AI-based multiple bone trauma criticality detection method

Abstract

The present application relates to the technical field of medical artificial intelligence, and more particularly to a method for detecting critical multiple bone trauma based on AI. The method integrates physiological monitoring and image data of the patient, extracts the fracture site and constructs a whole-body bleeding risk distribution map, performs correlation analysis on hemoglobin changes to locate the bleeding source, and predicts the decompensation time window through the coagulation state trajectory. The above features are input into the decision model to simulate the physiological evolution path of different treatment schemes and calculate the risk, and finally output the treatment scheme with the lowest adverse probability and the decision time node. The present application realizes dynamic quantitative evaluation and optimized decision support of the detection results of the bleeding risk of critical multiple bone trauma.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, and in particular to an AI-based method for detecting critical cases of multiple bone trauma. Background Technology

[0002] In the clinical treatment of critically ill patients with multiple bone injuries, rapid and accurate assessment of the patient's condition and development of a treatment plan are crucial for saving lives. Such injuries are often accompanied by complex fractures and severe bleeding. Clinical decision-making requires a comprehensive assessment of imaging findings and real-time physiological monitoring data to determine the patient's bleeding risk, coagulation status, and disease progression.

[0003] In current clinical practice, the assessment and decision-making for critically ill patients with multiple bone traumas mainly rely on the step-by-step analysis of discrete examination results and clinical experience by the medical team. This conventional model, which depends on manual integration and experience-based judgment, has significant shortcomings. On the one hand, the efficiency and accuracy of locating the bleeding source are insufficient. Imaging data only provides static anatomical structural information, making it difficult to dynamically quantify the extent of vascular network involvement in different fracture areas and the resulting bleeding risk level. Furthermore, changes in physiological indicators are systemic and lagging. Without a quantitative correlation model, it is difficult for doctors to quickly and accurately correlate the decline in hemoglobin with the specific, highest-risk anatomical site, resulting in a lack of targeted interventions and delays in treating critical bleeding points.

[0004] On the other hand, there is a lag in the early warning of disease deterioration. Coagulation function monitoring indicators are typically discrete, point-like values, requiring clinicians to rely on experience to interpret the coagulation status represented by these indicator combinations. Conventional methods struggle to continuously and dynamically depict the complete trajectory of the coagulation system's evolution from compensated to decompensated, nor can they quantify the remaining time window before decompensation occurs. This often leads to clinical decisions being made only after the patient's physiological indicators have significantly deteriorated, resulting in a reactive rather than proactive intervention based on prediction, thus missing the optimal opportunity to reverse the disease early. Summary of the Invention

[0005] This invention provides an AI-based method for detecting critical cases of multiple bone traumas, which can solve the problems in the prior art.

[0006] A first aspect of the present invention provides an AI-based method for detecting critical cases of multiple bone trauma, comprising:

[0007] Collect physiological monitoring data and imaging data from the target patients;

[0008] The fracture site is extracted from the image data and mapped to a three-dimensional anatomical coordinate system. The densely vascularized areas involved by the fracture are calculated. A local bleeding risk field is constructed for each area. A systemic bleeding risk distribution map is generated by the field strength superposition algorithm, and the centroid coordinates and diffusion radius of the risk area are marked.

[0009] Wavelet decomposition was performed on the hemoglobin concentration sequence of physiological monitoring data to obtain the trend component and calculate the rate of decrease. This component was then correlated with the bleeding risk distribution map to locate the bleeding source.

[0010] The coagulation index data of physiological monitoring data are projected onto the state space of coagulation cascade reaction, and the movement trajectory of coagulation state is tracked. When the trajectory crosses the compensated coagulation area, the shortest distance between it and the boundary of the decompensated area is calculated and converted into the remaining time window for the occurrence of decompensation.

[0011] The bleeding risk distribution map, bleeding source location, and remaining time window are used as features to input the decision tree model. The physiological parameter change trajectory is simulated for different treatment plans. The set of evolutionary paths is generated by Monte Carlo sampling and the adverse probability of each evolutionary path is calculated.

[0012] Extract the treatment plan and its target decision time point corresponding to the evolutionary path with the lowest adverse probability, and output it as the detection result.

[0013] The fracture site is extracted from the image data and mapped to a three-dimensional anatomical coordinate system. The densely vascularized areas affected by the fracture are calculated, and a local hemorrhage risk field is constructed. A systemic hemorrhage risk distribution map is generated using a field strength superposition algorithm, and the centroid coordinates and diffusion radius of high-risk areas are marked, including:

[0014] The image data is segmented to extract the fracture contour and mapped to a three-dimensional anatomical coordinate system. The corresponding vascular atlas is retrieved and a search radius is set with the fracture contour as the center. The vascular line density at each spatial location within the search radius is counted, and connected regions with vascular line density exceeding the density threshold are extracted as vascular dense regions.

[0015] Multiple sampling points are uniformly sampled on the boundary surface of each densely vascular region to generate sub-field sources, and the initial intensity of the sub-field sources is determined. The intensity of the sub-field sources is then corrected based on the degree of fracture displacement. The field intensity distribution of each sub-field source is described by a Gaussian decay function, and scalar superposition is performed to construct a local hemorrhage risk field.

[0016] A spatial grid is established in the three-dimensional anatomical coordinate system. The field strength value of all local bleeding risk fields is calculated for each grid node. The field strength gradient direction is calculated based on the field strength value of each local bleeding risk field. When the angle between the field strength gradient directions is less than the cooperative angle, the corresponding field strength value is enhanced and summed to obtain the total field strength of each grid node, which constitutes the whole-body bleeding risk distribution map.

[0017] Clustering is performed on the grid nodes in the systemic bleeding risk distribution map where the field strength value exceeds the risk threshold to form high-risk areas. The weighted average of the grid node coordinates in each high-risk area is calculated as the centroid coordinates, and the standard deviation of the distance between the grid node and the centroid coordinates in each high-risk area is calculated as the diffusion radius.

[0018] Multiple sampling points are uniformly sampled on the boundary surface of each densely vascularized region to generate sub-field sources. The initial intensity of the sub-field sources is determined and corrected based on the degree of fracture displacement to obtain the sub-field source intensity. A Gaussian decay function is used to describe the field intensity distribution of each sub-field source, and scalar superposition is performed to construct a local hemorrhage risk field, including:

[0019] Boundary extraction is performed on each densely vascular region, the normal vector distribution of the boundary is calculated, the concave and convex features are identified based on the normal vector distribution and the sampling priority is set, sampling points are generated on the boundary as sub-field sources according to the sampling priority, and the normal vector direction of each sub-field source is recorded as the field source direction.

[0020] For each sub-field source, the number and type distribution of blood vessels within the neighborhood are statistically analyzed, the density of blood vessels and the tissue fragility coefficient are calculated, and weighted fusion is performed to obtain the initial intensity of the sub-field source.

[0021] Fracture profiles are classified and identified. Based on the fracture classification, the displacement risk level is determined. The centroid coordinates of the fracture fragments are extracted. The displacement amplitude is determined based on the maximum distance between the centroid coordinates. The displacement risk level and displacement amplitude are mapped to a correction coefficient. The initial intensity of each subfield source is multiplied by the correction coefficient to obtain the subfield source intensity.

[0022] Anisotropic Gaussian decay functions are constructed with each sub-source as the center. The decay rate of the Gaussian decay function along the source direction is lower than the decay rate perpendicular to the source direction. The decay rate is determined by the sub-source intensity and the spatial scale of the densely vascularized region. The Gaussian decay functions are superimposed in a three-dimensional anatomical coordinate system to construct a local hemorrhage risk field.

[0023] Wavelet decomposition was performed on the hemoglobin concentration sequence of physiological monitoring data to obtain the trend component and calculate the rate of decrease. This component was then correlated with the bleeding risk distribution map to locate the bleeding source, including:

[0024] Wavelet basis functions are selected to perform multi-level decomposition of hemoglobin concentration sequence, and approximate coefficients are extracted as trend components and polynomial fitting is performed. The first derivative is calculated to obtain the rate of decline sequence. Local extrema are detected and the time interval and amplitude difference between adjacent extrema points are calculated. Multiplication is performed to obtain the fluctuation energy value. The time period when the fluctuation energy value exceeds the energy threshold is extracted as the active bleeding period.

[0025] Arterial blood pressure waveform data during the active bleeding period were extracted from physiological monitoring data and subjected to Fourier transform. The amplitude and phase of the dominant frequency component in the frequency domain feature were calculated and a blood pressure oscillation feature vector was constructed.

[0026] A three-dimensional grid index is established in the risk distribution map of systemic bleeding. The spatial gradient tensor of the field strength value is calculated by traversing each grid cell. The spatial gradient tensor is decomposed into eigenvalues ​​to extract the principal direction feature vector. The directional derivative of the field strength value is calculated along the principal direction feature vector to identify the location of the maximum rate of change.

[0027] A multiple regression model is established between the blood pressure oscillation feature vector and the location of the maximum rate of change. The explained variance of each location of the maximum rate of change on the blood pressure oscillation feature vector is calculated, and the location with the largest explained variance is extracted as the coordinates for locating the main blood source.

[0028] Coagulation index data from physiological monitoring are projected onto the state space of the coagulation cascade reaction. The trajectory of the coagulation state is tracked. When the trajectory crosses the compensated coagulation region, the shortest distance between it and the boundary of the decompensated region is calculated and converted into the remaining time window for the occurrence of decompensation, including:

[0029] The reaction time, coagulation time, and maximum amplitude parameters of thromboelastography were extracted from physiological monitoring data as coagulation index data, and a multidimensional index vector was constructed after standardization transformation.

[0030] A state space for the coagulation cascade reaction is established. Principal component analysis is used to reduce the dimensionality of the multidimensional index vector to obtain the state coordinates. A classification hyperplane for the normal region, compensatory region and decompensated region is constructed by support vector machine. The normal vector and intercept parameters of each hyperplane are extracted to define the boundary.

[0031] The continuously collected multidimensional index vectors are projected onto the state space according to the time series to form a trajectory point sequence and smoothed by a sliding window. The position derivative is calculated to obtain the velocity field. The velocity field is then decomposed into vector components along the boundary normal and tangential components along the boundary tangential.

