Intelligent decision-making method and system for perioperative multi-source data fusion based on internet of things

Abstract

The application provides a surgery period multi-source data fusion intelligent decision method and system based on Internet of Things, relates to a surgery data fusion technical field, and comprises the following steps: acquiring a multi-source heterogeneous data stream of a surgery period through an Internet of Things sensing layer, identifying a cross-source response event and extracting a surgery event anchor point, calculating a time phase difference to generate a time reference uniform surgery period data set through reverse compensation, performing a conditional independence test on the data set to construct a cause-effect dependence topology, identifying a cause-effect shielding node to generate a cause-effect decoupling surgery period feature through a cause-effect blocking test, constructing a surgery safety decision space containing an intervention and an injury boundary according to a surgery process, projecting the feature, generating a decision instruction through a constraint gradient search, and transmitting the decision instruction to a surgery equipment control unit to dynamically update a boundary parameter. The application realizes time synchronization and cause-effect decoupling of multi-source data, and improves the accuracy and safety of surgery decision.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of surgery data fusion, and in particular to a surgery period multi-source data fusion intelligent decision-making method and system based on Internet of Things. BACKGROUND

[0002] The surgery period decision-making method usually relies on a single data source (such as vital sign monitoring or image data) for real-time analysis, and triggers an alarm or intervention instruction through a preset threshold or rule base. The conventional method includes: collecting physiological parameters (such as heart rate, blood pressure), device status (such as electrotome power) and anesthesia depth data during the surgery, performing feature extraction based on a sliding window of time series, and then generating decision suggestions in combination with expert experience or machine learning models. These methods are widely used in clinical practice, especially for the auxiliary monitoring of standardized surgical procedures.

[0003] However, the conventional method has significant defects. The time benchmarks of multi-source data collection are not unified, resulting in time misalignment during data fusion. For example, the operation signal of surgical instruments (such as the start of an electrotome) and the physiological parameter changes (such as heart rate fluctuations) may have a millisecond-level phase difference due to differences in sampling frequency, transmission delay or system clock of different sensors. This time deviation distorts the causal relationship judgment between events, causing the decision model to mistakenly consider non-synchronous data as correlated events, thereby triggering false alarms or delaying intervention. In addition, there is a problem of redundant coupling between data sources. Different sensors (such as electrocardiogram and blood pressure) often show statistical dependence in surgical events, but existing methods usually directly fuse all features, ignoring the decoupling processing of causal relationships. This makes it difficult for the model to distinguish between direct causal relationships (such as electrotome cutting causing tissue impedance changes) and indirect correlations (such as anesthesia drug-induced blood pressure drops), thereby reducing the accuracy of decision instructions, especially in complex surgical stages, which may result in unnecessary intervention or missed critical risks. SUMMARY

[0004] The embodiments of the present application provide a surgery period multi-source data fusion intelligent decision-making method and system based on Internet of Things, which can solve the problems in the prior art.

[0005] In a first aspect, the embodiments of the present application provide a surgery period multi-source data fusion intelligent decision-making method based on Internet of Things, comprising:

[0006] acquiring a multi-source heterogeneous data stream of the surgery period through an Internet of Things perception layer;

[0007] identifying a cross-source response event triggered by the same surgical operation in the multi-source heterogeneous data stream, extracting a peak time of the cross-source response event as a surgery event anchor point, calculating a time phase difference between the surgery event anchor points of different data sources, performing reverse compensation on the collection timestamps of each data source based on the time phase difference, and generating a surgery period data set with unified time benchmarks.

[0008] performing a conditional independence test between data sources on the intraoperative data set, constructing a causal dependence topology between data sources based on the test result, identifying a causal shielding node in the causal dependence topology through a causal blocking test, and generating a causally decoupled intraoperative feature based on the causal shielding node;

[0009] constructing a surgical safety decision space containing an intervention boundary and an injury boundary according to a current surgical process stage, projecting the causally decoupled intraoperative feature into the surgical safety decision space, and generating a decision instruction meeting a surgical target through a constraint gradient search in the surgical safety decision space;

[0010] transmitting the decision instruction to a surgical equipment control unit for execution, and collecting surgical feedback data after the execution to dynamically update boundary parameters of the surgical safety decision space.

[0011] In an optional embodiment, a cross-source response event triggered by a same surgical operation is identified in the multi-source heterogeneous data stream, a peak time of the cross-source response event is extracted as a surgical event anchor point, a time phase difference between the surgical event anchor points of different data sources is calculated, a reverse compensation is performed on the collection time stamps of each data source based on the time phase difference, and a surgical period data set with uniform time reference is generated.

[0012] performing an adaptive sliding window scan on the multi-source heterogeneous data stream, calculating an energy change rate of each data source signal in the adaptive sliding window, marking a potential response point when the energy change rate exceeds a preset dynamic threshold, and extracting waveform segments before and after the potential response point to construct a response waveform library;

[0013] performing a morphological feature extraction on the waveform segments in the response waveform library, calculating a kurtosis, skewness and time domain envelope area of the waveform segments as a morphological feature vector, identifying a cross-source response event triggered by a same surgical operation through a similarity matching of the morphological feature vectors across data sources, and combining the waveform segments matched successfully into a response event group;

[0014] extracting a maximum amplitude time of each data source waveform segment from the response event group, marking a surgical event anchor point, constructing a time sequence correlation graph with the surgical event anchor points of each data source as nodes, and calculating a time offset of the surgical event anchor points relative to the surgical event anchor points of a reference data source in the time sequence correlation graph as a time phase difference;

[0015] establishing a time mapping relationship of each data source according to the time phase difference, performing a reverse compensation on the collection time stamps of each data source by applying a time offset opposite to the time phase difference, and rearranging the multi-source heterogeneous data streams of each data source according to a uniform time axis to generate a surgical period data set.

[0016] In an alternative embodiment, the conditional independence test is performed on the intraoperative dataset, and the causal dependency topology between the data sources is constructed based on the test results, comprising:

[0017] A data state snapshot of each data source at the action trigger moment of the surgical instrument is extracted from the intraoperative dataset, and a surgical data source graph structure is constructed with the data sources as nodes and the data state snapshots as node attributes;

[0018] Two data source nodes are selected from the surgical data source graph structure as a pair of nodes to be tested, and the other data source nodes except the pair of nodes to be tested are sequentially used as conditional variable nodes;

[0019] For each conditional variable node, the conditional mutual information of the pair of nodes to be tested under the condition of the conditional variable node is calculated, and when the conditional mutual information is lower than the preset independence threshold, it is determined that the pair of nodes to be tested is conditionally independent under the condition of the conditional variable node. The intermediate conduction node set is recorded, which is formed by all conditional variable nodes that make the pair of nodes to be tested conditionally independent;

[0020] When the intermediate conduction node set is empty, a causal dependency edge is established between the pair of nodes to be tested, and when the intermediate conduction node set is not empty, the conditional variable nodes in the intermediate conduction node set are marked as causal conduction path nodes between the pair of nodes to be tested;

[0021] The conditional independence test described above is performed on all pairs of data source nodes in the surgical data source graph structure, and all data source nodes and the causal dependency edges established therebetween are combined to construct the causal dependency topology between the data sources.

[0022] In an alternative embodiment, the causal shielding nodes are identified in the causal dependency topology by a causal blocking test, and the causal decoupling intraoperative features are generated based on the causal shielding nodes, comprising:

[0023] The target data source node corresponding to the surgical decision target and the source data source node corresponding to the surgical instrument control instruction are identified in the causal dependency topology, a reverse breadth-first search is performed from the target data source node to the source data source node along the causal dependency edge, all data source nodes passed by the reverse breadth-first search are recorded, and a causal conduction node set is constructed;

[0024] A causal blocking test is performed on each data source node in the causal conduction node set, the data source node and its connected causal dependency edges are temporarily removed from the causal dependency topology, the causal path search from the source data source node to the target data source node is re-executed, and when the causal path search fails, the data source node is marked as a causal shielding node. The causal shielding nodes and the causal dependency edges connected thereto are combined to form a causal backbone path;

[0025] Marking the data source nodes in the causal dependency topology that do not belong to the causal shielding nodes as bypass coupling nodes, and combining the bypass coupling nodes and the causal dependency edges connected thereto to form a bypass coupling path;

[0026] Extracting a data state snapshot corresponding to the causal shielding nodes as a causal decoupling surgery period feature.

