Blood transfusion demand and blood transfusion amount prediction model based on multi-task learning and establishment method

Abstract

The application discloses a blood transfusion demand and blood transfusion amount prediction model based on multi-task learning and a building method thereof. The building method comprises the following steps: obtaining historical perioperative data analysis scene data of a plurality of ATAAD data samples. A plurality of candidate variables are confirmed based on the historical perioperative data analysis scene data. The plurality of candidate variables are screened to obtain a plurality of target variables, which are used as inputs of a multi-task deep learning model. A multi-task deep learning model comprising a shared feature extraction layer, a blood transfusion demand prediction branch, a red blood cell resource consumption estimation branch and a massive hemorrhage risk prediction task branch is constructed. A training set, a validation set and a test set are constructed, the multi-task deep learning model is trained by using the training set, and the classification performance and the regression performance of the multi-task learning model are evaluated. After the model training is completed, the multi-task learning model for quantitatively analyzing and outputting blood transfusion related risk features is obtained.

Description

Technical Field

[0001] This invention relates to the field of information processing technology, and in particular to a model and method for predicting blood transfusion demand and volume based on multi-task learning. Background Technology

[0002] Acute type A aortic dissection (ATAAD) is an extremely high-risk cardiovascular emergency characterized by rapid onset, rapid progression, and high mortality. Data samples typically require a short period of time for surgical procedures centered on ascending aortic replacement, aortic arch repair, and proximal descending aortic reconstruction (such as the Sun procedure), often performed under complex cardiopulmonary bypass conditions including deep hypothermic circulatory arrest.

[0003] Despite improvements in surgical techniques and perioperative care, massive bleeding and the resulting demand for red blood cell (RBC) transfusions remain major challenges, leading to increased morbidity, mortality, and resource utilization. During ATAAD data collection, RBC transfusion was one of the most common remedial measures. Over-transfusion can increase the risk of infection, lung injury, coagulation disorders, and long-term poor outcomes, while under-transfusion can lead to insufficient tissue perfusion, circulatory instability, and postoperative organ dysfunction.

[0004] For high-risk ATAAD procedures, there is currently a lack of technology capable of simultaneously predicting transfusion events and transfusion volume. Therefore, technological innovation is needed to provide automated electronic decision-making for blood management, facilitating understanding, analysis, and verification by hospital blood managers, and assisting them in blood allocation, planning, and decision-making. Summary of the Invention

[0005] To address the problem in the existing technology that there is a lack of technical means to uniformly model multi-source historical perioperative data and simultaneously output transfusion-related risk characteristics and quantitative information on red blood cell resource consumption for high-risk ATAAD surgery, the purpose of this invention is to provide a transfusion demand and transfusion volume prediction model and establishment method based on multi-task learning.

[0006] In existing perioperative information systems, the data types involved in blood management are complex and from diverse sources. There are correlations and redundancies among different data characteristics. Traditional single-task modeling methods cannot simultaneously characterize the risk of transfusion-related events and the characteristics of blood resource consumption, resulting in insufficient stability and consistency of data analysis results, which is also not conducive to subsequent information understanding, verification and management.

[0007] To address the aforementioned issues, this invention proposes a method for establishing a transfusion-related data analysis model based on multi-task learning. By performing feature screening and joint modeling on perioperative related data records, the method achieves simultaneous analysis and output of transfusion-related risk characteristics and quantitative results of red blood cell resource consumption, thereby improving the systematicness, stability, and interpretability of perioperative blood management-related data analysis.

[0008] The objective of this invention is achieved through the following technical solution: In a first aspect, the present invention provides a method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning, the method comprising the following steps: Analyze scenario data by acquiring historical perioperative data from multiple ATAAD data samples; Based on the analysis of the historical perioperative data, multiple candidate variables were identified. Multiple candidate variables are filtered to obtain multiple target variables, which are used as inputs to a multi-task deep learning model; A multi-task deep learning model is constructed, which includes a shared feature extraction layer, a blood transfusion demand prediction branch, a red blood cell resource consumption estimation branch, and a major hemorrhage risk prediction task branch. Construct training, validation, and test sets. Use the training set to train the multi-task deep learning model, use the validation set to select hyperparameters, and use the test set to evaluate the classification and regression performance of the multi-task learning model. After the model training is completed, a multi-task learning model is obtained for quantitative analysis of transfusion-related risk characteristics.

[0009] In conjunction with the first aspect, the present invention also provides a first preferred embodiment of the first aspect, wherein when performing feature selection on the plurality of candidate variables, a two-stage feature selection method combining the minimum redundancy maximum correlation (mRMR) feature selection algorithm and the Boruta feature selection algorithm is adopted to reduce feature redundancy and enhance the stability of the model input features.

[0010] In conjunction with the first aspect, the present invention also provides a second preferred embodiment of the first aspect, wherein the target variables include gender, cardiopulmonary bypass (CPB) time, European Score I score, average tissue oxygen delivery (DO2-average), minimum D / V ratio, preoperative hematocrit-1, minimum bladder temperature, fibrinogen, activated partial thromboplastin time, prothrombin time, type of surgery, lowest intraoperative blood pressure, and peak lactate level.

[0011] In conjunction with the first aspect, the present invention also provides a third preferred embodiment of the first aspect, wherein the shared feature extraction layer of the multi-task learning model includes at least two fully connected network layers, and uses the LeakyReLU activation function and Dropout regularization for feature mapping to improve the generalization ability of the model.

[0012] In conjunction with the first aspect, the present invention also provides a fourth preferred embodiment of the first aspect, wherein the multi-task learning model is optimized using a weighted composite loss function, the weighted composite loss function including a transfusion-related event classification loss term and a red blood cell resource consumption regression loss term, and each loss term is weighted and combined through weight parameters to achieve collaborative optimization among multiple tasks. The multi-task learning model is optimized using a weighted composite loss function, the weighted composite loss function including a transfusion-related event classification loss term and a red blood cell resource consumption regression loss term; The weighted composite loss function The mathematical expression is: ; In the formula, N is the number of samples; First item = The mean squared error of the regression of red blood cell (RBC) volume is represented by, where, For weight parameters, =0.69; This represents the actual RBC value. This represents the predicted RBC value; Second item = The binary cross-entropy represents the classification of blood transfusions, where, For weight parameters, =0.69; Indicates the true blood transfusion label, This represents the probability value of a transfusion-related event. Third item The loss function representing the prediction of major bleeding risk is handled using Focal Loss. The mathematical expression for Focal Loss is as follows: ,in, Here, α is the probability value for predicting massive bleeding, α is the class weighting factor, α = 0.25, and γ is the focusing parameter, γ = 2.0.

[0013] In conjunction with the first aspect, the present invention also provides a fifth preferred embodiment of the first aspect, specifically, the blood transfusion demand prediction result of the blood transfusion demand and blood transfusion volume prediction model is generated by applying a Sigmoid activation function to the output of the blood transfusion demand prediction branch, and then the probability threshold is set to 0.5 to obtain a binary classification result.

[0014] In conjunction with the first aspect, the present invention also provides a sixth preferred embodiment of the first aspect, specifically, training a multi-task deep learning model using a training set, specifically employing a three-stage training strategy, including: Single-task pre-training was performed on the blood transfusion demand prediction branch, the red blood cell resource consumption estimation branch, and the major hemorrhage risk prediction task branch, respectively. The branches for predicting thawed blood transfusion demand, estimating red blood cell resource consumption, and predicting the risk of massive hemorrhage are jointly trained using fixed weights. The weights of each blood transfusion demand prediction branch, red blood cell resource consumption estimation branch, and major hemorrhage risk prediction task branch are dynamically adjusted using a dynamic weight adjustment strategy.

[0015] In conjunction with the first aspect, the present invention also provides a seventh preferred embodiment of the first aspect, specifically, the dynamic weight update formula of the dynamic weight adjustment strategy is as follows: ; in, Let be the weight of task a in the t-th round of training. The performance deviation of task a. For the learning rate, and When the value is greater than zero, the task weight decreases. When the value is less than zero, the task weight increases.

