Intelligent analysis and clinical decision support method for spinal operation real-time registration error

Abstract

The invention discloses an intelligent analysis and clinical decision support method for a spinal operation real-time registration error, and relates to the technical field of medical informatics and operation aided decision, and the method comprises the steps: obtaining multi-source heterogeneous data of a spinal operation region, and recognizing low-intensity error mode data based on local density and relative distance; and extracting a target error data fragment with time sequence and space relevance from the low-intensity error mode data by analyzing the periodicity, the space aggregation and the time sequence burstiness of the low-intensity error mode data. According to the method, the evolution trend and the spatial propagation path of the error in hundreds of milliseconds in the future can be predicted prospectively by constructing the real-time error prediction model and fusing the feature topological structure. Therefore, the system can give an early warning in advance and recommend a correction strategy before errors are accumulated to a dangerous level, a problem handling window is greatly advanced, post remedy is changed into beforehand prevention, and the active safety control capability of an operation is fundamentally improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical informatics and surgical decision support technology, in particular to an intelligent analysis and clinical decision support method for real-time registration error of spinal surgery. BACKGROUND

[0002] With the development of precision medicine and intelligent surgery, the minimally invasive and navigation dependence of spinal surgery are increasingly improved. The core of the surgical navigation system is to realize the real-time and high-precision registration and positioning of preoperative planning images and intraoperative patient anatomy. However, this process is continuously disturbed by various low-intensity and cumulative error sources, such as rhythmic micro-movement of the spine caused by respiration, soft tissue deformation caused by surgical instruments, signal drift of the navigation system itself, and noise in medical image acquisition. Although these errors are low in intensity, they have the characteristics of persistence, spatiotemporal correlation and cumulative effect. Traditional surgical navigation systems mainly focus on high-intensity and sudden errors for alarm and correction, and are difficult to effectively identify, quantify and predict the evolution of such low-intensity error patterns, thereby possibly leading to unpredictable navigation accuracy degradation in the later stage of surgery and increasing the risk of surgery. SUMMARY

[0003] The purpose of the present application is to provide an intelligent analysis and clinical decision support method for real-time registration error of spinal surgery to solve the problems in the background art.

[0004] In order to achieve the above-mentioned purpose, the present application provides the following technical scheme: an intelligent analysis and clinical decision support method for real-time registration error of spinal surgery, comprising: acquiring multi-source heterogeneous data of a spinal surgery area, and identifying low-intensity error pattern data therefrom based on local density and relative distance; from the low-intensity error pattern data, extracting target error data segments with spatiotemporal correlation by analyzing their periodicity, spatial aggregation and temporal burstiness; extracting key temporal error features based on the target error data segments, and training a real-time error prediction model for predicting error evolution; based on the distribution and correlation of error events in the low-intensity error pattern data on the anatomical structure, constructing a feature topology reflecting the error spatial propagation law; in the surgical navigation process, inputting the real-time acquired key temporal error features into the real-time error prediction model to predict future time series error features; fusing the future time series error features with the feature topology to generate a spatiotemporal fusion feature vector representing the spatiotemporal correlation of errors; based on the spatiotemporal fusion feature vector, performing comprehensive accuracy evaluation and risk analysis of the surgical area, and determining the warning level according to the evaluation result; According to the early warning level, surgical navigation decision information containing the early warning level and corresponding correction strategy is generated and output.

[0005] In a preferred embodiment, the multi-source heterogeneous data of the spinal surgery region is acquired, and low-intensity error mode data is identified therefrom based on local density and relative distance, including: Spinal movement data from an inertial measurement unit, image data from a medical imaging device, and instrument tracking data from a surgical instrument tracking system are collected; Based on a preset spinal movement intensity threshold, the spinal movement data is evaluated and analyzed to obtain a spinal movement data sequence; Based on a preset image quality intensity threshold, the image data is classified to obtain an image quality data sequence; Based on a preset instrument accuracy intensity threshold, the instrument tracking data is analyzed for reliability to obtain an instrument accuracy data sequence; The spinal movement data sequence, the image quality data sequence, and the instrument accuracy data sequence are time-aligned and data-fused to obtain a fused data matrix; A density estimation algorithm based on a Gaussian kernel function is used to calculate the local density of each data point in the fused data matrix; Based on the local density, the relative distance ratio between each data point and its nearest high-density point is calculated; A data subset with a relative distance ratio less than a set threshold is identified as the low-intensity error mode data, wherein the set threshold is a value dynamically determined based on the average distance between data points.

[0006] In a preferred embodiment, from the low-intensity error mode data, target error data segments with temporal and spatial correlation are extracted by analyzing their periodicity, spatial aggregation, and temporal burstiness, including: Based on the error data segments extracted from the low-intensity error mode data, the autocorrelation coefficient is calculated to obtain a periodicity characteristic value; A Gaussian kernel density estimation algorithm with fixed bandwidth is used to fit the spatial distribution density function of the error data segments, and the density kurtosis is used as an index of spatial aggregation degree; The standard deviation change rate of the error data segments is calculated by a sliding time window, and the peak value of the change rate is used as a temporal burstiness characteristic value; According to the periodicity characteristic value, the spatial aggregation degree index, and the temporal burstiness characteristic value, an error feature vector is constructed; The error feature vector is dimensionally reduced and reconstructed by using a deep autoencoder to obtain the target error data segment with time sequence and spatial correlation; wherein the deep autoencoder at least includes one encoder full connection layer and one decoder full connection layer.

[0007] In a preferred embodiment, the target error data segment includes a respiratory motion error data segment, a tissue deformation error data segment, a system drift error data segment, and an image noise error data segment.

[0008] In a preferred embodiment, the key time sequence error features are extracted based on the target error data segment, including: a respiratory cycle feature is calculated from the respiratory motion error data segment to obtain a respiratory motion frequency feature sequence; based on the respiratory motion frequency feature sequence, an image acquisition timestamp is obtained from the image data of the medical imaging device and an image registration error numerical sequence is calculated; based on the image registration error numerical sequence, the tissue deformation error data segment, the system drift error data segment, and the image noise error data segment are extracted and offset analyzed respectively to obtain a tissue deformation amplitude sequence, a system drift amplitude sequence, and an image noise amplitude sequence; the respiratory motion frequency feature sequence, the image registration error numerical sequence, the tissue deformation amplitude sequence, the system drift amplitude sequence, and the image noise amplitude sequence are standardized and feature screened to obtain the key time sequence error features.

[0009] In a preferred embodiment, the real-time error prediction model for predicting error evolution is trained, including: the key time sequence error features are divided by using a time series sliding window cross-validation method to obtain a training set and a validation set; an initial model based on a long short-term memory network is initialized as an initial model, and the input dimension of the initial model is the same as the dimension of the key time sequence error features; the initial model is trained by using the training set, and the validation set is input into the trained model to obtain a prediction sequence; a correlation coefficient between the prediction sequence and a real sequence corresponding to the validation set is calculated to obtain a prediction accuracy index; parameters of the initial model are optimized according to the prediction accuracy index to obtain the real-time error prediction model.