[0027] In an optional embodiment, constructing a surgery safety decision space including an intervention boundary and a damage boundary according to a current surgery process stage includes:

[0028] Obtaining a standard surgery operation sequence corresponding to the current surgery process stage from a surgery process database, parsing instrument action types and tissue contact ranges corresponding to each surgery operation step in the standard surgery operation sequence, and mapping the instrument action types and the tissue contact ranges to a standard operation space region formed in a surgery scene three-dimensional coordinate system;

[0029] Calculating physiological state deviation degrees of each anatomical structure in the current surgery scene based on the causal decoupling surgery period feature, marking anatomical structures with physiological state deviation degrees exceeding a preset safety threshold as high-risk tissue regions, and marking spatial boundaries of the high-risk tissue regions in the surgery scene three-dimensional coordinate system to construct a damage boundary;

[0030] Expanding a preset safety distance outward in the surgery scene three-dimensional coordinate system to form an operation allowable space region based on the standard operation space region, and defining a space region between the operation allowable space region and the damage boundary as an intervention boundary;

[0031] Defining a space region surrounded by the intervention boundary and the damage boundary in the surgery scene three-dimensional coordinate system as a surgery safety decision space.

[0032] In an optional embodiment, projecting the causal decoupling surgery period feature to the surgery safety decision space, and generating a decision instruction meeting a surgery target in the surgery safety decision space by constraint gradient search includes:

[0033] Extracting data state snapshots corresponding to each causal shielding node from the causal decoupling surgery period feature, converting instrument space coordinates and tissue deformation displacements in the data state snapshots into state point coordinates in the surgery safety decision space, and establishing a mapping relationship between the causal shielding nodes and the state points in the surgery safety decision space;

[0034] Constructing a boundary normal vector field based on the intervention boundary and the damage boundary, calculating intervention normal distances and damage normal distances of the state points, and generating a constraint potential field by nonlinear weighted fusion;

[0035] Starting from a target state point corresponding to a surgical target, gradient descent iteration is performed along the negative gradient direction of the constraint potential energy field, the constraint potential energy gradient vector of the current iteration point is calculated in each iteration, the next iteration point is moved by a preset step along the opposite direction of the constraint potential energy gradient vector, and the iteration point sequence passed in the iteration process is recorded to form a constraint guide trajectory;

[0036] The state point coordinates corresponding to each iteration point in the constraint guide trajectory are reversely mapped to the causal shielding node, and the instrument control parameters corresponding to the causal shielding node are extracted to generate a decision instruction.

[0037] In an optional embodiment, a boundary normal vector field is constructed based on the intervention boundary and the damage boundary, the intervention normal distance and the damage normal distance of the state point are calculated, and the constraint potential energy field is generated by nonlinear weighted fusion, including:

[0038] The boundary surface mesh nodes of the intervention boundary and the damage boundary are extracted, the unit normal vectors of each boundary surface mesh node pointing to the inside of the surgical safety decision space are calculated to construct a boundary normal vector field;

[0039] The Euclidean distance from the state point to the nearest mesh node of the intervention boundary is calculated in the surgical safety decision space, the intervention normal distance is determined, the Euclidean distance from the state point to the nearest mesh node of the damage boundary is calculated, and the damage normal distance is determined;

[0040] The intervention normal distance is subjected to scalar multiplication operation with the unit normal vector of the corresponding intervention boundary mesh node to obtain an intervention potential energy vector, the damage normal distance is subjected to scalar multiplication operation with the unit normal vector of the corresponding damage boundary mesh node to obtain a damage potential energy vector, and the intervention potential energy vector and the damage potential energy vector are subjected to nonlinear weighted fusion to generate a constraint potential energy vector;

[0041] The constraint potential energy vectors of all state points constitute a constraint potential energy field.

[0042] The second aspect of the embodiment of the application provides a surgical period multi-source data fusion intelligent decision system based on an Internet of Things, including:

[0043] A data synchronization unit is configured to acquire a multi-source heterogeneous data stream in a surgical period through an Internet of Things perception layer;

[0044] A causal analysis unit is configured to identify cross-source response events triggered by the same surgical operation in the multi-source heterogeneous data stream, extract peak moments of the cross-source response events as surgical event anchor points, calculate time phase differences between the surgical event anchor points of different data sources, perform reverse compensation on acquisition time stamps of each data source based on the time phase differences, and generate a surgical period data set with time reference unified;

[0045] A decision generation unit is configured to perform a condition independence test between data sources on the surgery period dataset, construct a causal dependence topology between the data sources based on a test result, identify a causal shielding node in the causal dependence topology through a causal blocking test, and generate a causally decoupled surgery period feature based on the causal shielding node;

[0046] A feedback update unit is configured to construct a surgery safety decision space containing an intervention boundary and a lesion boundary according to a current surgery process stage, project the causally decoupled surgery period feature into the surgery safety decision space, and generate a decision instruction meeting a surgery target through a constraint gradient search in the surgery safety decision space.

[0047] A processing unit is configured to transmit the decision instruction to a surgery equipment control unit for execution, and collect surgery feedback data after the execution to dynamically update boundary parameters of the surgery safety decision space.

[0048] In a third aspect, an electronic device is provided, and the electronic device comprises:

[0049] a processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to invoke the instructions stored in the memory to execute the method described above.

[0052] In a fourth aspect, a computer-readable storage medium is provided, and the computer-readable storage medium stores computer program instructions, and the computer program instructions are executed by a processor to implement the method described above.

[0053] In the embodiment of the present application, based on the multi-source heterogeneous data stream obtained by the Internet of Things perception layer, by identifying the cross-source response event and extracting the surgical event anchor point, the time deviation introduced by the different data sources due to the different collection time can be eliminated, the time reference unified surgical period data set is generated, and the alignment accuracy of the multi-source data in the time dimension is significantly improved; the condition independence test is performed on the unified data set and the causal dependence topology is constructed, the causal shielding node is identified through the causal blocking test, the redundant association and interference factors between the data sources can be effectively stripped, the surgical period characteristics with causal decoupling characteristics are generated, the feature dimension is greatly reduced and the interpretability of the feature is enhanced; according to the current surgical process stage, the surgical safety decision space containing the intervention boundary and the injury boundary is constructed, the causal decoupling characteristics are projected into the space, the decision instruction meeting the surgical target is generated through the constraint gradient search, the decision action can be ensured to be always in the safe and controllable range, the physical or physiological limit of the surgical operation is avoided to be exceeded, and the surgical efficiency and the patient safety are optimized; the boundary parameters of the surgical safety decision space are dynamically updated by collecting the surgical feedback data after execution, a closed-loop adaptive adjustment mechanism is formed, the decision system can respond to the surgical process change and the patient state fluctuation in real time, and the robustness and the precision of the decision are continuously improved. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 FIG. 1 is a flowchart of the surgical period multi-source data fusion intelligent decision method based on the Internet of Things according to the embodiment of the present application; Figure 2 FIG. 2 is a flowchart of the causal dependence topology analysis and feature extraction according to the embodiment of the present application. DETAILED DESCRIPTION

[0055] In order to make the purpose, technical scheme and advantages of the embodiment of the present application clearer, the technical scheme in the embodiment of the present application will be described clearly and completely below in combination with the drawings in the embodiment of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor belong to the protection scope of the present application.

[0056] The technical scheme of the present application will be described in detail below with specific embodiments. The following specific embodiments can be combined with each other, and the same or similar concepts or processes can not be described in some embodiments.

[0057] Figure 1 FIG. 1 is a flowchart of the surgical period multi-source data fusion intelligent method based on the Internet of Things according to the embodiment of the present application, as shown in the figure, the method comprises: Figure 1

[0058] The multi-source heterogeneous data stream of the surgical period is obtained through the Internet of Things perception layer; ​

[0059] identify cross-source response events triggered by the same surgical operation in the multi-source heterogeneous data stream, extract peak time points of the cross-source response events as surgical event anchors, calculate time phase differences between the surgical event anchors of different data sources, perform reverse compensation on the acquisition time stamps of each data source based on the time phase differences, and generate a surgical period data set with uniform time reference;

[0060] perform conditional independence test between data sources on the surgical period data set, construct a causal dependence topology between the data sources based on the test results, identify causal shielding nodes in the causal dependence topology through causal blocking test, and generate causal decoupling surgical period features based on the causal shielding nodes;

[0061] construct a surgical safety decision space containing intervention boundaries and damage boundaries according to the current surgical process stage, project the causal decoupling surgical period features into the surgical safety decision space, and generate decision instructions that meet the surgical target through constraint gradient search in the surgical safety decision space;

[0062] transmit the decision instructions to the surgical equipment control unit for execution, and dynamically update the boundary parameters of the surgical safety decision space by collecting surgical feedback data after execution.

