A system for predicting and intervening the risk of post-oral bleeding based on multimodal data fusion.

Abstract

This invention relates to the field of oral medical monitoring technology, specifically a postoperative oral bleeding risk prediction and intervention system based on multimodal data fusion. The system includes: collecting and standardizing preoperative coagulation function test data, intraoperative wound image data, and postoperative continuous physiological sign monitoring data; extracting three parameters from the intraoperative wound image: the proportion of exposed blood vessels, the density of tissue bleeding points, and the state of wound edge contraction; mapping these three parameters to the preoperative coagulation function data to construct a bleeding risk feature vector; training a bleeding risk classifier using this vector to obtain a risk prediction model; and inputting real-time postoperative physiological sign data into the model to obtain a bleeding risk probability value and trigger corresponding intervention measures. This system can accurately capture the real-time state of the wound, integrate multi-dimensional data, improve the accuracy of bleeding risk prediction, trigger timely intervention, and reduce postoperative bleeding complications.

Description

Technical Field

[0001] This invention relates to the field of oral medical monitoring technology, and in particular to a system for predicting and intervening in the risk of postoperative bleeding in the oral cavity based on multimodal data fusion. Background Technology

[0002] Postoperative bleeding after oral surgery is a common complication of oral surgery. If it is not predicted and intervened in a timely manner, it may lead to hematoma, infection, or even more serious systemic complications. Therefore, the prediction and intervention of postoperative bleeding risk is an important research direction in the field of oral medicine. Currently, the prediction of postoperative bleeding risk mainly relies on preoperative coagulation function test data. By analyzing indicators such as prothrombin time and activated partial thromboplastin time, bleeding tendency is judged. During the operation, medical staff mainly rely on visual observation of bleeding at the surgical wound. After the operation, the patient's blood pressure, heart rate, and other physiological signs are intermittently monitored to make a preliminary judgment on whether there is a bleeding risk.

[0003] In existing technologies, preoperative coagulation function testing can only reflect the patient's basic hemostasis status before surgery and cannot reflect the real-time damage of the surgical wound during surgery. Intraoperative wound observation relies on the clinical experience of medical staff and lacks quantitative evaluation indicators, making it difficult to accurately capture key bleeding-related information such as wound vessel exposure and the distribution of oozing points. Furthermore, existing technologies do not effectively correlate the real-time state of the intraoperative wound with preoperative coagulation function data. Relying on only one type of data for risk prediction cannot comprehensively reflect the patient's current hemostasis physiological state, resulting in insufficient accuracy in bleeding risk prediction. This often leads to the failure to detect potential bleeding risks in a timely manner, thus delaying intervention and increasing the probability of postoperative bleeding complications. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a system for predicting and intervening in the risk of post-oral bleeding based on multimodal data fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: a system for predicting and intervening in the risk of post-oral bleeding based on multimodal data fusion, comprising:

[0006] The data standardization module collects preoperative coagulation function test data, intraoperative wound image data of the surgical site, and continuous postoperative physiological sign monitoring data of the target patients, and converts all data into a digital information stream with standard timestamps.

[0007] The feature extraction module extracts tissue damage features from the intraoperative wound image data in the digital information stream, and generates parameters such as the proportion of exposed blood vessels, the distribution density of tissue bleeding points, and the wound edge contraction state of the surgical site wound.

[0008] The feature fusion module associates and maps the vascular exposure area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter with the preoperative coagulation function test data obtained from the digital information stream to construct a bleeding risk feature vector that reflects the patient's current hemostasis physiological state.

[0009] The model training module uses the bleeding risk feature vector to train a preset bleeding risk classifier to obtain the oral postoperative bleeding risk prediction model that can output risk probability values.

[0010] The risk intervention module inputs the real-time postoperative continuous physiological sign monitoring data of the target patient into the oral postoperative bleeding risk prediction model to obtain the bleeding risk probability value of the target patient at the current moment, and triggers corresponding intervention measures based on the bleeding risk probability value.

[0011] As a further aspect of the present invention, tissue damage features are extracted from the intraoperative wound image data in the digital information stream to generate parameters such as the proportion of exposed blood vessels, the distribution density of tissue bleeding points, and the wound edge contraction state of the surgical site wound, including:

[0012] The intraoperative wound image data is denoised and enhanced to improve the contrast between blood vessels and soft tissues in the intraoperative wound image data, resulting in a preprocessed wound image.

[0013] Based on a pre-set vascular tissue colorimetric model, the color space of the pre-processed wound image is converted, and a region growing algorithm is used to separate the vascular tissue region and the non-vascular tissue region from the pre-processed wound image.

[0014] Calculate the proportion of the separated vascular tissue region in the total area of ​​the preprocessed wound image, and determine the proportion as the proportion of the exposed vascular area;

[0015] In the preprocessed wound image, a fixed pixel sampling window is set, and each pixel sampling window in the preprocessed wound image is traversed to count the number of pixels with grayscale anomalies. The number of pixels with grayscale anomalies is divided by the total number of pixels in the preprocessed wound image to calculate the tissue bleeding point distribution density parameter. The tissue bleeding point distribution density parameter is obtained by counting and normalizing the bleeding points in the intraoperative wound image data using an image segmentation algorithm.

[0016] The edge contour of the preprocessed wound image is extracted, and the wound edge contraction state parameter is calculated by calculating the perimeter change rate of the edge contour within two adjacent sampling periods. The wound edge contraction state parameter is used to represent the dynamic trend of wound closure.

[0017] As a further aspect of the present invention, the parameters of the proportion of exposed blood vessel area, the distribution density of tissue bleeding points, and the wound edge contraction state are correlated and mapped with the preoperative coagulation function test data obtained from the digital information stream to construct a bleeding risk feature vector reflecting the patient's current hemostatic physiological state, including:

[0018] The preoperative coagulation function test data were analyzed to extract the values ​​of prothrombin time, activated partial thromboplastin time, and fibrinogen concentration.

[0019] The values ​​of the prothrombin time index, the activated partial thromboplastin time index, and the fibrinogen concentration index are time-aligned with the vascular exposure area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter to ensure that each index corresponds to the same postoperative time point.

[0020] According to the pre-established clinical weight mapping rules, weight factors for the coagulation dysfunction dimension are assigned to the prothrombin time index, the activated partial thromboplastin time index, and the fibrinogen concentration index. At the same time, weight factors for the wound healing dimension are assigned to the vascular exposure area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter.

[0021] The weighted values ​​of the six indicators are concatenated in sequence and normalized to eliminate differences in dimensions, generating a multi-dimensional bleeding risk feature vector.