[0010] In a preferred embodiment, the feature topology reflecting the error space propagation law is constructed based on the distribution and association of error events on the anatomical structure in the low-intensity error mode data, including: Low-intensity error events are acquired, and the spatial coordinates of each node are obtained from a predefined spinal topology based on the anatomical node identifiers. The occurrence frequency of the low-intensity error events at each anatomical node is calculated to obtain the event frequency matrix; The event frequency matrix is ​​processed by network structuring to obtain an initial network structure with dissected nodes as vertices and error correlations between nodes as edges; A random walk algorithm with a fixed step size is used to traverse the initial structure of the network, and the frequency of visits between nodes is counted to construct a node transition probability matrix. The spatial distance between nodes is calculated based on the node transition probability matrix, and the nodes are aggregated and an error propagation path is established based on the spatial distance. The error propagation path is traversed using Dijkstra's shortest path search algorithm, and the path is quantified and filtered using a scoring function to obtain the characteristic topology; wherein, the scoring function is a weighted sum of the geometric mean of the transition probabilities of each edge on the path and the reciprocal of the Euclidean distance of the path.

[0011] In a preferred embodiment, during surgical navigation, the key time-series error features acquired in real time are input into the real-time error prediction model to predict future time-series error features, including: The key time-series error features obtained based on real-time multi-source heterogeneous data are input into the real-time error prediction model. The real-time error prediction model outputs predicted error feature values ​​for multiple consecutive future time points based on the temporal dependency of the key time-series error features, thus forming the future time-series error features.

[0012] In a preferred embodiment, fusing the future time series error features with the feature topology to generate a spatiotemporal fusion feature vector characterizing the spatiotemporal correlation of errors includes: The wavelet transform multi-scale decomposition of the future time series error features is performed to obtain the energy proportion at different time scales, which constitutes the time series weight vector. The connection strength of each node in the feature topology is normalized using softmax based on biomechanical constraints to obtain a spatial weight vector. The correlation weight matrix between the temporal weight vector and the spatial weight vector is calculated using a scaled dot product attention mechanism. Based on the correlation weight matrix, the future time series error features and the features of the feature topology are weighted and concatenated, and dimensionality is reduced by principal component analysis to obtain the spatiotemporal fusion feature vector. Using spinal anatomical nodes as vertices, error propagation paths in the aforementioned characteristic topology as edges, and the associated weights as edge weights, a weighted directed graph is constructed as an error event evolution map.

[0013] In a preferred embodiment, the comprehensive accuracy assessment and risk analysis of the surgical area based on the spatiotemporal fusion feature vector are performed, and a warning level is determined according to the assessment results; based on the warning level, surgical navigation decision information including the warning level and the corresponding correction strategy is generated and output, including: Based on the spatiotemporal fusion feature vector, calculate the comprehensive accuracy evaluation value of the current surgical area; Based on the aforementioned error event evolution map, the critical paths of error propagation and cumulative risks are identified; Based on the comprehensive accuracy assessment value and the cumulative risk, predict the accuracy trend at multiple future time points; The warning level is determined based on the accuracy trend and multiple preset accuracy thresholds; Based on the warning level, a corresponding correction strategy is matched from a preset strategy library; The surgical navigation decision information includes the warning level, the correction strategy, and the corresponding accuracy prediction results.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: This invention constructs a real-time error prediction model and integrates characteristic topology structures to proactively predict the evolution trend and spatial propagation path of errors within hundreds of milliseconds. This enables the system to issue early warnings and recommend correction strategies before errors accumulate to dangerous levels, significantly advancing the problem-solving window and transforming post-event remediation into pre-event prevention, fundamentally improving the proactive safety control capabilities of surgery.

[0015] This invention utilizes an identification method based on local density and relative distance to keenly capture these hidden error patterns from multi-source heterogeneous data. Furthermore, by analyzing periodicity, clustering, and suddenness, and mapping error events onto anatomical topological structures, it achieves the separation, quantification, and source localization of various low-intensity errors, laying a data foundation for precise intervention. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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, 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] Example 1, please refer to Figure 1 As shown in this embodiment, the intelligent analysis and clinical decision support method for real-time registration error in spinal surgery includes: S1. Acquire multi-source heterogeneous data of the spinal surgical area and identify low-intensity error pattern data based on local density and relative distance; S2. From the low-intensity error pattern data, by analyzing its periodicity, spatial clustering and temporal burstiness, extract the target error data fragments with temporal and spatial correlation; S3. Extract key time-series error features based on the target error data segment, and train a real-time error prediction model for predicting error evolution. S4. Based on the distribution and correlation of error events in the anatomical structure in the low-intensity error pattern data, construct a feature topology that reflects the spatial propagation law of error. S5. During surgical navigation, the key time-series error features acquired in real time are input into the real-time error prediction model to predict future time-series error features. S6. The future time series error features are fused with the feature topology to generate a spatiotemporal fusion feature vector representing the spatiotemporal correlation of errors; S7. Based on the spatiotemporal fusion feature vector, perform a comprehensive accuracy assessment and risk analysis of the surgical area, and determine the warning level based on the assessment results; S8. Based on the warning level, generate and output surgical navigation decision information containing the warning level and the corresponding correction strategy.

[0020] In one possible implementation, step S1 specifically includes: S11, Simultaneous Acquisition of Multi-Source Heterogeneous Data: During surgical navigation, data streams from the following three types of devices are simultaneously acquired via an integrated data interface at a fixed interval of 100 milliseconds: Spinal movement data: from an inertial measurement unit attached to the target spinal segment of the patient (such as the spinous process of the target vertebra), the data includes triaxial acceleration and triaxial angular velocity, and after attitude calculation, the six-degree-of-freedom displacement and rotation sequence of the segment is obtained.

[0021] Image data: acquired in real time from intraoperative imaging equipment (such as O-arm, C-arm) via the DICOM protocol using X-ray or CT scan sequences. In addition to the images themselves, the system simultaneously captures the acquisition timestamp of each frame and image quality parameters automatically calculated by the system, such as signal-to-noise ratio, contrast, and sharpness.

[0022] Instrument tracking data: Real-time acquisition of the three-dimensional spatial coordinate sequence of the tips of registered surgical instruments (such as pedicle probes and drills) from optical or electromagnetic surgical navigation systems.

[0023] S12, Data Preprocessing and Sequence Generation: The original data stream is cleaned and features are extracted to generate a standardized evaluation sequence. For spinal movement data, the magnitude of the displacement vector (i.e., the total displacement) within each sampling period is calculated and compared and evaluated with a preset first intensity threshold (e.g., displacement > 0.5 mm) to generate a binary spinal movement data sequence (1 indicates exceeding the threshold, 0 indicates not exceeding the threshold).

[0024] For image data, each frame of image is classified according to a preset second intensity threshold (e.g., signal-to-noise ratio > 20dB and contrast > 30) to generate an image quality data sequence (1 represents high quality and 0 represents low quality).