[0032] The system determines the current trajectory point's location in real time. When entering a compensatory region, it calculates the vertical distance from the current trajectory point to the hyperplane of the uncompensated region as the shortest distance. It extracts the magnitude of the normal component as the approach rate, calculates the base time based on the shortest distance, calculates the stability coefficient based on the magnitude of the tangential component, and uses the stability coefficient to perform weighted correction on the base time. The output is the remaining time window for the occurrence of uncompensation.

[0033] The bleeding risk distribution map, bleeding source location, and remaining time window are used as features input to a decision tree model. The model simulates the physiological parameter changes for different treatment plans, generates an evolutionary path set through Monte Carlo sampling, and calculates the adverse probability of each path, including:

[0034] Spatial features are extracted from the bleeding risk distribution map, the spatial distribution density and risk gradient value of each high-risk area are calculated, the spatial coordinates and risk intensity of the bleeding source are obtained based on the bleeding source location, the maximum intervention delay time is calculated based on the remaining time window, and the features are combined to form a feature vector.

[0035] The feature vector is input into the decision tree model, which generates multiple candidate treatment plans based on a preset treatment plan library. Each candidate treatment plan includes intervention start time and intervention intensity parameters. A physiological response function describing the evolution law of hemodynamic parameters is constructed for each candidate treatment plan, and the physiological parameter change trajectory corresponding to each candidate treatment plan is calculated.

[0036] Monte Carlo sampling was performed on the trajectory of changes in various physiological parameters, and random perturbation sampling was performed in the parameter space to generate evolutionary paths with random perturbations. Repeated sampling was used to obtain a set of evolutionary paths.

[0037] Determine whether the physiological parameters of each evolutionary path in the set of evolutionary paths exceed the safety threshold range, mark the time when the safety threshold range is exceeded as the occurrence of adverse events, count the number or duration of adverse events in each evolutionary path, and calculate the adverse probability corresponding to each evolutionary path.

[0038] Monte Carlo sampling was performed on the trajectories of changes in various physiological parameters, and random perturbation sampling was performed within the parameter space to generate evolutionary paths with random perturbations. Repeated sampling yielded a set of evolutionary paths, including:

[0039] Parameters describing the treatment response characteristics are extracted from the trajectory of each physiological parameter change, and the influence of the unit change of each parameter on the final state of the physiological parameter change trajectory is calculated. The parameter set is sorted by sensitivity, and parameters with sensitivity higher than the sensitivity threshold are selected to form a subset of perturbation target parameters.

[0040] Based on the statistical distribution characteristics of each parameter in the subset of disturbance target parameters, the disturbance center value and the disturbance amplitude range are determined. Random disturbance quantities are generated within the disturbance amplitude range, and the disturbance parameter values ​​are calculated. All disturbance parameter values ​​are combined to form a disturbance parameter vector.

[0041] The perturbation parameter vector is substituted into the physiological response function to recalculate the time evolution process of hemodynamic parameters, resulting in an evolution path with random perturbation. The physiological parameter values ​​at each moment in the evolution path with random perturbation are recorded.

[0042] The process of generating random perturbations, superimposing, and recalculating is repeated. Each time the process is repeated, a new random perturbation is generated independently to produce a different perturbation parameter vector. All evolution paths with random perturbations obtained from multiple repetitions are summarized to form an evolution path set.

[0043] A second aspect of the present invention provides an AI-based critical care system for multiple bone traumas, comprising:

[0044] The data acquisition unit is used to collect physiological monitoring data and imaging data of the target patient;

[0045] The risk modeling unit is used to extract the fracture site from the image data and map it to a three-dimensional anatomical coordinate system, calculate the densely vascularized area involved by the fracture, construct a local bleeding risk field for each area, generate a systemic bleeding risk distribution map through a field strength superposition algorithm, and mark the centroid coordinates and diffusion radius of the risk area.

[0046] The bleeding localization unit is used to perform wavelet decomposition on the hemoglobin concentration sequence of physiological monitoring data, obtain the trend component and calculate the rate of decrease, and perform correlation calculation with the bleeding risk distribution map to locate the bleeding source.

[0047] The coagulation analysis unit is used to project the coagulation index data of physiological monitoring data into the state space of the coagulation cascade reaction, track the movement trajectory of the coagulation state, and calculate the shortest distance between the trajectory and the boundary of the decompensated region when the trajectory crosses the compensated coagulation region and convert it into the remaining time window for the occurrence of decompensation.

[0048] The decision simulation unit is used to input the bleeding risk distribution map, bleeding source location and remaining time window as features into the decision tree model, simulate the change trajectory of physiological parameters for different treatment plans, generate a set of evolution paths through Monte Carlo sampling and calculate the adverse probability of each evolution path;

[0049] The scheme output unit is used to extract the treatment scheme and its target decision time node corresponding to the evolution path with the lowest adverse probability, and output it as the detection result.

[0050] A third aspect of the present invention provides an electronic device, comprising:

[0051] processor;

[0052] Memory used to store processor-executable instructions;

[0053] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0054] 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.

[0055] This method enables precise dynamic assessment and quantitative prediction of bleeding risk and coagulation status in critically ill patients with multiple bone traumas. By mapping fracture sites to a three-dimensional anatomical coordinate system and calculating densely vascularized areas, a local bleeding risk field can be constructed. Then, by superimposing the field strength, a systemic bleeding risk distribution map is generated, quantifying the potential bleeding threat in different anatomical regions. Combined with wavelet decomposition trend analysis of hemoglobin concentration sequences, the dynamic changes of physiological indicators can be correlated with the spatial risk map, thereby achieving rapid localization of active bleeding sources and significantly improving the early identification capability of occult bleeding.

[0056] This method enables continuous tracking and prospective early warning of a patient's coagulation function status. By projecting coagulation index data onto the state space of the coagulation cascade reaction and tracking its trajectory, the compensatory state of the coagulation system can be dynamically characterized. When the trajectory approaches the decompensation boundary, its shortest distance is calculated and converted into a remaining time window, thereby quantitatively predicting the time of occurrence of coagulation decompensation.

[0057] This method generates individualized optimal treatment plans and decision-making timings based on multi-source quantitative features. By inputting features such as bleeding risk, bleeding source location, and coagulation decompensation time window into a decision tree model, and simulating the evolution path of physiological parameters for different treatment plans, the corresponding treatment plans and key decision-making time nodes are extracted as outputs. This process realizes the simulation and quantitative comparison of treatment strategies, providing clinicians with data-driven, precise decision support, and helping to select the safest and most effective interventions and implementation timings in complex and critical situations. Attached Figure Description

[0058] Figure 1 This is a flowchart illustrating the AI-based method for detecting critical cases of multiple bone trauma according to an embodiment of the present invention.

[0059] Figure 2 This is a flowchart illustrating the method for constructing a local bleeding risk field according to an embodiment of the present invention. Detailed Implementation

[0060] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0061] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0062] Figure 1 This is a flowchart illustrating the AI-based method for detecting critically ill patients with multiple bone injuries according to an embodiment of the present invention. Figure 1 As shown, AI-based methods for detecting critical cases of multiple bone trauma include:

[0063] Collect physiological monitoring data and imaging data from the target patients;

[0064] The fracture site is extracted from the image data and mapped to a three-dimensional anatomical coordinate system. The densely vascularized areas involved by the fracture are calculated. A local bleeding risk field is constructed for each area. A systemic bleeding risk distribution map is generated by the field strength superposition algorithm, and the centroid coordinates and diffusion radius of the risk area are marked.

[0065] Wavelet decomposition was performed on the hemoglobin concentration sequence of physiological monitoring data to obtain the trend component and calculate the rate of decrease. This component was then correlated with the bleeding risk distribution map to locate the bleeding source.

[0066] The coagulation index data of physiological monitoring data are projected onto the state space of coagulation cascade reaction, and the movement trajectory of coagulation state is tracked. When the trajectory crosses the compensated coagulation area, the shortest distance between it and the boundary of the decompensated area is calculated and converted into the remaining time window for the occurrence of decompensation.

[0067] The bleeding risk distribution map, bleeding source location, and remaining time window are used as features to input the decision tree model. The physiological parameter change trajectory is simulated for different treatment plans. The set of evolutionary paths is generated by Monte Carlo sampling and the adverse probability of each evolutionary path is calculated.

[0068] Extract the treatment plan and its target decision time point corresponding to the evolutionary path with the lowest adverse probability, and output it as the detection result.

[0069] Figure 2 This is a flowchart illustrating the method for constructing a local hemorrhage risk field according to an embodiment of the present invention. In one optional implementation, the fracture site is extracted from the image data and mapped to a three-dimensional anatomical coordinate system. The densely vascularized area involved in the fracture is calculated, and a local hemorrhage risk field is constructed. A systemic hemorrhage risk distribution map is generated using a field strength superposition algorithm, and the centroid coordinates and diffusion radius of high-risk areas are marked, including:

[0070] The image data is segmented to extract the fracture contour and mapped to a three-dimensional anatomical coordinate system. The corresponding vascular atlas is retrieved and a search radius is set with the fracture contour as the center. The vascular line density at each spatial location within the search radius is counted, and connected regions with vascular line density exceeding the density threshold are extracted as vascular dense regions.

[0071] Multiple sampling points are uniformly sampled on the boundary surface of each densely vascular region to generate sub-field sources, and the initial intensity of the sub-field sources is determined. The intensity of the sub-field sources is then corrected based on the degree of fracture displacement. The field intensity distribution of each sub-field source is described by a Gaussian decay function, and scalar superposition is performed to construct a local hemorrhage risk field.

[0072] A spatial grid is established in the three-dimensional anatomical coordinate system. The field strength value of all local bleeding risk fields is calculated for each grid node. The field strength gradient direction is calculated based on the field strength value of each local bleeding risk field. When the angle between the field strength gradient directions is less than the cooperative angle, the corresponding field strength value is enhanced and summed to obtain the total field strength of each grid node, which constitutes the whole-body bleeding risk distribution map.