[0063] In an alternative embodiment, identifying cross-source response events triggered by the same surgical operation in the multi-source heterogeneous data stream, extracting peak time points of the cross-source response events as surgical event anchors, calculating time phase differences between the surgical event anchors of different data sources, performing reverse compensation on the acquisition time stamps of each data source based on the time phase differences, and generating a surgical period data set with uniform time reference includes:

[0064] perform adaptive sliding window scanning on the multi-source heterogeneous data stream, calculate the energy change rate of each data source signal in the adaptive sliding window, and mark as a potential response point when the energy change rate exceeds a preset dynamic threshold, extract waveform segments before and after the potential response point to construct a response waveform library;

[0065] perform morphological feature extraction on the waveform segments in the response waveform library, calculate the kurtosis, skewness and time domain envelope area of the waveform segments as morphological feature vectors, identify cross-source response events triggered by the same surgical operation through similarity matching of morphological feature vectors across data sources, and combine the waveform segments that match successfully into a response event group;

[0066] extract the maximum amplitude time point of each data source waveform segment from the response event group, mark the surgical event anchor, construct a time sequence correlation graph with the surgical event anchors of each data source as nodes, and calculate the time offset of the surgical event anchor relative to the surgical event anchor of the reference data source as the time phase difference in the time sequence correlation graph;

[0067] According to the time phase difference, a time mapping relationship of each data source is established, reverse compensation is performed on the collection time stamp of each data source by applying a time offset opposite to the time phase difference, and the multi-source heterogeneous data streams of each data source are rearranged according to a unified time axis to generate a surgery period data set.

[0068] In a specific embodiment, when adaptive sliding window scanning is performed on the multi-source heterogeneous data streams, the window length is not fixed but is dynamically adjusted according to the signal frequency characteristics of each data source. For high-frequency physiological signals (such as electroencephalogram and electromyogram), the window length is relatively short to capture the transient response of rapid changes; for low-frequency device state signals (such as instrument pressure and temperature), the window length is appropriately lengthened to ensure a sufficient energy accumulation period. In each sliding window, the energy change rate is calculated, which is the ratio of the sum of the square of the amplitudes of all sampling points in the current window to the corresponding value of the previous window. When the preset dynamic threshold is exceeded, the time point is marked as a potential response point. The dynamic threshold is not a global fixed value but is adaptively estimated according to the historical baseline noise level of each data source signal. Specifically, the current threshold is the average of the energy change rates of a number of recent windows plus several times the standard deviation, thereby effectively suppressing false positives caused by electromagnetic interference in the surgical environment or slight shaking of the device.

[0069] After identifying the potential response point, waveform segments within a certain time range before and after the point are extracted to construct a response waveform library. The forward intercept length of the waveform segment covers the response rising edge, and the backward intercept length covers the complete process of the response decaying to the baseline. For different data sources, the intercept length is independently set according to the typical response duration of each signal to ensure that the waveform segment fully reflects the response process triggered by a single surgical operation. The response waveform library is organized and stored in the form of a triple of data source identifier, potential response point time, and waveform segment, providing a unified data input for subsequent cross-source matching.

[0070] For each waveform segment in the response waveform library, morphological feature extraction is performed, and kurtosis , skewness , and time domain envelope area are selected as the three components of the morphological feature vector. Kurtosis reflects the sharpness of the waveform amplitude distribution, skewness describes the asymmetry of the waveform amplitude distribution, and the time domain envelope area is obtained by integrating the absolute values of the waveform segment amplitudes, reflecting the total amount of response energy. The three components together constitute the morphological feature vector It can characterize the overall shape of a waveform from different dimensions, so that even if the data sources come from different physical dimensions, their response waveforms are still comparable in shape.

[0071] Cross-data source morphological feature vector similarity matching employs a normalized Euclidean distance metric. Before matching, z-score normalization is performed on the morphological feature vectors of each data source according to the historical statistical distribution of their respective feature components to eliminate scale differences caused by different physical dimensions. After normalization, the Euclidean distance between the morphological feature vectors of waveform segments from different data sources is calculated. ,when Below the matching similarity threshold When two waveform segments are determined to be triggered by the same surgical operation, they are grouped into the same response event group. The matching process employs a greedy strategy to progressively expand the response event group: using a potential response point from a certain baseline data source as a seed, it sequentially searches for candidate segments in other data sources that have the highest similarity to the seed waveform segment and satisfy the distance constraint, until all participating data sources are covered or no candidate segments satisfy the conditions are found. Potential response points that fail to match are considered local noise responses and are not included in the subsequent time alignment process.

[0072] The maximum amplitude moment of each data source waveform segment is extracted from each response event group and marked as the surgical event anchor point of that data source in this response event. The extraction of the maximum amplitude moment is based on the sampling point accuracy, and sub-sampling accuracy anchor point positioning is achieved by performing quadratic curve fitting on the waveform segment near the peak, thereby reducing the anchor point discretization error caused by sampling rate differences. Using the surgical event anchor points of all participating data sources as nodes, a time series correlation graph is constructed. Each edge in the graph connects anchor point nodes from different data sources, and the weight of the edge is the difference between the corresponding anchor point moments of the two nodes, i.e., the time phase difference. .

[0073] In the timing correlation diagram, a data source is designated as the baseline data source (usually the one with the highest sampling rate and best clock stability). The time offset of the surgical event anchor point of each of the other data sources relative to the surgical event anchor point of the baseline data source is calculated, and this offset is taken as the time phase difference of each data source. subscript Identify the specific data source number. When the same data source participates in matching in multiple response event groups, the result from multiple calculations is used. The median is used as the final time phase difference estimate to improve the robustness of the estimate and avoid time alignment deviations caused by individual abnormal response events.

[0074] Based on the time phase difference of each data source Establish time mapping relationships between various data sources. each acquisition timestamp of the plurality of data sources , performing reverse compensation, which is modified as , that is, applying a time offset in the opposite direction of the time phase difference and equal in size on the original timestamp. The physical meaning of reverse compensation is that if the response event anchor point of a certain data source lags behind the reference data source , then all the timestamps of the data source need to be moved forward to align with the time axis of the reference data source. The compensation operation directly acts on the timestamp field without any modification to the signal amplitude, ensuring the amplitude integrity of the original physiological and device signals.

[0075] After completing the reverse compensation of the timestamps of each data source, the multi-source heterogeneous data streams of all data sources are rearranged according to the unified time axis. The rearrangement process takes the compensated timestamp as the key, performs global sorting on the data points of each data source, and performs equal-interval resampling on the time axis. Linear interpolation is used to fill in the missing data points of each data source at the time instant, and finally a time reference unified perioperative data set is generated. In this data set, the signals of each data source are strictly aligned in the time dimension, eliminating the time deviation introduced by device clock drift, signal transmission delay and sampling rate difference, providing a reliable time synchronization basis for subsequent conditional independence test and causal dependence topology construction between data sources. The perioperative data set is indexed by the unified time axis, retains the original physical quantity identifier of each data source, supports flexible retrieval by time interval and data source type, and meets the input format requirements of subsequent multi-source fusion analysis.

[0076] In an optional embodiment, the conditional independence test between data sources is performed on the perioperative data set, and the causal dependence topology between data sources is constructed based on the test results, including:

[0077] Extracting the data state snapshot of each data source at the action trigger time of the surgical instrument from the perioperative data set, and constructing a surgical data source graph structure with data sources as nodes and data state snapshots as node attributes;

[0078] Selecting two data source nodes from the surgical data source graph structure as a node pair to be tested, and sequentially selecting other data source nodes as conditional variable nodes except the node pair to be tested;

[0079] For each conditional variable node, calculate the conditional mutual information of the node pair to be tested under the condition of the conditional variable node. When the conditional mutual information is lower than the preset independence threshold, it is determined that the node pair to be tested is conditionally independent under the condition of the conditional variable node, and all conditional variable nodes that make the node pair to be tested conditionally independent are recorded to form an intermediate conduction node set;

[0080] When the intermediate conduction node set is empty, a causal dependence edge is established between the pair of nodes to be verified, and when the intermediate conduction node set is not empty, the conditional variable nodes in the intermediate conduction node set are marked as the causal conduction path nodes between the pair of nodes to be verified;

[0081] The conditional independence test described above is performed on all pairs of data source nodes in the operation data source graph structure, and a causal dependence topology between the data sources is constructed by combining all data source nodes and the causal dependence edges established therebetween.