[0016] In conjunction with the first aspect, the present invention also provides an eighth preferred embodiment of the first aspect, specifically, the method further includes an enhancement step for the multi-task learning model, comprising: The historical perioperative data is used to identify the data acquisition stage and output several data acquisition stage labels. Based on the data acquisition stage labels, select the data sub-regions or feature sets corresponding to the data acquisition stage and map them as regions of interest; Multimodal data, including image features, operational features, and physiological indicators, are extracted from the region of interest. Consistency analysis and conflict detection are performed on the extracted multimodal features. When high-risk data is identified, the corresponding scenario sub-process is triggered. By combining scene identifiers and multimodal data from scene sub-processes, the multi-task deep learning model is enhanced by adjusting the feature weights or task weights within the model, and the enhanced multi-task deep learning model is output.

[0017] In conjunction with the first aspect, the present invention also provides a ninth preferred embodiment of the first aspect, wherein the method of establishing further includes: A backbone time series prediction model is constructed, which takes perioperative baseline data as input and outputs baseline prediction results. The credibility of the basic prediction results is calculated, and the credibility result is output. The computing power allocation decision is executed based on the credibility result. The computing power allocation decision includes calling a multi-task deep learning model to perform quantitative analysis and output of transfusion-related risk characteristics when the credibility value is low and the risk result is high.

[0018] In conjunction with the first aspect, the present invention also provides a tenth preferred embodiment of the first aspect, which calculates the credibility of the basic prediction results, specifically including: The credibility of the basic prediction results is calculated, and the credibility value is output. The credibility value includes the prediction stability index within the time window, the internal uncertainty index of the model, and the historical sample coverage index. A lightweight phased risk screening is performed on the reliability of the prediction, and the risk results are output.

[0019] Secondly, the present invention also provides a transfusion demand and transfusion volume prediction model based on multi-task learning. The model is established by the above method and can perform joint analysis and output of transfusion-related risk characteristics and quantitative information on red blood cell resource consumption based on perioperative related data records.

[0020] Compared with the prior art, the present invention has at least the following beneficial effects: This invention uses a multi-task learning framework to uniformly model perioperative related data, enabling simultaneous analysis and output of transfusion-related risk characteristics and quantitative results of red blood cell resource consumption, thus avoiding the problem of information fragmentation in traditional single-task models. By introducing a shared feature extraction layer and a task branching structure, the model's ability to express the features of complex historical perioperative data was improved, and the stability and consistency of the analysis results were enhanced. By employing a two-stage feature selection and a weighted composite loss function, the risk of feature redundancy and model overfitting is effectively reduced, and the generalization performance of the model on different datasets is improved. The model output of this invention is a standardized risk probability value and a continuous quantitative result, which facilitates subsequent information understanding, analysis, verification and management, and is applicable to data processing and information analysis scenarios related to perioperative blood management. Attached Figure Description

[0021] Figure 1 This is a flowchart illustrating a method for constructing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to the present invention. Figure 2 This is a model architecture diagram of a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to the present invention. Figure 3 This invention provides a performance evaluation of a multi-task learning-based blood transfusion demand and transfusion volume prediction model in blood transfusion demand prediction. Figure 4 This invention relates to the regression performance of a multi-task learning-based blood transfusion demand and transfusion volume prediction model in predicting transfusion volume. Figure 5 Shap was used to perform interpretability analysis of a multi-task deep learning model for predicting blood transfusion demand and volume based on multi-task learning in this invention. Figure 6 This invention presents the performance of a multi-task learning-based blood transfusion demand and transfusion volume prediction model in predicting the blood transfusion demand of multiple cohorts. Detailed Implementation

[0022] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0023] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0024] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor devices and / or microcontroller devices.

[0025] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0026] The user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this manual are all information and data authorized by the user or fully authorized by all parties. The collection, use and processing of related data shall comply with the relevant laws, regulations and standards of the relevant regions, and corresponding operation portals shall be provided for users to choose to authorize or refuse.

[0027] For example, all relevant data mentioned later in the embodiments of this specification, such as historical perioperative data, are derived from anonymized datasets that have been explicitly authorized by the data subject or reviewed and approved for use. The data collection, storage, and processing processes strictly comply with relevant laws and regulations to ensure the legality, compliance, and security of data use.

[0028] For high-risk ATAAD procedures, there is currently a lack of technology that can simultaneously predict transfusion events and transfusion volume. Therefore, accurately predicting whether a data sample needs a transfusion and how much to transfuse is of great significance for historical perioperative data analysis.

[0029] To this end, this invention provides a blood transfusion demand and transfusion volume prediction model and method based on multi-task learning. By constructing a blood transfusion demand and transfusion volume prediction model based on multi-task deep learning, the blood transfusion decision-making shifts from relying on experience to being data-driven, realizing the integration and intelligent prediction of blood transfusion events and blood transfusion dosage. This can effectively optimize the allocation of blood resources and improve the overall management level of the perioperative period, and has important historical perioperative data analysis scenarios and application value.

[0030] The transfusion demand and transfusion volume prediction model constructed in the embodiments of this specification is essentially a data analysis and decision support tool. This model analyzes and calculates historical perioperative data to generate predictive information and risk quantification assessment reports, falling under the category of information processing and methods. The model itself does not possess, and will not perform, any form of clinical intervention.

[0031] Example 1 like Figure 1 As shown, Figure 1 This is a flowchart of a method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to the present invention. The method can be executed by a PC, which can be implemented in hardware and / or software form, such as a PC and an application deployed on the PC.

[0032] like Figure 1 As shown, the method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning includes the following steps: Step S110: Obtain historical perioperative data analysis scenario data from multiple ATAAD data samples.

[0033] In one example, before establishing a model for predicting transfusion demand and volume, it is necessary to first collect historical perioperative data analysis scenario data of a large number of acute Stanford type A aortic dissection (ATAAD) data samples.

[0034] For example, historical perioperative data analysis scenarios cover the entire process of ATAAD data samples from preoperative to postoperative periods, including but not limited to the data samples' age, gender, height, weight, underlying diseases, data collection process-related indicators (e.g., cardiopulmonary bypass time, data collection time, etc.), intraoperative monitoring indicators (e.g., blood gas analysis, urine output, etc.), and postoperative recovery status. Data sources can include the hospital's electronic medical record system, surgical records, laboratory test reports, etc.

[0035] Understandably, historical perioperative data analysis is fundamental for building predictive models. This data contains various potential influencing factors related to the transfusion needs and volume of the data samples. By collecting comprehensive and accurate historical perioperative data analysis, rich information can be provided for subsequent feature selection and model training, thereby improving the model's predictive performance and generalization ability.

[0036] In one specific implementation, this embodiment uses data samples from patients undergoing surgery for acute type A aortic dissection (ATAAD) as the research object, collecting multimodal historical perioperative data analysis scenario parameters. The historical perioperative data analysis scenario data includes, but is not limited to: demographic information, preoperative laboratory indicators, cardiopulmonary bypass (CPB) operating parameters, oxygen supply and perfusion indicators, surgical procedure information, body temperature, and metabolic status.

[0037] In a typical embodiment, the data comes from a continuous data sample of over 300 cases undergoing Sun surgery at the same center. The data is chronologically ordered and accurately reflects the surgical procedure and changes in the status of the data samples. The acquired data undergoes standardized formatting to remove missing records that cannot be used for model training.

[0038] Step S120: Based on the historical perioperative data analysis scenario data, identify multiple candidate variables.

[0039] In one example, from the collected historical perioperative data analysis scenario data, several candidate variables that may be related to transfusion needs and transfusion volume are initially screened. The selection of these candidate variables should be based on experience in historical perioperative data analysis scenarios and medical research, for example, preoperative hemoglobin levels, intraoperative blood loss, cardiopulmonary bypass time, surgical type, and underlying diseases of the data samples. The number of candidate variables can be adjusted according to the completeness and complexity of the data, but should cover as many factors as possible that may affect transfusion decisions.

[0040] Understandably, identifying candidate variables aims to narrow down the data pool, focusing on factors that may significantly impact transfusion demand and volume. Preliminary screening of candidate variables reduces data noise and improves the efficiency and accuracy of model training. Furthermore, appropriate candidate variable selection ensures the model captures key transfusion-related information, providing a solid foundation for subsequent feature selection and model building.

[0041] Step S130: Filter multiple candidate variables to obtain multiple target variables, which are used as inputs to a multi-task deep learning model.