[0073] Clustering is performed on the grid nodes in the systemic bleeding risk distribution map where the field strength value exceeds the risk threshold to form high-risk areas. The weighted average of the grid node coordinates in each high-risk area is calculated as the centroid coordinates, and the standard deviation of the distance between the grid node and the centroid coordinates in each high-risk area is calculated as the diffusion radius.

[0074] Based on imaging data acquired from medical imaging equipment such as CT or MRI, a deep learning segmentation network is used to accurately identify the fracture area. The segmentation network employs a variant of the U-Net architecture, trained to recognize discontinuities in the cortical bone, disordered trabecular bone structure, and radiographic features of the fracture line. After segmentation, the three-dimensional contour information of the fracture is extracted. This contour is represented as a point cloud, including the spatial coordinates of the fracture ends, the normal vector of the fracture surface, and the distribution of fracture fragments. The extracted fracture contour is mapped to a standard three-dimensional anatomical coordinate system. This coordinate system is established based on the patient's anatomical landmarks, with the origin typically set at the midpoint of the anterior superior border of the sacrum. The X-axis points to the patient's right side, the Y-axis points to the head, and the Z-axis points to the ventral side, ensuring a unified spatial reference system for imaging data from different patients.

[0075] After mapping, a pre-built human vascular atlas database is retrieved. This database contains information such as the three-dimensional course, diameter, and blood flow velocity of major arterial trunks and their branches. A search radius is set with the geometric center of each fracture contour as the sphere center for vascular retrieval. The search radius is dynamically determined based on the fracture type: 50mm for long bone fractures, 80mm for periarticular fractures, and 120mm for pelvic fractures. Within the search radius, the three-dimensional space is divided into voxel units with a side length of 2mm. The total length of the blood vessel passing through each voxel unit is calculated. This length divided by the voxel volume yields the vascular linear density, expressed in mm / mm². 3 After statistical analysis, voxel units with vessel line density exceeding a density threshold were extracted. The density threshold was set differently based on the anatomical location, with 0.15 mm / mm set for the trunk. 3 The proximal end of the limbs is set at 0.10 mm / mm. 3 The distal end of the limbs is set to 0.08 mm / mm. 3 Three-dimensional connectivity analysis was performed on voxel units that met the conditions. The six-connectivity criterion was used to aggregate adjacent voxels to form several spatially continuous vascular dense regions.

[0076] For each identified densely vascular region, a sub-field source is generated by uniformly sampling its three-dimensional boundary surface. The sampling process employs the Poisson disk sampling algorithm to ensure a minimum spacing of 5mm between adjacent sampling points, avoiding computational redundancy caused by excessively dense field sources. Each sampling point serves as a sub-field source, and its initial intensity is determined comprehensively based on the vascular linear density, vascular diameter, and hemodynamic parameters at that location. Specifically, the initial intensity of the sub-field source... With blood vessel line density Proportional to the average diameter of blood vessels It is proportional to the square of the value, and a blood flow velocity correction factor is introduced. The calculation formula is: The proportionality coefficient Determined through retrospective data fitting of clinical bleeding events.

[0077] Based on the initial strength, the subfield source strength is corrected according to the degree of fracture displacement. The degree of fracture displacement is quantified by measuring the relative displacement of the fracture ends, including three dimensions: axial displacement, lateral displacement, and rotation angle. The shortest distance between the fracture ends and the boundary of the densely vascularized region is calculated. When displacement causes the fractured bone fragments to pierce into areas with dense blood vessels, When the value is negative, the subfield source intensity is significantly enhanced. The corrected subfield source intensity. Calculated as The enhancement coefficient The value is 2.5, which is the attenuation coefficient. The value is 0.1, and the unit is per millimeter. When the fracture ends are far from areas with dense blood vessels, If the value is positive and relatively large, the exponential decay term approaches zero, and the correction effect weakens.

[0078] After determining the intensity of the sub-sources, a Gaussian decay function is used to describe the field intensity distribution of each sub-source in space, based on the sub-source location. Centered on any point in space field strength Represented as , where diffusion parameters Based on tissue type, the value is set to 15mm for muscle tissue, 20mm for adipose tissue, and 8mm for bone tissue. For densely vascularized regions containing multiple sub-field sources, a scalar superposition method is used to construct the local hemorrhage risk field, that is, the local field strength at any point in space is equal to the sum of the field strengths generated at that point by all sub-field sources in that region.

[0079] A spatial mesh is established in a standard three-dimensional anatomical coordinate system. The mesh resolution is adaptively adjusted according to the patient's body size, typically set to 5mm × 5mm × 5mm. Each mesh node is traversed, and the field strength contribution of all local bleeding risk fields at that node is calculated. Because different densely vascularized regions are spatially close, their local field strengths will have a superposition effect. To accurately reflect the synergistic effect of multi-source bleeding risks, the gradient direction of each local field strength is calculated at the mesh nodes. Place, No. Local field strength gradient direction The direction pointing to the direction of the fastest increase in electric field intensity. When the angle between the gradient directions of two local electric fields... Less than the cooperative angle threshold At that time, it was assumed that the two risk sources had a synergistic enhancement effect, and the synergistic angle threshold was set at 60 degrees. Under this condition, the corresponding field strength value was nonlinearly enhanced, and the enhancement coefficient was... The enhanced field strength values ​​are summed to obtain the total field strength value of the grid node. After traversing all grid nodes to complete the calculation, the systemic bleeding risk distribution map is constructed. This distribution map is stored in the form of a three-dimensional scalar field, recording the quantitative values ​​of bleeding risk at various spatial locations throughout the patient's body.

[0080] The generated systemic bleeding risk distribution map is segmented by thresholding, and grid nodes with field strength values ​​exceeding the risk threshold are extracted. The risk threshold is set according to the clinically acceptable bleeding probability, usually selecting the field strength value corresponding to a 10% bleeding incidence rate as the threshold. This threshold is determined by the receiver operating characteristic curve analysis of historical case data. The extracted high-risk grid nodes are spatially dispersed, and a density-based clustering algorithm is used to aggregate them into several high-risk regions. The clustering algorithm sets the neighborhood radius to 10 mm and the minimum number of points to 8 to ensure that the formed high-risk regions have spatial continuity and contain enough risk nodes.

[0081] For each identified high-risk area, its centroid coordinates and diffusion radius are calculated for quantitative description. The centroid coordinates are calculated using a weighted average method, with the weight of each grid node being its field strength value. The calculation formula is as follows: ,in Let be the total number of grid nodes contained in the j-th high-risk area. For the first The field strength value at each node, Its spatial coordinates. The diffusion radius is used to characterize the spatial diffusion degree of high-risk areas, and is obtained by calculating the standard deviation of the distance from each grid node to the centroid coordinates. Specifically, the calculation is as follows: This parameter reflects the spatial dispersion of bleeding risk. A larger diffusion radius indicates a wider range of risk diffusion, requiring a more comprehensive treatment intervention strategy. The centroid coordinates and diffusion radius are used as characteristic parameters of high-risk areas to provide accurate spatial positioning information and risk quantification indicators for subsequent clinical decisions.

[0082] In one optional implementation, multiple sampling points are uniformly sampled at the boundary surface of each densely vascularized region to generate sub-field sources. The initial intensity of the sub-field sources is determined, and the intensity is corrected based on the degree of fracture displacement to obtain the sub-field source intensity. A Gaussian decay function is used to describe the field intensity distribution of each sub-field source, and scalar superposition is performed to construct a local hemorrhage risk field, including:

[0083] Boundary extraction is performed on each densely vascular region, the normal vector distribution of the boundary is calculated, the concave and convex features are identified based on the normal vector distribution and the sampling priority is set, sampling points are generated on the boundary as sub-field sources according to the sampling priority, and the normal vector direction of each sub-field source is recorded as the field source direction.

[0084] For each sub-field source, the number and type distribution of blood vessels within the neighborhood are statistically analyzed, the density of blood vessels and the tissue fragility coefficient are calculated, and weighted fusion is performed to obtain the initial intensity of the sub-field source.

[0085] Fracture profiles are classified and identified. Based on the fracture classification, the displacement risk level is determined. The centroid coordinates of the fracture fragments are extracted. The displacement amplitude is determined based on the maximum distance between the centroid coordinates. The displacement risk level and displacement amplitude are mapped to a correction coefficient. The initial intensity of each subfield source is multiplied by the correction coefficient to obtain the subfield source intensity.

[0086] Anisotropic Gaussian decay functions are constructed with each sub-source as the center. The decay rate of the Gaussian decay function along the source direction is lower than the decay rate perpendicular to the source direction. The decay rate is determined by the sub-source intensity and the spatial scale of the densely vascularized region. The Gaussian decay functions are superimposed in a three-dimensional anatomical coordinate system to construct a local hemorrhage risk field.

[0087] After obtaining the 3D spatial location of the densely vascularized region, fine-grained sampling is required on the boundary surfaces of each region to construct the sub-field source distribution. The Marching Cubes algorithm is used to extract the boundary of the 3D volume data of the densely vascularized region. This algorithm separates the interior and exterior of the region by setting a density threshold, generating triangular mesh patches at the threshold boundaries to form closed boundary surfaces. A normal vector is calculated for each triangular mesh patch on the extracted boundary surface. The normal vector is calculated using the cross product operation of the patch vertices to ensure that the normal vector points outward. The rate of change of the normal vector at each point on the boundary surface is statistically analyzed. When the angle between the normal vectors of adjacent points exceeds 45 degrees, they are identified as concave or convex feature points. Concave regions are easily compressed during fracture displacement due to geometric constraints; therefore, the sampling priority of concave feature points is set to high priority, convex feature points to medium priority, and flat regions to low priority. Adaptive sampling is performed according to the sampling priority: the sampling density is 3 sampling points per square millimeter for high-priority regions, 2 sampling points per square millimeter for medium-priority regions, and 1 sampling point per square millimeter for low-priority regions. The sampling points are generated using the Poisson disk sampling method to ensure minimal spacing between them and avoid excessive clustering. The three-dimensional coordinates of each sampling point and its normal vector direction are recorded; this normal vector direction serves as the dominant direction for subsequent field diffusion.