[0082] In a specific embodiment, data state snapshots of each data source at the action trigger time of the surgical instrument are extracted from the operation period data set, which is the starting point of constructing the causal dependence topology. The action trigger time of the surgical instrument corresponds to the uniform time reference node after the alignment of the surgical event anchor, and at this time, the state information such as the signal amplitude, frequency distribution and statistical moment of each data source is completely retained, forming a set of high-dimensional state vectors. Each data source is regarded as a node in the graph structure, and the corresponding data state snapshot is regarded as the attribute vector of the node, thereby constructing the operation data source graph structure. The number of nodes in the graph structure is equal to the total number of data sources participating in fusion, and the attribute dimension of the node is determined by the number of features of the state snapshot. There is no edge connection between the nodes in the graph in the initial state, and the subsequent conditional independence test process will gradually fill the causal dependence relationship between the nodes.

[0083] After the operation data source graph structure is established, two nodes are selected from all data source nodes to form a pair of nodes to be verified, and the remaining other data source nodes are sequentially set as conditional variable nodes. Let the pair of nodes to be verified correspond to data source and data source , and the current conditional variable node corresponds to data source , the conditional mutual information between and needs to be calculated under the condition of . The conditional mutual information measures the residual statistical dependence between two data sources under the premise of knowing the information of the third party data source. If the association between and is completely mediated by , then under the condition of , the two tend to be conditionally independent, and the value of will tend to zero. The calculation of the conditional mutual information is based on the estimation of the joint probability distribution and the marginal probability distribution of the state snapshots of each data source, and the kernel density estimation or the non-parametric estimation method based on the neighborhood is used to approximate the probability density, so as to avoid strong hypothesis constraints on the form of data distribution.

[0084] Let the preset independence threshold be When the calculated conditional mutual information satisfy At that time, determine the data source With data source In condition variables Conditional independence under given conditions. Threshold. The settings need to comprehensively consider the data sampling volume, the dimension of the state snapshot, and the tolerance of the surgical scenario for misjudgment. They are typically calibrated using permutation tests or statistical significance tests based on asymptotic distributions to ensure the reliability of the independence determination. For the current node pair to be tested... Iterate through all condition variable nodes and perform the conditional mutual information calculation and independence determination on each one. Summarize all condition variable nodes that satisfy the conditional independence determination to form a set of intermediate transit nodes for that node pair. .

[0085] intermediate transmission node set The state directly determines how the causal relationship between the pairs of nodes to be tested is modeled. When it is an empty set, it means that no matter what third-party data source is introduced as a condition, and The consistent statistical dependence between them indicates a direct causal relationship, which is reflected in the nodes of the surgical data source graph structure. With nodes Establish causal dependencies between them. The directionality of this edge can be further determined through subsequent temporal relationships or intervention experiment results. In the initial graph construction stage, bidirectional dependencies can be represented by undirected edges. When not empty, it means and The statistical dependency between them can be explained by one or more condition variable nodes in the set. The relationship between them is not a direct causal one, but rather transmitted through the set of intermediate nodes. The nodes in the process are passed. At this point, [the process will be...]. All condition variable nodes are marked as nodes With nodes The nodes in the causal transmission path between these nodes will be used as important structural information in subsequent causal shielding analysis.

[0086] Conditional independence tests are performed sequentially on all data source node pairs in the surgical data source graph structure according to the above procedure. The traversal order of node pairs follows the principle of no repeated enumeration, and the total number of tests is [number missing]. ,in is the total number of data source nodes. After the verification of each pair of nodes, i.e., whether to add a causal dependence edge between the pair of nodes according to the state of the intermediate conduction node set, and record the corresponding conduction path node information. After the verification of all node pairs is completed, all data source nodes and all causal dependence edges established during the verification process are integrated to form a complete causal dependence topology between data sources. The topology directly expresses the direct causal relationship structure between each data source in the form of a graph. The node pair connected by the causal dependence edge represents the existence of direct physiological or operational causal action, and the node pair connected only through the conduction path node represents that its association is mediated by a specific intermediate data source.

[0087] In an actual surgical scene, the causal relationship between different data sources often has obvious physiological significance. For example, there may be a direct causal dependence edge between the surgical operation force sensor data and the tissue impedance data, while the association between the surgical instrument position data and the patient vital sign data may need to be explained through the intermediate conduction node of the surgical operation force. The construction of the causal dependence topology makes the subsequent causal shielding node identification have a structured graph theory foundation, avoiding the false association interference introduced by relying only on correlation analysis. At the same time, the verification method based on conditional mutual information has a relatively relaxed assumption on data distribution, which can effectively handle the non-Gaussian, nonlinear statistical dependence structure commonly existing in multi-source data during the surgery period, and improves the accuracy and robustness of the causal dependence topology.

[0088] In a surgical real-time scene with limited computing resources, the above verification process can be processed in a layered acceleration manner. The node pairs with the absolute value of the Pearson correlation coefficient between the state snapshots exceeding a preset screening threshold are preferentially executed for complete conditional mutual information verification, and the node pairs with a correlation coefficient lower than are directly determined as conditionally independent and skipped for detailed verification, thereby significantly reducing the computational overhead while ensuring the integrity of the main causal structure. The value of the screening threshold needs to be pre-calibrated according to the statistical characteristics of historical surgical data to ensure that key causal dependence relationships are not missed. The finally constructed causal dependence topology will serve as a structural prior for causal decoupling surgery period feature extraction, supporting the accurate modeling of the subsequent surgical safety decision space.

[0089] As shown in Figure 2 , a causal dependence topology analysis and feature extraction flowchart is shown.

[0090] In an optional embodiment, the causal shielding nodes are identified in the causal dependence topology through a causal blocking test, and the causal decoupling surgery period features are generated based on the causal shielding nodes, including:

[0091] identify a target data source node corresponding to a surgical decision target and a source data source node corresponding to a surgical instrument control instruction in the causal dependency topology, perform a reverse breadth-first search from the target data source node to the source data source node along the causal dependency edges, record all data source nodes passed by the reverse breadth-first search, and construct a causal conduction node set;

[0092] perform a causal blocking test on each data source node in the causal conduction node set, temporarily remove the data source node and its connected causal dependency edges from the causal dependency topology, re-perform the causal path search from the source data source node to the target data source node, and when the causal path search fails, mark the data source node as a causal shielding node, and combine the causal shielding node and its connected causal dependency edges to form a causal backbone path;

[0093] mark data source nodes in the causal dependency topology that do not belong to the causal shielding node as bypass coupling nodes, and combine the bypass coupling nodes and their connected causal dependency edges to form a bypass coupling path;

[0094] extract a data state snapshot corresponding to the causal shielding node as a causal decoupling surgery period feature.

[0095] In a specific embodiment, after the construction of the causal dependency topology between data sources is completed, it is necessary to accurately locate the nodes that play a key conduction role in the surgery decision from the topology, and to generate a causal decoupling surgery period feature that removes redundant coupling information on this basis. The whole process is divided into four links of construction of a causal conduction node set, identification of a causal shielding node, stripping of a bypass coupling path, and feature extraction, which are logically progressive and interlocking.

[0096] In the causal dependency topology, first, the semantic positioning of two types of key nodes is clarified. The target data source node corresponding to the surgical decision target represents the output end that needs to be finally predicted or evaluated in the topology, such as an intraoperative tissue damage risk score or a patient vital sign comprehensive state node. The source data source node corresponding to the surgical instrument control instruction represents the input end that applies intervention in the topology, such as an ultrasonic knife power instruction node or a mechanical arm pose instruction node. Starting from the target data source node, a reverse breadth-first search is performed along the directed causal dependency edges in the causal dependency topology, and the source is traced layer by layer upwards until the source data source node or the leaf node that cannot continue to extend is reached. During the search process, all intermediate data source nodes passed are uniformly included in the causal conduction node set . The reverse breadth-first search is implemented using a queue structure, and each time a node is taken out of the queue, all predecessor nodes (i.e., nodes pointing to the node in the directed graph) are traversed. If the predecessor node has not been visited and is not a node specified by the search termination condition, it is added to the queue and recorded. Finally, All data source nodes involved in the possible causal conduction between source and target are included in the candidate set for subsequent causal blocking test.