[0042] In one example, multiple candidate variables are screened, specifically using a two-stage screening process combining the mRMRe feature selection algorithm and the Boruta feature selection algorithm. To avoid feature redundancy and collinearity that could lead to model instability, this embodiment employs a two-stage feature selection mechanism: (1) First stage: mRMRe (minimum redundancy maximum correlation) screening, used to eliminate collinear variables and retain the features that are most relevant to blood transfusion needs but are not redundant.

[0043] (2) Second stage: Boruta algorithm screens robust feature selection methods based on random forest, and further extracts features with real predictive value from candidate variables.

[0044] In one specific implementation, the phase feature screening mechanism ultimately retains 13 target variables, including: gender, cardiopulmonary bypass (CPB) time, European Score I score, average tissue oxygen delivery (DO2-average), minimum D / V ratio, preoperative hematocrit-1 (Hct-1), minimum bladder temperature, fibrinogen, activated partial thromboplastin time (APPT), prothrombin time, surgical type, intraoperative lowest blood pressure, and peak lactate level. These variables are used as the final input features of the multi-task model.

[0045] Understandably, a hybrid feature selection strategy is employed to further screen the initially identified candidate variables. Specific methods may include the Minimum Redundancy Maximum Relevance (mRMRe) algorithm and the Boruta algorithm. The mRMRe algorithm maximizes the correlation between features and the target variable while minimizing redundancy between features, thus selecting features with high predictive value and that are independent of each other. The Boruta algorithm, based on feature importance evaluation using random forests, verifies the importance of each feature by comparing it with randomly generated "shadow features." Finally, the results of both algorithms are combined to determine a set of target variables as input to the multi-task deep learning model.

[0046] Ultimately, the target variables for the output include gender, cardiopulmonary bypass (CPB) time, European Score I score, average tissue oxygen delivery (DO2-average), minimum D / V ratio, preoperative hematocrit-1 (Hct-1), minimum bladder temperature, fibrinogen, activated partial thromboplastin time (APLT), prothrombin time, type of surgery, lowest intraoperative blood pressure, and peak lactate level.

[0047] Step S140: Construct a multi-task deep learning model, which includes a shared feature extraction layer, a blood transfusion demand prediction branch, a red blood cell resource consumption estimation branch, and a major hemorrhage risk prediction task branch.

[0048] In this technique, the constructed multi-task deep learning model adopts a hierarchical structure, including a shared feature extraction layer, a transfusion demand prediction branch, a red blood cell resource consumption estimation branch, and a major hemorrhage risk prediction task branch. The shared feature extraction layer is responsible for extracting general feature representations from the input target variables. These feature representations capture common information related to transfusion demand and transfusion volume. Subsequently, the model is divided into three task-specific branches: the transfusion demand prediction branch is used for binary classification tasks, predicting whether a data sample needs a transfusion; the red blood cell resource consumption estimation branch is used for regression tasks, predicting the amount of transfusion required for a data sample; and the three branches share the output of the feature extraction layer but independently perform task-specific predictions.

[0049] For example, the multi-task deep learning structure includes a shared feature extraction layer + a transfusion demand prediction branch + a red blood cell resource consumption estimation branch + a major hemorrhage risk prediction task branch, to achieve joint learning of classification tasks (whether to transfuse) and regression tasks (transfusion amount): (1) Shared Representation: This layer consists of a multi-layer fully connected network used for high-dimensional representation learning of the seven input features. LeakyReLU activation, Dropout, and batch normalization are employed to enhance the model's generalization ability.

[0050] (2) Transfusion demand prediction branch (classification branch): Receives the shared feature representation and outputs the transfusion probability. The Sigmoid activation function is used, and the loss is the binary cross-entropy.

[0051] (3) Red blood cell resource consumption estimation branch (regression branch): It also receives the shared feature representation and outputs the red blood cell transfusion volume (mL) after passing through several fully connected layers. The output layer is linearly activated, and the loss is the mean squared error (MSE).

[0052] (4) Major Hemorrhage Risk Prediction Branch: Based on the output of the shared feature extraction layer, after processing by the residual fully connected network, the output is the probability of major hemorrhage activated by Sigmoid. For samples ∈[0,1], Focal Loss is used to handle sample imbalance, and hard sample mining techniques are combined to improve the model's learning ability for a small number of positive samples.

[0053] In one specific embodiment, the shared feature extraction layer of the multi-task deep learning model includes a first shared fully connected layer and a second shared fully connected layer connected in sequence. Together, they constitute a shared feature extraction network for extracting high-dimensional nonlinear feature representations from multiple input target variables.

[0054] Specifically, each shared fully connected layer uses the LeakyReLU activation function to avoid the "neuron death" problem caused by traditional ReLU in the low gradient region; after each layer, a Dropout regularization module is set, with a random deactivation ratio of 0.2 to 0.5, to reduce model overfitting and improve generalization ability; through the step-by-step mapping of the above two shared feature extraction layers, the original input features are transformed into deep feature vectors that can serve both classification and regression tasks.

[0055] In a preferred embodiment, the multi-task learning model is optimized using a weighted composite loss function, which includes a loss term for transfusion-related event classification and a loss term for red blood cell resource consumption regression. The weighted composite loss function The mathematical expression is: ; In the formula, N is the number of samples; First item = The mean squared error of the regression of red blood cell (RBC) volume is represented by, where, For weight parameters, =0.69; This represents the actual RBC value. This represents the predicted RBC value; Second item = The binary cross-entropy represents the classification of blood transfusions, where, For weight parameters, =0.69; Indicates the true blood transfusion label, This represents the probability value of a transfusion-related event. Third item The loss function representing the prediction of major bleeding risk is handled using Focal Loss. The mathematical expression for Focal Loss is as follows: ,in, Here, α is the probability value for predicting massive bleeding, α is the class weighting factor, α = 0.25, and γ is the focusing parameter, γ = 2.0.

[0056] In one specific implementation, the blood transfusion demand prediction result of the blood transfusion demand and blood transfusion volume prediction model is generated by applying the Sigmoid activation function to the output of the blood transfusion demand prediction branch, and then the probability threshold is set to 0.5 to obtain a binary classification result.

[0057] Understandably, in one specific embodiment, the transfusion demand prediction branch in the transfusion demand and transfusion volume prediction model is used to output a prediction result of whether the target data sample needs a transfusion. This branch adopts a binary classification neural network structure, and its output layer is set with a Sigmoid activation function to map the real values ​​output by the model to the interval [0, 1], thereby generating a continuous value representing the "probability of needing a transfusion".

[0058] For example, the transfusion demand prediction branch first receives a deep feature representation from a shared feature extraction layer, then outputs a one-dimensional prediction value z after passing through at least one fully connected layer. This prediction value is mapped using a sigmoid activation function: Where p represents the probability that a data sample needs a blood transfusion. To obtain the final binary classification decision, this embodiment sets the probability threshold to 0.5, that is: when p≥0.5, the output prediction result is that a blood transfusion is needed; when p<0.5, the output prediction result is that a blood transfusion is not needed.

[0059] Step S150: Construct training, validation, and test sets. Use the training set to train the multi-task deep learning model, use the validation set to select hyperparameters, and use the test set to evaluate the classification and regression performance of the multi-task deep learning model.

[0060] In practice, the collected data samples are divided into training, validation, and test sets according to chronological order or randomization. The training set is used to train the multi-task deep learning model, adjusting the model's parameters by optimizing the loss function. The validation set is used to evaluate the model's performance under different hyperparameter settings, thereby selecting the optimal hyperparameter combination. The test set is used to finally evaluate the model's classification and regression performance, ensuring that the model has good generalization ability on unseen data. During training, methods such as cross-validation can be used to further improve the model's stability and reliability.

[0061] Specifically, to align with real-world application scenarios for analyzing and predicting historical perioperative data, this embodiment divides the data into three sets based on the data collection time: training set: 60%, validation set: 20%, and test set: 20%. This avoids the risk of "future information leakage" caused by random data splitting.

[0062] Specifically, during training, the optimizer used is Adam. The learning rate is set to 1e−4 to 1e−2 using grid search. The batch size is 32. Early stopping is used to prevent overfitting, and the optimal hyperparameters are selected on the validation set.

[0063] In practical implementation, the performance evaluation uses the following metrics for classification: AUC, accuracy, recall, and sensitivity. For regression, the metrics are: R², MAE, and RMSE. In typical implementation results: AUC ≈ 0.93 for classification tasks, and R² ≈ 0.60 for regression tasks.