[0025] For the instrument tracking data, calculate the standard deviation of the positions of the 10 most recent sampling points as an indicator of positional stability. Based on a preset third intensity threshold (e.g., positional standard deviation > 0.3 mm), generate an instrument accuracy data sequence (1 indicates poor stability, 0 indicates good stability).

[0026] S13, Constructing a fusion matrix and density clustering identification: The three time series of the same length are time-stamped and concatenated column by column to form an N×3 fused data matrix (N is the total number of sampling points within the time window, and the three columns represent the status of motion, image quality, and instrument accuracy, respectively).

[0027] A density estimation algorithm based on the Gaussian kernel function is used to calculate the local density of each data point in the matrix. Then, for each data point, the Euclidean distance to the nearest data point with a higher local density is calculated to obtain the relative distance ratio.

[0028] A dynamic threshold is set based on the average distance between all data points (e.g., 0.6 times the average distance). Finally, a subset of data points whose relative distance ratios are less than this dynamic threshold is identified as low-intensity error pattern data. These data points typically correspond to latent, compound error states caused by breathing, micro-tremors, etc., that are not triggered by a single-dimensional threshold.

[0029] In one possible implementation, step S2 specifically includes: S21, Define the analysis window and initial cut: The system sets a sliding time window of 5 seconds, with a step size of 1 second, and slides over the low-intensity error pattern data sequence obtained in S1. Each slide extracts the original multi-source data (i.e., the displacement, image quality parameters, instrument position coordinates, etc. collected in step S11) corresponding to all data points within the window, forming an error data segment to be analyzed.

[0030] S22, Multi-dimensional Feature Extraction: For each error data segment, three types of feature indicators are computed in parallel to quantify its characteristics in different dimensions: Periodic analysis (calculating periodic eigenvalues): For the spinal displacement sequence of this segment, its autocorrelation function is calculated. The autocorrelation function describes the similarity between the sequence and itself at different time delays.

[0031] Within a delay range of 0.5 to 3 seconds (covering the typical respiratory cycle range), identify the first significant peak in the autocorrelation function. Record the delay time (in seconds) corresponding to this peak as the periodic characteristic value of that segment. For example, if a peak occurs at a delay of 1.8 seconds, it indicates that the segment exhibits a quasi-periodic fluctuation of approximately 1.8 seconds.

[0032] Spatial clustering analysis (calculating spatial clustering indexes): All data points within this segment are mapped to a three-dimensional feature space, with the three dimensions being: displacement amplitude, image signal-to-noise ratio, and instrument position stability.

[0033] A Gaussian kernel density estimation algorithm with fixed bandwidth (e.g., bandwidth parameter h=0.5) is used to fit the probability density distribution function of the data points in this three-dimensional space.

[0034] Calculate the kurtosis of the density distribution function. A higher kurtosis value indicates that the data points are more concentrated in a specific region in the feature space (e.g., concentrated in states with small displacement, moderate image quality, and moderate instrument stability), i.e., the stronger the spatial clustering. Record this kurtosis value as an indicator of the spatial clustering degree of this segment.

[0035] Temporal burst analysis (calculating temporal burst characteristic values): Within this segment, a shorter 500-millisecond sliding sub-window is set up to calculate the standard deviation of the data points within the sub-window.

[0036] Track the curve of the standard deviation over the entire 5-second segment and calculate the first difference (i.e. the rate of change) of the curve.

[0037] The maximum absolute value of the rate of change sequence is taken as the temporal burst characteristic value. The larger the value, the more severe the sudden increase or decrease in error intensity within that segment.

[0038] S23, Feature Fusion and Pattern Reconstruction: The periodic characteristic value, spatial clustering index, and temporal burst characteristic value obtained from the above calculations are combined into a three-dimensional error characteristic vector V=[period, clustering, burst].

[0039] This feature vector is then input into a pre-trained deep autoencoder network. The network structure is as follows: The encoder consists of a fully connected layer with a 3-dimensional input and a 128-dimensional output, using ReLU as the activation function.

[0040] Bottleneck layer: 16-dimensional.

[0041] The decoder consists of a fully connected layer with 16-dimensional input and 128-dimensional output, using ReLU activation, and an output layer that ultimately reconstructs the 3-dimensional structure.

[0042] This autoencoder is trained on a large amount of historical surgical data. Its goal is to learn the ability to reconstruct the original feature vector from the original feature vector to a low-dimensional bottleneck layer encoding (16 dimensions). After this process, the 16-dimensional bottleneck layer encoding output by the encoder is a digital representation of the target error data fragment extracted from the original data, after removing noise.

[0043] S24, Pattern Classification (Optional but Beneficial Implementation): The system can use this 16-dimensional encoding and a simple classifier (such as a support vector machine) to map it to common error types, for example: Respiratory motion error data segments typically exhibit a periodic characteristic value that is stable between 1.5 and 3 seconds, moderate clustering, and low suddenness.

[0044] Tissue deformation error data segments: typically without significant periodicity, clustering may vary, and suddenness is moderate (occurring with each step of the operation).

[0045] System drift error data segment: periodic characteristic values ​​are extremely low or invalid, the degree of clustering changes slowly, and the suddenness is low but persistent.

[0046] Image noise error data segments: periodic feature values ​​are invalid, clustering is low (data points are scattered), and burstiness is high and irregular.

[0047] In one possible implementation, step S3 specifically includes: Step S3: Extract key time-series error features based on the target error data segment, and train a real-time error prediction model for predicting error evolution.

[0048] This step aims to transform the structured error pattern fragments obtained in S2 into engineering features that can be used for time series forecasting, and to train a model capable of proactively predicting error evolution. Its implementation includes the following two core stages: Phase 1: Extraction of temporal feature sequences from pattern fragments; This stage involves in-depth analysis of the target error data segments (especially respiratory motion error data segments) obtained from S2 classification, deriving a set of interconnected key temporal error features.

[0049] Generate respiratory motion frequency feature sequence: The system identifies each data segment classified as respiratory motion error and parses the dominant periodic feature value T, in seconds, from its 16-dimensional encoding.

[0050] Calculate the instantaneous respiratory rate F = 60 / T (breaths / minute) for this segment. Assign the calculated respiratory rate value to the center timestamp of this segment, ultimately generating a respiratory frequency characteristic sequence that varies over time. breathe.

[0051] Generate a numerical sequence of image registration errors: by Using the respiratory sequence as a reference, specific phase points (such as end of inspiration and end of expiration) are located within each respiratory cycle.

[0052] Near the timestamps corresponding to these key phase points, the images actually acquired by the medical imaging equipment are queried. By comparing the actual instrument position provided by the instrument tracking system at that moment with the planned instrument target position in the preoperative CT / MR images, the three-dimensional spatial registration error (Euclidean distance) is calculated.

[0053] The calculated registration error value (unit: mm) is associated with its corresponding timestamp to generate an image registration error numerical sequence. Registration. This sequence directly reflects the periodic effect of the respiratory cycle on the accuracy of surgical navigation.

[0054] Other error magnitude sequences can be derived: by Using the registration sequence as a reference time axis and error baseline, the system further analyzes other types of error data segments: Organizational deformation amplitude sequence Deformation: in The time points at which registration shows a non-periodic abnormal increase are analyzed to determine the coding intensity of tissue deformation error data segments and quantify the additional offset caused by them.