[0097] For each data source node in , a causal blocking test is performed sequentially. The test procedure adopts a three-step operation paradigm of "temporary removal-research-recovery". Specifically, for the th candidate node in , the node and all causal dependency edges connected to it are temporarily removed from the causal dependency topology, forming a reduced temporary topology. On this temporary topology, a causal path search from the source data source node to the target data source node is re-executed. If the search result is that the path does not exist (i.e. the source node and the target node are no longer connected in the temporary topology), it indicates that the existence of node is a necessary condition to maintain the source-target causal conduction link, which is marked as a causal shielding node and added to the causal shielding node set ; if the search result shows that the path still exists, it indicates that even if node is removed, causal signals can still be transmitted from the source to the target through other paths, and this node is not indispensable to the causal backbone conduction. It is not marked as a causal shielding node. After each test, the temporarily removed node and its edges are immediately restored to the topology to ensure that the subsequent node test is performed on the complete topology structure, avoiding test bias caused by cumulative removal. After completing the test of all candidate nodes, all causal shielding nodes in and their causal dependency edges between each other are combined to form the causal backbone path . represent the most critical information transmission backbone between the surgical instrument control instructions and the surgical decision target, and the absence of each node will cause the causal link to be broken.

[0098] The data source nodes in the causal dependency topology that belong to but do not belong to are marked as bypass coupling nodes, and these bypass coupling nodes and the causal dependency edges connected to them are combined to form the bypass coupling path . The nodes in the bypass coupling path have a directed connection relationship with the backbone nodes in the causal dependency topology, but they are not indispensable to the source-target causal conduction, and more reflect the covariation relationship or indirect association between data sources, rather than the real causal backbone conduction. In the surgical decision scenario, the bypass coupling path often corresponds to the device noise propagation link, the indirect reflection path of patient physiological parameters, or the diffusion path of surgical environment interference. Stripping the information of these bypass coupling nodes from the decision features helps to eliminate the interference of confounding variables on the decision model, and improves the causal explainability and clinical reliability of the decision instruction.

[0099] After the identification of the set of causal shielding nodes is completed , for each causal shielding node in , its corresponding data state snapshot in the current time reference unified surgery period data set is extracted. The extraction of the data state snapshot is indexed by the time-synchronized acquisition timestamp, and the data segment within the fixed time window centered on the current surgery event anchor point is intercepted, covering all sampling values of the data source corresponding to the node within the time window. For continuous data source nodes (such as vital sign waveforms, instrument torque sensor signals), the data state snapshot is a time series segment; for discrete data source nodes (such as surgery stage labels, instrument status codes), the data state snapshot is a category encoding vector. The data state snapshots of all causal shielding nodes in are spliced in node number order to form the causal decoupling surgery period feature vector . Since only contains key node information on the causal backbone path, while eliminating the redundant coupling components introduced by bypass coupling nodes, it has stronger causal explanation ability for surgery decision targets, and higher robustness to surgery environment noise and device individual differences.

[0100] In actual surgery scenarios, the structure of the causal dependency topology may change locally with the switching of the surgery process stage. For example, in the suturing stage and the resection stage, the correspondence between the source data source node and the target data source node may be adjusted, resulting in different compositions of the set of causal transmission nodes . Therefore, whenever the surgery process stage switches, the reverse breadth-first search and the causal blocking test need to be re-executed to dynamically update the composition of and the causal decoupling surgery period feature vector , ensuring that the feature always matches the causal structure of the current surgery stage. In addition, when surgery feedback data triggers boundary parameter updates of the surgery safety decision space, if the weight of a causal dependency edge in the causal dependency topology changes significantly, the causal blocking test needs to be re-executed for the affected nodes to verify whether their shielding node identity still holds, ensuring the continuous effectiveness of the causal decoupling surgery period feature.

[0101] In an optional embodiment, constructing a surgery safety decision space including intervention boundaries and damage boundaries according to the current surgery process stage comprises:

[0102] Obtaining the standard surgery operation sequence corresponding to the current surgery process stage from the surgery procedure database, parsing the instrument action type and tissue contact range corresponding to each surgery operation step in the standard surgery operation sequence, and mapping the instrument action type and tissue contact range to the standard operation space region formed in the surgery scene three-dimensional coordinate system;

[0103] The physiological state deviation degree of each anatomical structure in the current operation scene is calculated based on the causally decoupled intraoperative feature, and the anatomical structure with a physiological state deviation degree exceeding a preset safety threshold is marked as a high-risk tissue area, and a spatial boundary of the high-risk tissue area is labeled in a three-dimensional coordinate system of the operation scene to construct an injury boundary.

[0104] With the standard operation space area as a reference, an operation allowable space area is formed by extending a preset safety distance outward in the three-dimensional coordinate system of the operation scene, and a space area between the operation allowable space area and the injury boundary is defined as an intervention boundary.

[0105] The space area surrounded by the intervention boundary and the injury boundary in the three-dimensional coordinate system of the operation scene is defined as a surgical safety decision space.

[0106] In a specific embodiment, when a standard operation sequence matching the current operation process stage is obtained from the operation flow database, the operation stage identifier currently located needs to be determined according to the preoperative planning information and the real-time operation process recognition result, and then the identifier is used as an index to retrieve the corresponding standard operation sequence record from the database. The standard operation sequence includes a plurality of ordered operation steps, each step being associated with an instrument action type field and a tissue contact range field. The instrument action type describes the motion mode of the surgical instrument in the step, such as cutting, suturing, coagulation, and traction. The tissue contact range describes the tissue area expected to be contacted by the instrument tip in the anatomical space, usually given as a three-dimensional coordinate range with reference to anatomical landmarks. The above two types of information are jointly mapped to the three-dimensional coordinate system of the operation scene to obtain the standard operation space area of the operation stage. Specifically, for each operation step , the instrument action type determines the set of motion direction constraints of the step in the three-dimensional coordinate system, and the tissue contact range determines the set of spatial position constraints of the step. The intersection of the two sets constitutes the standard operation sub-area of the step , and the union of the standard operation sub-areas of all steps constitutes the complete standard operation space area , that is .

[0107] Based on the generated causally decoupled intraoperative feature vector , the current physiological state of each anatomical structure in the operation scene is quantitatively evaluated. For an anatomical structure numbered , a physiological indicator sub-vector related to the structure is extracted from , and a standard physiological state reference vector of the anatomical structure in the current operation stage is obtained from the operation flow database. The physiological state deviation degree is defined as the weighted distance between the current physiological indicator sub-vector and the standard reference vector, i.e. wherein is the physiological indicator weight matrix for the anatomical structure , used to highlight the physiological indicator dimensions that have greater impact on the safety of the structure. Let the preset safety threshold be , when , the anatomical structure is marked as a high-risk tissue region. The spatial boundary of the high-risk tissue region is obtained by applying an inflation operation to the three-dimensional anatomical model of the anatomical structure, and the inflation radius is positively correlated with the physiological state deviation , the higher the deviation, the larger the boundary range, to ensure that the protection margin of the high-risk region increases dynamically with the risk level. The spatial boundaries of all high-risk tissue regions are merged in the three-dimensional coordinate system of the surgical scene to form an injury boundary , and the spatial region inside the injury boundary is a forbidden region where instruments are absolutely prohibited from entering.

[0108] Take the standard operation space region as the reference, and uniformly expand the preset safety distance outward along the outward normal direction of in the three-dimensional coordinate system of the surgical scene to obtain the operation allowable space region . The value of the preset safety distance is determined according to the operation accuracy requirement of the current surgical stage and the position control error range of the instrument tip, and is usually in the order of millimeters. For the fine operation stage, it can be appropriately reduced to maintain the operation accuracy, and for the large-range pulling stage, it can be appropriately increased to accommodate the instrument movement fluctuations. The operation allowable space region represents the maximum operation offset range allowed on the basis of the standard operation, and the instrument movement beyond this region will trigger a warning response. The intervention boundary is defined as the spatial region between the outer boundary of the operation allowable space region and the injury boundary , i.e. the outer boundary surface of , used to represent the buffer zone of the transition of the instrument movement from the safe region to the dangerous region. When the instrument tip enters the spatial region corresponding to the intervention boundary, a warning-level intervention instruction is triggered to prompt the operator to adjust the instrument movement direction or amplitude to avoid further approaching the injury boundary.