[0064] In another preferred embodiment, the multi-task deep learning model is trained using a training set, specifically employing a three-stage training strategy, including: Step S151: Perform single-task pre-training on the blood transfusion demand prediction branch, the red blood cell resource consumption estimation branch, and the major hemorrhage risk prediction task branch, respectively.

[0065] Single-task pre-training specifically involves: freezing all other branches and training only the transfusion demand prediction branch; optimizing classification performance using the training set to enable the model to learn the mapping relationship between perioperative characteristics and transfusion demand; freezing all other branches and training only the red blood cell resource consumption estimation branch; optimizing regression performance using the training set to enable the model to accurately predict transfusion volume; and freezing all other branches and training only the major hemorrhage risk prediction branch; using Focal Loss to address sample imbalance and combining it with hard sample mining techniques to enable the model to effectively identify high-risk major hemorrhage patients.

[0066] Step S152: The thawing transfusion demand prediction branch, the red blood cell resource consumption estimation branch, and the major hemorrhage risk prediction task branch are jointly trained using fixed weights.

[0067] By unfreezing all branches, the three-task model is trained simultaneously. During joint training, fixed weights are assigned to the three-task branches (e.g., transfusion volume prediction λ1=0.4, transfusion demand prediction λ2=0.3, and major hemorrhage risk prediction λ3=0.3). The parameters of the shared feature extraction layer and each branch are iteratively optimized using the training set. Hyperparameters are selected using the validation set to optimize the overall performance of the three tasks, enabling the model to achieve good classification and regression metrics across different tasks.

[0068] Step S153: Use a dynamic weight adjustment strategy to dynamically adjust the weights of each blood transfusion demand prediction branch, red blood cell resource consumption estimation branch, and major hemorrhage risk prediction task branch.

[0069] By introducing a Dynamic Feature Weighting (DFW) strategy, the weights of the three task branches are adjusted in real time based on the performance changes of each task on the validation set. For example, the weights of lagging tasks are increased to enhance the model's learning effort for those tasks; the weights of leading tasks are appropriately reduced to avoid overfitting. This embodiment achieves a balance between the three task objectives through dynamic weight optimization, while ensuring high accuracy in predicting high-risk massive bleeding patients, transfusion needs, and transfusion volume.

[0070] Through the above training, the following results were achieved: prediction of major bleeding risk: AUC ≥ 0.85, recall ≥ 0.85; prediction of transfusion demand: AUC ≥ 0.90, F1-score ≥ 0.85; prediction of transfusion volume: R² ≥ 0.60, MAE < 100 mL. The three-stage training strategy combined with dynamic weight adjustment improved the prediction accuracy of high-risk patients by ≥ 3%, while maintaining the model's generalization ability and stability.

[0071] In some specific implementations, the dynamic weight update formula of the dynamic weight adjustment strategy is as follows: ;in, Let be the weight of task a in the t-th round of training. The performance deviation of task a. For the learning rate, and When the value is greater than zero, the task weight decreases. When the value is less than zero, the task weight increases.

[0072] Step S160: After the multi-task deep learning model is trained, a blood transfusion demand and blood transfusion volume prediction model is obtained.

[0073] In practice, after the model training is completed, a data analysis model based on multi-task learning can be obtained.

[0074] The model can receive a single historical perioperative data record as input and output corresponding data analysis results based on the input data, including: (1) Quantitative analysis values ​​corresponding to the probability of transfusion-related events; (2) Continuous numerical estimation results related to red blood cell resource consumption.

[0075] The output results can be used to quantitatively describe the transfusion-related risk characteristics of different data samples within a preset time window. The model supports feature contribution analysis based on the SHAP method to improve the interpretability of the model output results and facilitate the understanding, verification, and management of the analysis results.

[0076] Combination Figure 3 , Figure 3 The (ac) ROC curves show that the model has significant discriminative power on the training, validation, and test sets, and the corresponding AUC values ​​are labeled. The (df) confusion matrix quantifies the classification results of the three cohorts. The (gi) predicted probability distribution plot clearly presents the effective risk stratification between the transfusion and non-transfusion groups. The (jl) decision curve analysis shows that, within different threshold probability ranges, this model demonstrates better practicality and net benefit compared to the default strategy for historical perioperative data analysis scenarios.

[0077] Combination Figure 4 , Figure 4 The scatter plots (ac) show the predicted and actual transfusion volumes in the training (a), validation (b), and test (c) cohorts. Solid lines represent the regression fit between predicted and actual values, with the coefficient of determination (R²) and Pearson correlation coefficient (r) labeled for each cohort. (df) shows the distribution of the prediction residuals (predicted values ​​minus actual values) across the three cohorts, indicating the absence of significant systematic bias. (gi) The Brand-Altman plot assesses the agreement between predicted and actual values. The central solid line represents the mean difference (bias), while the upper solid and dashed lines represent the 95% agreement threshold.

[0078] like Figure 2 As shown, the present invention provides a blood transfusion demand and blood transfusion volume prediction model based on multi-task learning, which is obtained by the above-mentioned method for establishing a blood transfusion demand and blood transfusion volume prediction model based on multi-task learning.

[0079] In practice, the blood transfusion demand prediction results of the blood transfusion demand and blood transfusion volume prediction model are generated by applying the Sigmoid activation function to the output of the blood transfusion demand prediction branch, and then the probability threshold is set to 0.5 to obtain the binary classification result.

[0080] Combination Figure 5 and Figure 6 , Figure 5 To perform interpretability analysis of multi-task deep learning models using Shap Figure 5 The diagram (a, d) shows the global feature importance for predicting transfusion demand (a) and transfusion volume (d), sorted by mean absolute Shap value. (b, e) are Shap summary plots for the classification task (b) and regression task (e), showing the direction of each feature's influence on the model output. (c, f) are instance-level explanations of representative data samples presented in waterfall plots, illustrating how each feature contributes to the baseline values ​​to arrive at the final predictions of transfusion probability (c) and transfusion volume (f). Positive contributions (increased risk / volume) are shown in red, and negative contributions (decreased risk / volume) are shown in blue.

[0081] Example 2 The method described in this embodiment involves state identification and pattern analysis of perioperative related data. It is only used to enhance the data processing workflow and does not involve disease diagnosis, treatment decisions, or medical procedures for any individual. This second embodiment, based on the multi-task learning-based transfusion demand and transfusion volume prediction model described in the first embodiment, further provides a prediction enhancement method that combines Typical Surgical Pattern (TSP) identification. This method targets complex dynamic processes and enhances the prediction of the multi-task deep learning model by identifying the data acquisition stage, key anatomical structures, and multimodal risk characteristics, thereby improving the foresight and accuracy of transfusion demand and transfusion volume predictions.

[0082] The content described in this second embodiment is only a part of the embodiments of this application. Those skilled in the art can design other implementation schemes without departing from the technical solution of this invention, and these should also be included within the protection scope of this application.

[0083] Embodiment 2 of this application provides an enhanced method for predicting blood transfusion demand and volume based on key scenario identification. This embodiment, building upon the multi-task deep learning prediction model described in Embodiment 1, enhances the model's prediction process by introducing the identification of key stages and structural risk scenarios, thereby improving the accuracy and reliability of the prediction results in complex data acquisition environments. Obviously, the content described in this embodiment is only a part of the embodiments of this application. Other embodiments that can be designed by those skilled in the art without departing from the technical solution of this application should also be included within the protection scope of this application.

[0084] In this application embodiment, the key scene identification includes the stage determination of the data acquisition process and the binding of high-risk areas. By monitoring key operation stages and potentially high-risk anatomical structures, feature information related to bleeding risk is extracted, thereby assisting the multi-task deep learning model in predicting blood transfusion demand and transfusion volume.

[0085] Specifically, it includes the following steps: Step S210 Data Acquisition Stage Determination: Identify the data acquisition stage of the historical perioperative data and output several data acquisition stage labels.

[0086] In Embodiment 2 of this application, during the perioperative period, the staged data states in the historical perioperative data processing flow are identified, including but not limited to the clamp placement or release stage, suturing or anastomosis stage, rewarming stage, cardiopulmonary bypass stabilization stage, and complex multi-structure operation stage. These staged data states are distinguished based on timeline information, trends in cardiopulmonary bypass-related parameters, and preset rules, and are used to characterize data feature patterns within different time periods. In specific implementation, the method for determining the data acquisition stage can be as follows: (1) Time window method: Automatically infer the stage based on the start time of data collection and the time period already executed.