[0088] For each generated sub-field source, its initial field strength needs to be calculated, reflecting the potential risk of bleeding at that location. The radius of the neighborhood for each sub-field source is defined as 8 mm, and the number of blood vessels within this neighborhood is counted. Blood vessel identification is achieved by performing blood vessel enhancement filtering on the image data. The Frangi blood vessel enhancement algorithm is used to calculate the Hessian matrix eigenvalues ​​of the image data. Based on the ratio of eigenvalues, tubular structures are identified, the centerline of the blood vessel is extracted, and skeletonization is performed. The number of blood vessel skeleton pixels within the neighborhood is counted as the blood vessel quantity index. Simultaneously, blood vessel types are identified, classified into arteries, veins, and capillaries based on their diameter. Vessels with a diameter greater than 2 mm are identified as arteries, those between 0.5 mm and 2 mm are identified as veins, and those less than 0.5 mm are identified as capillaries. Different blood vessel types contribute differently to the bleeding risk; arterial rupture leads to rapid and large-volume bleeding, therefore, arteries are assigned a weight coefficient of 1.5, veins a weight coefficient of 1.0, and capillaries a weight coefficient of 0.6. The blood vessel density index D is calculated. v The calculation method is as follows: multiply the number of arteries in the neighborhood by 1.5, add the number of veins by 1.0, add the number of capillaries by 0.6, and then divide by the neighborhood volume. Tissue fragility coefficient F t The calculation is based on the gray-level distribution characteristics of image data, extracting the gray-level standard deviation within the neighborhood as an indicator of tissue heterogeneity. The larger the gray-level standard deviation, the more complex the tissue structure and the higher the fragility. The tissue fragility coefficient F t The normalized result is obtained by dividing the gray standard deviation by the reference gray standard deviation of normal tissue in that area. The vascular density D... v With tissue fragility coefficient F t Weighted fusion is performed with weights set to 0.6 and 0.4 respectively to obtain the initial intensity I0 of the sub-field source, calculated as I0 = 0.6 × D. v +0.4×F t .

[0089] Fracture classification was performed on the fracture area using a deep learning-based fracture classification network. This network, with ResNet50 as the backbone, was pre-trained on the ImageNet dataset and fine-tuned on annotated fracture image datasets. The fracture classification followed the AO / OTA classification standard, dividing fractures into three types: Type A (simple fracture), Type B (wedge fracture), and Type C (complex fracture). Type A fractures were assigned a displacement risk level of 1, Type B level 2, and Type C level 3. Three-dimensional segmentation of the fracture fragments was performed using a region growing algorithm, expanding from the fracture line and grouping connected bone tissue pixels into the same segment. The centroid coordinates of each fracture fragment were calculated by weighted averaging of the coordinates of all pixels within the segment, with the weights being the pixel's grayscale value. All centroid coordinate pairs were iterated, and the Euclidean distance between each pair was calculated. The maximum distance was selected as the displacement magnitude. Displacement amplitude reflects the degree of spatial separation between fracture fragments; a larger displacement amplitude indicates more severe damage to surrounding tissues. A correction factor is calculated based on the displacement risk level and displacement amplitude. The calculation formula is: ,in The displacement risk level is determined by the initial intensity of each sub-field source. With correction factor Multiplying them together yields the corrected subfield source intensity. ,Right now .

[0090] When constructing the field strength distribution of each sub-field source, an anisotropic Gaussian decay function is used to simulate the diffusion characteristics of bleeding risk in space. Since tissue damage typically spreads faster along the normal vector direction, a lower decay rate is set along the field source direction (i.e., the normal vector direction), and a higher decay rate is set perpendicular to the field source direction. The field source coordinates are defined as follows: The unit vector of the source direction is The coordinates of any point in space are Calculate the vector from this point to the source. . Vector Decomposed into components parallel to the source direction. and components perpendicular to the source direction ,in , Define the attenuation coefficient along the source direction. Attenuation coefficient in the vertical direction The attenuation coefficient is determined by the subfield source intensity. Spatial scale of densely vascularized areas Jointly determined. Spatial scale. The equivalent radius of a densely vascularized region is calculated by taking the cube root of the region's volume divided by the volume of a sphere. Attenuation coefficient. Set as , Set as This ensures a slower decay rate along the source direction. The expression for the Gaussian decay function is: .

[0091] A regular mesh is established in a three-dimensional anatomical coordinate system with a resolution of 1 mm, covering the entire densely vascularized area and its surrounding extended region. For each voxel point in the mesh, the field strength value of that voxel point is calculated for all sub-field sources. The field strengths generated by each sub-field source at that point are then superimposed using scalar methods to obtain the total field strength at that point. ,in The total number of sub-field sources, For the first Individual venue source at point The field strength at a given location is determined. After scalar superposition, the field strength distribution data for the entire region is obtained, which represents the local hemorrhage risk field. The field strength distribution is normalized, mapping the field strength values ​​to the range of 0 to 1, facilitating subsequent fusion with field strengths from other regions. A threshold segmentation method is used to extract high-risk areas, setting the threshold to 0.7 of the normalized field strength; areas with field strength values ​​greater than 0.7 are marked as high-risk areas. The centroid coordinates of the high-risk areas are calculated by weighted averaging of the coordinates of all voxels within the high-risk area, with the weight being the field strength value of each voxel. The diffusion radius of the high-risk area is calculated, defined as the maximum distance from each voxel within the high-risk area to the centroid. The centroid coordinates and diffusion radius are recorded as feature parameters of this local hemorrhage risk field, used for subsequent construction of a systemic hemorrhage risk distribution map and hemorrhage source localization analysis.

[0092] In one optional implementation, wavelet decomposition is performed on the hemoglobin concentration sequence of physiological monitoring data to obtain a trend component and calculate the rate of decrease. This component is then correlated with the bleeding risk distribution map to locate the bleeding source, including:

[0093] Wavelet basis functions are selected to perform multi-level decomposition of hemoglobin concentration sequence, and approximate coefficients are extracted as trend components and polynomial fitting is performed. The first derivative is calculated to obtain the rate of decline sequence. Local extrema are detected and the time interval and amplitude difference between adjacent extrema points are calculated. Multiplication is performed to obtain the fluctuation energy value. The time period when the fluctuation energy value exceeds the energy threshold is extracted as the active bleeding period.

[0094] Arterial blood pressure waveform data during the active bleeding period were extracted from physiological monitoring data and subjected to Fourier transform. The amplitude and phase of the dominant frequency component in the frequency domain feature were calculated and a blood pressure oscillation feature vector was constructed.

[0095] A three-dimensional grid index is established in the risk distribution map of systemic bleeding. The spatial gradient tensor of the field strength value is calculated by traversing each grid cell. The spatial gradient tensor is decomposed into eigenvalues ​​to extract the principal direction feature vector. The directional derivative of the field strength value is calculated along the principal direction feature vector to identify the location of the maximum rate of change.

[0096] A multiple regression model is established between the blood pressure oscillation feature vector and the location of the maximum rate of change. The explained variance of each location of the maximum rate of change on the blood pressure oscillation feature vector is calculated, and the location with the largest explained variance is extracted as the coordinates for locating the main blood source.

[0097] In scenarios involving multiple bone traumas, the dynamic changes in hemoglobin concentration contain temporal information about hemorrhage activity. Wavelet decomposition technology can decompose non-stationary physiological signals into components at different time scales, thereby effectively separating trend changes from transient fluctuations. Specifically, the db4 wavelet basis function from the Daubechies wavelet family is selected as the decomposition basis. This wavelet basis possesses good time-frequency localization and orthogonality, making it suitable for handling abrupt changes in physiological signals. A five-level wavelet decomposition is performed on the acquired hemoglobin concentration sequence, yielding five detail coefficient sequences. and one approximate coefficient sequence Approximation coefficient The low-frequency trend component representing the signal reflects the overall trend of hemoglobin concentration changes. (This refers to the approximation coefficient sequence.) Perform a cubic polynomial fit, the fitted function is in the form of: ,in This indicates the trend of hemoglobin concentration after fitting. For time variables, The fitting coefficients are used. The first derivative of the fitting function is calculated to obtain the descent rate sequence. The rate of decrease sequence reflects the instantaneous rate of change in hemoglobin concentration; the higher the value, the more intense the bleeding activity.

[0098] For the descent rate sequence Local extremum detection is performed using the sliding window method to identify local maxima. The window length is set to 15 minutes, and the search is conducted within the window to find points that meet the specified conditions. and Time point As local maxima, extract all local maxima to form a set. Calculate the time interval between adjacent extreme points. and the corresponding amplitude differences The time interval reflects the periodic characteristics of blood activity, and the amplitude difference reflects the range of change in blood intensity. This is achieved through multiplication. The resulting fluctuation energy value, which comprehensively considers both time span and intensity variation, effectively quantifies the activity level of bleeding. An energy threshold is set based on statistical data from clinical experience. grams per minute per liter per minute squared, when At that time, determine the corresponding time period. This is the period of most active bleeding.

[0099] After identifying the active bleeding period, further analysis of the blood pressure waveform characteristics within this period is needed to establish a correlation with the location of the bleeding source. Arterial blood pressure waveform data during the active bleeding period is extracted from physiological monitoring data, typically recorded at a sampling frequency of 125 Hz. The extracted blood pressure waveform sequence... Perform a Fast Fourier Transform (FFT) to obtain the frequency domain representation. ,in This represents a frequency variable. The frequency component with the largest amplitude in the frequency domain is identified as the dominant frequency component. Record its amplitude and phase The dominant frequency component reflects the dominant oscillation mode of blood pressure fluctuations. Hemorrhage at different anatomical locations will cause blood pressure waveforms to exhibit different frequency characteristics. A blood pressure oscillation feature vector is constructed. This vector contains three key feature dimensions of blood pressure oscillation: amplitude, phase, and frequency.

[0100] The systemic bleeding risk distribution map provides risk information in the spatial domain. It requires spatial analysis techniques to extract key feature points of the risk field. In a three-dimensional anatomical coordinate system, the systemic bleeding risk distribution map is divided into regular three-dimensional grid cells, with the grid cell size set to [size missing]. This approach ensures computational accuracy while avoiding the computational burden caused by excessive refinement. A three-dimensional mesh index structure is established to record the center coordinates of each mesh cell. and the corresponding field strength value Iterate through all grid cells and calculate the spatial gradient tensor of the field strength for each cell. The spatial gradient tensor is defined as follows: ,in Indicates the field strength value along The partial derivatives along the axial direction are similarly defined. and The partial derivatives are calculated using the central difference method, for example... ,in The differential step size is half the size of the mesh cell.

