Method and device for predicting bleeding risk in spine surgery based on machine learning

Abstract

The invention discloses a machine learning-based intra-operative bleeding risk prediction method and device for spinal surgery. The machine learning-based intraoperative bleeding risk prediction method for spinal surgery comprises the following steps: acquiring information of a patient to be predicted; obtaining a trained hemorrhage risk prediction model; and inputting the information of the patient to be predicted into the trained massive hemorrhage risk prediction model so as to obtain a prediction result. According to the method, the high-precision prediction model is trained through large-scale patient data (including basic information, operation parameters, blood indexes and the like), so that the accuracy and the stability of intraoperative SBL risk prediction are improved. And a Cell Saver use suggestion based on a risk threshold is provided, and blood resource allocation is optimized. Clinical decision-making efficiency is improved through an automatic tool, blood transfusion related complications (such as infection and immune response) are reduced, and patient prognosis is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a spine surgery intraoperative bleeding risk prediction method based on machine learning and a spine surgery intraoperative bleeding risk prediction device based on machine learning. BACKGROUND

[0002] Current massive bleeding in spine surgery is an important factor affecting patient prognosis and blood transfusion management. The risk prediction model constructed by traditional single factor or multiple factor regression method has the problems of limited prediction index, single feature selection, and difficulty in considering individual differences. In recent years, machine learning technology has shown high prediction performance and stability in medical big data. Using multi-model integration and explanatory tools (such as SHAP value) to select and refine model features can provide support for accurate risk prediction. At the same time, the decision of using Cell Saver intraoperative autologous blood transfusion based on risk level can help optimize resource allocation and reduce unnecessary allogeneic blood transfusion operations, thereby improving the overall blood management level in clinic. There is a lack of a comprehensive method in the prior art that uses multi-model comparison, feature cross selection, and decision curve analysis to construct a spine surgery intraoperative massive bleeding risk prediction and corresponding Cell Saver intraoperative autologous blood transfusion auxiliary decision system. At present, there is no spine surgery intraoperative massive bleeding prediction model, and there is no network calculator based on such a model.

[0003] Therefore, it is desirable to have a technical solution to overcome or at least alleviate at least one of the above-mentioned deficiencies of the prior art.

[0004] SUMMARY

[0005] The purpose of the present application is to provide a spine surgery intraoperative bleeding risk prediction method based on machine learning to overcome or at least alleviate at least one of the above-mentioned deficiencies of the prior art.

[0006] To achieve the above-mentioned purpose, the present application provides a spine surgery intraoperative bleeding risk prediction method based on machine learning, which comprises:

[0007] Obtaining patient information to be predicted;

[0008] Obtaining a trained massive bleeding risk prediction model;

[0009] Inputting the patient information to be predicted into the trained massive bleeding risk prediction model to obtain a prediction result.

[0010] Optionally, the spine surgery intraoperative bleeding risk prediction method based on machine learning further comprises:

[0011] training the large hemorrhage risk prediction model;

[0012] The training of the large hemorrhage risk prediction model comprises:

[0013] obtaining a historical data set comprising historical basic data of a plurality of patients;

[0014] obtaining a plurality of different types of machine learning models;

[0015] training each machine learning model according to the historical data set and optimizing hyperparameters by 10-fold cross-validation, thereby obtaining each trained machine learning model;

[0016] cross-screening 12 independent prediction factors from the plurality of models by using SHAP value analysis to evaluate the importance of each model feature in the historical basic data;

[0017] selecting one of the trained machine learning models as the trained large hemorrhage risk prediction model.

[0018] Optionally, the historical basic data comprises age characteristics, gender characteristics, height characteristics, weight characteristics, diabetes characteristics, hypertension characteristics, coronary heart disease characteristics, cerebrovascular disease characteristics, liver and kidney function abnormalities characteristics, spinal degenerative disease characteristics, spinal deformity characteristics, spinal infection characteristics, spinal tumor characteristics, surgical method characteristics, fusion vertebrae number characteristics, surgical site characteristics, preoperative coagulation function characteristics, preoperative blood routine characteristics, intraoperative Cell Saver use characteristics, intraoperative and postoperative red blood cell transfusion volume characteristics;

[0019] The different types of machine learning models comprise Ranger models, cforest models, randomForest models, ridge models, and glmnet models.