[0087] (2) Action recognition method: Identify the operation type by acquiring video or sensor signals.

[0088] (3) Parameter threshold method: the stage is determined based on changes in extracorporeal circulation flow, pressure and temperature.

[0089] This step outputs the current data acquisition stage label (Stage ID), providing a reference for subsequent ROI binding.

[0090] Step S220 Key structural region binding: Based on the data acquisition stage label, select the data sub-region or feature set corresponding to the data acquisition stage and map it as the region of interest.

[0091] In Embodiment 2 of this application, based on the data acquisition stage, the data sub-regions or feature sets related to that stage are mapped to Regions of Interest (ROIs) according to the stage-specific data status, for subsequent multimodal data feature extraction. The ROIs may include the aortic root, anastomosis, low pericardial cavity region, or other complex operational areas. By binding ROIs, the bleeding risk in key areas can be monitored in a focused manner, and target regions can be provided for multimodal feature extraction.

[0092] For example, a relevant high-risk anatomical region is predefined as the ROI for each data acquisition stage: (1) Clip release stage → Aortic root, proximal ascending aorta; (2) Suturing stage → Suturing the anastomosis and the low position of the pericardial cavity; (3) Rewarming stage → Global low-level tissue area; (4) Complex multi-structure operation stage → Multiple regions operating simultaneously In one implementation, the ROI can be bound by: data acquisition navigation system calibrating coordinates; real-time segmentation for medical imaging (CT, ultrasound, intraoperative photography); and automatic mapping combining data acquisition plans and operation records.

[0093] In this step, the ROI coordinate set is output for subsequent multimodal feature extraction.

[0094] Step S230 Multimodal feature extraction: Extract multimodal data including image features, operational features and physiological indicator features from the region of interest.

[0095] In Embodiment 2 of this application, multimodal data, including image features, operational features, and physiological indicators, are extracted for the bound Region of Interest (ROI). Image features may include bleeding velocity, bleeding area, and changes in oozing morphology; operational features may include action repeatability and operational stability; physiological indicators may include changes in hemoglobin, changes in tissue oxygen supply, and perfusion indicators. These multimodal features provide input information for subsequent scene determination and prediction enhancement.

[0096] In some examples, various image features can be obtained as follows: bleeding velocity (calculated through changes in consecutive image frames), bleeding area (based on image segmentation techniques), and oozing morphology (dripping blood, diffuse oozing, spurting blood). Blood detection within a region of interest (ROI) can be achieved using deep learning image segmentation models (such as the U-Net series). Specifically, the operational stability of operational features mainly includes the smoothness of the instrument trajectory and the amplitude of jitter.

[0097] In some specific implementations, various features are extracted and then standardized (Z-score or min-max) to form a vector X. ROI Input a multi-task model.

[0098] Step S240 Multimodal Consistency and Conflict Analysis: Perform consistency analysis and conflict detection on the extracted multimodal features. When high-risk data is identified, trigger the corresponding scenario sub-process.

[0099] In Embodiment 2 of this application, consistency analysis and conflict detection are performed on the extracted multimodal features to determine whether the information between different modalities matches. This step can identify potential abnormal bleeding states, providing a basis for triggering high-risk data patterns.

[0100] For example, this can be achieved by checking whether image features, operational characteristics, and physiological indicators are consistent. For instance, does a decrease in hemoglobin correspond to the amount of bleeding detected in the image? Does frequent manipulation correspond to an increased risk of local bleeding? Inconsistencies are then flagged to form an abnormal signal F. conflict .

[0101] Specifically, anomalous signals serve as risk-enhancing inputs to the model, triggering high-risk data patterns.

[0102] Step S250: When the preset data mode conditions are met, the corresponding high-risk data status indicator is triggered.

[0103] In Embodiment 2 of this application, based on the data acquisition stage, ROI characteristics, and multimodal analysis results, it is determined whether the preset triggering conditions for key high-risk data patterns are met. When the conditions are met, the corresponding scenario sub-process is triggered. Key high-risk data patterns include, but are not limited to, the following five categories: Type 1: Sudden bleeding scenario during clamp release or release phase: Triggered when the bleeding rate in the ROI area increases rapidly after clamp release or release.

[0104] Type 2: Structural leakage during suturing or anastomosis: Triggered when repeated adjustments are made to the suture area and accompanied by continuous low-speed bleeding.

[0105] Type 3: Bleeding scenarios related to insufficient coagulation function during the rewarming phase: triggered when multiple ROIs show slow-speed bleeding and physiological indicators show a declining trend in coagulation ability.

[0106] Type 4: Accumulated hidden bleeding related to extracorporeal circulation: Triggered when repeated bleeding occurs in low-lying ROIs and hemoglobin levels continue to decline.

[0107] Type 5: High-risk scenarios in complex multi-structured operation phases: triggered when multiple ROIs bleed simultaneously and operation stability decreases.

[0108] For example, based on the outputs of S210–S240, determine whether any key high-risk data pattern condition is met. Five preset scenario trigger conditions are provided (quantitative examples): Type 1: Bleeding rate at the ROI during the clamp placement phase > 50 mL / min; Type 2: Continuous bleeding during the suturing stage > 3 times, and the area > 5 cm²; Type 3: Multiple ROI bleeding during the rewarming phase + hemoglobin decrease >10 g / L / h; Type 4: Accumulated bleeding at low-level ROI during cardiopulmonary bypass >100 mL; Type 5: Complex multi-structure operation stage with ≥ 3 bleeding ROIs and operation stability score < 0.8; Based on the above identification, and according to the data acquisition stage, ROI characteristics, and multimodal analysis results, the Scene ID and related risk characteristics are output.

[0109] Step S260 Prediction Enhancement: Combine scene identifiers and multimodal data from the scene sub-process to enhance the multi-task deep learning model, adjust the feature weights or task weights within the model, and output the enhanced multi-task deep learning model.

[0110] In Embodiment 2 of this application, when any key high-risk data pattern is triggered, the scene identifier and related risk characteristics are input as enhancement signals to the multi-task deep learning model described in Embodiment 1. By adjusting the feature weights or task weights within the model through the enhancement signals, a forward-looking enhancement of the prediction of transfusion demand and transfusion volume is achieved, improving the prediction accuracy of the model in complex or abnormal data acquisition environments.

[0111] For example, the Scene ID is associated with the corresponding multimodal risk feature vector X. scene Input a multi-task deep learning model. Specifically, model enhancement strategies may include: (1) Feature weight adjustment: weighting the shared feature layer output according to the scenario risk. (2) Task weight adjustment: increasing the regression task weight α in high-risk scenarios. (3) Attention mechanism: adding attention weights to features within the ROI.

[0112] In practice, the enhanced feature vectors are used to update the classification output (whether a blood transfusion is needed) and the regression output (estimated blood transfusion amount).

[0113] Step S270: Output the prediction results.

[0114] In Embodiment 2 of this application, after prediction enhancement, the multi-task deep learning model outputs the predicted transfusion demand and transfusion volume of the target data sample. By combining key high-risk data pattern recognition, the prediction results can reflect the dynamically changing risk information during data collection, providing data support for decision-making in preoperative planning and historical perioperative data analysis scenarios.

[0115] By combining the output results with dynamic high-risk data patterns, real-time risks can be reflected, assisting in preoperative planning and decision-making based on historical perioperative data analysis.

[0116] It is understood that Example 2 is an enhancement method proposed based on the multi-task deep learning prediction model described in Example 1. Example 2 directly uses the historical perioperative data analysis scenario features selected in Example 1 as the basic input, and introduces the data acquisition stage, key anatomical area of ​​interest (ROI), and multimodal risk features as enhancement signals, thereby improving the prediction accuracy and foresight of the model in complex and dynamic surgical environments.

[0117] The two are logically closely linked: Implementation 1 provides global multi-task prediction capabilities, while Implementation 2 dynamically enhances the model through real-time high-risk data pattern perception, enabling the model to respond to bleeding risks at different stages and in different regions. This design allows the two to form a complete prediction system of "basic prediction model + dynamic scenario-enhanced prediction".