[0101] For the calculated spatial gradient tensor Eigenvalue decomposition yields three eigenvalues. and the corresponding feature vectors Largest eigenvalue Corresponding feature vector The direction representing the most drastic change in electric field strength is the principal direction eigenvector. The directional derivative of the electric field strength value is calculated along this principal direction eigenvector. The directional derivative quantifies the rate of change of the electric field along the direction of maximum change. By traversing all grid cells and comparing the magnitudes of the directional derivatives, the location where the directional derivative reaches its maximum value is identified. These locations of maximum rate of change typically correspond to the edges or core regions of the hemorrhage risk field and are potential candidate locations for hemorrhage sources. Extract all locations that meet the criteria... The positions constitute the candidate position set .

[0102] Establishing the correlation between blood pressure oscillation feature vectors and candidate bleeding source locations requires a multiple regression model, assuming that the blood pressure oscillation feature vectors... It can be represented as a linear combination of candidate positions. ,in For regression coefficients, This represents the residual term. The least squares method is used to estimate the regression coefficients, minimizing the objective function. The regression coefficient vector is obtained through matrix operations. Calculate the explained variance of the blood pressure oscillation feature vector for each candidate location. The explained variance is defined as... This indicator quantifies the first The contribution of each candidate location to the blood pressure oscillation feature is analyzed. The explained variance of all candidate locations is compared, and the location with the largest explained variance is extracted. This coordinate serves as the location coordinate for the primary bleeding source. It precisely marks the most likely location of the bleeding source in the three-dimensional anatomical coordinate system, providing a spatial basis for subsequent treatment decisions.

[0103] In one optional implementation, coagulation index data from physiological monitoring is projected onto the state space of the coagulation cascade reaction, the trajectory of the coagulation state is tracked, and when the trajectory crosses the compensated coagulation region, the shortest distance between it and the boundary of the decompensated region is calculated and converted into the remaining time window for the occurrence of decompensation, including:

[0104] The reaction time, coagulation time, and maximum amplitude parameters of thromboelastography were extracted from physiological monitoring data as coagulation index data, and a multidimensional index vector was constructed after standardization transformation.

[0105] A state space for the coagulation cascade reaction is established. Principal component analysis is used to reduce the dimensionality of the multidimensional index vector to obtain the state coordinates. A classification hyperplane for the normal region, compensatory region and decompensated region is constructed by support vector machine. The normal vector and intercept parameters of each hyperplane are extracted to define the boundary.

[0106] The continuously collected multidimensional index vectors are projected onto the state space according to the time series to form a trajectory point sequence and smoothed by a sliding window. The position derivative is calculated to obtain the velocity field. The velocity field is then decomposed into vector components along the boundary normal and tangential components along the boundary tangential.

[0107] The system determines the current trajectory point's location in real time. When entering a compensatory region, it calculates the vertical distance from the current trajectory point to the hyperplane of the uncompensated region as the shortest distance. It extracts the magnitude of the normal component as the approach rate, calculates the base time based on the shortest distance, calculates the stability coefficient based on the magnitude of the tangential component, and uses the stability coefficient to perform weighted correction on the base time. The output is the remaining time window for the occurrence of uncompensation.

[0108] In this specific embodiment, several key parameters of thromboelastography are extracted from continuously collected physiological monitoring data, including reaction time, coagulation time, and maximum amplitude. The reaction time parameter characterizes the time from blood placement to initial fibrin formation, typically recorded in milliseconds, reflecting the speed of the coagulation initiation phase. The coagulation time parameter measures the time span from fibrin formation to reaching a fixed amplitude threshold, characterizing the clot formation rate. The maximum amplitude parameter represents the maximum strength of the clot, numerically corresponding to the peak width of the thromboelastography curve, reflecting the combined effect of platelet function and fibrin aggregation. These parameters are derived from routine clinical thromboelastography equipment, with data typically collected every fifteen to thirty minutes during dynamic monitoring of patients with multiple bone trauma.

[0109] After acquiring the raw coagulation index data, a standardization transformation was performed to address the dimensional differences of different parameters. The Z-score standardization method was used to calculate the mean and standard deviation of each parameter in the historical dataset. The mean was subtracted from the real-time collected parameter values, and then the result was divided by the standard deviation to eliminate the influence of orders of magnitude between different indicators. The standardized values ​​typically fall within the range of -3 to 3, exhibiting statistical characteristics of zero mean and unit variance. The standardized reaction time parameter, coagulation time parameter, and maximum amplitude parameter were arranged sequentially to form a three-dimensional multidimensional index vector. This vector serves as the basic mathematical expression describing the state of the coagulation system in subsequent processing, with each dimension corresponding to a key physiological step in the coagulation process.

[0110] A state space for the coagulation cascade reaction is established, and principal component analysis (PCA) is used to reduce the dimensionality of the multidimensional index vectors. PCA finds the projection direction with the largest data variance by calculating the eigenvalues ​​and eigenvectors of the covariance matrix. The covariance matrix of historical sample data is calculated, and eigenvalue decomposition is performed. The eigenvectors corresponding to the two largest eigenvalues ​​are selected as the projection basis. The inner product operation is performed on the real-time acquired three-dimensional multidimensional index vectors and these two eigenvectors to obtain the two-dimensional state coordinates. The dimensionality-reduced state space retains approximately 85% to 95% of the information in the original data, while facilitating visualization and geometric calculations on a plane. The horizontal and vertical axes in the state space correspond to the first and second principal components, respectively, and each state coordinate point represents the comprehensive state of the coagulation system at a specific moment.

[0111] In the state space, classification boundaries for three functional regions are constructed, corresponding to normal coagulation, compensated coagulation, and decompensated coagulation states, respectively. Support Vector Machine (SVM) techniques are used to train these multi-class boundaries, employing a large amount of historically accumulated patient coagulation data as training samples. Each sample contains state coordinates and a clinical diagnostic label. The SVM searches for the optimal classification hyperplane that maximizes the margin between samples of different classes. In the two-dimensional state space, the hyperplane degenerates into a straight line, and each line can be represented as a combination of a normal vector and an intercept parameter. The boundary hyperplane between the normal and compensated regions is obtained by solving a constrained optimization problem to obtain its normal vector and intercept parameter. The boundary hyperplane between the compensated and decompensated regions is also obtained through SVM training to obtain its corresponding geometric parameters. The normal vector defines the vertical direction of the boundary, and the intercept parameter determines the specific location of the boundary in the state space.

[0112] Multidimensional index vectors, continuously acquired over time, are projected one by one onto a two-dimensional state space to form a series of discrete trajectory points. These trajectory points are arranged in chronological order of acquisition, constituting a trajectory point sequence describing the evolution of coagulation state. Due to measurement noise and physiological fluctuations during physiological signal acquisition, the original trajectory point sequence exhibits some random jitter. A sliding window smoothing technique is applied to denoise the trajectory point sequence. A time window of five consecutive sampling points is selected, and the weighted average of the trajectory point coordinates within the window is calculated. The weight is highest at the center of the window and decreases towards both sides, using a triangular weight distribution. The sliding window moves along the time axis, smoothing the trajectory points at each time step to obtain a smoothed trajectory point sequence. This sequence significantly reduces the impact of high-frequency noise while preserving the main evolutionary trends.

[0113] The velocity field of the coagulation state is calculated based on the smoothed trajectory point sequence. Two trajectory points at adjacent time points are selected, and their coordinate difference in the state space is calculated. Dividing the difference by the corresponding time interval yields the position derivative at that time. Geometrically, the position derivative is represented as a vector pointing from the trajectory point at the previous time point to the trajectory point at the current time point. Its direction represents the evolution direction of the coagulation state, and its magnitude represents the evolution rate. The entire trajectory point sequence is calculated point by point to construct a complete velocity field distribution. The velocity field describes the dynamic behavior characteristics of the coagulation system in the state space and provides a quantitative basis for subsequent determination of the evolution trend.

[0114] For each velocity vector in the velocity field, a geometric decomposition is performed, decomposing it into a component normal to the boundary of the decompensated region and a component tangential to the boundary. The normal vector of the hyperplane at the boundary of the decompensated region is extracted and normalized to a unit vector. The dot product of the velocity vector and the normal unit vector is calculated to obtain the projection length of the velocity vector onto the normal direction; this value is the magnitude of the normal component. The sign of the normal component reflects the angle between the velocity direction and the boundary normal; a positive value indicates movement towards the decompensated region, and a negative value indicates movement away from the decompensated region. The tangential component is obtained by subtracting the normal component from the velocity vector, characterizing the motion characteristics of the coagulation state along the direction parallel to the boundary. The magnitude of the tangential component reflects the degree to which the state lingers or wanders near the boundary; a larger value indicates higher instability of the evolutionary trajectory.

[0115] The system monitors the current trajectory point's region affiliation in the state space in real time. This is determined by calculating the relative positional relationship between the trajectory point's coordinates and each classification hyperplane. The current state coordinates are substituted into the expressions of each hyperplane, and the region affiliation is determined based on the sign and numerical range of the calculation results. When a trajectory point is determined to have entered the compensated coagulation region, the decompensation risk assessment process is initiated. The vertical distance from the current trajectory point to the boundary hyperplane of the decompensation region is calculated using the point-to-line distance formula. The trajectory point coordinates are substituted into the normal vector expression of the boundary hyperplane, and the absolute value is divided by the magnitude of the normal vector. Geometrically, the vertical distance represents the shortest path length between the current coagulation state and the critical decompensation state; a smaller distance indicates a closer risk of decompensation.

[0116] The magnitude of the normal component in the velocity field at the current moment is extracted as the approach rate. This rate quantifies how quickly the coagulation state moves towards the decompensated region. Dividing the vertical distance by the approach rate yields the base time value, which physically represents the time required for the trajectory point to reach the boundary of the decompensated region under the assumption that the current velocity remains constant. The base time provides a preliminary time estimate of when decompensation occurs, but does not consider the stability factors of the evolving trajectory.