[0022] As a further aspect of the present invention, a pre-set bleeding risk classifier is trained using the bleeding risk feature vector to obtain the post-oral bleeding risk prediction model capable of outputting risk probability values, including:

[0023] The bleeding risk classifier is pre-built based on a historical multimodal data sample set;

[0024] Retrieve multiple historical multimodal data sample sets stored in the historical case database. Each historical multimodal data sample set contains preoperative coagulation function test data, intraoperative wound image data, and postoperative physiological sign monitoring data corresponding to the historical case.

[0025] For each of the aforementioned historical multimodal data sample sets, the feature extraction and feature vector construction steps are repeated to generate a standard bleeding risk feature vector corresponding to the historical case, and the vector is marked as a positive or negative sample.

[0026] The standard bleeding risk feature vector is divided into a training subset and a validation subset. The gradient boosting decision tree algorithm is used to iteratively train the initial classification model on the training subset. After each iteration, the recall and precision of the model are calculated using the validation subset.

[0027] When the product of recall and precision reaches a preset performance threshold, iterative training is stopped, the classification model at this time is solidified into the oral postoperative bleeding risk prediction model, and the output layer of the oral postoperative bleeding risk prediction model is set to output the bleeding risk probability value in the form of a floating-point number between zero and one.

[0028] As a further aspect of the present invention, the real-time postoperative continuous physiological sign monitoring data of the target patient is input into the oral postoperative bleeding risk prediction model to obtain the bleeding risk probability value of the target patient at the current moment, and corresponding intervention measures are triggered based on the bleeding risk probability value, including:

[0029] The system receives real-time data streams from a blood oxygen saturation probe attached to the patient's neck and a photoelectric volumetric sensor attached to the patient's wrist, and extracts two indicators, heart rate coefficient of variation and blood oxygen saturation fluctuation amplitude, as the postoperative continuous physiological sign monitoring data.

[0030] The current heart rate variability coefficient, the blood oxygen saturation fluctuation amplitude, the latest vascular exposed area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter are combined to generate the real-time bleeding risk feature vector.

[0031] The real-time bleeding risk feature vector is input into the oral surgery bleeding risk prediction model to calculate the bleeding risk probability value of the patient at the current moment.

[0032] Determine whether the bleeding risk probability value exceeds a preset warning threshold. If the bleeding risk probability value is greater than the warning threshold, retrieve the intervention protocol that matches the current risk level and activate the physical execution mechanism corresponding to the intervention protocol.

[0033] As a further aspect of the present invention, the step of retrieving an intervention protocol that matches the current risk level and activating the physical execution mechanism corresponding to the intervention protocol includes:

[0034] Establish a lookup table containing multiple risk intervals, each risk interval corresponding to a preset risk level, and each risk level is bound to a set of intervention measures protocols;

[0035] The bleeding risk probability value is mapped to the lookup table to determine its target risk range and corresponding target risk level.

[0036] Read the intervention protocol bound to the target risk level, which includes parameters for the duration of pressure cold compress, dosage parameters for hemostatic drugs, and setting parameters for the alarm level of medical staff;

[0037] Activate the cold compress device control module corresponding to the execution duration parameter of the pressure cold compress, and send a control command corresponding to the dosage parameter of the hemostatic drug to the drug infusion pump. At the same time, a warning window corresponding to the setting parameter of the alarm level pops up on the medical terminal interface.

[0038] As a further aspect of the present invention, the activation of the cold compress device control module corresponding to the execution duration parameter of the pressure cold compress includes:

[0039] Read the execution duration parameter of the pressure cold compress and convert it into a working cycle instruction that the cold compress device can recognize;

[0040] Based on the wound location coordinates of the surgical site, the robotic arm is driven to move the cold compress head to a position aligned with the wound location coordinates;

[0041] Start the cooling cycle of the cold compress device and control the cold compress head to contact the wound skin surface intermittently according to the duty cycle set by the work cycle instruction;

[0042] Before the execution duration parameter of the pressure cold compress is exhausted, the temperature change at the location coordinates of the wound is continuously monitored. If the temperature is lower than the freezing point threshold, the duration of each contact is automatically shortened to protect the surrounding tissue.

[0043] As a further aspect of the present invention, the step of sending a control command corresponding to the dosage parameters of the hemostatic drug to the drug infusion pump includes:

[0044] The dosage parameters of the hemostatic drug were analyzed and converted into the flow rate set value and total infusion time of the drug infusion pump.

[0045] Verify the current patient's weight data and drug allergy history record to confirm that the flow rate setting is within the safe dosage range of the drug instructions and that the hemostatic drug is not in the allergy history record;

[0046] A start command is sent to the drug infusion pump via a serial communication interface. The start command includes the converted flow rate setpoint and the total infusion time.

[0047] During the operation of the drug infusion pump, the status feedback code from the drug infusion pump is monitored in real time. If a status feedback code indicating pipeline blockage is received, the infusion is stopped immediately and a pipeline maintenance alarm is issued.

[0048] As a further aspect of the present invention, a model update module is also included, used for postoperative bleeding trend retrospection and model updating based on time series:

[0049] Collect full-cycle multimodal data of all patients who have completed the intervention process, and arrange the full-cycle multimodal data according to time series to form the postoperative bleeding evolution trajectory of the patients;

[0050] The postoperative bleeding evolution trajectory is compared with the prediction output of the oral postoperative bleeding risk prediction model at various historical time points, and the deviation value between the two is calculated.

[0051] The sample segments whose deviation values ​​exceed the allowable error range are selected, and the sample segments and their corresponding true outcomes are labeled as high-value corrected samples.

[0052] The high-value corrected samples are added to the original historical multimodal data sample set, and incremental training is performed based on the original model parameters to generate a new version of the oral postoperative bleeding risk prediction model with better prediction performance and complete the version iteration.

[0053] As a further aspect of the present invention, the step of adding the high-value corrected samples to the original historical multimodal data sample set, performing incremental training based on the original model parameters, generating a new version of the oral postoperative bleeding risk prediction model with better prediction performance, and completing version iteration includes:

[0054] Load the network structure and weight parameters of the currently deployed oral postoperative bleeding risk prediction model from the model parameter server as the initial model for incremental training;

[0055] The newly added high-value corrected samples are extracted from the high-value corrected sample library, which contains the labeled sample fragments and their corresponding true outcomes.

[0056] The high-value corrected samples are mixed with a portion of randomly sampled samples from the historical multimodal data sample set to form a training batch for this incremental training, and the high-value corrected samples are ensured to account for a preset proportion in the batch.

[0057] Based on freezing the weight parameters of all layers of the initial model except the last fully connected layer, the initial model is subjected to supervised incremental training using the training batch. During the training process, a smaller learning rate than that used in the original training is employed to fine-tune the initial model.