[0118] In the embodiments of this application, the combination of the two demonstrates significant technological innovation. First, by combining multi-task deep learning with the data acquisition phase and the identification of key anatomical regions, dynamic enhanced prediction is achieved, which has not yet been realized in the prior art. Through data acquisition phase determination, ROI binding, multimodal feature extraction, and consistency analysis, the model can predict transfusion demand and transfusion volume before or just after bleeding occurs, achieving prospective prediction.

[0119] Secondly, the innovative multimodal features and key high-risk data pattern triggering mechanism serve as enhanced signal input models, enabling dynamic adjustment of the weights of shared feature layers and task branches. In high-risk scenarios, the model becomes more sensitive to blood transfusion volume prediction, and the collaborative optimization of classification and regression tasks improves the model's accuracy and stability under complex data acquisition conditions.

[0120] Furthermore, the dual-layer prediction system formed by combining these two approaches not only improves prediction accuracy but also enhances interpretability and applicability to historical perioperative data analysis scenarios. The enhanced signal has a clear source, traceable to the surgical stage, key anatomical structures, and physiological indicators, which helps in intraoperative decision-making and preoperative planning. This method, which combines static baseline prediction with dynamic enhanced prediction, technically fills the gap in existing multi-task transfusion prediction models that cannot respond in real time to complex, high-risk data patterns.

[0121] Example 3 This embodiment decouples the functions and schedules computing resources of the historical perioperative data processing flow from the perspective of system architecture. This is used to improve the efficiency and stability of complex data analysis tasks under different computing power conditions, without involving any medical operations or treatment decisions.

[0122] In Example 1, a multi-task deep learning model was constructed that can simultaneously output transfusion demand judgment results and transfusion volume prediction results based on multi-dimensional historical perioperative data analysis scenario variables. In Example 2, a further technical idea was proposed to dynamically correct the prediction results by identifying key high-risk data patterns (TSPs). However, in actual applications of historical perioperative data analysis scenarios, the following problems still exist: (1) Computing power and stability constraints: When the prediction model is running throughout the high-risk data collection process, it may face computing power bottlenecks or stability problems caused by long-term operation; (2) Risk differences in high-risk data collection stages: The bleeding risk and prediction sensitivity differ significantly in different high-risk data collection stages, and a single model is difficult to maintain high accuracy throughout the entire process; (3) Short time window error concentration: Prediction error is often concentrated in a short period of time when the risk of bleeding changes rapidly; (4) Lack of credibility characterization: The credibility of a single prediction result is not quantified, which is not conducive to decision-making in the context of historical perioperative data analysis.

[0123] Based on the above problems, this embodiment three proposes a key scenario-triggered multimodal enhancement method based on the transfusion prediction backbone model. It decouples and refines the prediction process from the system architecture level to achieve real-time, high-precision, and forward-looking prediction.

[0124] Specifically, the key scenario-triggered multimodal enhancement method based on the transfusion prediction backbone model includes the following steps: Step S100: Basic data collection and prediction using the backbone time series prediction model.

[0125] In practical implementation, a backbone time series prediction model is first constructed. For example, the backbone time series prediction model can use LSTM or Temporal Fusion Transformer (TFT) for time series modeling, relying only on low-dimensional numerical features, suitable for real-time operation with low computing power. Specifically, step S100 includes: S101: During the operation of the data processing system, perioperative relevant data are periodically collected at a preset time step for input into the main time series prediction model.

[0126] In step S100, the system continuously collects intraoperative baseline data at fixed time steps during high-risk data acquisition, including extracorporeal circulation parameters (pump flow rate, body temperature, arterial blood pressure and their rate of change), vital signs and metabolic indicators (hemoglobin concentration, blood oxygen delivery and derived metabolic indicators), and intraoperative timeline information (extracorporeal circulation start time, elapsed time, and position of high-risk data acquisition process). After time alignment, normalization, and outlier suppression, the collected data is input into the transfusion prediction backbone model for dynamic modeling, generating the basic transfusion demand probability and basic transfusion volume prediction values, and outputting prediction stability indicators. The purpose of this step is to provide real-time, low-computing-power, and continuous baseline predictions, providing reliable backbone prediction support for subsequent multimodal augmentation analysis, and ensuring continuous monitoring of transfusion risk throughout the data acquisition process.

[0127] The basic data includes, but is not limited to: Extracorporeal circulation related parameters: pump flow rate, body temperature, arterial blood pressure and instantaneous rate of change, used to characterize the circulatory status and blood volume change trend of the data sample; Vital signs and metabolic indicators: hemoglobin concentration (Hb), oxygen delivery (DO2), and derived indicators related to metabolic state; Intraoperative timeline information: cardiopulmonary bypass initiation time, elapsed time, and relative position of data acquisition progress.

[0128] In some embodiments, before step S101, the steps may include: performing time alignment, normalization, and outlier suppression on data from different sources to map them to a unified time axis and eliminate dimensional differences and noise.

[0129] S102: Input the preprocessed basic data into the backbone time series prediction model to generate the basic transfusion demand probability p0, the basic transfusion volume prediction value r0, and the prediction stability or confidence index c0.

[0130] S103: Runs the backbone time series prediction model resident in a low-computing-power manner, without calling image, audio or anatomical structure analysis modules, and outputs basic prediction results.

[0131] Step S200: Predictive stability judgment and light-weight stage coarse screening.

[0132] In this step, the credibility of the basic prediction results is calculated, and the credibility result is output. Specifically, the calculation indicators include, but are not limited to: the magnitude and variance of the basic prediction value within the sliding time window; internal uncertainty indicators of the model, such as the width of the output distribution after multiple random samplings and the dispersion of attention weights; and the historical sample coverage indicator for the current time period. The above indicators are combined to output the credibility result Conf_score, which is used to evaluate the stability and reliability of the basic prediction results.

[0133] In step S200, the system assesses the credibility of the basic prediction results output in S100, including prediction stability indicators within the time window, internal model uncertainty indicators, and historical sample coverage indicators. Subsequently, the system performs a lightweight risk screening based on prediction credibility and the current data acquisition stage. By quickly determining prediction stability, model confidence, and risks at the data acquisition stage, the system can perform tiered evaluations of the stability, uncertainty, and sample coverage of the model output results to determine whether to invoke a more complex data analysis process. The purpose of this step is to optimize computational resource allocation: in stages where predictions are stable and low-risk, only the backbone model is used to complete the prediction; while in stages where predictions are uncertain or high-risk, it provides a decision-making basis for triggering subsequent innovative branches, thereby improving the overall efficiency of the system and reducing unnecessary computational burden.

[0134] In specific implementation, this step S200 includes: S201: Calculate the confidence level of the basic prediction results of the backbone time series prediction model, including: For each time step, the basic prediction results are mapped to the confidence index space, and the calculation includes: Prediction stability index within a time window: Based on the sliding window, the magnitude, variance, or amplitude of the predicted value change is calculated to measure the continuity and stability of the prediction in the short term; Internal uncertainty metrics: The width of the output distribution is evaluated through multiple random samplings, Dropout, or attention weight distribution, reflecting the model's confidence in the current input sample; Historical sample coverage index: Determines whether the input features of the current time period are sparse in the training sample distribution or deviate from the training data distribution, and is used to identify time periods where the model may face extrapolation risks; The above indicators are weighted or normalized to output an overall confidence value Conf_scoreConf, which is used to quantify the reliability of the basic prediction results.

[0135] In one specific implementation, the basic prediction results are stratified according to complexity from low to high, including: the first layer is rapid stability assessment; the second layer is model uncertainty assessment; and the third layer is historical sample coverage verification. When the prediction reliability is low, a lightweight risk screening stage is entered, which judges whether it is close to a high-risk stage based on the trend of extracorporeal circulation parameters, the timeline of the data collection stage, and the directionality of the basic prediction.

[0136] In specific implementation, the credibility calculation of the basic prediction results includes: The credibility of the basic prediction results is calculated and a credibility value is output. The credibility value includes the prediction stability index within the time window, the internal uncertainty index of the model, and the historical sample coverage index.

[0137] A lightweight phased risk screening is performed on the reliability of the prediction, and the risk results are output.