[0117] A stability coefficient is introduced to correct the baseline time. This coefficient is calculated based on the tangential component magnitude of the velocity field. A larger tangential component magnitude indicates strong lateral drift in the coagulation state near the boundary, suggesting an unstable trajectory. The stability coefficient is defined as the ratio of the tangential component magnitude to the normal component magnitude after a nonlinear transformation; a larger ratio corresponds to a larger coefficient. A stability coefficient greater than one indicates that the lateral drift of the trajectory exceeds the normal approach velocity, requiring an upward correction of the baseline time to reflect the increased uncertainty and its prolonging effect. A stability coefficient less than one indicates that the trajectory primarily moves along the normal direction, allowing the baseline time to be maintained or slightly shortened. Multiplying the baseline time by a weighting factor of the stability coefficient yields the corrected time value, which serves as the remaining time window for decompensation. This remaining time window provides clinicians with an early warning time for coagulation function deterioration, supporting early intervention with coagulation factor supplementation or other corrective measures to prevent patients from entering an irreversible state of coagulation failure.

[0118] In one optional implementation, the bleeding risk distribution map, bleeding source location, and remaining time window are used as features input to a decision tree model. The model simulates the physiological parameter change trajectory for different treatment plans, generates an evolutionary path set through Monte Carlo sampling, and calculates the adverse probability of each evolutionary path, including:

[0119] Spatial features are extracted from the bleeding risk distribution map, the spatial distribution density and risk gradient value of each high-risk area are calculated, the spatial coordinates and risk intensity of the bleeding source are obtained based on the bleeding source location, the maximum intervention delay time is calculated based on the remaining time window, and the features are combined to form a feature vector.

[0120] The feature vector is input into the decision tree model, which generates multiple candidate treatment plans based on a preset treatment plan library. Each candidate treatment plan includes intervention start time and intervention intensity parameters. A physiological response function describing the evolution law of hemodynamic parameters is constructed for each candidate treatment plan, and the physiological parameter change trajectory corresponding to each candidate treatment plan is calculated.

[0121] Monte Carlo sampling was performed on the trajectory of changes in various physiological parameters, and random perturbation sampling was performed in the parameter space to generate evolutionary paths with random perturbations. Repeated sampling was used to obtain a set of evolutionary paths.

[0122] Determine whether the physiological parameters of each evolutionary path in the set of evolutionary paths exceed the safety threshold range, mark the time when the safety threshold range is exceeded as the occurrence of adverse events, count the number or duration of adverse events in each evolutionary path, and calculate the adverse probability corresponding to each evolutionary path.

[0123] After obtaining the bleeding risk distribution map, bleeding source location, and remaining time window, these key data need to be transformed into structured features that can be processed by the decision-making model. For the bleeding risk distribution map, the three-dimensional space is divided into several voxel units, each voxel unit corresponding to a risk value. By statistically analyzing the number of voxels with risk values ​​exceeding a preset threshold and their spatial distribution locations, the spatial distribution density of high-risk areas is calculated. Specifically, if in a radius of... Within the spherical region If there are 1 high-risk voxels, then the spatial distribution density of the region can be expressed as: This density value reflects the degree of clustering of high-risk areas. Simultaneously, it is necessary to calculate the risk gradient value for each high-risk area. This gradient value describes the rate of change of risk intensity in space. The risk value difference between adjacent voxels is calculated along each of the three coordinate axes, and the vector magnitude is taken as the risk gradient value at that location. ,in Indicates the risk value. , , These represent the directions of the three-dimensional coordinate axes.

[0124] Extract the spatial coordinates of the bleeding source from the bleeding source localization results. and the corresponding risk intensity This risk level is determined by comprehensively considering vascular density, blood flow velocity, and the degree of tissue damage at the location of the bleeding source. Spatial coordinates are normalized to a standard coordinate range to ensure that data from different patients are compared on the same scale. For the remaining time window... Calculate the maximum intervention delay duration This duration represents the maximum time that treatment intervention can be delayed without decompensation. By analyzing the rate of change of the distance between the coagulation state trajectory and the decompensation boundary, combined with the patient's physiological reserve parameters, the duration is estimated. ,in The physiological reserve coefficient typically ranges from 0.6 to 0.8. This coefficient is determined based on the patient's age, underlying disease status, and current vital sign stability.

[0125] The spatial distribution density extracted above Risk gradient value Hemorrhage source spatial coordinates Risk intensity and maximum intervention delay duration The feature vectors are formed by combining them in a fixed order. This feature vector contains key information about the patient's current state, reflecting both the spatial distribution of bleeding risk and temporal constraints.

[0126] The constructed feature vector is input into a decision tree model, which has learned the mapping relationship between treatment plans and prognoses from a large number of historical cases during the training phase. The model contains multiple decision nodes, each branching based on a specific dimension of the feature vector. The decision tree model is connected to a pre-defined treatment plan library, which contains common treatment strategies for critical cases of multiple bone traumas, such as emergency hemostasis surgery, vascular interventional embolization, blood transfusion and fluid resuscitation, and coagulation factor supplementation. Based on the specific values ​​of the feature vector, the decision tree model selects multiple candidate treatment plans from the library. Each candidate treatment plan includes a clearly defined intervention initiation time. With intervention intensity parameters superscript Indicates the first There are several candidate intervention options. The intervention start time indicates the time delay from the current moment before initiating the treatment plan; its value must be less than or equal to the maximum intervention delay time. The intervention intensity parameter quantifies the aggressiveness of the treatment measures, such as transfusion rate, surgical scope, and drug dosage, and is usually represented by a standardized value between 0 and 1.

[0127] For each candidate treatment option, a physiological response function describing the evolution of hemodynamic parameters needs to be constructed. This function, based on physiological principles, describes the changes over time in key physiological parameters such as blood pressure, heart rate, hemoglobin concentration, and coagulation function indicators under a specific treatment intervention. The general form of the physiological response function can be expressed as follows: ,in Indicates time A certain physiological parameter value, As the initial value, This indicates the changes in parameters caused by the treatment intervention. This represents the parameter changes resulting from the natural progression of the disease. Parameter changes induced by therapeutic intervention typically exhibit exponential response characteristics and can be modeled as... ,in This indicates the maximum effect range of the treatment regimen. The response time constant describes the speed at which the treatment effect reaches a steady state. Parameter changes caused by natural disease progression are modeled based on factors such as blood loss rate and metabolic consumption, typically exhibiting a linear decrease or exponential decay trend.

[0128] The physiological parameter change trajectory corresponding to each candidate treatment plan is calculated using the physiological response function. This trajectory is a time series, recording the changes from the current moment to the preset termination moment. During this period, the expected changes in various physiological parameters are recorded. The preset termination time is usually set at 24 or 48 hours, covering the acute treatment period for critically ill patients. For each candidate treatment plan, a trajectory of parameter changes in multiple dimensions, including systolic blood pressure, diastolic blood pressure, heart rate, hemoglobin concentration, prothrombin time, and activated partial thromboplastin time, is generated.

[0129] Due to individual patient differences and various uncertainties during treatment, a single deterministic trajectory cannot fully reflect the actual clinical situation. Therefore, Monte Carlo sampling is performed on the trajectory of changes in various physiological parameters to introduce randomness to simulate real-world scenarios. Specifically, key parameters of the physiological response function are randomly perturbed within the parameter space, including the maximum effect amplitude. Response time constant Parameters such as blood loss rate are assumed to follow a specific probability distribution, such as a normal or log-normal distribution. Random sampling is performed based on their mean and standard deviation. Each sampling yields a set of perturbed parameter values, which are then substituted into the physiological response function for recalculation, resulting in an evolutionary path with random perturbation. This sampling process is repeated, typically thousands to tens of thousands of times, generating a set of evolutionary paths containing a large number of paths. Each path in this set represents the evolution of the patient's physiological state under specific parameter perturbations.

[0130] After obtaining the set of evolutionary pathways, a safety assessment needs to be performed on each pathway. Safe threshold ranges for each physiological parameter are defined, based on clinical guidelines and statistical analysis of extensive case data. For example, the safe range for systolic blood pressure is typically 90 to 180 mmHg, for heart rate it's 60 to 120 beats per minute, and for hemoglobin concentration it's 70 g / L. All moments within each evolutionary pathway are traversed to determine if the physiological parameter at that moment exceeds the corresponding safe threshold range. If a physiological parameter exceeds the safe range at a given moment, that moment is marked as an adverse event.

[0131] In one optional implementation, Monte Carlo sampling is performed on the trajectory of each physiological parameter change, and random perturbation sampling is performed within the parameter space to generate evolutionary paths with random perturbations. Repeated sampling yields a set of evolutionary paths, including:

[0132] Parameters describing the treatment response characteristics are extracted from the trajectory of each physiological parameter change, and the influence of the unit change of each parameter on the final state of the physiological parameter change trajectory is calculated. The parameter set is sorted by sensitivity, and parameters with sensitivity higher than the sensitivity threshold are selected to form a subset of perturbation target parameters.

[0133] Based on the statistical distribution characteristics of each parameter in the subset of disturbance target parameters, the disturbance center value and the disturbance amplitude range are determined. Random disturbance quantities are generated within the disturbance amplitude range, and the disturbance parameter values ​​are calculated. All disturbance parameter values ​​are combined to form a disturbance parameter vector.

[0134] The perturbation parameter vector is substituted into the physiological response function to recalculate the time evolution process of hemodynamic parameters, resulting in an evolution path with random perturbation. The physiological parameter values ​​at each moment in the evolution path with random perturbation are recorded.

[0135] The process of generating random perturbations, superimposing, and recalculating is repeated. Each time the process is repeated, a new random perturbation is generated independently to produce a different perturbation parameter vector. All evolution paths with random perturbations obtained from multiple repetitions are summarized to form an evolution path set.

[0136] After obtaining the trajectory of changes in various physiological parameters, the uncertainty of the evolutionary path of the treatment plan needs to be quantitatively assessed using Monte Carlo sampling techniques. The core of this process lies in identifying key parameters that significantly affect the treatment outcome and performing systematic perturbation sampling within their uncertainty range, thereby constructing a set of evolutionary paths that can reflect the true clinical variability.

