower bound of evidence loss function. :

[0121]

[0122] In the formula: the first term is the reconstructed likelihood expectation, and the second term is the KL divergence between the variational distribution and the prior distribution, which guides the variational posterior to shrink towards the prior, preventing overfitting. Gradient estimation is achieved through reparameterization techniques:

[0123]

[0124] Step S502: BNN probability prediction inference;

[0125] During the inference phase, for input features The posterior prediction distribution of the BNN is approximated using Monte Carlo sampling. First, from the variational posterior... Weights of S groups of repeated sampling Calculate the probability prediction of the five event types under each sample:

[0126]

[0127] in, The weights of the s-th group Below, the network processes the input. The calculated output. The output value under the weight of the s-th group is calculated by the Softmax function to obtain the category probability distribution, which is a vector of length 6, with each element corresponding to the probability of an epidural puncture event type.

[0128] Based on this, the mean of the posterior prediction distribution is calculated, which serves as the model's final probability prediction for that sample.

[0129]

[0130] Step S503: Construct a multi-source evidence fusion mechanism for the false resistance disappearance judgment;

[0131] Since the interspinous ligament space, ligamentum flavum space, paravertebral space, and intermuscular space can all produce the illusion of resistance disappearance, relying solely on the feature vector at the current moment can easily lead to misjudgment. Therefore, this invention designs a multi-source evidence fusion mechanism based on Dempster-Shafer evidence theory, fusing the following two types of evidence:

[0132] (1) Posterior prediction probability of BNN ;

[0133] (2) DTW preliminary classification probability ;

[0134] Transform the two types of probability vectors mentioned above into basic probability assignment functions. , Evidence fusion was performed according to Dempster's combination rule:

[0135]

[0136] in, The target set for calculating trust after fusion is the set of five event types corresponding to epidural puncture. It serves as a normalization factor to eliminate probability distortion caused by conflicts. This represents the degree of conflict between the two types of evidence. The higher the value, the more obvious the contradiction between the evidence, and the system will output a high-risk warning.

[0137] The final probability of confirming the epidural space after fusion is That is, the fusion probability of the disappearance of the true resistance at the normal ligamentum flavum-epidural interface when category c=1.

[0138] Step S60: Establish a multi-level risk assessment and real-time feedback output mechanism for clinical decision-making.

[0139] For step S60, based on the multi-source evidence fusion probability in step S50... A three-tiered clinical risk assessment and output decision-making mechanism is constructed as follows:

[0140] Decision conditions Output decision Confirm epidural space Suspected epidural space, requires manual confirmation. Non-epidural space, high risk of accidental puncture

[0141] In the formula, , These are the high and low thresholds for epidural confirmation probability, respectively. These thresholds were determined offline by maximizing the Youden index on an independent validation dataset.

[0142] Step S70: Establish an online adaptive update mechanism to achieve continuous learning and individualized calibration of model parameters.

[0143] For step S70, in order to enable the system to continuously adapt to different operator habits, individual patient differences and sensor calibration drift, this invention establishes a model adaptive update mechanism based on Bayesian online learning.

[0144] Step S701: After the puncture is completed, the anesthesiologist marks the result of this puncture. If the actual resistance disappears, mark c=1 directly; if the spurious resistance disappears, mark it according to experience. Together with the enhanced feature vector sequence from this puncture Save as a new sample.

[0145] Step S702: Add new training samples to the set at fixed intervals of 10 new samples. The objective function for updating the BNN variational parameters is obtained by minimizing the evidence lower bound increment using the following variational method. Perform online fine-tuning:

[0146]

[0147] The variational posterior from the last update serves as the pseudo-prior for the current update. This is the KL divergence regularization coefficient, used to prevent catastrophic forgetting caused by online updates.