[0058] After each training cycle, the fine-tuned model is evaluated using an independent validation set. If the product of recall and precision on the validation set does not improve for several consecutive cycles, training is terminated early.

[0059] The final model parameters that have reached the termination condition are solidified to generate a new version of the oral postoperative bleeding risk prediction model. The version number, training time, and performance indicators of the new version of the oral postoperative bleeding risk prediction model are recorded in the model version management log to complete the version iteration.

[0060] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0061] By extracting tissue damage features from intraoperative wound image data, parameters such as the proportion of exposed blood vessels, the distribution density of tissue bleeding points, and the state of wound edge contraction are generated. Through quantified wound feature parameters, the real-time damage status of the wound during surgery can be accurately captured. This breaks through the limitations of conventional techniques that rely on medical staff's visual observation and subjective judgment of the wound bleeding status. It transforms wound-related bleeding risk factors from qualitative descriptions to quantitative data, which can more accurately reflect the hemostasis status of the wound, reduce the bias caused by subjective judgment, and make key information related to bleeding risk more objective and traceable.

[0062] By mapping parameters such as the proportion of exposed blood vessel area, the distribution density of tissue oozing points, and the wound edge contraction status with preoperative coagulation function test data obtained from digital information flow, a bleeding risk feature vector reflecting the patient's current hemostatic physiological state is constructed. This feature vector is then used to train a pre-set bleeding risk classifier to obtain a bleeding risk prediction model. This model integrates real-time quantitative data of the wound during surgery with basic preoperative coagulation data, breaking through the limitations of conventional techniques that rely solely on preoperative coagulation data or subjective observation during surgery for risk prediction. It can comprehensively and holistically reflect the patient's hemostatic physiological state, making bleeding risk prediction more closely aligned with the patient's actual situation, improving the accuracy of risk prediction, and enabling early detection of potential bleeding risks. This provides a precise direction for timely intervention measures and reduces the occurrence of postoperative bleeding complications. Attached Figure Description

[0063] Figure 1 This is a time sequence diagram of the oral postoperative bleeding risk prediction and intervention system based on multimodal data fusion described in this invention;

[0064] Figure 2 A flowchart illustrating the process of extracting tissue damage features;

[0065] Figure 3 A flowchart illustrating the process of constructing a bleeding risk feature vector;

[0066] Figure 4 Performance trend chart for training bleeding risk prediction model;

[0067] Figure 5 A roadmap for implementing interventions to address the risk of post-oral bleeding (biaxial efficacy-risk curve). Detailed Implementation

[0068] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0069] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0070] See Figure 1The data standardization module is responsible for collecting preoperative coagulation function test data, intraoperative wound image data of the surgical site, and postoperative continuous physiological sign monitoring data of the target patients. This module unifies data from different sources and with varying formats into a digital information stream with standard timestamps, providing a basis for temporal alignment for subsequent processing. The feature extraction module focuses on processing the intraoperative wound image data in the digital information stream, extracting tissue damage features that can quantitatively reflect the wound state through image analysis technology, including parameters such as the proportion of exposed blood vessels, the distribution density of tissue bleeding points, and the state of wound edge contraction. The feature fusion module associates and maps the parameters extracted from the images with the preoperative coagulation function test data parsed from the digital information stream. Based on clinical knowledge, this module aligns coagulation function indicators with local wound features in the time dimension, assigns weights, and finally splices and normalizes them to generate a multi-dimensional bleeding risk feature vector that comprehensively reflects the patient's current hemostasis physiological state. The model training module uses a sample set of bleeding risk feature vectors constructed from a large number of historical cases to conduct supervised training of a pre-set bleeding risk classifier. The training process continuously optimizes the model parameters until the model performance meets preset standards, resulting in a predictive model for post-oral bleeding risk that can output continuous risk probability values. The risk intervention module is the terminal for the system's interaction with the physical world. It inputs real-time post-operative continuous physiological monitoring data of the target patient into the trained post-oral bleeding risk prediction model, calculating and outputting the patient's bleeding risk probability value at the current moment. This module has built-in decision logic that can automatically match and trigger corresponding tiered intervention measures based on the calculated risk probability value, achieving proactive management of bleeding risk.

[0071] In one embodiment of the present invention, tissue damage features are extracted from intraoperative wound image data in a digital information stream to generate parameters such as the proportion of exposed blood vessels, the distribution density of bleeding points, and the state of wound edge contraction. The intraoperative wound image data originates from a real-time video stream acquired by an oral surgical microscope. An example scenario is the alveolar socket wound after mandibular molar extraction, with an image resolution of 1280x960 pixels and a bit depth of 24 bits. Data comparison shows that the contrast between blood vessels and soft tissue is low in the original image, but significantly improved after processing. In some embodiments, see [reference needed]. Figure 2The intraoperative wound image data was denoised and enhanced to improve the contrast between blood vessels and soft tissues, resulting in a preprocessed wound image. The process employed a median filter to remove salt-and-pepper noise and a histogram equalization algorithm to enhance the red channel information of the vascular region. Example data comparisons show that the signal-to-noise ratio of the denoised image improved by approximately 30%, while the average red component of the vascular region increased by 15%. In practice, based on a pre-defined vascular tissue chromaticity model, the preprocessed wound image underwent color space conversion. A region growing algorithm was then used to separate the vascular and non-vascular tissue regions from the preprocessed wound image. The vascular tissue chromaticity model defines the hue and saturation range of vascular tissue in the HSV color space. Example data comparisons show that after conversion, the hue values ​​of the vascular region are concentrated in the 0-10 degree range, while the hue values ​​of the non-vascular tissue are dispersed in the 20-60 degree range. Optionally, the seed point of the region growing algorithm is selected based on the pixel with the highest red component in the image. The growth criterion is that the color difference between pixels is less than a threshold of 5. Data comparisons show that the vascular tissue region segmentation integrity reaches over 95% under this threshold. The proportion of the isolated vascular tissue region in the total area of ​​the preprocessed wound image is calculated and determined as the vascular exposure area ratio parameter. In the example scenario, the pixel count of the vascular tissue region is 15,000, while the total pixel count of the preprocessed wound image is 1,228,800. The vascular exposure area ratio parameter is calculated to be 0.0122. Data comparison shows that the value of this parameter varies between 0.005 and 0.05 for different patients.