[0137] Parameters describing the treatment response characteristics were extracted from the trajectory of various physiological parameters. These parameters cover multiple dimensions, including hemodynamic response rate, drug clearance rate, tissue perfusion recovery constant, vascular resistance regulation coefficient, cardiac output change rate, and fluid redistribution time constant. For each parameter, local sensitivity analysis was used to calculate the degree of influence of its unit change on the endpoint state of the physiological parameter trajectory. In practice, a small perturbation was applied near the parameter's baseline value, such as increasing or decreasing the parameter value by 1%, and the physiological parameter trajectory was recalculated to observe the magnitude of change in key endpoint indicators. Key indicators include the stable value of mean arterial pressure, the final level of hemoglobin concentration, lactate clearance rate, and cumulative urine output, among other clinically relevant physiological endpoints. The magnitude of the endpoint state change caused by each parameter was normalized to obtain a dimensionless sensitivity index. All parameters were sorted from high to low according to their sensitivity index, and a sensitivity threshold of 0.15 was set, meaning that when the sensitivity index of a parameter is greater than 0.15, the parameter is considered to have a significant impact on the treatment outcome. A subset of perturbation target parameters is selected based on the criteria; this subset typically contains 5 to 12 parameters. This sensitivity-based selection strategy significantly reduces the computational burden of sampling while retaining the most critical sources of uncertainty affecting the results.

[0138] For each parameter in the subset of target parameters to be perturbed, it is necessary to determine its perturbation center value and perturbation range. The perturbation center value is usually set as the expected value of the parameter under the current treatment regimen. This expected value can be obtained from statistical analysis of historical clinical data or calculated theoretically based on pharmacological models. Determining the perturbation range requires comprehensive consideration of the parameter's physiological variability and measurement uncertainty. For parameters with clear clinical data support, the perturbation range can be determined based on the parameter's distribution characteristics in similar patient populations, for example, using the mean plus or minus 1.5 times the standard deviation as the upper and lower bounds of the perturbation range. For parameters lacking sufficient statistical data, a reasonable range of variation can be set based on expert experience, typically 15% to 30% of the parameter's center value. After determining the perturbation range, an appropriate random sampling strategy is selected based on the parameter's distribution characteristics. For normally distributed parameters, a Gaussian random number generator is used to generate random perturbations within the perturbation range. For parameters with skewed distribution characteristics, a log-normal or gamma distribution is used for sampling. For parameters with physical constraints, such as physiological constants that cannot be negative, truncation is required after sampling to ensure that the parameter value remains within the physiologically reasonable range after perturbation.

[0139] After generating random perturbations, they are superimposed onto the center values ​​of the corresponding parameters in the perturbed target parameter subset to calculate the perturbed parameter values. All perturbed parameter values ​​are then combined according to parameter type to form a perturbed parameter vector. The dimension of this vector is equal to the number of parameters in the perturbed target parameter subset, and each element in the vector corresponds to a perturbed parameter value. It is important to note that when generating the perturbed parameter vector, it is necessary to check whether there are physiological constraints between the parameters. For example, there is an inverse regulatory relationship between cardiac output and peripheral vascular resistance, and a positive correlation between blood volume and intravascular pressure. When related parameters are detected to be perturbed simultaneously, the perturbations need to be coordinated and corrected according to their physiological correlations to avoid generating parameter combinations that violate physiological laws. This coordination and correction can be achieved by constructing a correlation matrix between parameters, introducing a multivariate normal distribution or a Copula function during random sampling to maintain the correlation structure between parameters.

[0140] The perturbation parameter vector is substituted into the physiological response function to recalculate the temporal evolution of hemodynamic parameters. The physiological response function describes the dynamic changes in a patient's physiological state over time under a given treatment intervention, typically expressed as a system of differential equations. The perturbation parameter values ​​are replaced with the corresponding parameters in the physiological response function, and the system of differential equations is solved using numerical integration to obtain the evolution trajectory of hemodynamic parameters over time under that parameter combination. Numerical integration can employ the fourth-order Runge-Kutta method, with a time step set from 0.5 to 2 minutes, and the integration time span covering the entire expected time window of the treatment plan, typically 6 to 48 hours. During integration, physiological parameter values ​​at each moment are recorded, including key indicators such as heart rate, blood pressure, central venous pressure, cardiac output, tissue oxygenation index, and lactate concentration, forming a complete evolutionary path with random perturbations. This evolutionary path reflects the trajectory of the patient's physiological state under specific parameter perturbation conditions.

[0141] The process of generating, superimposing, and recalculating random perturbations is repeated, with each repetition generating a new random perturbation to produce a different perturbation parameter vector. In each iteration, all parameters in the subset of perturbation target parameters are resampled randomly to ensure that each generated perturbation parameter vector is statistically independent and conforms to a preset distribution characteristic of each parameter. The evolution path with random perturbations obtained in each iteration is stored in the evolution path set. The number of repetitions needs to be determined by balancing computational efficiency and statistical reliability, typically ranging from 500 to 2000 sampling iterations. Convergence diagnostic methods can be used to dynamically determine the number of sampling iterations. For example, monitoring whether the statistics of the endpoint state indicators in the evolution path set tend to stabilize, and when the change in each statistic is less than 2% after 100 consecutive sampling iterations, the sampling is considered to have converged, and the iteration process can be terminated.

[0142] All evolutionary paths with random perturbations obtained through repeated iterations are aggregated to form an evolutionary path set. This set contains hundreds to thousands of independent evolutionary paths, each corresponding to a physiological response pattern. The evolutionary path set can quantify the range of response variation of treatment plans under parameter uncertainty, providing a basis for subsequent adverse probability assessment. During the construction of the evolutionary path set, the perturbation parameter vector corresponding to each path also needs to be recorded to facilitate subsequent analysis of the correlation between key parameter perturbations and adverse outcomes. Statistical analysis of the evolutionary path set allows for the extraction of the distribution characteristics of physiological parameters at various time points, such as calculating the median, quartile range, and 95% confidence interval of blood pressure at each time point. This statistical information can intuitively demonstrate the expected effect of the treatment plan and its range of uncertainty, providing a quantitative risk assessment basis for clinical decision-making.

[0143] A second aspect of the present invention provides an AI-based critical care system for multiple bone traumas, comprising:

[0144] The data acquisition unit is used to collect physiological monitoring data and imaging data of the target patient;

[0145] The risk modeling unit is used to extract the fracture site from the image data and map it to a three-dimensional anatomical coordinate system, calculate the densely vascularized area involved by the fracture, construct a local bleeding risk field for each area, generate a systemic bleeding risk distribution map through a field strength superposition algorithm, and mark the centroid coordinates and diffusion radius of the risk area.

[0146] The bleeding localization unit is used to perform wavelet decomposition on the hemoglobin concentration sequence of physiological monitoring data, obtain the trend component and calculate the rate of decrease, and perform correlation calculation with the bleeding risk distribution map to locate the bleeding source.

[0147] The coagulation analysis unit is used to project the coagulation index data of physiological monitoring data into the state space of the coagulation cascade reaction, track the movement trajectory of the coagulation state, and calculate the shortest distance between the trajectory and the boundary of the decompensated region when the trajectory crosses the compensated coagulation region and convert it into the remaining time window for the occurrence of decompensation.

[0148] The decision simulation unit is used to input the bleeding risk distribution map, bleeding source location and remaining time window as features into the decision tree model, simulate the change trajectory of physiological parameters for different treatment plans, generate a set of evolution paths through Monte Carlo sampling and calculate the adverse probability of each evolution path;

[0149] The scheme output unit is used to extract the treatment scheme and its target decision time node corresponding to the evolution path with the lowest adverse probability, and output it as the detection result.

[0150] A third aspect of the present invention provides an electronic device, comprising:

[0151] processor;

[0152] Memory used to store processor-executable instructions;

[0153] The processor is configured to invoke instructions stored in the memory to execute the aforementioned method.

[0154] 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.

[0155] 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.

[0156] 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 AI-based method for detecting critical cases of multiple bone traumas, characterized in that, include: Collect physiological monitoring data and imaging data from the target patients; The fracture site is extracted from the image data and mapped to a three-dimensional anatomical coordinate system. The densely vascularized areas involved by the fracture are calculated. A local bleeding risk field is constructed for each area. A systemic bleeding risk distribution map is generated by the field strength superposition algorithm, and the centroid coordinates and diffusion radius of the risk area are marked. Wavelet decomposition was performed on the hemoglobin concentration sequence of physiological monitoring data to obtain the trend component and calculate the rate of decrease. This component was then correlated with the bleeding risk distribution map to locate the bleeding source. The coagulation index data of physiological monitoring data are projected onto the state space of coagulation cascade reaction, and the movement trajectory of coagulation state is tracked. When the trajectory crosses the compensated coagulation area, the shortest distance between it and the boundary of the decompensated area is calculated and converted into the remaining time window for the occurrence of decompensation. The bleeding risk distribution map, bleeding source location, and remaining time window are used as features to input the decision tree model. The physiological parameter change trajectory is simulated for different treatment plans. The set of evolutionary paths is generated by Monte Carlo sampling and the adverse probability of each evolutionary path is calculated. Extract the treatment plan and its target decision time point corresponding to the evolutionary path with the lowest adverse probability, and output it as the detection result.