[0055] System drift amplitude sequence Drift: A moving average calculated over a long time window (e.g., 30 seconds) from... The slowly changing trend term is separated during registration and used as an estimate of the system's cumulative drift.

[0056] Image noise amplitude sequence Noise: Correlation calculations are performed within time intervals where image quality data sequences are marked as low quality. The instantaneous fluctuation variance of registration is the magnitude of random error introduced by image noise.

[0057] Feature standardization and screening: The above five sequences [ breathe, Registration deformation, drift, [Noise] is Z-score normalized to eliminate the influence of dimensions.

[0058] Principal component analysis or tree-based feature importance assessment are used to screen features that are important for predicting the future. The most critical features of the registration error together constitute the final key time-series error feature matrix, with the dimension of [time step t × number of features n].

[0059] Phase Two: Training and Optimization of the Real-Time Error Prediction Model; In this stage, a prediction model based on a long short-term memory network is trained using key temporal error feature matrices extracted from historical surgical data.

[0060] Data partitioning: A time-series sliding window cross-validation method is employed. A sliding window of length L = 30 time steps (e.g., corresponding to 30 seconds) is set as the model input to predict the future H = 5 time steps (e.g., the next 5 seconds). Registration value.

[0061] All time-series data were divided into training and validation sets in an 8:2 ratio to ensure that the validation set was located after the training set on the timeline, in order to simulate real-time prediction scenarios.

[0062] Model building and training: Initialize a prediction model based on a Long Short-Term Memory (LSTM) network as the initial model. Its input layer dimension is equal to the number of key temporal error features n. The network contains two stacked LSTM layers, each with 64 hidden units, and is finally connected to a fully connected output layer with an output dimension of H (prediction stride).

[0063] The model is trained using a sliding window of samples from the training set, with the mean squared error as the loss function. After training, the validation set data is input into the model to obtain the corresponding predicted sequences.

[0064] Model optimization and solidification: The Pearson correlation coefficient between the predicted sequence and the true sequence on the validation set is calculated and used as the core indicator of prediction accuracy.

[0065] If the metric is lower than the preset target (e.g., correlation coefficient > 0.85), optimization can be achieved by adjusting network hyperparameters (such as learning rate, Dropout rate) or increasing the number of training epochs.

[0066] Once the model achieves stable and excellent performance on the validation set, its parameters are solidified and deployed as the final real-time error prediction model, ready for real-time use during surgery.

[0067] In one possible implementation, step S4 specifically includes: S41, Spatial mapping and initial statistics of error events: Define anatomical nodes: The patient's spinal anatomy is pre-digitized into a topological map, where each vertebra (e.g., from thoracic T1 to lumbar L5, a total of 17 nodes) is defined as a unique anatomical node and its three-dimensional spatial coordinates based on preoperative CT images are stored.

[0068] Event Mapping: For each low-intensity error pattern data point identified in S1, its timestamp is used to associate it with the 3D coordinates of the instrument tip reported by the surgical instrument tracking system at that moment. Using a nearest neighbor search algorithm, these spatial coordinates are mapped to the nearest spinal anatomical node, recording the occurrence of a low-intensity error event at that node.

[0069] Generate an event frequency matrix: Count the total number of low-intensity error events occurring at each anatomical node within a sufficiently long observation window (such as a key stage of the entire surgery), forming a 17×1 event frequency vector F. Further, calculate the co-occurrence frequency or temporal correlation frequency between pairs of nodes, forming a 17×17 event frequency (correlation) matrix M, where elements... This represents the error correlation strength from node i to node j.

[0070] S42, Constructing the initial network structure and random walk learning: Network initialization: Using 17 dissected nodes as vertices, and the normalized value of the event frequency matrix M (or the initial weights calculated based on the three-dimensional Euclidean distance and co-occurrence relationship between nodes) as the initial weights of the edges, a weighted, undirected (or initially directed) network initial structure G0 is constructed.

[0071] Random walk simulation of propagation: A fixed-step random walk algorithm is used to simulate propagation on G0. The walk step size is set to S=1000, and multiple walks are performed starting from each node. In each walk, the probability of moving to the next node is proportional to the weights of the edges connected to the current node. The frequency of consecutive visits between all node pairs is statistically analyzed throughout the simulation to capture the potential one-step propagation tendency of the error along the anatomical structure.

[0072] Constructing the transition probability matrix: Based on the access frequency statistics in the previous step, a 17×17 node transition probability matrix P is calculated.

[0073] Elements in matrix P This represents the one-step transition probability of the error's influence propagating from cone i to cone j under low-intensity error conditions. This matrix is ​​the core mathematical expression of the propagation law in the learning space.

[0074] S43, Path Discovery, Quantization, and Topology Generation: Calculating propagation distance and node aggregation: Based on the transition probability matrix P, the propagation distance from node i to j is defined as follows: -log (The lower the probability, the greater the distance). Using this distance metric, a hierarchical clustering algorithm was used to aggregate the 17 nodes, clustering closely related vertebrae (such as adjacent segments) into larger functional groups, and initially identifying spinal regions that may be synchronously affected.

[0075] Critical path search: Using the transition probability matrix P as edge weights (here, probability is considered as propagation rate, and higher probability means higher weight), the Dijkstra shortest path search algorithm is used to calculate the optimal propagation path between any two dissected nodes in the network. This path is defined as the shortest path in the sense of the path along which error is most easily propagated.

[0076] Path Quantification and Screening: A scoring function is designed to quantitatively evaluate and screen all potential propagation paths found. This function comprehensively considers the propagation efficiency and spatial rationality of the paths. Example of a scoring function: Score(path) = a * GeoMean( )+ b * (1 / EuclideanLength(path)); in, Score(path) Representing a pathpath The rating; Let be the set of transition probabilities for all edges on this path; GeoMean(·) It is a geometric mean function used to calculate the overall propagation probability of a path; EuclideanLength(path) The Euclidean distance between the start and end points of the path in three-dimensional space (unit: millimeters) is the reciprocal of which penalizes excessively long paths that do not conform to the common sense of anatomical connections. α and β are weighting coefficients used to balance the weights of transmission efficiency and spatial rationality, and satisfy α+β=1; GeoMean( ) : The geometric mean of the transition probabilities of all edges on the path, reflecting the overall propagation difficulty of the path (the higher the value, the better).

[0077] Generating Feature Topology: All critical paths with scores above a threshold are selected and, together with anatomical nodes and the transition probability matrix, constitute the final feature topology. This structure is a weighted directed graph that clearly and quantitatively reveals the spatial propagation patterns of low-intensity errors on specific spinal anatomy structures of the patient. For example, it may clearly show the probability path and intensity of error propagation from the L3 vertebra to the L1 vertebra, which is much higher than that of direct propagation to the T12 vertebra.