[0072] In specific implementation, a fixed pixel sampling window is set in the preprocessed wound image. Each pixel sampling window in the preprocessed wound image is traversed, and the number of pixels with grayscale anomalies is counted. This number is then divided by the total number of pixels in the preprocessed wound image to calculate the tissue bleeding point distribution density parameter. This parameter is obtained by counting and normalizing bleeding points in the intraoperative wound image data using an image segmentation algorithm. It can be understood that pixels with grayscale anomalies are defined as pixels whose grayscale value is more than 20 points lower than the average grayscale value of the surrounding area. Example data comparison shows that the grayscale value of the bleeding point area is typically 25-40 points lower than the surrounding area. In some embodiments, the formula for calculating the tissue bleeding point distribution density parameter is:

[0073]

[0074] in: This parameter represents the density of tissue bleeding points. This indicates the number of pixels with grayscale anomalies. This represents the total number of pixels in the preprocessed wound image, as shown in the example scene. For 5000 The value is 1228800, calculated as follows: The value is 0.00407. In specific implementation, the edge contour of the preprocessed wound image is extracted. By calculating the perimeter change rate of the edge contour within two adjacent sampling periods, the wound edge contraction state parameter is deduced. The wound edge contraction state parameter is used to represent the dynamic trend of wound closure. In the example scenario, the time interval between adjacent sampling is 5 minutes. The perimeter of the edge contour in the first sampling is 120 pixels, and the perimeter of the edge contour in the second sampling is 115 pixels, with a perimeter change rate of -0.0417. Optionally, the edge contour extraction uses the Canny edge detection algorithm, and the standard deviation of the Gaussian filter is set to 1.5. Data comparison shows that the edge continuity is better than other settings under this standard deviation. It can be understood that a negative value of the wound edge contraction state parameter indicates wound contraction, and a positive value indicates wound expansion. Example data comparison shows that this parameter is often negative in the early postoperative period, but may turn positive in the case of infection.

[0075] In one embodiment of the present invention, parameters such as the proportion of exposed blood vessel area, the distribution density of tissue bleeding points, and the wound edge contraction status are correlated and mapped with preoperative coagulation function test data obtained from a digital information stream to construct a bleeding risk feature vector reflecting the patient's current hemostatic physiological state. In the example scenario, the feature vector construction process for patient A 30 minutes postoperatively uses preoperative coagulation function test data with a prothrombin time of 12.5 seconds, an activated partial thromboplastin time of 35.2 seconds, and a fibrinogen concentration of 2.8 g / L. For specific implementation details, please refer to... Figure 3The preoperative coagulation function test data was analyzed to extract values ​​for prothrombin time, activated partial thromboplastin time, and fibrinogen concentration. Data comparison showed that the normal range for prothrombin time was between 11 and 13.5 seconds for different patients. In some embodiments, the values ​​of prothrombin time, activated partial thromboplastin time, and fibrinogen concentration were time-aligned with parameters for the proportion of exposed vascular area, the distribution density of tissue bleeding points, and the wound edge contraction status, respectively, to ensure that each indicator corresponds to the same postoperative time point. In the example scenario, the parameters for the proportion of exposed vascular area, the distribution density of tissue bleeding points, and the wound edge contraction status of patient A were all collected and calculated 30 minutes after surgery, and were time-series aligned with the preoperative coagulation function test data. In practice, based on pre-established clinical weighting mapping rules, weighting factors for coagulation dysfunction were assigned to the prothrombin time, activated partial thromboplastin time, and fibrinogen concentration parameters. Simultaneously, weighting factors for wound healing were assigned to the vascular exposure area ratio, tissue bleeding point distribution density, and wound edge contraction status parameters. These clinical weighting mapping rules were set based on expert experience. In the example, the weighting factor for prothrombin time was 0.2, for activated partial thromboplastin time it was 0.3, for fibrinogen concentration it was 0.1, for vascular exposure area ratio it was 0.15, for tissue bleeding point distribution density it was 0.2, and for wound edge contraction status it was 0.05. It can be understood that weighting involves multiplying each indicator value by its corresponding weighting factor. Data comparison shows that after weighting, the prothrombin time value changed from 12.5 to 2.5. In practice, the weighted values ​​of the six indicators are concatenated sequentially and normalized to eliminate dimensional differences, generating a multi-dimensional bleeding risk feature vector. In the example, the weighted values ​​of patient A are concatenated as [2.5, 10.56, 0.28, 0.00183, 0.000814, -0.002085], and the normalized feature vector is [0.45, 0.62, 0.12, 0.05, 0.03, 0.01]. Optionally, the normalization process uses the min-max normalization method, linearly transforming the values ​​of each dimension to the [0,1] interval, as shown in the formula:

[0076]

[0077] in: Represents the normalized eigenvalues. Represents the original feature values. This represents the minimum value of the feature dimension in the training set. This indicates the maximum value of the feature dimension in the training set. In the example, for the prothrombin time index dimension, the minimum value in the training set is 1.8 and the maximum value is 5.0.

[0078] In the specific implementation, a pre-set bleeding risk classifier is trained using bleeding risk feature vectors to obtain a postoperative oral bleeding risk prediction model capable of outputting risk probability values. The bleeding risk classifier is pre-built based on a historical multimodal data sample set, which contains 3000 historical case records with labeled outcomes. In some embodiments, multiple historical multimodal data sample sets stored in a historical case database are retrieved. Each historical multimodal data sample set contains preoperative coagulation function test data, intraoperative wound image data, and postoperative physiological sign monitoring data corresponding to the historical case. In the example, the historical case database spans three years. In the specific implementation, for each historical multimodal data sample set, the feature extraction and feature vector construction steps are repeatedly performed to generate a standard bleeding risk feature vector corresponding to the historical case, and it is marked as a positive sample or a negative sample. Positive samples correspond to clinically confirmed cases of postoperative bleeding. In the example, 450 out of 3000 samples are marked as positive samples. Optionally, the standard bleeding risk feature vector is divided into a training subset and a validation subset. A gradient boosting decision tree algorithm is used to iteratively train the initial classification model on the training subset. After each iteration, the recall and precision of the model are calculated using the validation subset. The training subset contains 2400 samples, and the validation subset contains 600 samples. The learning rate of the gradient boosting decision tree algorithm is set to 0.1, and the maximum tree depth is 5. In specific implementation, when the product of recall and precision reaches a preset performance threshold, iterative training is stopped, and the classification model at this point is solidified as a post-oral bleeding risk prediction model. The output layer of the post-oral bleeding risk prediction model is set to output bleeding risk probability values ​​in floating-point form between zero and one. The preset performance threshold is 0.85. In the example training process, the product of recall and precision reached 0.852 at the 150th iteration.