[0109] The closed spatial region surrounded by the intervention boundary and the injury boundary in the three-dimensional coordinate system of the surgical scene is defined as the surgical safety decision space . The surgical safety decision space ​​​In a physical sense, this represents the three-dimensional spatial range within which the instrument tip can move safely during the current surgical phase, and its internal boundaries are subject to intervention. It is further divided into two levels: a core safety zone and a buffer warning zone. The core safety zone corresponds to the standard operating space area. Within and around this area, when the device moves within this area, the decision-making instructions maintain the current operation; the buffer warning area corresponds to... Outer boundary to intervention boundary Within the designated area, when the device enters this area, corrective decision instructions are generated to guide the device back to the core safety zone; when the device exceeds the intervention boundary... And the damage boundary has not yet been reached. When an emergency stop command is generated, it is triggered; when the instrument reaches the damage boundary. At that time, the highest priority forced rollback command is triggered. The three-dimensional geometric representation of the surgical safety decision space is stored using a signed distance field. For any point in the three-dimensional coordinate system of the surgical scene... Define its signed distance field value This is the signed distance from the point to the nearest boundary surface. The point indicates that it is located inside the safe zone. The point indicates that it is located inside the prohibited area. The corresponding boundary surface location. In the subsequent constraint gradient search process, this will be directly utilized. The gradient information guides the direction of decision command generation, ensuring that the generated decision commands always meet the spatial constraints of the surgical safety decision space.

[0110] In practical applications, the establishment of a three-dimensional coordinate system for the surgical scene relies on the registration results of intraoperative image navigation data and instrument tracking data. The origin of the three-dimensional coordinate system is usually set at an anatomical reference point in the surgical area, and the coordinate axes are aligned with the patient's body axis to ensure that the spatial description of the anatomical structures has clear clinical significance. Standard operating space area Damage boundary and intervention boundaries All data is constructed and stored within this unified coordinate system, ensuring consistent processing and decision inference of spatial information from different data sources within the same reference framework. Each time a surgical stage changes, the aforementioned boundary parameters are recalculated based on the standard operating sequence corresponding to the new stage and the current causal decoupling surgical period characteristics, achieving dynamic adaptive updating of the surgical safety decision space as the surgical progresses.

[0111] In one optional embodiment, projecting causal decoupled perioperative features onto a surgical safety decision space, and generating decision instructions that satisfy surgical objectives within the surgical safety decision space through constrained gradient search, includes:

[0112] extracting the data state snapshot corresponding to each pair of causal shielding nodes from the causally decoupled intraoperative feature vector, converting the instrument spatial coordinates and tissue deformation displacement in the data state snapshot into state point coordinates in the surgical safety decision space, and establishing a mapping relationship between the causal shielding nodes and the state points in the surgical safety decision space;

[0113] Based on the intervention boundary and the damage boundary, a boundary normal vector field is constructed, the intervention normal distance and the damage normal distance of the state point are calculated, and a constraint potential field is generated by nonlinear weighted fusion;

[0114] Starting from the target state point corresponding to the surgical target, gradient descent iteration is performed along the negative gradient direction of the constraint potential field, the constraint potential gradient vector of the current iteration point is calculated in each iteration, and the next iteration point is moved by a preset step along the opposite direction of the constraint potential gradient vector. The sequence of iteration points passed in the iteration process is recorded to form a constraint guide trajectory;

[0115] The state point coordinates of each iteration point in the constraint guide trajectory are reversely mapped to the causal shielding nodes, and the instrument control parameters corresponding to the causal shielding nodes are extracted to generate decision instructions.

[0116] In a specific embodiment, from the causally decoupled intraoperative feature vector , for each causal shielding node, a data state snapshot corresponding to the current time is extracted. The data state snapshot contains two types of core information: instrument spatial coordinates and tissue deformation displacement. The instrument spatial coordinates are derived from the pose sensor data of the surgical robot end effector, represented in a three-dimensional Cartesian coordinate system; the tissue deformation displacement is derived from the real-time tracking results of soft tissue deformation by intraoperative ultrasound or optical coherence tomography. The instrument spatial coordinates are denoted as , the tissue deformation displacement vector is denoted as , and the two are concatenated to form the state point coordinates , completing the mapping from the causal shielding node feature space to the state point in the surgical safety decision space . This mapping relationship is realized through a pre-calibrated coordinate transformation matrix , ensuring the consistency between the sensor physical coordinate system and the decision space coordinate system. For each causal shielding node, a corresponding state point is established, forming a state point set, and the subsequent constraint potential field construction and gradient search are carried out on this state point set.

[0117] In the surgical safety decision space , the corresponding boundary normal vector fields are constructed based on the intervention boundary and the damage boundary . For any state point in the space, the boundary normal vector is determined using the gradient direction of the signed distance field : intervention boundary normal vector defined as the unit normal vector pointing outward at the unit normal vector pointing outward at the injury boundary defined as the unit normal vector pointing outward at the unit normal vector pointing outward at the intervention boundary is the state point along the direction to the signed distance to the signed distance to along the direction to the signed distance to When the state point is inside the operational admissible space region

[0118] constraint potential field is generated by nonlinearly weighting and fusing the intervention potential component and the injury potential component. The intervention potential component is defined by an exponential repulsive potential function, which sharply increases when the state point approaches to prevent the instrument from exceeding the allowed operation range; the injury potential component is defined by a more steep exponential repulsive potential function, which corresponds to the hard constraint characteristic of the injury risk. The weights and of the nonlinear weighting and fusing are determined by the risk priority of the current surgical process stage, and the injury boundary weight is given a higher value in the high-precision operation stage to strengthen the protection of the key anatomical structures. The fused constraint potential field is expressed as , which forms a continuous and differentiable scalar field in , and its gradient has a clear physical meaning at each state point: pointing to the direction of potential energy increase, i.e., the direction away from the safe region.

[0119] The constraint gradient search starts from the target state point corresponding to the surgical target, and performs iterative descent along the negative gradient direction of the constraint potential field. In the th iteration, the current iteration point is denoted as , and the constraint potential gradient vector at this point is calculated . The iterative update rule is , where is a preset step size that controls the moving amplitude of each iteration. To prevent the state point from crossing the boundary and entering the forbidden region during the iteration process, a feasibility projection is performed after each iteration update: if falls into the forbidden region Otherwise, project it back. The closest feasible point on the boundary. Iteration termination conditions fall into two categories: one is the displacement norm between adjacent iteration points. Below the convergence threshold This indicates that the search has converged to a local stable point; secondly, the number of iterations has reached the preset maximum iteration limit. This prevents the search from getting stuck in an infinite loop. Records from... The sequence of all iteration points passed to the convergence point This sequence constitutes a constrained guiding trajectory. This represents the optimal path starting from the target state while satisfying both the intervention boundary and the damage boundary constraints.

[0120] Constrained Guiding Trajectory The coordinates of the state points at each iteration point are transformed by the coordinate transformation matrix. The inverse transformation maps back to the physical feature space corresponding to the causal shielding node. For each iteration point... The instrument space coordinate components Directly corresponds to the target pose of the surgical instrument end effector, tissue deformation displacement components This is used to verify whether the current tissue deformation is within the expected range. The instrument control parameters corresponding to each iteration point are extracted from the causal shielding node, including the joint angle command sequence. ( (Joint degrees of freedom) and end effector force control parameters and instrument opening / closing status indicators The above control parameters are arranged according to the trajectory timing and encapsulated into structured decision instructions. This decision instruction sequence comprehensively describes the instrument motion control strategy from the current surgical state to the target state, satisfying both the surgical operation objectives and strictly adhering to the safety constraints of intervention and damage boundaries. After the decision instructions are generated, they are transmitted to the surgical equipment control unit, which executes the instrument actions step by step in sequence, continuously collecting feedback data during execution for dynamic updates of subsequent boundary parameters.