2. The method according to claim 1, characterized in that, The fracture site is extracted from the image data and mapped to a three-dimensional anatomical coordinate system. The densely vascularized areas affected by the fracture are calculated, and a local hemorrhage risk field is constructed. A systemic hemorrhage risk distribution map is generated using a field strength superposition algorithm, and the centroid coordinates and diffusion radius of high-risk areas are marked, including: The image data is segmented to extract the fracture contour and mapped to a three-dimensional anatomical coordinate system. The corresponding vascular atlas is retrieved and a search radius is set with the fracture contour as the center. The vascular line density at each spatial location within the search radius is counted, and connected regions with vascular line density exceeding the density threshold are extracted as vascular dense regions. Multiple sampling points are uniformly sampled on the boundary surface of each densely vascular region to generate sub-field sources, and the initial intensity of the sub-field sources is determined. The intensity of the sub-field sources is then corrected based on the degree of fracture displacement. The field intensity distribution of each sub-field source is described by a Gaussian decay function, and scalar superposition is performed to construct a local hemorrhage risk field. A spatial grid is established in the three-dimensional anatomical coordinate system. The field strength value of all local bleeding risk fields is calculated for each grid node. The field strength gradient direction is calculated based on the field strength value of each local bleeding risk field. When the angle between the field strength gradient directions is less than the cooperative angle, the corresponding field strength value is enhanced and summed to obtain the total field strength of each grid node, which constitutes the whole-body bleeding risk distribution map. Clustering is performed on the grid nodes in the systemic bleeding risk distribution map where the field strength value exceeds the risk threshold to form high-risk areas. The weighted average of the grid node coordinates in each high-risk area is calculated as the centroid coordinates, and the standard deviation of the distance between the grid node and the centroid coordinates in each high-risk area is calculated as the diffusion radius.

3. The method according to claim 2, characterized in that, Multiple sampling points are uniformly sampled on the boundary surface of each densely vascularized region to generate sub-field sources. The initial intensity of the sub-field sources is determined and corrected based on the degree of fracture displacement to obtain the sub-field source intensity. A Gaussian decay function is used to describe the field intensity distribution of each sub-field source, and scalar superposition is performed to construct a local hemorrhage risk field, including: Boundary extraction is performed on each densely vascular region, the normal vector distribution of the boundary is calculated, the concave and convex features are identified based on the normal vector distribution and the sampling priority is set, sampling points are generated on the boundary as sub-field sources according to the sampling priority, and the normal vector direction of each sub-field source is recorded as the field source direction. For each sub-field source, the number and type distribution of blood vessels within the neighborhood are statistically analyzed, the density of blood vessels and the tissue fragility coefficient are calculated, and weighted fusion is performed to obtain the initial intensity of the sub-field source. Fracture profiles are classified and identified. Based on the fracture classification, the displacement risk level is determined. The centroid coordinates of the fracture fragments are extracted. The displacement amplitude is determined based on the maximum distance between the centroid coordinates. The displacement risk level and displacement amplitude are mapped to a correction coefficient. The initial intensity of each subfield source is multiplied by the correction coefficient to obtain the subfield source intensity. Anisotropic Gaussian decay functions are constructed with each sub-source as the center. The decay rate of the Gaussian decay function along the source direction is lower than the decay rate perpendicular to the source direction. The decay rate is determined by the sub-source intensity and the spatial scale of the densely vascularized region. The Gaussian decay functions are superimposed in a three-dimensional anatomical coordinate system to construct a local hemorrhage risk field.

4. The method according to claim 1, characterized in that, Wavelet decomposition was performed on the hemoglobin concentration sequence of physiological monitoring data to obtain the trend component and calculate the rate of decrease. This component was then correlated with the bleeding risk distribution map to locate the bleeding source, including: Wavelet basis functions are selected to perform multi-level decomposition of hemoglobin concentration sequence, and approximate coefficients are extracted as trend components and polynomial fitting is performed. The first derivative is calculated to obtain the rate of decline sequence. Local extrema are detected and the time interval and amplitude difference between adjacent extrema points are calculated. Multiplication is performed to obtain the fluctuation energy value. The time period when the fluctuation energy value exceeds the energy threshold is extracted as the active bleeding period. Arterial blood pressure waveform data during the active bleeding period were extracted from physiological monitoring data and subjected to Fourier transform. The amplitude and phase of the dominant frequency component in the frequency domain feature were calculated and a blood pressure oscillation feature vector was constructed. A three-dimensional grid index is established in the risk distribution map of systemic bleeding. The spatial gradient tensor of the field strength value is calculated by traversing each grid cell. The spatial gradient tensor is decomposed into eigenvalues ​​to extract the principal direction feature vector. The directional derivative of the field strength value is calculated along the principal direction feature vector to identify the location of the maximum rate of change. A multiple regression model is established between the blood pressure oscillation feature vector and the location of the maximum rate of change. The explained variance of each location of the maximum rate of change on the blood pressure oscillation feature vector is calculated, and the location with the largest explained variance is extracted as the coordinates for locating the main blood source.

5. The method according to claim 1, characterized in that, Coagulation index data from physiological monitoring are projected onto the state space of the coagulation cascade reaction. The trajectory of the coagulation state is tracked. When the trajectory crosses the compensated coagulation region, the shortest distance between it and the boundary of the decompensated region is calculated and converted into the remaining time window for the occurrence of decompensation, including: The reaction time, coagulation time, and maximum amplitude parameters of thromboelastography were extracted from physiological monitoring data as coagulation index data, and a multidimensional index vector was constructed after standardization transformation. A state space for the coagulation cascade reaction is established. Principal component analysis is used to reduce the dimensionality of the multidimensional index vector to obtain the state coordinates. A classification hyperplane for the normal region, compensatory region and decompensated region is constructed by support vector machine. The normal vector and intercept parameters of each hyperplane are extracted to define the boundary. The continuously collected multidimensional index vectors are projected onto the state space according to the time series to form a trajectory point sequence and smoothed by a sliding window. The position derivative is calculated to obtain the velocity field. The velocity field is then decomposed into vector components along the boundary normal and tangential components along the boundary tangential. The system determines the current trajectory point's location in real time. When entering a compensatory region, it calculates the vertical distance from the current trajectory point to the hyperplane of the uncompensated region as the shortest distance. It extracts the magnitude of the normal component as the approach rate, calculates the base time based on the shortest distance, calculates the stability coefficient based on the magnitude of the tangential component, and uses the stability coefficient to perform weighted correction on the base time. The output is the remaining time window for the occurrence of uncompensation.

6. The method according to claim 1, characterized in that, The bleeding risk distribution map, bleeding source location, and remaining time window are used as features input to a decision tree model. The model simulates the physiological parameter changes for different treatment plans, generates an evolutionary path set through Monte Carlo sampling, and calculates the adverse probability of each path, including: Spatial features are extracted from the bleeding risk distribution map, the spatial distribution density and risk gradient value of each high-risk area are calculated, the spatial coordinates and risk intensity of the bleeding source are obtained based on the bleeding source location, the maximum intervention delay time is calculated based on the remaining time window, and the features are combined to form a feature vector. The feature vector is input into the decision tree model, which generates multiple candidate treatment plans based on a preset treatment plan library. Each candidate treatment plan includes intervention start time and intervention intensity parameters. A physiological response function describing the evolution law of hemodynamic parameters is constructed for each candidate treatment plan, and the physiological parameter change trajectory corresponding to each candidate treatment plan is calculated. Monte Carlo sampling was performed on the trajectory of changes in various physiological parameters, and random perturbation sampling was performed in the parameter space to generate evolutionary paths with random perturbations. Repeated sampling was used to obtain a set of evolutionary paths. Determine whether the physiological parameters of each evolutionary path in the set of evolutionary paths exceed the safety threshold range, mark the time when the safety threshold range is exceeded as the occurrence of adverse events, count the number or duration of adverse events in each evolutionary path, and calculate the adverse probability corresponding to each evolutionary path.

7. The method according to claim 6, characterized in that, Monte Carlo sampling was performed on the trajectories of changes in various physiological parameters, and random perturbation sampling was performed within the parameter space to generate evolutionary paths with random perturbations. Repeated sampling yielded a set of evolutionary paths, including: Parameters describing the treatment response characteristics are extracted from the trajectory of each physiological parameter change, and the influence of the unit change of each parameter on the final state of the physiological parameter change trajectory is calculated. The parameter set is sorted by sensitivity, and parameters with sensitivity higher than the sensitivity threshold are selected to form a subset of perturbation target parameters. Based on the statistical distribution characteristics of each parameter in the subset of disturbance target parameters, the disturbance center value and the disturbance amplitude range are determined. Random disturbance quantities are generated within the disturbance amplitude range, and the disturbance parameter values ​​are calculated. All disturbance parameter values ​​are combined to form a disturbance parameter vector. The perturbation parameter vector is substituted into the physiological response function to recalculate the time evolution process of hemodynamic parameters, resulting in an evolution path with random perturbation. The physiological parameter values ​​at each moment in the evolution path with random perturbation are recorded. The process of generating random perturbations, superimposing, and recalculating is repeated. Each time the process is repeated, a new random perturbation is generated independently to produce a different perturbation parameter vector. All evolution paths with random perturbations obtained from multiple repetitions are summarized to form an evolution path set.

8. An AI-based critical care system for multiple bone traumas, used to implement the method as described in any one of claims 1-7, characterized in that, include: The data acquisition unit is used to collect physiological monitoring data and imaging data of the target patient; The risk modeling unit is used to extract the fracture site from the image data and map it to a three-dimensional anatomical coordinate system, calculate the densely vascularized area involved by the fracture, construct a local bleeding risk field for each area, generate a systemic bleeding risk distribution map through a field strength superposition algorithm, and mark the centroid coordinates and diffusion radius of the risk area. The bleeding localization unit is used to perform wavelet decomposition on the hemoglobin concentration sequence of physiological monitoring data, obtain the trend component and calculate the rate of decrease, and perform correlation calculation with the bleeding risk distribution map to locate the bleeding source. The coagulation analysis unit is used to project the coagulation index data of physiological monitoring data into the state space of the coagulation cascade reaction, track the movement trajectory of the coagulation state, and calculate the shortest distance between the trajectory and the boundary of the decompensated region when the trajectory crosses the compensated coagulation region and convert it into the remaining time window for the occurrence of decompensation. The decision simulation unit is used to input the bleeding risk distribution map, bleeding source location and remaining time window as features into the decision tree model, simulate the change trajectory of physiological parameters for different treatment plans, generate a set of evolution paths through Monte Carlo sampling and calculate the adverse probability of each evolution path; The scheme output unit is used to extract the treatment scheme and its target decision time node corresponding to the evolution path with the lowest adverse probability, and output it as the detection result.

9. 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 to execute the method according to any one of claims 1 to 7.

10. 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 method described in any one of claims 1 to 7.

Images