[0079] In one embodiment of the present invention, real-time postoperative continuous physiological sign monitoring data of the target patient is input into the oral postoperative bleeding risk prediction model to obtain the bleeding risk probability value of the target patient at the current moment, and corresponding intervention measures are triggered based on the bleeding risk probability value. In the example scenario, patient B triggers the system risk assessment process two hours after surgery. In some embodiments, data streams from a blood oxygen saturation probe attached to the patient's neck and a photoplethysmography sensor attached to the patient's wrist are received in real time. Two indicators, heart rate variability coefficient and blood oxygen saturation fluctuation amplitude, are extracted from these data as postoperative continuous physiological sign monitoring data. In the example, the raw data stream from patient B's blood oxygen saturation probe contains a sampling value once per second. The heart rate variability coefficient is calculated to be 85 milliseconds by calculating the standard deviation of the RR interval within the last five minutes, and the blood oxygen saturation fluctuation amplitude is calculated to be 4% by calculating the difference between the maximum and minimum blood oxygen saturation values ​​within the last three minutes. In practice, the current heart rate variability coefficient, blood oxygen saturation fluctuation amplitude, and the latest parameters for the proportion of exposed blood vessels, the distribution density of tissue bleeding points, and the wound edge contraction status are combined to generate a real-time bleeding risk feature vector. In the example, at the assessment time, patient B's latest parameters for the proportion of exposed blood vessels are 0.0105, the distribution density of tissue bleeding points is 0.0061, and the wound edge contraction status is -0.0302. These parameters are combined with the heart rate variability coefficient of 85 and the blood oxygen saturation fluctuation amplitude of 4 to form a vector [0.0105, 0.0061, -0.0302, 85, 4], which is then normalized. In practice, the real-time bleeding risk feature vector is input into the oral surgery bleeding risk prediction model to calculate the patient's bleeding risk probability value at the current moment. In the example, after inputting the vector, the model outputs a floating-point number of 0.72 through forward propagation. In practice, it is determined whether the bleeding risk probability value exceeds the preset warning threshold. If the bleeding risk probability value is greater than the warning threshold, the system retrieves the intervention measure protocol that matches the current risk level and activates the physical execution mechanism corresponding to the intervention measure protocol. In the example, the preset warning threshold is 0.6, and the calculated bleeding risk probability value of 0.72 is greater than 0.6, so the system triggers the intervention retrieval and execution logic.

[0080] In specific implementation, the retrieval and initiation process includes establishing a lookup table containing multiple risk intervals. Each risk interval corresponds to a preset risk level, and each risk level is bound to a set of intervention protocols. In some embodiments, the lookup table defines three risk intervals. In this example, interval one (0, 0.4) corresponds to a low risk level, interval two (0.4, 0.6) corresponds to a medium risk level, and interval three (0.6, 1.0) corresponds to a high risk level. In specific implementation, the calculated bleeding risk probability value is mapped to the lookup table to determine its target risk interval and corresponding target risk level. In this example, the bleeding risk probability value of 0.72 falls into interval three (0.6, 1.0), so the target risk level is determined to be high risk. In specific implementation, the intervention protocol bound to the target risk level is read. The intervention protocol includes parameters for the execution duration of pressure cold compress, the dosage parameters of hemostatic drugs, and the alarm level for medical staff. In the example, the intervention protocol corresponding to the high-risk level sets the parameters as follows: the duration of pressure cold compress is 20 minutes, the dosage of the hemostatic drug is 1 gram of tranexamic acid, and the alarm level for medical staff is set to red level one alarm. The system activates the cold compress device control module corresponding to the pressure cold compress duration parameter and sends a control command corresponding to the hemostatic drug dosage parameter to the drug infusion pump. Simultaneously, an alert window corresponding to the alarm level setting parameter pops up on the medical staff terminal interface. It can be understood that for the high-risk level, the control command instructs the drug infusion pump to infuse the hemostatic drug at a rate of 100 ml per hour. In practical implementation, the system automatically executes subsequent physical operations based on the search results. The risk value mapping relationship is as follows:

[0081]

[0082] in: This represents the discretized risk level number. This represents the calculated probability value of bleeding risk. This represents the lower bound of the current risk range. This represents the upper limit of the current risk range, indicated by the symbol. This represents the floor function. In the example, for a probability value of 0.72, within the interval (0.6, 1.0], , Calculated This corresponds to a high-risk level.

[0083] In one embodiment of the present invention, when the intervention protocol is initiated, the cold compress device control module corresponding to the execution duration parameter of the pressure cold compress is activated. The execution duration parameter of the pressure cold compress is read and converted into a work cycle instruction recognizable by the cold compress device. In the example scenario, the execution duration parameter of the pressure cold compress read from the high-risk level intervention protocol is 20 minutes. The converted work cycle instruction includes a total duration of 1200 seconds, a work cycle of 30 seconds, and a duty cycle of 50%. In some embodiments, the work cycle instruction is sent to the cold compress device control module in the form of a structured data packet. The data packet contains parameters such as the number of cycles, the working duration, and the interval duration. Refer to Table 1 for the correspondence of the specific parameters.

[0084] Table 1: Example Table of Working Cycle Command Parameters

[0085]

[0086] In practical implementation, based on the wound location coordinates at the surgical site, the robotic arm drives the cold compress head to move to an alignment with the wound location coordinates. In this example, the wound location coordinates at the surgical site are defined in three-dimensional space as (150.2, 80.5, 25.0) mm. The end effector of the robotic arm carries the cold compress head from the initial position to the target coordinates, with a positioning accuracy error of less than 0.5 mm. In practical implementation, the cooling cycle of the cold compress device is activated, and according to the duty cycle set by the work cycle command, the cold compress head is controlled to contact the wound skin surface intermittently. In this example, after the cooling cycle of the cold compress device is activated, the temperature of the cold compress head drops from room temperature to 4 degrees Celsius within 60 seconds, and operates in an intermittent mode of 15 seconds of contact and 15 seconds of removal. Optionally, the formula for calculating the duration of a single contact is:

[0087]

[0088] in: Indicates the duration of a single contact. Indicates duty cycle, This indicates the duration of a single cycle; in the example, it represents the duty cycle. The value is 0.5, and the duration of a single cycle is... The duration of a single contact is calculated to be 30 seconds. The duration is 15 seconds. Before the execution time parameter of the pressure cold compress is exhausted, the temperature change at the location coordinates of the wound is continuously monitored. If the temperature is lower than the freezing point threshold, the single contact time is automatically shortened to protect the surrounding tissue. In the example, the freezing point threshold is set to 2 degrees Celsius. When the monitored temperature drops to 1.8 degrees Celsius, the system dynamically adjusts the duty cycle from 50% to 30%, so that the single contact time is shortened from 15 seconds to 9 seconds.