[0121] In one optional embodiment, a boundary normal vector field is constructed based on the intervention boundary and the damage boundary. The intervention normal distance and the damage normal distance of the state point are calculated. A constraint potential energy field is generated by nonlinear weighted fusion, including:

[0122] Extract the boundary surface mesh nodes of the intervention boundary and the damage boundary, and calculate the unit normal vectors of each boundary surface mesh node pointing into the surgical safety decision space to construct the boundary normal vector field;

[0123] In the surgical safety decision space, the state points obtained by mapping the causal shielding nodes are calculated, and the Euclidean distance from the state point to the nearest grid node of the intervention boundary is calculated to determine the intervention normal distance. The Euclidean distance from the state point to the nearest grid node of the damage boundary is also calculated to determine the damage normal distance.

[0124] The intervention potential energy vector is obtained by performing a scalar multiplication operation between the intervention normal distance and the unit normal vector of the corresponding intervention boundary grid node, and the damage potential energy vector is obtained by performing a scalar multiplication operation between the damage normal distance and the unit normal vector of the corresponding damage boundary grid node. The intervention potential energy vector and the damage potential energy vector are then subjected to nonlinear weighted fusion to generate the constraint potential energy vector.

[0125] The constraint potential vectors of all state points constitute the constraint potential field.

[0126] In one specific implementation, extracting the boundary surface mesh nodes of the intervention boundary and damage boundary is the starting point for constructing the boundary normal vector field. Intervention boundary With damage boundary Geometrically, all surfaces are discretized and stored as triangular mesh surfaces, each consisting of several mesh nodes and their connections. For the intervention boundary surface, all its triangular faces are traversed, and the cross product direction of the three edge vectors of each triangular face is used as the original normal vector of that face. This normal vector is then normalized to a unit vector, and its orientation towards the surgical safety decision space is verified. Internal directional consistency—if the cross product points outwards, it is inverted. For each grid node, the unit normal vectors of all its adjacent triangular faces are weighted and averaged (weights are the areas of each face), and then the mean vector is normalized to obtain the unit normal vector at that grid node. Damage boundary The same process is used to create normal vector index tables for all grid nodes of both the intervention and damage boundaries, thereby constructing the boundary normal vector field. This normal vector field is stored in discrete form and retrieved in subsequent calculations using spatial nearest neighbor queries.

[0127] In the surgical safety decision space In the context of causal decoupling, the feature vector during the surgical period... via calibration coordinate transformation matrix After projection, the corresponding state points are obtained. For each state point At the intervention boundary Mesh node set and damage boundary Perform a nearest neighbor search within the set of grid nodes. For the intervention boundary, calculate... The Euclidean distance between all grid nodes on the intervention boundary is used to determine the grid node with the minimum Euclidean distance on the intervention boundary. This minimum Euclidean distance is then denoted as the intervention normal distance. Simultaneously, record the unit normal vector at the nearest grid node. Similarly, performing the same operation on the damage boundary yields the damage normal distance. and the unit normal vector at the nearest mesh node of the damage boundary To improve the efficiency of nearest neighbor search, KD-tree index structures are pre-built for the grid nodes of intervention and damage boundaries, reducing the time complexity of each query from linear to logarithmic, thus meeting the stringent requirements of real-time surgical decision-making for response speed.

[0128] Obtain the normal distance of the intervention and damage normal distance After obtaining the corresponding unit normal vector, scalar multiplication is performed to generate the potential energy vector. Intervention potential energy vector Intervention normal distance The unit normal vector of the nearest grid node to the intervention boundary Multiplying them together yields the result, i.e. Damage potential vector Damage normal distance The unit normal vector of the nearest mesh node to the damage boundary Multiplying them together yields the result, i.e. The physical meaning of the above scalar multiplication is that distance information is encoded into the direction vector, so that the potential energy vector simultaneously carries the proximity of the state point to the boundary and the direction information of the constraint to be applied. When the state point is closer to the boundary, the magnitude of the potential energy vector is smaller, and the corresponding constraint guidance effect is stronger; when the state point is farther from the boundary, the magnitude of the potential energy vector increases, and the constraint guidance tends to be gentler.

[0129] Intervention potential vector and damage potential vector Perform nonlinear weighted fusion to generate a constraint potential vector. The core of nonlinear weighting lies in the two sets of weight coefficients. and These all change dynamically with the distance between the state point and the corresponding boundary, rather than being a fixed constant. Specifically, Follow The decrease rather than the nonlinear increase reflects the sensitivity to the approaching response to the intervention boundary; Follow The decrease exhibits a steeper nonlinear increase, reflecting the priority given to protecting the damage boundary. The fusion result is... ,in and The nonlinear form can be implemented using an inverse proportional function or an exponentially decaying function, ensuring that the weights approach infinity as the distance approaches zero, thus generating a sufficiently strong constraint repulsive force near the boundary to prevent state points from crossing the boundary. In practical implementations, to avoid numerical singularities where the denominator is zero, [the following is omitted as it is not explicitly stated in the original text]. and Each variable is regularized by introducing a small positive bias.

[0130] The final construction of the confined potential field is achieved through the surgical safety decision space. The above process is repeated for all state points. The intervention potential vector is calculated independently for each state point. and damage potential vector And generate the corresponding constraint potential vector according to the above nonlinear weighted fusion rule. The constraint potential energy vectors of all state points are aggregated to form a constraint potential energy field covering the entire surgical safety decision space. This potential energy field is stored in the form of a vector field, where each state point corresponds to a three-dimensional constraint potential energy vector, reflecting the comprehensive guiding direction and intensity of that point under the dual constraints of the intervention boundary and the damage boundary. During the constraint gradient search phase, the constraint gradient at each step of the gradient descent iteration is queried from this potential energy field or calculated in real time, driving the trajectory of the state point to remain within the constraints while satisfying the surgical objectives. Within the safe zone.

[0131] This constraint potential field construction scheme exhibits excellent geometric adaptability, capable of handling situations where the shapes of the intervention and damage boundaries are complex and where local curvature changes drastically. Since the normal vector field is established based on discrete nodes of the actual mesh surface, its directional information remains consistent with the geometric relationships of the real anatomical structure and instrument motion space, avoiding directional deviations caused by simplified geometric assumptions. Simultaneously, the nonlinear weighted fusion mechanism ensures that the constraint weight of the damage boundary is never lower than that of the intervention boundary under any circumstances, thereby prioritizing surgical safety at the decision-making level.

[0132] A second aspect of this invention provides an intelligent decision-making system based on Internet of Things (IoT) for fusion of multi-source data during surgery, comprising:

[0133] The data synchronization unit is used to acquire multi-source heterogeneous data streams during the surgical period through the IoT sensing layer.

[0134] The causal analysis unit is used to identify cross-source response events triggered by the same surgical operation in multi-source heterogeneous data streams, extract the peak time of the cross-source response events as the surgical event anchor point, calculate the time phase difference between surgical event anchor points from different data sources, perform reverse compensation on the collection timestamps of each data source based on the time phase difference, and generate a surgical period dataset with a unified time reference.

[0135] The decision generation unit is used to perform conditional independence tests between data sources on the surgical dataset, construct a causal dependency topology between data sources based on the test results, identify causal blocking nodes in the causal dependency topology through causal blocking tests, and generate causal decoupling surgical features based on the causal blocking nodes.

[0136] The feedback update unit is used to construct a surgical safety decision space containing intervention and damage boundaries based on the current surgical process stage, project the causal decoupled surgical period features onto the surgical safety decision space, and generate decision instructions that meet the surgical objectives through constrained gradient search within the surgical safety decision space.

[0137] The processing unit is used to transmit decision instructions to the surgical equipment control unit for execution, and to collect surgical feedback data after execution to dynamically update the boundary parameters of the surgical safety decision space.

[0138] A third aspect of the present invention provides an electronic device, comprising:

[0139] processor;

[0140] Memory used to store processor-executable instructions;

[0141] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0142] A fourth aspect of the present invention provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the aforementioned method.