[0138] S202: Execute computing power allocation decisions based on credibility results, and perform tiered computing power scheduling: When the confidence value is high and the risk result is low, only the backbone model is used to output the basic prediction result, and the multi-task deep learning model is not triggered to perform quantitative analysis and output of transfusion-related risk characteristics. When the confidence value is low or the risk result is medium, the data collection frequency can be increased or the lightweight risk verification module can be enabled. When the confidence value is low and the risk result is high, the system triggers a multi-task deep learning model to perform quantitative analysis and output of transfusion-related risk characteristics, including prediction of major bleeding risk, prediction of transfusion demand, and prediction of transfusion volume. By employing the aforementioned computing power distribution strategy, real-time continuous basic predictions can be achieved under low computing power conditions, while ensuring that high-complexity analysis can be invoked as needed during high-risk phases, thereby enhancing prediction accuracy and foresight.

[0139] In a preferred embodiment, when the confidence value is low and the risk result is high, the multi-task deep learning model is triggered and a new branch analysis is created, proceeding to S300.

[0140] Step S300: Innovative branch triggering and data collection stage precise judgment.

[0141] In step S300, when the system determines that the prediction confidence is low and the stage is high-risk, it triggers an innovative branch to refine the data acquisition stage. This stage integrates information such as cardiopulmonary bypass parameters, arterial blood pressure trends, and intraoperative voice prompts, and uses a time-series model and rule-based fusion method to accurately determine the current high-risk data acquisition stage (Phase_t, such as the clip placement stage, suturing stage, rewarming stage, cardiopulmonary bypass stabilization stage, etc.). The purpose of this step is to provide reliable stage information for subsequent anatomical structure binding and key scenario determination, making multimodal analysis more targeted and ensuring a high degree of matching between the predicted enhancement signal and the data acquisition process.

[0142] In specific implementation, this step S300 includes: S301: After triggering the innovative branch, the system integrates information such as cardiopulmonary bypass parameters, arterial blood pressure trends, and intraoperative voice prompts, and uses time series models and rule fusion methods to determine the current fine data acquisition phase Phase_t, including the clamping phase, suturing phase, rewarming phase, and cardiopulmonary bypass stabilization phase.

[0143] S302: Phase_t serves as a key conditional variable for subsequent anatomical structure binding and key scene determination.

[0144] Step S400: Anatomical structure binding and high-sensitivity monitoring area determination.

[0145] In step S400, the system uses a deep learning model to perform 3D segmentation of key anatomical structures based on preoperative imaging (CTA / MRI) data, and spatially registers the preoperative images with intraoperative videos to identify highly sensitive monitoring areas (ROI_struct), such as the aortic root, anastomosis, and low pericardial cavity. Subsequent multimodal analysis is performed only within the ROI_struct, significantly reducing computational load and improving analytical focus. The purpose of this step is to accurately map complex surgical fields into key monitoring areas, ensuring that multimodal feature extraction and risk assessment focus on high-risk hemorrhage structures, thereby improving prediction accuracy.

[0146] In specific implementation, this step S400 includes: S401: Based on preoperative CTA or MRI data, a deep learning model is used to perform three-dimensional segmentation of key anatomical structures of the heart and major blood vessels.

[0147] S402: Spatial registration of preoperative images and intraoperative videos is performed to map anatomical structures into the surgical field images and determine the ROI_struct, a high-sensitivity monitoring region at the structural level, including the aortic root, anastomosis, and low pericardial cavity.

[0148] S403: Subsequent multimodal analysis is performed only within ROI_struct, reducing computational load and improving the focus of the analysis.

[0149] Step S500: Fine extraction of multimodal features under structural constraints.

[0150] In step S500, the system extracts multimodal features within the ROI_struct, including image features (slow-speed bleeding, blood accumulation, recurring patterns), motion features (surgical instrument / hand pose estimation, operational stability and repeatability analysis), audio features (suction device, electrosurgical unit, and intraoperative speech analysis), and physiological features (Hb change trends, DO2 fluctuations, and coagulation index modeling). Through refined modeling of multimodal information, the system can capture occult bleeding signals and abnormal operational signals during surgery. The purpose of this step is to improve the perception of complex bleeding risks and provide rich feature inputs for triggering key scenarios.

[0151] In practice, multimodal features are extracted in parallel within the ROI_struct, including: image features: slow bleeding, blood accumulation, recurring patterns, etc.; motion features: instrument / hand pose estimation, operation stability, repeatability analysis; audio features: suction device, electrosurgical sounds and intraoperative speech analysis; and physiological features: Hb change trends, DO2 fluctuations and coagulation-related index modeling.

[0152] Step S600: Multimodal conflict analysis and hidden risk modeling.

[0153] In step S600, the system inputs multimodal features into the modal conflict analysis module to construct a modal conflict tensor and analyze the consistency and inconsistency between visual, action, audio, and physiological signals. This step can identify hidden risk scenarios, such as images showing bleeding but vital signs temporarily stable, abnormal operation accompanied by local bleeding, or physiological indicators lagging behind visual or audio signals. The purpose of this step is to identify potential hidden bleeding risks through multimodal cross-validation, providing crucial evidence for determining key high-risk data patterns and enhancing the predictive foresight.

[0154] In specific implementation, this step S600 includes: S601: The system inputs multimodal features into the modal conflict analysis module, constructs a conflict tensor, and analyzes the consistency and inconsistency between modes.

[0155] S602: Identify hidden risks, including: images showing bleeding while vital signs are temporarily stable; abnormal operation coupled with local bleeding; physiological indicators lagging behind visual or audio signals.

[0156] Step S700: Determine the combination of key high-risk data patterns (TSP1–TSP5).

[0157] In step S700, based on the results of Phase_t, ROI_struct, and modal conflict analysis, the system performs conditional combination judgment on five key high-risk data patterns (TSP1–TSP5): TSP1: sudden bleeding during clamping / unclamping; TSP2: structural leakage during suturing / anastomosis; TSP3: bleeding related to mid-section coagulation insufficiency during rewarming; TSP4: occult bleeding accumulation related to extracorporeal circulation; TSP5: high-risk scenario coupled with multiple ruptures and action deviations. Each scenario outputs a risk confidence level, potential outbreak time, and scenario identifier, which are fed back to the main model as an enhancement signal. The purpose of this step is to systematically identify high-risk bleeding scenarios in different stages and structural regions, achieving dynamic and forward-looking enhancement of transfusion demand and volume prediction.

[0158] In practical implementation, the key high-risk data patterns (TSP1–TSP5) include: TSP1 is a sudden hemorrhage during the clamping / unclamping phase. Specifically, Phase_t = clamping / unclamping; ROI_struct = aortic root and proximal anastomosis. The condition for TSP1 is a rapid increase in the hemorrhage rate v(t). Output: Scenario identifier TSP1, sudden structural risk; backflow to S8 to dynamically adjust the weights of the main trunk model.

[0159] TSP2 identifies structural leakage during the suture / anastomosis phase. Specifically, Phase_t = suture / anastomosis; ROI_struct = suture site and suture course. TSP2 conditions: repeated tightening action + low-speed suture bleeding; output: TSP2, signs of structural closure failure. This is redirected to S8 to improve the sensitivity of transfusion volume prediction.

[0160] TSP3 indicates bleeding related to mid-stage rewarming coagulation insufficiency. Specifically, Phase_t = mid-stage rewarming; ROI_struct = multiple treated but exposed structures. TSP3 conditions: low-velocity bleeding in multiple ROIs + elevated DO2 fluctuations; output: TSP3, systemic hemostasis insufficiency; return to S8, increasing the predictive weight of transfusion / coagulation factors.

[0161] TSP4 is related to occult bleeding accumulation in cardiopulmonary bypass, specifically: Phase_t = stable cardiopulmonary bypass; ROI_struct = low pericardial cavity, aortic root. Conditions for TSP4: recurrent bleeding + decreasing Hb trend; Output: TSP4, occult cumulative risk. Return to S8 input delayed burst time τ.

[0162] TSP5 is a high-risk scenario involving multiple rupture points and coupled with action deviations. Specifically, Phase_t = multi-structure operation phase; ROI_struct = multiple high-risk structures. The conditions for TSP5 are: simultaneous bleeding in multiple ROIs + unstable actions + rapid increase in tissue oxygen supply; outputs include: TSP5, precursors to systemic loss of control; and return to S8 to trigger the highest Gate weight and intervention assessment.