[0078] In one possible implementation, step S5 specifically includes: S51, Real-time Data Stream Processing and Feature Extraction: During surgical navigation, the system executes a real-time data processing pipeline in parallel: Continuously execute step S1, collect multi-source data at 100ms intervals, and identify the latest low-intensity error pattern data stream in real time.

[0079] Simultaneously execute step S2 to perform sliding window analysis on the newly identified low-intensity error data and output the latest target error data fragments and their classification results (such as respiratory motion fragments and tissue deformation fragments) in real time.

[0080] The feature extraction process in stage one of the linked execution steps S3, based on the latest error fragment, continuously calculates and updates the five key time series sequences. breathe, Registration deformation, drift, (noise), and then perform the same standardization and screening process on it.

[0081] S52, Construct the model input vector: The system maintains a first-in-first-out queue as a real-time feature buffer. Whenever a new feature is calculated (i.e., a new time step begins), a new feature is added. The system will: Take out from -29 to All key timing error features of the most recent 30 consecutive time steps (corresponding to a historical window of approximately 3 seconds).

[0082] Arrange these 30 sets of features in chronological order to construct a model input tensor of dimension [30, n]. , where n is the number of features (e.g., 5). This tensor characterizes the latest and most relevant error evolution observed by the system.

[0083] S53, performs real-time forward prediction: The constructed The tensor is input into the real-time error prediction model (i.e., an LSTM-based neural network) that has been trained and deployed in S3.

[0084] The two-layer LSTM unit inside the model processes the 30-step historical information based on the long-term temporal dependencies it has learned, updates its internal hidden state, and remembers the current error evolution pattern.

[0085] The fully connected output layer of the model directly outputs the results for the next 5 consecutive time steps (e.g., the future hidden state) based on the final hidden state. +1 to +5 (corresponding to the predicted value 0.5 seconds in the future). The output is a vector of dimension [5]. Each value represents a predicted value of the core error feature at the corresponding future time (usually...). Predicted registration error, in millimeters.

[0086] S54, forming future time series error characteristics: The model output A vector, aligned with timestamps for the next 5 points in time.

[0087] Simultaneously, the system will collaboratively predict other relevant features (such as...) based on the model's internal state and historical trends. deformation, (Drift) The possible value at the same time in the future. For example, if Registration is predicted to increase, so what is associated with it? The magnitude of deformation can also be used to calculate the increase.

[0088] All predicted values ​​for various error features at the next five time points are organized to form a structured set of future time series error features. For example, a complete prediction output might be shown in Table 1 below: Table 1. Prediction of Error Characteristics in Future Time Series

[0089] Execute in a loop: Steps S51 to S54 above are automatically triggered once at the end of each new data sampling period (100ms), thereby achieving rolling, online, and continuous error prediction, providing the surgical navigation system with a risk preview within half a second ahead of the actual occurrence of the error.

[0090] In one possible implementation, step S6 specifically includes: S61, multi-scale analysis in the time dimension, generating time-series weight vectors: The future time series error features obtained from S5 (based on the core) The registration prediction sequence is used as the main input, and continuous wavelet transform is performed. The Morlet wavelet is selected as the mother wavelet and decomposed into multiple time scales, such as three scales corresponding to instantaneous jitter (0.1-0.5 seconds), short-period fluctuation (0.5-2 seconds), and trend term (2-5 seconds).

[0091] Calculate the energy percentage of the wavelet coefficients at each time scale. For example, a typical decomposition might result in: instantaneous jitter energy accounting for 15%, short-period fluctuation energy accounting for 70%, and trend term energy accounting for 15%. This distribution reveals the main driving forces behind future error evolution.

[0092] The energy proportions at these three scales are constructed into a time-series weight vector. =[0.15, 0.70, 0.15]. This vector quantifies the importance of dynamics at different time scales in future early warnings.

[0093] S62, normalization of biomechanical constraints in the spatial dimension, generating a spatial weight vector: Read the feature topology constructed by S4, especially the node transition probability matrix P and the critical propagation path. Extract the connection strength (i.e., transition probability) of all edges. This serves as the initial spatial association.

[0094] Biomechanical constraints are introduced to correct the initial connection strength. Constraint factors are defined based on the anatomical connectivity and biomechanical coupling of the spine. For example, the connection between adjacent vertebrae (e.g., L4-L5) is multiplied by a strengthening factor of 1.2; the connection between two vertebrae (e.g., L4-L3) is multiplied by a factor of 1.0; and the connection between vertebrae at greater distances is multiplied by a weakening factor of 0.8.

[0095] All the corrected connection strength values ​​are normalized using a Softmax function to ensure that the sum of all values ​​is 1. The result constitutes a spatial weight vector. Its dimension is equal to the number of all considered node pairs or paths. This vector reflects the prior probability distribution of error propagation at different locations in space under biomechanical constraints.

[0096] S63, Spatiotemporal correlation weight calculation (via attention mechanism): Time weight vector and spatial weight vector These are respectively considered as query and key.

[0097] The correlation between the two is calculated using a scaled dot product attention mechanism. Specifically, firstly, the correlation between the two is calculated using a scaled dot product attention mechanism. and The scaled dot product score is then calculated by projecting different learnable linear layers onto the same dimension. The formula is: (Q, K, V) = softmax( )·V , where Q= · K= · V= · , This is the scaling factor.

[0098] in, and The temporal weight vector and spatial weight vector generated in the preceding steps are used as inputs to the attention mechanism. , , This is the linear projection weight matrix learned during model training; Q, K, V These are the query, key, and value matrices obtained after projection from the input, where... This represents the transpose operation on the key matrix K; The dimension of the key vector is used as a scaling factor to stabilize training. softmax This is a normalized exponential function used to generate attention weights; (Q, K, V) This scales the output of the dot product attention function; The calculated attention weight matrix A is the correlation weight matrix. The physical meaning of the element A[i][j] in the matrix is: the correlation strength between the error dynamics at the i-th time scale and the j-th spatial propagation path. For example, a high value in the matrix may indicate that short-period fluctuations (i=2) are highly correlated with the path from L3 to L4 (j=a certain index).

[0099] S64, Feature-weighted concatenation and dimensionality reduction (generating spatiotemporal fusion feature vectors): Weighted concatenation: Using the correlation weight matrix A as the weight, future temporal features (wavelet coefficients at multiple scales) and spatial topological features (such as critical path scores and node clustering coefficients) are weighted and fused to form a high-dimensional temporary feature vector.

[0100] Principal Component Analysis (PCA) Dimensionality Reduction: Due to the high dimensionality and potential redundancy of the concatenated features, PCA is used for dimensionality reduction. The top k principal components with a cumulative contribution rate exceeding 95% are selected, mapping the high-dimensional features to a low-dimensional, orthogonal, and information-condensed vector space.

[0101] The final output spatiotemporal fusion feature vector is a k-dimensional vector (e.g., k=32). This vector is the system's ultimate digital summary of the future spatiotemporal error state, simultaneously encoding the future intensity of the error, its evolution rate, and the most likely propagation path.

[0102] S65, construct an error event evolution graph and visualize the output: As a direct result of this step, the system also generates a dynamic error event evolution map: Vertex: Anatomical node of the spine.