[0143] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0144] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An intelligent decision-making system for multi-source data fusion during surgery based on the Internet of Things, characterized in that: include: The data synchronization unit is used to acquire multi-source heterogeneous data streams during the surgical period through the IoT sensing layer; The causal analysis unit is used to identify cross-source response events triggered by the same surgical operation in multi-source heterogeneous data streams, extract the peak time of the cross-source response events as the surgical event anchor point, calculate the time phase difference between surgical event anchor points from different data sources, perform reverse compensation on the collection timestamps of each data source based on the time phase difference, and generate a surgical period dataset with a unified time reference. The decision generation unit is used to perform conditional independence tests between data sources on the surgical dataset, construct a causal dependency topology between data sources based on the test results, identify causal blocking nodes in the causal dependency topology through causal blocking tests, and generate causal decoupling surgical features based on the causal blocking nodes. The feedback update unit is used to construct a surgical safety decision space containing intervention and damage boundaries based on the current surgical process stage, project the causal decoupled surgical period features onto the surgical safety decision space, and generate decision instructions that meet the surgical objectives through constrained gradient search within the surgical safety decision space. The processing unit is used to transmit decision instructions to the surgical equipment control unit for execution, and collect surgical feedback data after execution to dynamically update the boundary parameters of the surgical safety decision space; The decision generation unit performs conditional independence tests between data sources on the surgical period dataset and constructs a causal dependency topology between data sources based on the test results, including: Extract data state snapshots of each data source at the moment of surgical instrument action from the surgical period dataset, and construct a surgical data source graph structure with data sources as nodes and data state snapshots as node attributes; Two data source nodes are selected from the surgical data source graph structure as the node pair to be tested, and the other data source nodes are selected as condition variable nodes in turn. For each condition variable node, calculate the conditional mutual information of the node pair to be tested under the condition variable node. When the conditional mutual information is lower than the preset independence threshold, determine that the node pair to be tested is conditionally independent under the condition variable node. Record all condition variable nodes that make the node pair to be tested conditionally independent to form an intermediate transmission node set. When the intermediate transmission node set is empty, causal dependency edges are established between the pairs of nodes to be tested. When the intermediate transmission node set is not empty, the condition variable nodes in the intermediate transmission node set are marked as causal transmission path nodes between the pairs of nodes to be tested. Perform the above conditional independence test on all data source node pairs in the surgical data source graph structure, and combine all data source nodes and the causal dependency edges established between them to construct the causal dependency topology between data sources.

2. The system according to claim 1, characterized in that, In multi-source heterogeneous data streams, cross-source response events triggered by the same surgical operation are identified. The peak time of these cross-source response events is extracted as the surgical event anchor point. The time phase difference between surgical event anchor points from different data sources is calculated. Based on the time phase difference, reverse compensation is performed on the acquisition timestamps of each data source to generate a surgical period dataset with a unified time reference, including: An adaptive sliding window scan is performed on the multi-source heterogeneous data stream. The energy change rate of each data source signal is calculated within the adaptive sliding window. When the energy change rate exceeds a preset dynamic threshold, it is marked as a potential response point. Waveform segments before and after the potential response point are extracted to construct a response waveform library. Morphological feature extraction is performed on waveform segments in the response waveform library. The kurtosis, skewness, and temporal envelope area of ​​the waveform segments are calculated as morphological feature vectors. By matching the morphological feature vectors across data sources, cross-source response events triggered by the same surgical operation are identified, and the successfully matched waveform segments are grouped into response event groups. Extract the maximum amplitude moment of each data source waveform segment from the response event group, mark the surgical event anchor point, construct a time series correlation graph with the surgical event anchor point of each data source as the node, and calculate the time offset of the surgical event anchor point relative to the surgical event anchor point of the reference data source in the time series correlation graph as the time phase difference. Establish time mapping relationships between various data sources based on time phase differences, apply a time offset opposite to the time phase difference to the collection timestamps of each data source to perform reverse compensation, and rearrange the multi-source heterogeneous data streams of each data source according to a unified time axis to generate a surgical period dataset.

3. The system according to claim 1, characterized in that, Causal shielding nodes are identified in causal dependency topologies through causal blocking tests. Based on these causal shielding nodes, causal decoupling features are generated during the operative period, including: In the causal dependency topology, identify the target data source node corresponding to the surgical decision target and the source data source node corresponding to the surgical instrument control command. Starting from the target data source node, perform a reverse breadth-first search along the causal dependency edge to the source data source node, record all data source nodes traversed by the reverse breadth-first search, and construct a set of causal transmission nodes. For each data source node in the causal transmission node set, perform a causal blocking test, temporarily remove the data source node and its connected causal dependency edges from the causal dependency topology, re-execute the causal path search from the source data source node to the target data source node, and when the causal path search fails, mark the data source node as a causal blocking node, and combine the causal blocking node and its connected causal dependency edges to form a causal backbone path. In the causal dependency topology, data source nodes that are not causal shielding nodes are marked as bypass coupling nodes, and bypass coupling nodes and their connected causal dependency edges are combined to form bypass coupling paths. Extract the data state snapshot corresponding to the causal shielding node as the causal decoupling operation period feature.

4. The system according to claim 1, characterized in that, The surgical safety decision space, which includes intervention and damage boundaries, is constructed based on the current stage of the surgical procedure. Obtain the standard surgical operation sequence corresponding to the current stage of the surgical process from the surgical procedure database, analyze the instrument movement type and tissue contact range corresponding to each surgical operation step in the standard surgical operation sequence, and map the instrument movement type and tissue contact range to the three-dimensional coordinate system of the surgical scene to form a standard operation space area. Based on the causal decoupling of the surgical period features, the physiological state deviation of each anatomical structure in the current surgical scene is calculated. Anatomical structures whose physiological state deviation exceeds the preset safety threshold are marked as high-risk tissue areas. The spatial boundary of the high-risk tissue area is marked in the three-dimensional coordinate system of the surgical scene to construct the damage boundary. Based on the standard operating space area, a preset safe distance is extended outward in the three-dimensional coordinate system of the surgical scene to form an operating allowable space area. The space area between the operating allowable space area and the damage boundary is defined as the intervention boundary. The spatial region enclosed by the intervention boundary and the damage boundary in the three-dimensional coordinate system of the surgical scene is defined as the surgical safety decision space.

5. The system according to claim 1, characterized in that, Projecting causal decoupling of perioperative features onto the surgical safety decision space, and generating decision instructions that satisfy surgical objectives through constrained gradient search within the surgical safety decision space, including: Extract the data state snapshots corresponding to each causal shielding node from the causal decoupling surgical period features, convert the instrument space coordinates and tissue deformation displacement in the data state snapshots into state point coordinates in the surgical safety decision space, and establish the mapping relationship between causal shielding nodes and state points in the surgical safety decision space. Based on the intervention boundary and the damage boundary, a boundary normal vector field is constructed, the intervention normal distance and the damage normal distance of the state point are calculated, and a constraint potential energy field is generated by nonlinear weighted fusion. Starting from the target state point corresponding to the surgical target, gradient descent iteration is performed along the negative gradient direction of the constraint potential energy field. In each iteration, the constraint potential energy gradient vector of the current iteration point is calculated, and the preset step size is moved to the next iteration point in the opposite direction of the constraint potential energy gradient vector. The sequence of iteration points passed during the iteration process constitutes the constraint guidance trajectory. The coordinates of the state points corresponding to each iteration point in the constraint-guided trajectory are reverse-mapped to the causal shielding node, and the instrument control parameters corresponding to the causal shielding node are extracted to generate decision instructions.

6. The system according to claim 5, characterized in that, Based on the intervention boundary and damage boundary, a boundary normal vector field is constructed. The intervention normal distance and damage normal distance of the state point are calculated. The constraint potential energy field is generated through nonlinear weighted fusion, including: Extract the boundary surface mesh nodes of the intervention boundary and the damage boundary, and calculate the unit normal vectors of each boundary surface mesh node pointing into the surgical safety decision space to construct the boundary normal vector field; In the surgical safety decision space, the state points obtained by mapping the causal shielding nodes are calculated, and the Euclidean distance from the state point to the nearest grid node of the intervention boundary is calculated to determine the intervention normal distance. The Euclidean distance from the state point to the nearest grid node of the damage boundary is also calculated to determine the damage normal distance. The intervention potential energy vector is obtained by performing a scalar multiplication operation between the intervention normal distance and the unit normal vector of the corresponding intervention boundary grid node, and the damage potential energy vector is obtained by performing a scalar multiplication operation between the damage normal distance and the unit normal vector of the corresponding damage boundary grid node. The intervention potential energy vector and the damage potential energy vector are then subjected to nonlinear weighted fusion to generate the constraint potential energy vector. The constraint potential vectors of all state points constitute the constraint potential field.

7. An electronic device, characterized in that, include: processor; Memory used to store processor-executable instructions; The processor is configured to invoke instructions stored in the memory, and when the instructions are executed by the processor, they implement the system according to any one of claims 1 to 6.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that, When the computer program instructions are executed by the processor, they implement the system according to any one of claims 1 to 6.

Images