[0148] Example 2

[0149] This embodiment provides an automated epidural anesthesia confirmation system. Through high-precision pressure curve analysis and physiological negative pressure capture and identification, it achieves objective and accurate determination of the epidural space location, reducing the rate of false puncture. The automated system includes:

[0150] The data acquisition and preprocessing module is used to receive real-time contact pressure signals acquired by miniature pressure sensors and displacement signals acquired by optical encoders. It uses an adaptive piecewise linear baseline estimation algorithm, a sliding median filter, and a low-pass Butterworth filter to perform baseline removal and smoothing and noise reduction on the original signals, thereby constructing and outputting a pressure-displacement two-dimensional time series curve with a high signal-to-noise ratio.

[0151] The anatomical layer segmentation module is used to input the two-dimensional time-series curve into a pre-trained hidden Markov-Gaussian mixture model, calculate the observation probability of each anatomical tissue layer where the needle tip is located, and perform maximum a posteriori decoding through the Viterbi algorithm to automatically segment the continuous long time-series signal into sub-series segments that map specific anatomical structures such as ligaments and epidural spaces.

[0152] The multi-dimensional feature extraction module is used to extract time-domain statistical features, pressure-displacement joint features, sudden drop features of resistance disappearance events, and frequency-domain features extracted based on short-time Fourier transform for the segmented subsequence segments. The multi-dimensional features are then concatenated to construct an enhanced feature vector that characterizes the current tissue resistance state.

[0153] The template matching and initial classification module is used to extract the pressure subsequence within the current sliding window, calculate the elastic distance between it and the built-in multi-class real or false resistance disappearance curve template using the dynamic time warping algorithm, transform the distance into a preliminary classification probability distribution using the Softmax function, and add it as an auxiliary feature to the enhanced feature vector.

[0154] The Bayesian probability inference and fusion decision module is used to input the enhanced feature vector after adding auxiliary features into the Bayesian neural network, calculate the mean posterior prediction probability of each category through Monte Carlo sampling to quantify uncertainty, and use Dempster-Shafer evidence theory to fuse the posterior prediction probability of the Bayesian neural network with the aforementioned preliminary classification probability through multi-source evidence fusion to eliminate judgment conflicts and output the final true resistance disappearance fusion probability.

[0155] The risk assessment and feedback output module is used to dynamically compare the final fusion probability with the system's preset high and low thresholds for epidural confirmation, and trigger the human-computer interaction interface based on the comparison results to issue corresponding three-level clinical decision instructions: confirmation has been reached, suspected cases require manual confirmation, or non-epidural areas are accompanied by high risk of mis-puncture.

[0156] The online adaptive update module receives the actual results labeled by clinicians after puncture and stores them in a local buffer. When the accumulated samples reach a preset number, the objective function is updated incrementally by variationally minimizing the lower bound of evidence and combined with KL divergence regularization to fine-tune and calibrate the parameters of the Bayesian neural network online, solving the problem of decreased accuracy caused by differences in physician techniques and sensor drift.

Claims

1. A method for assisting epidural anesthesia based on pressure curve characteristics, characterized in that, include: (1) Collect real-time pressure signals during the puncture process and construct a pressure time series dataset; (2) Based on prior knowledge of anatomy, construct a puncture layer segmentation model to automatically identify the tissue layer where the puncture needle is located; (3) Extract the multidimensional time-frequency features of the pressure curve and construct the puncture state feature vector; (4) Based on the dynamic time warping algorithm, a multi-category pressure curve template library is constructed to perform similarity measurement and preliminary classification of hierarchical transformation patterns; (5) Construct a Bayesian neural network probabilistic inference model to determine the location of the epidural space; (6) Establish a multi-level risk assessment and real-time feedback output mechanism for clinical decision-making; (7) Establish an online adaptive update mechanism to continuously learn and individually calibrate model parameters.