[0103] Edges: Key error propagation paths in the feature topology.

[0104] Edge weight / visual attributes: determined by the weight value corresponding to the path in the association weight matrix A, usually represented by the thickness and color intensity of the edge.

[0105] Vertex status: Based on the magnitude of the error in future predictions, the risk level that each node will face is represented by color (e.g., blue-yellow-red).

[0106] This figure visually illustrates the complete spatiotemporal prediction conclusion that, similar to the prediction that the error at the L4 vertebra will reach a high risk (red) after 0.3 seconds, and that there is a high probability that it will spread to the L3 and L5 vertebrae along the bolded path in the figure.

[0107] In one possible implementation, step S7 specifically includes: S71, Comprehensive accuracy assessment based on spatiotemporal feature vectors; The system inputs a k-dimensional spatiotemporal fusion feature vector (e.g., k=32) into a pre-trained comprehensive accuracy evaluation regression model. This model typically employs support vector regression or gradient boosting decision tree regression algorithms.

[0108] Model Functionality: This model learns to map complex spatiotemporal feature patterns to a single, global accuracy metric. Its output is a scalar called the comprehensive accuracy assessment value E, which can be intuitively understood as the expected registration error of the core surgical area (such as the pedicle screw track) under the current and predicted spatiotemporal error conditions, in millimeters.

[0109] Calculation example: For a specific spatiotemporal feature vector as input, the model might output E=1.8mm. This means that based on all analyses, the system predicts that, under the current circumstances, the accuracy of the critical surgical site will deviate from the ideal position by approximately 1.8 mm.

[0110] S72, Risk analysis based on error event evolution map; At the same time, the system performs real-time analysis on the error event evolution map generated by S6 to identify and quantify specific risks in the spatial dimension.

[0111] Identifying critical propagation paths: The algorithm traverses all edges in the graph and selects the top 2-3 paths with the highest edge weights (association weights) as the most active critical error propagation paths. For example, the system may identify paths P1: L4→L3 and P2: L4→L5 as critical paths.

[0112] Calculate the cumulative risk value: For each critical path, calculate a path risk score: = (Average prediction error of each vertex on the path) × (Average weight of the path edges) × (Path length factor).

[0113] The risk scores of all critical paths are weighted and summed to obtain a global cumulative risk value R. The higher the R value, the stronger and wider the tendency of error to spread across the anatomical structure, and its harm may exceed that of local high error values.

[0114] Example of risk tracing: Combining atlas and calculations, the system may generate the following risk analysis conclusion: The main risk is concentrated in the L4 vertebral body, with a prediction error of 1.9 mm (red). High risk is spreading at an intensity of 0.85 along two main pathways: one towards the cephalic L3 (expected to increase risk by 30%), and the other towards the caudal L5 (expected to increase risk by 25%).

[0115] S73, integrating assessment and risk assessment early warning levels; Ultimately, the system combines quantitative accuracy assessment with qualitative risk analysis, and determines the final warning level through a multi-threshold decision function.

[0116] Decision function inputs: The core inputs are the comprehensive accuracy assessment value E and the cumulative risk value R.

[0117] Multi-level threshold determination: The system presets the following two-parameter threshold table to map to different warning levels, as shown in Table 2 below: Table 2 Warning Level Thresholds

[0118] Judgment logic: The system compares the (E, R) pairs calculated in real time with the table above. For example, if E=1.8mm and R=2.8, then both the Level II accuracy condition and the risk condition are met, so it is judged as a Level II (yellow) warning.

[0119] Predicting future trends: Based on the E and R values ​​of the most recent periods, the system will also use simple linear extrapolation or filtering algorithms to predict the accuracy and risk trends of the next 2-3 periods (about 0.2-0.3 seconds), and add trend arrows (such as ↗ to indicate continued deterioration) after the warning level to provide more forward-looking information.

[0120] In one possible implementation, step S8 specifically includes: S81, intelligent strategy library and matching engine; The system has a built-in structured clinical correction strategy library. This library uses warning levels as the primary index and specific contexts (such as main error types and risk locations) as the secondary index, storing standardized decision items jointly developed by senior spine surgeons and engineers.

[0121] Example of a strategy library structure: Level I (Green) Strategy: {Display: Navigation accuracy is good, maintained at <1.0mm. Recommendation: Continue current operation.} Level II (Yellow) Strategy: Subdivided according to the dominant error source: Mode A (Periodic Error Dominant): {Display: Warning: Respiratory motion has increased the periodic error to [E] mm. Recommendation: It is recommended to perform critical steps at the trough of the next respiratory cycle (approximately [T] seconds later).} Mode B (Deformation Error Dominant): {Display: Warning: Tissue deformation in region L4 has resulted in an error of [E] mm. Recommendation: Slightly withdraw the instrument and wait for tissue relaxation before repositioning; or consider enabling the real-time deformation compensation algorithm.} Mode C (Diffusion Risk Dominant): {Display: Warning: Error at L3 shows a tendency to diffuse towards L2 / L4. Recommendation: It is recommended to suspend operations at L3 and prioritize stabilizing or correcting this node to prevent the error range from expanding.} Level III (Red) Strategy: {Display: High-risk alert! The overall error has reached [E] mm and is spreading. Recommendation: Immediately suspend all operations.}

[0122] Suggested sequence: Verify and re-register the device; Check the field of view of the optical tracker; Consider performing a local O-arm scan for re-registration. Dynamic matching: The matching engine receives the warning level, dominant error type, key risk location (such as L4), comprehensive accuracy value E, and diffusion path from the S7 as input. The engine searches the policy library and replaces the placeholders (such as [E], [T]) in the general policy template with the specific values ​​calculated in real time (such as 1.8mm, 2.3 seconds), generating a personalized correction policy text.

[0123] S82, Synthesis and output of decision information: The generated decision information is a structured data packet, which is presented on the surgical navigation interface in a multimodal manner.

[0124] Information packet structure: The final output surgical navigation decision information should contain at least the following core fields: Warning level: such as Level II (yellow).

[0125] Prediction accuracy: For example, the current / prediction error is 1.8mm (↗).

[0126] Risk Summary: Main risk source: L4 vertebral body; Key diffusion path: L4→L3.

[0127] Correction strategy: Detailed suggestion text generated from matching with the strategy library.

[0128] Timestamp: The time when the decision was generated.

[0129] Multimodal interface output: Visual layer (overlay on the main screen): A colored warning bar (green / yellow / red) is continuously displayed on the side or corner of the navigation screen, clearly showing the warning level and a summary of the most important correction strategy (such as accuracy warning - waiting for the breathing trough).

[0130] On the 3D anatomical model, the error event evolution map is directly overlaid and rendered: high-risk vertebrae are highlighted with corresponding colors, and key diffusion paths are displayed as luminous arrow animations.

[0131] Auditory layer: A Level II warning triggers a medium-pitched alert tone.

[0132] A Level III alert triggers a repetitive, high-pitched alarm until a doctor confirms the action.