[0089] In some embodiments, a control command corresponding to the dosage parameters of the hemostatic drug is sent to the drug infusion pump. The dosage parameters of the hemostatic drug are parsed and converted into the flow rate setpoint and total infusion time of the drug infusion pump. In the example, the dosage parameter of the hemostatic drug parsed from the intervention protocol is "tranexamic acid 1 gram". Combined with the drug concentration of 10 mg / mL, the total volume of drug solution to be infused is calculated to be 100 mL. The total infusion time is set to 60 minutes, and the flow rate setpoint is calculated to be 100 mL / hour. In practice, the patient's weight data and drug allergy history are checked to confirm that the flow rate setting is within the safe dosage range of the drug's instructions and that the hemostatic drug is not listed in the allergy history. In the example, the patient weighs 70 kg, and the drug instructions state that the upper limit of the safe infusion rate of tranexamic acid is 15 mg / kg / hour. The calculated maximum safe dose for this patient is 1050 mg / hour, which is equivalent to 105 ml / hour. The set flow rate of 100 ml / hour is within the safe range. Furthermore, a check of the patient's allergy history reveals no allergy to tranexamic acid. It can be understood that a start command is sent to the drug infusion pump via a serial communication interface. The start command includes the converted flow rate setting and the total infusion time. In the example, the communication protocol for the start command is RS-485, and the data frame includes a command header "0xA0", a flow rate value "100", and a duration value "60". In practice, during the operation of the drug infusion pump, the status feedback codes from the drug infusion pump are monitored in real time. If a status feedback code indicating pipeline blockage is received, the infusion is stopped immediately and a pipeline maintenance alarm is issued. In the example, the drug infusion pump provides a status code once per second. The status code "0xE1" indicates pipeline blockage. After receiving three consecutive "0xE1" status codes, the system sends a stop command "0xB0" to the infusion pump and triggers audible and visual alarms locally and at the nurse station.

[0090] See Figure 4During the training and optimization phases of the bleeding risk prediction model, the three core performance indicators—accuracy, recall, and F1 score—showed a significant synergistic improvement trend with the number of iterations, reflecting the complete evolution of the model from initial convergence to performance saturation. In the early training phase (iterations 0-10), all three indicators showed rapid growth: accuracy increased from approximately 0.76 to 0.87, recall from approximately 0.72 to 0.77, and F1 score from approximately 0.73 to 0.79. This indicates that the model effectively learned the mapping relationship between bleeding risk features and the true outcome in the early iterations, and its overall discrimination ability rapidly improved. Entering the middle training phase (iterations 10-15), the performance growth rate slowed but remained stable: accuracy fluctuated between 0.87 and 0.89, recall increased to 0.81-0.83, and F1 score climbed to 0.83-0.85. This indicates that the model's generalization ability on complex samples gradually improved, while a slight overfitting tendency began to appear, manifested as small fluctuations in accuracy and recall. In the later stages of training (15-20 iterations), the three metrics stabilized and reached their peaks: accuracy stabilized at 0.87-0.89, recall remained at 0.82-0.83, and the F1 score peaked at approximately 0.88. At this point, the overall performance of the model was close to saturation, and the marginal effect of further iterations on performance improvement significantly decreased, consistent with the convergence characteristics of gradient boosting decision tree algorithms in incremental training. From the perspective of metric synergy, accuracy consistently exceeded recall, indicating that the model has a stronger ability to distinguish negative samples. The continuous increase in the F1 score, as the harmonic average of accuracy and recall, validates that while maintaining high specificity, the model gradually improved its sensitivity in identifying positive samples with high bleeding risk. This is of great significance for avoiding missed diagnoses of high-risk patients in clinical settings.

[0091] In one embodiment of the present invention, the model update module is used for time-series-based retrospective analysis of postoperative bleeding trends and model updates. An example scenario involves the first periodic incremental update of the oral postoperative bleeding risk prediction model after one month of system operation. In a specific implementation, full-cycle multimodal data of all patients who have completed the intervention process is collected. This full-cycle multimodal data is arranged according to time series to form the postoperative bleeding evolution trajectory of each patient. In the example, full-cycle multimodal data of 300 patients who completed the intervention in the past month are collected. Each patient's data includes feature vectors recorded every 5 minutes from the end of surgery to 24 hours postoperatively, model predicted probability values, and actual clinical observation outcomes. In some embodiments, the full-cycle multimodal data is stored in a time-series database using patient ID and time as a joint primary key. The data is sorted from earliest to latest by timestamp, forming a data sequence where each record contains a time point, feature vector, predicted value, and true label—that is, the postoperative bleeding evolution trajectory. Data comparison shows that patient C's postoperative bleeding evolution trajectory contains 288 consecutive records, while patient D's data only contains 200 records due to data interruption. In practice, the postoperative bleeding trajectory is compared with the prediction outputs of the oral postoperative bleeding risk prediction model at various historical time points. The deviation value between the two is calculated. For each time point in the postoperative bleeding trajectory, the predicted probability value output by the model at that time is compared with the corresponding actual clinical outcome label, which is 0 (no bleeding) or 1 (bleeding). Optionally, the deviation value... The calculation formula is:

[0092]

[0093] in: This represents the deviation value at time point i. This represents the predicted probability value output by the oral surgery bleeding risk prediction model at time point i. This represents the true outcome label recorded by clinical observation at time point i. In the example, it represents the model-predicted probability value at time point 180 minutes after patient E's surgery. The value is 0.25, while the true ending tag is... The value is 1 (bleeding occurs), and the deviation value is calculated. It is 0.75.

[0094] In practice, sample segments with deviation values ​​exceeding the allowable error range are selected. These sample segments and their corresponding true outcomes are labeled as high-value corrected samples. In this example, the allowable error range is set to 0.3. Therefore, sample segments with deviation values ​​greater than 0.3 are selected. For patient E, four consecutive records from the 170th to 190th minute post-surgery all have deviation values ​​greater than 0.3. These four records and their corresponding true outcome labels are collectively labeled as a high-value corrected sample. In some embodiments, the high-value corrected sample library is stored in units of sample segments. Each entry contains the patient ID, start and end timestamps, feature vector sequences of all time points within that time period, and a unified true outcome label corresponding to that time period. Data comparison shows that in one update cycle, 45 high-value corrected samples were selected from the data of 300 patients. In the specific implementation, high-value corrected samples are added to the existing historical multimodal data sample set. Incremental training is performed based on the original model parameters to generate a new version of the oral surgery bleeding risk prediction model with better prediction performance, and the version iteration is completed. The original historical multimodal data sample set contains 3000 samples. After adding 45 high-value corrected samples, the total number of samples becomes 3045. Optionally, the network structure and weight parameters of the currently deployed oral surgery bleeding risk prediction model are loaded from the model parameter server as the initial model for incremental training. The version number of the currently deployed model is V1.2, and its network structure includes one input layer, three hidden layers, and one output layer. In the specific implementation, the latest high-value corrected samples are extracted from the high-value corrected sample library. The high-value corrected samples contain labeled sample segments and their corresponding true outcomes. In the example, the 45 high-value corrected samples extracted this time each sample segment contains an average of 3.2 consecutive time point feature vector data.