2. The method for assisting epidural anesthesia according to claim 1, characterized in that, In step (1), the puncture process continuously acquires the real-time contact pressure signal p(t) between the tip of the puncture needle and the tissue interface through a miniature high-precision pressure sensor connected to the tail end of the puncture needle. At the same time, an optical encoder acquires the puncture needle advance displacement signal d(t). After all sensor data are synchronized with timestamps, they are converted from analog to digital and uploaded to the embedded signal processing unit to form a two-dimensional time-series observation pair. ; where t k For the k-th sampling time, K is the total number of sampling points in the current puncture operation.

3. The method for assisting epidural anesthesia according to claim 2, characterized in that, The pretreatment process for the puncture includes: (1) Baseline drift correction: An adaptive piecewise linear baseline estimation algorithm is used to subtract the estimated baseline from the original pressure signal to obtain the baseline-de-baseline pressure signal. ; (2) Sliding median filtering: with window length W m right Apply sliding median filtering to suppress impulse noise; (3) Low-pass Butterworth filtering: The median-filtered signal is smoothed using a zero-phase low-pass Butterworth filter to obtain the final preprocessed pressure sequence. .

4. The method for assisting epidural anesthesia according to claim 3, characterized in that, After pretreatment, the displacement is advanced using a puncture needle. With pressure as the horizontal axis Construct a two-dimensional time series curve of pressure-displacement with the vertical axis as the ordinate. This serves as the foundational data for subsequent feature extraction and recognition.

5. The method for assisting epidural anesthesia according to claim 1, characterized in that, In step (2), the puncture level segmentation model uses a hidden Markov model to constrain the hierarchical order, including: (1) Based on anatomical priors, the state transition matrix of the hidden Markov model is ordered because the puncture needle can only move forward in one direction. (2) The stress observation distribution at each organizational level is defined as a Gaussian mixture model; (3) After obtaining the stress observation probabilities at different organizational levels, the Viterbi algorithm is used to perform maximum a posteriori decoding on the observation sequence to obtain the optimal hierarchical labeling sequence; (4) Based on the optimal hierarchical labeling sequence results, the pressure-displacement curve is divided into sub-sequence segments corresponding to each anatomical level.

6. The method for assisting epidural anesthesia according to claim 1, characterized in that, In step (3), the process of constructing the puncture state feature vector specifically includes: (1) Extracting time-domain statistical features of pressure: For the segmented subsequence of the current level, calculate the time-domain statistical features; (2) Extract pressure-displacement joint features: Define unit displacement pressure and pressure-displacement curve slope sequence, and further calculate the slope mean and slope standard deviation from the pressure-displacement curve slope sequence to obtain the number of slope sign changes that characterize the frequency of pressure fluctuations; (3) Extracting features of resistance disappearance events: Define a local pressure drop detection function, and further extract the drop magnitude, drop rate and the current hidden Markov model annotation level for the detected resistance disappearance events; (4) Extracting the frequency domain and time-frequency features of the pressure curve: Perform short-time Fourier transform on the segmented subsequence of the current level to extract the power spectral density features, characterize the frequency components of pressure oscillation, and distinguish between mechanical resistance features and physiological noise components. (5) Based on all the extracted features, construct a complete puncture state feature vector.

7. The method for assisting epidural anesthesia according to claim 1, characterized in that, In step (4), the process of performing similarity measurement and preliminary classification of hierarchical transformation patterns includes: (1) Constructing a pressure curve template library: From the clinical annotation dataset, pressure-displacement sub-curve segments were extracted according to five event types to construct a template library; the five event types include the disappearance of true resistance at the normal ligamentum flavum-epidural interface, the disappearance of false resistance in the interspinous ligament space, the disappearance of false resistance in the ligamentum flavum space, the disappearance of false resistance in the paravertebral space, and the disappearance of false resistance in the muscle space; for each event type, the existing pressure-displacement curves accurately annotated by clinicians were used as the original templates, and the average template sequence for that event type was calculated using the dynamic time warping centroid averaging algorithm; (2) Real-time similarity measurement based on DTW distance: For the currently collected real-time data points, take the current sampling point as the center, take a certain length of sliding window forward to obtain the pressure subsequence, and calculate the DTW distance between it and the average sequence of various templates; (3) Preliminary classification decision based on DTW distance: Based on the calculated DTW distance vector, the preliminary category probability distribution is obtained by Softmax normalization. This distribution is used as an auxiliary feature and concatenated to the extracted puncture state feature vector to form an enhanced feature vector, which is then input into the subsequent Bayesian neural network for probability inference.