[0163] Step S800: Innovation results feedback and prediction fusion.

[0164] In step S800, the system inputs the status identifiers and risk characteristics of the key scenarios identified in S700 (output from the high-complexity analysis process) into the backbone model via gating fusion. This is used to modulate the model's internal parameters or weight configuration to generate the final data analysis output. The purpose of this step is to fuse the enhanced signals from multimodal scenarios with the low-computing-power backbone prediction, achieving real-time and accurate transfusion prediction while maintaining computational efficiency, thus providing reliable data support for decision-making in historical perioperative data analysis scenarios.

[0165] In specific implementation, this step S800 includes: S801: The system feeds back the scene identifiers, risk characteristics, and potential outbreak times output by TSP1–TSP5 to the original backbone model.

[0166] S802: Dynamically modulates the basic prediction results of the backbone model (feature weights, task weights, etc.) through the gating fusion module, and outputs the final blood transfusion demand probability and blood transfusion volume prediction values.

[0167] S803: The final prediction results reflect the dynamic changes in risk during the operation, providing data support for decision-making in preoperative planning and historical perioperative data analysis scenarios.

[0168] The core of Example 3 is built upon the multi-task deep learning backbone model of Example 1. Example 1 provides real-time, low-computational-power prediction capabilities for transfusion demand and volume. By continuously collecting extracorporeal circulation parameters, vital signs, and data acquisition timeline information, it outputs a basic transfusion probability p0 and a predicted transfusion volume r0, and provides prediction stability or confidence indicators. Example 3 completely adopts the basic prediction logic of Example 1, ensuring continuous and stable low-computational-power prediction output throughout the entire data acquisition process, serving as the basic input for high-computational-power triggered analysis. This inheritance ensures the system's real-time performance and low-computational-power operation capabilities, while providing a reliable data foundation for subsequent enhancement modules.

[0169] Furthermore, Example 3 uses a feedback mechanism to feed the results of subsequent multimodal augmentation analysis back to the backbone model, enabling dynamic adjustment of the basic prediction results. This mechanism is built upon the basic prediction framework provided in Example 1, allowing the system to improve prediction accuracy while maintaining the continuous operation of the backbone model.

[0170] Understandably, the combination of these three technologies forms a complete technical system from low-computing-power real-time basic prediction (Example 1) → key high-risk data pattern enhancement (Example 2) → trigger-based multimodal fine enhancement (Example 3). By controlling the allocation of computing power through credibility judgment and stage risk screening, combined with anatomical structure binding, multimodal conflict analysis, and TSP1–TSP5 scenario determination, it achieves high-precision and forward-looking prediction of complex intraoperative massive hemorrhage, and dynamically adjusts the output of the backbone model. It innovatively solves the problem of balancing the utilization of multimodal information and prediction accuracy under computing power constraints, and significantly improves the scientificity and operability of transfusion decisions.

[0171] The above embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of protection of the present invention. For those skilled in the art, it will be understood that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning, characterized in that, The establishment method includes the following steps: Analyze scenario data by acquiring historical perioperative data from multiple ATAAD data samples; Based on the analysis of the historical perioperative data, multiple candidate variables were identified. Multiple candidate variables are filtered to obtain multiple target variables, which are used as inputs to a multi-task deep learning model; A multi-task deep learning model is constructed, which includes a shared feature extraction layer, a blood transfusion demand prediction branch, a red blood cell resource consumption estimation branch, and a major hemorrhage risk prediction task branch. Construct training, validation, and test sets. Use the training set to train the multi-task deep learning model, use the validation set to select hyperparameters, and use the test set to evaluate the classification and regression performance of the multi-task learning model. After the model training is completed, a multi-task learning model is obtained for quantitative analysis of transfusion-related risk characteristics.

2. The method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to claim 1, characterized in that: The feature selection includes a two-stage feature selection method using the minimum redundancy maximum correlation mRMR feature selection algorithm and the Boruta feature selection algorithm; The target variables include gender, cardiopulmonary bypass (CPB) time, European Score I score, average tissue oxygen delivery (DO2-average), minimum D / V ratio, preoperative hematocrit-1 (Hct-1), minimum bladder temperature, fibrinogen, activated partial thromboplastin time (APLT), prothrombin time, type of surgery, intraoperative lowest blood pressure, and peak lactate level.

3. The method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to claim 1, characterized in that: The multi-task learning model is optimized using a weighted composite loss function, which includes a loss term for transfusion-related event classification and a loss term for red blood cell resource consumption regression. The weighted composite loss function The mathematical expression is: ； In the formula, N is the number of samples; , , These are the weights of the three task branches; First item = The mean squared error of the regression of red blood cell (RBC) volume is represented by, where, For weight parameters, =0.69; This represents the actual RBC value. This represents the predicted RBC value; Second item = The binary cross-entropy represents the classification of blood transfusions, where, For weight parameters, =0.69; Indicates the true blood transfusion label, This represents the probability value of a transfusion-related event. Third item The loss function representing the prediction of major bleeding risk is handled using Focal Loss. The mathematical expression for Focal Loss is as follows: ,in, Here, α is the probability value for predicting massive hemorrhage, α is the class weight factor, α = 0.25; and γ is the focusing parameter, γ = 2.

0. The loss term for the classification of transfusion-related events is a binary cross-entropy loss, and the loss term for the regression of red blood cell resource consumption is a mean squared error loss. The loss terms are weighted and combined using weight parameters.

4. The method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to claim 1, characterized in that, The multi-task deep learning model is trained using the training set, specifically employing a three-stage training strategy, including: Single-task pre-training was performed on the blood transfusion demand prediction branch, the red blood cell resource consumption estimation branch, and the major hemorrhage risk prediction task branch, respectively. The branches for predicting thawed blood transfusion demand, estimating red blood cell resource consumption, and predicting the risk of massive hemorrhage are jointly trained using fixed weights. The weights of each blood transfusion demand prediction branch, red blood cell resource consumption estimation branch, and major hemorrhage risk prediction task branch are dynamically adjusted using a dynamic weight adjustment strategy.

5. The method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to claim 7, characterized in that: The dynamic weight update formula of the dynamic weight adjustment strategy is as follows: ; in, Let be the weight of task a in the t-th round of training. The performance deviation of task a. For the learning rate, and When the value is greater than zero, the task weight decreases. When the value is less than zero, the task weight increases.

6. The method for establishing a transfusion demand and transfusion volume prediction model based on multi-task learning according to claim 1, characterized in that, The method for establishing the model also includes enhancement steps for the multi-task learning model, including: The historical perioperative data is used to identify the data acquisition stage and output several data acquisition stage labels. Based on the data acquisition stage labels, select the data sub-regions or feature sets corresponding to the data acquisition stage and map them as regions of interest; Multimodal data, including image features, operational features, and physiological indicators, are extracted from the region of interest. Consistency analysis and conflict detection are performed on the extracted multimodal features. When high-risk data is identified, the corresponding scenario sub-process is triggered. By combining scene identifiers and multimodal data from scene sub-processes, the multi-task deep learning model is enhanced by adjusting the feature weights or task weights within the model, and the enhanced multi-task deep learning model is output.

7. The method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to claim 1, characterized in that, The establishment method also includes: A backbone time series prediction model is constructed, which takes perioperative baseline data as input and outputs baseline prediction results. The credibility of the basic prediction results is calculated, and the credibility result is output. The computing power allocation decision is executed based on the credibility result. The computing power allocation decision includes calling a multi-task deep learning model to perform quantitative analysis and output of transfusion-related risk characteristics when the credibility value is low and the risk result is high.

8. The method for establishing a blood transfusion demand and transfusion volume prediction model based on multi-task learning according to claim 7, characterized in that, The credibility calculation of the basic prediction results specifically includes: The credibility of the basic prediction results is calculated, and the credibility value is output. The credibility value includes the prediction stability index within the time window, the internal uncertainty index of the model, and the historical sample coverage index. A lightweight phased risk screening is performed on the reliability of the prediction, and the risk results are output.

9. A blood transfusion demand and transfusion volume prediction model based on multi-task learning, characterized in that, The model is established by the method described in any one of claims 1 to 8, and includes a shared feature extraction layer, a transfusion-related event probability analysis branch, a red blood cell resource consumption estimation branch, and a major hemorrhage risk prediction task branch.

Images