[0095] It is understandable that high-value corrected samples are mixed with a portion of randomly sampled samples from the historical multimodal data set to form training batches for this incremental training. The high-value corrected samples are ensured to account for a predetermined proportion in each batch; in this example, the predetermined proportion is 20%. Each training batch is 256 in size, meaning each batch contains approximately 51 high-value corrected samples and approximately 205 samples randomly sampled from the original historical sample set. In practice, after freezing the weight parameters of all layers in the initial model except the last fully connected layer, supervised incremental training of the initial model is performed using the training batches. A smaller learning rate than the original training rate is used during training to fine-tune the initial model; in this example, the learning rate for the original model training is 0.001, and the learning rate for incremental training is 0.0001. It is understandable that after each training cycle, the performance of the fine-tuned model is evaluated using an independent validation set. If the product of recall and precision on the validation set does not improve for several consecutive cycles, training is terminated early. In this example, the validation set contains 500 historical samples that did not participate in training. The early termination condition is set as the improvement of the product of recall and precision being less than 0.001 within 5 consecutive training cycles. In practice, the final model parameters that meet the termination condition are fixed, generating a new version of the oral surgery bleeding risk prediction model. The version number, training time, and performance metrics of the new version of the oral surgery bleeding risk prediction model are recorded in the model version management log, completing the version iteration. In this example, the newly generated model is assigned version number V1.3, the training time is October 26, 2023, and its final recall on the validation set is 0.88, its final precision is 0.91, and its product is 0.8008. This information is written to a new record in the version management log.

[0096] See Figure 5In the roadmap for intervention measures to manage post-oral bleeding risk, the effectiveness of the intervention process and the risk control effect can be quantitatively evaluated using a biaxial curve. Specifically, the left vertical axis represents the execution time (in minutes) of the intervention measures, the right vertical axis represents the real-time bleeding risk value, and the horizontal axis sequentially presents the complete intervention stages from risk identification to model update. The intervention process begins in the "risk identification" stage, at which point the real-time bleeding risk value reaches a peak of 0.8, while the execution time is only 5 minutes. As intervention measures such as "cold compress initiation" and "drug infusion" are executed sequentially, the execution time gradually increases to 30 minutes, while the real-time bleeding risk value decreases simultaneously to 0.3. In the "vital signs monitoring" and "effect evaluation" stages, the execution time of the intervention measures continues to increase, while the real-time bleeding risk value further decreases to below 0.1. Finally, in the "model update" stage, the execution time reaches 75 minutes, and the real-time bleeding risk value is controlled at an extremely low level of 0.03. The synergistic changes of the biaxial curves intuitively reflect the efficiency of the intervention and the effectiveness of risk control: the upward trend of the execution time curve and the downward trend of the real-time bleeding risk value curve are highly negatively correlated, indicating that the bleeding risk is effectively controlled as the intervention process deepens, and also demonstrating the important role of model updates in improving intervention effectiveness.

[0097] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A system for predicting and intervening in the risk of post-oral bleeding based on multimodal data fusion, characterized in that, include: The data standardization module collects preoperative coagulation function test data, intraoperative wound image data of the surgical site, and continuous postoperative physiological sign monitoring data of the target patients, and converts all data into a digital information stream with standard timestamps. The feature extraction module extracts tissue damage features from the intraoperative wound image data in the digital information stream, generating parameters such as the proportion of exposed blood vessels, the distribution density of tissue bleeding points, and the wound edge contraction state of the surgical site wound, including: The intraoperative wound image data is denoised and enhanced to improve the contrast between blood vessels and soft tissues in the intraoperative wound image data, resulting in a preprocessed wound image. Based on a pre-set vascular tissue colorimetric model, the color space of the pre-processed wound image is converted, and a region growing algorithm is used to separate the vascular tissue region and the non-vascular tissue region from the pre-processed wound image. Calculate the proportion of the separated vascular tissue region in the total area of ​​the preprocessed wound image, and determine the proportion as the proportion of the exposed vascular area; In the preprocessed wound image, a fixed pixel sampling window is set, and each pixel sampling window in the preprocessed wound image is traversed to count the number of pixels with grayscale anomalies. The number of pixels with grayscale anomalies is divided by the total number of pixels in the preprocessed wound image to calculate the tissue bleeding point distribution density parameter. The tissue bleeding point distribution density parameter is obtained by counting and normalizing the bleeding points in the intraoperative wound image data using an image segmentation algorithm. The edge contour of the preprocessed wound image is extracted, and the wound edge contraction state parameter is calculated by calculating the perimeter change rate of the edge contour within two adjacent sampling periods. The wound edge contraction state parameter is used to represent the dynamic trend of wound closure. The feature fusion module associates and maps the vascular exposure area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter with the preoperative coagulation function test data obtained from the digital information stream to construct a bleeding risk feature vector that reflects the patient's current hemostasis physiological state. The model training module uses the bleeding risk feature vector to train a preset bleeding risk classifier to obtain the oral postoperative bleeding risk prediction model that can output risk probability values. The risk intervention module inputs the real-time postoperative continuous physiological sign monitoring data of the target patient into the oral postoperative bleeding risk prediction model to obtain the bleeding risk probability value of the target patient at the current moment, and triggers corresponding intervention measures based on the bleeding risk probability value.

2. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 1, characterized in that, The parameters of the proportion of exposed blood vessel area, the distribution density of tissue bleeding points, and the wound edge contraction status are correlated and mapped with the preoperative coagulation function test data obtained from the digital information stream to construct a bleeding risk feature vector reflecting the patient's current hemostatic physiological state, including: The preoperative coagulation function test data were analyzed to extract the values ​​of prothrombin time, activated partial thromboplastin time, and fibrinogen concentration. The values ​​of the prothrombin time index, the activated partial thromboplastin time index, and the fibrinogen concentration index are time-aligned with the vascular exposure area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter to ensure that each index corresponds to the same postoperative time point. According to the pre-established clinical weight mapping rules, weight factors for the coagulation dysfunction dimension are assigned to the prothrombin time index, the activated partial thromboplastin time index, and the fibrinogen concentration index. At the same time, weight factors for the wound healing dimension are assigned to the vascular exposure area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter. The weighted values ​​of the six indicators are concatenated in sequence and normalized to eliminate differences in dimensions, generating a multi-dimensional bleeding risk feature vector.

3. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 2, characterized in that, The oral cavity postoperative bleeding risk prediction model is obtained by training a pre-set bleeding risk classifier using the bleeding risk feature vector, and is capable of outputting risk probability values. The model includes: The bleeding risk classifier is pre-built based on a historical multimodal data sample set; Retrieve multiple historical multimodal data sample sets stored in the historical case database. Each historical multimodal data sample set contains preoperative coagulation function test data, intraoperative wound image data, and postoperative physiological sign monitoring data corresponding to the historical case. For each of the aforementioned historical multimodal data sample sets, the feature extraction and feature vector construction steps are repeated to generate a standard bleeding risk feature vector corresponding to the historical case, and the vector is marked as a positive or negative sample. The standard bleeding risk feature vector is divided into a training subset and a validation subset. The gradient boosting decision tree algorithm is used to iteratively train the initial classification model on the training subset. After each iteration, the recall and precision of the model are calculated using the validation subset. When the product of recall and precision reaches a preset performance threshold, iterative training is stopped, the classification model at this time is solidified into the oral postoperative bleeding risk prediction model, and the output layer of the oral postoperative bleeding risk prediction model is set to output the bleeding risk probability value in the form of a floating-point number between zero and one.

4. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 3, characterized in that, The process involves inputting real-time postoperative continuous physiological monitoring data of the target patient into the oral postoperative bleeding risk prediction model to obtain the bleeding risk probability value of the target patient at the current moment, and triggering corresponding intervention measures based on the bleeding risk probability value, including: The system receives real-time data streams from a blood oxygen saturation probe attached to the patient's neck and a photoelectric volumetric sensor attached to the patient's wrist, and extracts two indicators, heart rate coefficient of variation and blood oxygen saturation fluctuation amplitude, as the postoperative continuous physiological sign monitoring data. The current heart rate variability coefficient, the blood oxygen saturation fluctuation amplitude, the latest vascular exposed area ratio parameter, the tissue bleeding point distribution density parameter, and the wound edge contraction state parameter are combined to generate the real-time bleeding risk feature vector. The real-time bleeding risk feature vector is input into the oral postoperative bleeding risk prediction model to calculate the bleeding risk probability value of the patient at the current moment. Determine whether the bleeding risk probability value exceeds a preset warning threshold. If the bleeding risk probability value is greater than the warning threshold, retrieve the intervention protocol that matches the current risk level and activate the physical execution mechanism corresponding to the intervention protocol.

5. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 4, characterized in that, The process of retrieving intervention protocols that match the current risk level and activating the physical execution mechanism corresponding to the intervention protocol includes: Establish a lookup table containing multiple risk intervals, each risk interval corresponding to a preset risk level, and each risk level is bound to a set of intervention measures protocols; The bleeding risk probability value is mapped to the lookup table to determine its target risk range and corresponding target risk level. Read the intervention protocol bound to the target risk level, which includes parameters for the duration of pressure cold compress, dosage parameters for hemostatic drugs, and setting parameters for the alarm level of medical staff; Activate the cold compress device control module corresponding to the execution duration parameter of the pressure cold compress, and send a control command corresponding to the dosage parameter of the hemostatic drug to the drug infusion pump. At the same time, a warning window corresponding to the setting parameter of the alarm level pops up on the medical terminal interface.

6. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 5, characterized in that, The activation of the cold compress device control module corresponding to the execution duration parameter of the pressure cold compress includes: Read the execution duration parameter of the pressure cold compress and convert it into a working cycle instruction that the cold compress device can recognize; Based on the wound location coordinates of the surgical site, the robotic arm is driven to move the cold compress head to a position aligned with the wound location coordinates; Start the cooling cycle of the cold compress device and control the cold compress head to contact the wound skin surface intermittently according to the duty cycle set by the work cycle instruction; Before the execution duration parameter of the pressure cold compress is exhausted, the temperature change at the location coordinates of the wound is continuously monitored. If the temperature is lower than the freezing point threshold, the duration of each contact is automatically shortened to protect the surrounding tissue.

7. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 6, characterized in that, Sending control commands to the drug infusion pump corresponding to the dosage parameters of the hemostatic drug includes: The dosage parameters of the hemostatic drug were analyzed and converted into the flow rate set value and total infusion time of the drug infusion pump. Verify the current patient's weight data and drug allergy history record to confirm that the flow rate setting is within the safe dosage range of the drug instructions and that the hemostatic drug is not in the allergy history record; A start command is sent to the drug infusion pump via a serial communication interface. The start command includes the converted flow rate setpoint and the total infusion time. During the operation of the drug infusion pump, the status feedback code from the drug infusion pump is monitored in real time. If a status feedback code indicating pipeline blockage is received, the infusion is stopped immediately and a pipeline maintenance alarm is issued.

8. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 7, characterized in that, It also includes a model update module for time-series-based retrospective analysis of postoperative bleeding trends and model updates: Collect full-cycle multimodal data of all patients who have completed the intervention process, and arrange the full-cycle multimodal data according to time series to form the postoperative bleeding evolution trajectory of the patients; The postoperative bleeding evolution trajectory is compared with the prediction output of the oral postoperative bleeding risk prediction model at various historical time points, and the deviation value between the two is calculated. The sample segments whose deviation values ​​exceed the allowable error range are selected, and the sample segments and their corresponding true outcomes are labeled as high-value corrected samples. The high-value corrected samples are added to the original historical multimodal data sample set, and incremental training is performed based on the original model parameters to generate a new version of the oral postoperative bleeding risk prediction model with better prediction performance and complete the version iteration.

9. The oral surgery bleeding risk prediction and intervention system based on multimodal data fusion according to claim 8, characterized in that, The process of adding the high-value corrected samples to the existing historical multimodal data sample set, incrementally training based on the original model parameters, generating a new version of the oral postoperative bleeding risk prediction model with better prediction performance, and completing the version iteration includes: Load the network structure and weight parameters of the currently deployed oral postoperative bleeding risk prediction model from the model parameter server as the initial model for incremental training; The newly added high-value corrected samples are extracted from the high-value corrected sample library, which contains the labeled sample fragments and their corresponding true outcomes. The high-value corrected samples are mixed with a portion of randomly sampled samples from the historical multimodal data sample set to form a training batch for this incremental training, and the high-value corrected samples are ensured to account for a preset proportion in the batch. Based on freezing the weight parameters of all layers of the initial model except the last fully connected layer, the initial model is subjected to supervised incremental training using the training batch. During the training process, a smaller learning rate than that used in the original training is employed to fine-tune the initial model. After each training cycle, the fine-tuned model is evaluated using an independent validation set. If the product of recall and precision on the validation set does not improve for several consecutive cycles, training is terminated early. The final model parameters that have reached the termination condition are solidified to generate a new version of the oral postoperative bleeding risk prediction model. The version number, training time, and performance indicators of the new version of the oral postoperative bleeding risk prediction model are recorded in the model version management log to complete the version iteration.

Images