8. The method for assisting epidural anesthesia according to claim 1, characterized in that, In step (5), determining the location of the epidural space includes constructing a BNN network structure, BNN probability prediction and inference, and constructing a multi-source evidence fusion mechanism for identifying the disappearance of false resistance.

9. The method for assisting epidural anesthesia according to claim 1, characterized in that, In step (7), the process of establishing the online adaptive update mechanism includes: (1) After the puncture is completed, the anesthesiologist will mark the results of the puncture. If the real resistance disappears, it will be marked directly. If the false resistance disappears, it will be marked according to experience. The enhanced feature vector sequence of this puncture will be stored as a new sample. (2) With a fixed number of new samples as the interval, the variational parameters are fine-tuned online by updating the objective function by minimizing the lower bound of evidence in the newly added training sample set.

10. An epidural anesthesia support system based on pressure curve characteristics, characterized in that, include: Data acquisition and preprocessing module: Receives real-time contact pressure signals acquired by miniature pressure sensors and displacement signals acquired by optical encoders, and uses an adaptive piecewise linear baseline estimation algorithm, a sliding median filter, and a low-pass Butterworth filter to perform baseline removal and smoothing and noise reduction on the original signals, outputting a pressure-displacement two-dimensional time series curve with a high signal-to-noise ratio. Anatomical layer segmentation module: Input the two-dimensional time series curve into the pre-trained hidden Markov-Gaussian mixture model, calculate the observation probability of each anatomical tissue layer where the needle tip is located, and perform maximum a posteriori decoding through the Viterbi algorithm to automatically segment the continuous long time series signal into sub-sequence segments of anatomical structures such as ligaments and epidural space. Multidimensional feature extraction module: For the segmented subsequence segments, extract time-domain statistical features, pressure-displacement joint features, sudden drop features of resistance disappearance event, and frequency-domain features extracted based on short-time Fourier transform. Concatenate the multidimensional features to form an enhanced feature vector to characterize the current tissue resistance state. Template matching and initial classification module: Extract the pressure subsequence within the current sliding window, use the dynamic time warping algorithm to calculate the elastic distance between it and the built-in multi-class real or false resistance disappearance curve template, use the Softmax function to transform the distance into a preliminary classification probability distribution, and add it as an auxiliary feature to the enhanced feature vector; Bayesian Probability Inference and Fusion Decision Module: The enhanced feature vector with additional auxiliary features is input into the Bayesian neural network. The mean posterior prediction probability of each category is calculated through Monte Carlo sampling to quantify uncertainty. The Dempster-Shafer evidence theory is used to fuse the posterior prediction probability of the Bayesian neural network with the preliminary classification probability through multi-source evidence fusion, eliminate judgment conflicts, and output the final true resistance disappearance fusion probability. Risk assessment and feedback output module: Dynamically compares the final fusion probability with the system's preset high and low thresholds for epidural confirmation, and triggers the human-computer interaction interface based on the comparison results to issue corresponding three-level clinical decision instructions: confirmation has been reached, suspected case requires manual confirmation, or non-epidural area with high risk of mis-puncture. Online adaptive update module: Receives the actual results labeled by clinicians after puncture and stores them in a local buffer. When the accumulated samples reach a preset number, the objective function is updated by variational minimization of the lower bound of evidence and combined with KL divergence regularization to perform online fine-tuning and parameter calibration of the Bayesian neural network.

Images