[0133] Log layer: All decision information, along with all spatiotemporal fusion feature data at the time of triggering, is encrypted and automatically stored in the surgical black box log for postoperative review, research, and system optimization.

[0134] S83, example of closed-loop operation; Assuming the system issues a Level II (yellow) warning, the dominant mode is respiratory motion error, the critical site is L4, the prediction error is 1.8 mm, and the predicted respiratory trough is 2.3 seconds later, then the doctor will see / hear the following on the navigation screen: A yellow warning bar illuminates at the edge of the screen, displaying the text: Level II Warning: Respiratory movement is affected. It is recommended to place the pin when breathing dips after 2.3 seconds.

[0135] The L4 vertebral body is highlighted in yellow on the 3D model.

[0136] At the same time, the system may announce via voice: "Caution, breathing may be affected."

[0137] Based on this information, the doctor pauses the current action, observes the respiratory cycle indicator that may be displayed simultaneously on the screen, and performs the pin placement operation at the trough time about 2.3 seconds later, thus effectively avoiding the predicted error peak.

[0138] To fully disclose the algorithm model involved in this invention, the training and construction process of the core model is described in detail below.

[0139] Deep autoencoder training example: Training data source and preparation: Approximately 150,000 samples of labeled "error feature vectors" were extracted from a database of over 300 historical spinal navigation surgeries, processed in steps S1-S2. Among them, the four types of samples—respiratory motion, tissue deformation, system drift, and image noise—were basically balanced.

[0140] Network Structure and Training: The aforementioned structure was adopted (128-dimensional encoder, 16-dimensional bottleneck layer, 128-dimensional decoder). The Adam optimizer was used with an initial learning rate of 0.001 and a batch size of 256. Training was performed iteratively for 500 epochs on the training set. The loss function was mean squared error (MSE).

[0141] Training results: The model achieved approximately 95% reconstruction accuracy on the independent test set (i.e., the MSE loss between input and output is less than 0.05). The 16-dimensional encoded vector processed by this autoencoder improved the average classification accuracy of four classes of errors to 92% in subsequent classification tasks (such as SVM classification), which is significantly higher than the results of directly using the original 3D feature vector (average accuracy of approximately 78%), demonstrating its powerful feature extraction and noise reduction capabilities.

[0142] Training example of a real-time error prediction model (LSTM): Training data generation: Based on 50 surgical records containing complete multi-source data streams, execute steps S1-S5 of this invention to generate a "key temporal error feature matrix" and its corresponding sequence of real error values ​​for the next 5 steps (500 milliseconds), which constitutes a supervised learning dataset.

[0143] Model training and validation: A two-layer stacked LSTM (64 hidden units per layer) is used, followed by a 5-dimensional fully connected output layer. Strict time-series sliding window cross-validation is employed to ensure no temporal overlap between the training and validation sets. Mean squared error is used as the loss function, and early stopping is implemented to prevent overfitting.

[0144] Optimization and Consolidation: The final model's Pearson correlation coefficient between the predicted and actual sequences on the validation set stabilizes above 0.87, and the mean absolute error (MAE) is less than 0.15 mm. When the correlation coefficient is not lower than 0.85 for 10 consecutive validation periods, the model is considered converged, and its parameters are consolidated as the final deployment model.

[0145] Training example for comprehensive accuracy evaluation model: Data annotation: Three senior spinal surgeons jointly reviewed and annotated hundreds of "high-precision moments" (expected error <1.0mm) and "low-precision moments" (expected error ≥2.0mm) based on surgical videos and navigation logs, which served as the true value labels for model training.

[0146] Model selection and training: The Gradient Boosting Decision Tree (GBDT) algorithm is used. The spatiotemporal fusion feature vector (k=32) generated in step S6 is used as input, and the labeled accuracy level is used as output for training.

[0147] Performance: The trained GBDT model achieved a classification accuracy of 94% for precision states in cross-validation, and can reliably map spatiotemporal fusion features into a quantified comprehensive accuracy evaluation value E.

[0148] Example of experimental effect verification; To demonstrate the beneficial effects of the present invention, the following comparative experiments were conducted.

[0149] Experimental setup: Control group: The threshold alarm method of the traditional surgical navigation system was used. That is, an alarm was triggered when the real-time registration error exceeded a fixed threshold (2.0 mm).

[0150] Experimental group: The intelligent analysis and decision support method of this invention was adopted.

[0151] Test data: The test was conducted using a novel spinal surgery simulation data stream of 20 cases that were not involved in any model training.

[0152] Evaluation indicators: ① Warning lead time: the time difference between the system's first warning and the actual error exceeding the 2.0mm safety threshold; ② False alarm rate; ③ Peak error suppression rate.

[0153] Experimental results: Early warning lead time: The average early warning lead time for the experimental group was 520 ± 120 milliseconds. In contrast, the control group only triggered an alarm after the error exceeded the limit, with an early warning lead time of 0 milliseconds. This invention provides a significant early warning window.

[0154] Early warning accuracy: While providing advanced early warnings, the experimental group kept the false alarm rate below 5%, which was comparable to the false alarm rate of the control group (about 8%), demonstrating the reliability of the prediction.

[0155] Error control effect: In the experimental group, where warnings were issued and corrective strategies (such as respiratory synchronization) were implemented, the peak registration error of subsequent operation cycles was reduced by an average of approximately 35%. In contrast, the control group, lacking predictive ability, was unable to achieve this type of active inhibition.

[0156] Value of risk tracing: The "error event evolution map" generated by this invention successfully identified the main paths of subsequent error propagation in 100% of the test cases, proving the effectiveness of spatial topology analysis.

[0157] Experimental conclusion: Comparative experimental data fully demonstrates that, compared with traditional threshold alarm methods, this invention, through spatiotemporal fusion intelligent predictive analysis, can provide effective early warnings before errors accumulate to dangerous levels, and actively suppress error growth through precise correction strategies, achieving a qualitative leap from "passive response" to "active prevention," significantly improving the proactive safety control capabilities of surgical navigation.

[0158] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An intelligent analysis and clinical decision support method for real-time registration error in spinal surgery, characterized in that, include: Acquire multi-source heterogeneous data of the spinal surgical area and identify low-intensity error pattern data based on local density and relative distance; From the low-intensity error pattern data, by analyzing its periodicity, spatial clustering and temporal burstiness, target error data segments with temporal and spatial correlations are extracted; Key time-series error features are extracted based on the target error data fragments, and a real-time error prediction model for predicting error evolution is trained. Based on the distribution and correlation of error events in the anatomical structure in the low-intensity error pattern data, a characteristic topology structure reflecting the spatial propagation law of error is constructed. During surgical navigation, the key time-series error features acquired in real time are input into the real-time error prediction model to predict future time-series error features. The future time series error features are fused with the feature topology to generate a spatiotemporal fusion feature vector characterizing the spatiotemporal correlation of errors; Based on the spatiotemporal fusion feature vector, a comprehensive accuracy assessment and risk analysis of the surgical area are performed, and the warning level is determined according to the assessment results; Based on the warning level, surgical navigation decision information containing the warning level and the corresponding correction strategy is generated and output.

2. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 1, characterized in that: The acquisition of multi-source heterogeneous data of the spinal surgical area, and the identification of low-intensity error pattern data based on local density and relative distance, includes: It collects spinal motion data from the inertial measurement unit, image data from medical imaging equipment, and instrument tracking data from the surgical instrument tracking system; Based on a preset spinal movement intensity threshold, the spinal movement data is evaluated and analyzed to obtain a spinal movement data sequence. Based on a preset image quality intensity threshold, the image data is classified to obtain an image quality data sequence. Based on a preset instrument accuracy and strength threshold, the instrument tracking data is subjected to reliability analysis to obtain an instrument accuracy data sequence. The spinal movement data sequence, the image quality data sequence, and the instrument precision data sequence are time-aligned and fused to obtain a fused data matrix. A density estimation algorithm based on Gaussian kernel function is used to calculate the local density of each data point in the fused data matrix; Based on the local density, calculate the ratio of the relative distance between each data point and its nearest high-density point; The subset of data whose relative distance ratio is less than a set threshold is identified as the low-intensity error pattern data, wherein the set threshold is a value dynamically determined based on the average distance between the data points.

3. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 2, characterized in that: The step of extracting target error data segments with temporal and spatial correlations from the low-intensity error pattern data by analyzing its periodicity, spatial clustering, and temporal burstiness includes: Based on the error data segments extracted from the low-intensity error pattern data, their autocorrelation coefficients are calculated to obtain periodic characteristic values; The spatial distribution density function of the error data segment is fitted by a Gaussian kernel density estimation algorithm with fixed bandwidth, and density kurtosis is used as an index of spatial clustering. The standard deviation rate of change of the error data segment is calculated by a sliding time window, and the peak value of the rate of change is used as the time-series burst characteristic value. Based on the periodic characteristic value, spatial clustering index, and temporal burst characteristic value, an error feature vector is constructed; A deep autoencoder is used to reduce the dimensionality and reconstruct the error feature vector to obtain the target error data segment with temporal and spatial correlation; wherein the deep autoencoder contains at least one encoder fully connected layer and one decoder fully connected layer.

4. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 3, characterized in that: The target error data segments include respiratory motion error data segments, tissue deformation error data segments, system drift error data segments, and image noise error data segments.

5. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 4, characterized in that: The extraction of key time-series error features based on the target error data segment includes: Calculate the respiratory cycle features obtained from the respiratory motion error data segment to obtain a respiratory motion frequency feature sequence; Based on the respiratory motion frequency feature sequence, the image acquisition timestamp is obtained from the image data of the medical imaging device, and the image registration error numerical sequence is calculated. Based on the image registration error numerical sequence, the tissue deformation error data segment, system drift error data segment, and image noise error data segment are extracted and offset analysis is performed respectively to obtain the tissue deformation amplitude sequence, system drift amplitude sequence, and image noise amplitude sequence; The key temporal error features are obtained by standardizing and feature filtering the respiratory motion frequency feature sequence, image registration error numerical sequence, tissue deformation amplitude sequence, system drift amplitude sequence, and image noise amplitude sequence.

6. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 1, characterized in that: The training yields a real-time error prediction model for predicting error evolution, including: The key time-series error features are divided using a time-series sliding window cross-validation method to obtain a training set and a validation set; Initialize a prediction model based on a long short-term memory network as an initial model, wherein the input dimension of the initial model is the same as the dimension of the key temporal error feature; The initial model is trained using the training set, and the validation set is input into the trained model to obtain a prediction sequence; Calculate the correlation coefficient between the predicted sequence and the real sequence corresponding to the validation set to obtain the prediction accuracy index; The parameters of the initial model are optimized based on the prediction accuracy index to obtain the real-time error prediction model.

7. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 1, characterized in that: The method of constructing a feature topology reflecting the spatial propagation law of errors based on the distribution and correlation of error events in the anatomical structure of the low-intensity error pattern data includes: Low-intensity error events are acquired, and the spatial coordinates of each node are obtained from a predefined spinal topology based on the anatomical node identifiers. The occurrence frequency of the low-intensity error events at each anatomical node is calculated to obtain the event frequency matrix; The event frequency matrix is ​​processed by network structuring to obtain an initial network structure with dissected nodes as vertices and error correlations between nodes as edges; A random walk algorithm with a fixed step size is used to traverse the initial structure of the network, and the frequency of visits between nodes is counted to construct a node transition probability matrix. The spatial distance between nodes is calculated based on the node transition probability matrix, and the nodes are aggregated and an error propagation path is established based on the spatial distance. The error propagation path is traversed using Dijkstra's shortest path search algorithm, and the path is quantified and filtered using a scoring function to obtain the characteristic topology; wherein, the scoring function is a weighted sum of the geometric mean of the transition probabilities of each edge on the path and the reciprocal of the Euclidean distance of the path.

8. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 5, characterized in that: During surgical navigation, the key time-series error features acquired in real time are input into the real-time error prediction model to predict future time-series error features, including: The key time-series error features obtained based on real-time multi-source heterogeneous data are input into the real-time error prediction model. The real-time error prediction model outputs predicted error feature values ​​for multiple consecutive future time points based on the temporal dependency of the key time-series error features, thus forming the future time-series error features.

9. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 1, characterized in that: The step of fusing the future time series error features with the feature topology to generate a spatiotemporal fusion feature vector characterizing the spatiotemporal correlation of errors includes: The wavelet transform multi-scale decomposition of the future time series error features is performed to obtain the energy proportion at different time scales, which constitutes the time series weight vector. The connection strength of each node in the feature topology is normalized using softmax based on biomechanical constraints to obtain a spatial weight vector. The correlation weight matrix between the temporal weight vector and the spatial weight vector is calculated using a scaled dot product attention mechanism. Based on the correlation weight matrix, the future time series error features and the features of the feature topology are weighted and concatenated, and dimensionality is reduced by principal component analysis to obtain the spatiotemporal fusion feature vector. Using spinal anatomical nodes as vertices, error propagation paths in the aforementioned characteristic topology as edges, and the associated weights as edge weights, a weighted directed graph is constructed as an error event evolution map.

10. The intelligent analysis and clinical decision support method for real-time registration error in spinal surgery according to claim 9, characterized in that, The procedure involves a comprehensive accuracy assessment and risk analysis of the surgical area based on the spatiotemporal fusion feature vector, and a warning level is determined based on the assessment results. Based on the warning level, surgical navigation decision information containing the warning level and the corresponding correction strategy is generated and output, including: Based on the spatiotemporal fusion feature vector, calculate the comprehensive accuracy evaluation value of the current surgical area; Based on the aforementioned error event evolution map, the critical paths of error propagation and cumulative risks are identified; Based on the comprehensive accuracy assessment value and the cumulative risk, predict the accuracy trend at multiple future time points; The warning level is determined based on the accuracy trend and multiple preset accuracy thresholds; Based on the warning level, a corresponding correction strategy is matched from a preset strategy library; The surgical navigation decision information includes the warning level, the correction strategy, and the corresponding accuracy prediction results.

Images