[0020] Optionally, the historical basic data comprises original clinical data, and the original clinical data comprises age characteristics, gender characteristics, height characteristics, weight characteristics, BMI characteristics, diabetes characteristics, hypertension characteristics, coronary heart disease characteristics, cerebrovascular disease characteristics, liver and kidney function abnormalities characteristics, surgical method characteristics, fusion vertebrae number characteristics, surgical site characteristics, surgical estimated duration characteristics, intraoperative Cell Saver use characteristics, preoperative coagulation function characteristics, preoperative blood routine characteristics, preoperative liver and kidney function characteristics, preoperative electrolyte characteristics, intraoperative and postoperative red blood cell transfusion volume characteristics.

[0021] The preoperative coagulation function time series data comprises prothrombin time characteristics, activated partial thromboplastin time characteristics, fibrinogen characteristics, D-dimer characteristics, and international normalized ratio characteristics.

[0022] The three-dimensional anatomical parameters of the surgical site include the radius of curvature feature of the vertebral body, the diameter feature of the spinal canal, the intervertebral space height feature, the bone density feature, the vertebral artery running feature, the vascular density feature around the vertebral body, the intercostal artery or lumbar artery position feature, the epidural venous plexus density feature, the interspinous ligament thickness feature, the yellow ligament thickness feature, the epidural fat thickness feature, the surgical approach angle feature, the exposure range feature, and the instrument implantation depth feature.

[0023] The different types of machine learning models include a fractal adaptive Ranger model, a spatio-temporal convolution network, and a causal graph neural network.

[0024] Optionally, training the fractal adaptive Ranger model includes:

[0025] generating an engineering feature matrix from a historical data set and dividing the engineering feature matrix into an engineering feature matrix training set, an engineering feature matrix validation set, and an engineering feature matrix test set;

[0026] performing fractal adaptive feature weighting according to the engineering feature matrix training set to obtain a fractal adaptive weight matrix;

[0027] performing game theory-driven regularization tree construction according to the fractal adaptive weight matrix to obtain a regularization decision tree, a hyperparameter combination, and an iterative process performance curve;

[0028] performing cellular automaton hyperparameter optimization according to the hyperparameter combination and the iterative process performance curve to obtain a globally optimal hyperparameter combination.

[0029] Optionally, the fractal adaptive feature weighting according to the engineering feature matrix training set to obtain a fractal adaptive weight matrix includes:

[0030] performing fractal dimension calculation on each dimension of features in the engineering feature matrix to obtain a fractal dimension vector;

[0031] assigning a weight to each dimension of features in the fractal dimension vector through fractal dimension-driven weight adjustment to generate an adaptive weight matrix;

[0032] performing a surgical type enhancement operation on the adaptive weight matrix to obtain a fractal adaptive weight matrix.

[0033] Optionally, the game theory-driven regularization tree construction according to the fractal adaptive weight matrix to obtain a regularization decision tree, an initial hyperparameter combination, and an iterative process performance curve includes:

[0034] Step 101: Obtain a strategy probability distribution table;

[0035] Step 102: randomly selecting a hyperparameter combination according to the strategy probability distribution table, so as to obtain the hyperparameter combination of the current round;

[0036] Step 103: calculating the adjusted importance of the fractal adaptive weight matrix, so as to obtain the adjusted feature importance vector;

[0037] Step 104: generating a regularization coefficient according to the learning rate;

[0038] Step 105: calculating the split criterion of the adjusted feature importance vector according to the regularization coefficient, so as to obtain the split criterion value of each feature in the feature importance vector;

[0039] Step 106: constructing a regularization decision tree according to the feature importance vector and the split criterion value of each feature;

[0040] Step 107: verifying the validation set performance index of the decision tree model through the validation set;

[0041] Step 108: obtaining the return value of the current hyperparameter combination according to the validation set performance index;

[0042] Step 109: updating the strategy probability according to the return value of the current hyperparameter combination;

[0043] Step 110: repeating the steps 101 to 109 until the maximum iteration number is reached or the convergence condition is reached, and obtaining the iteration process performance curve and the hyperparameter combination after the maximum iteration number is reached or the convergence condition is reached.

[0044] Optionally, the cell automaton hyperparameter optimization according to the hyperparameter combination and the iteration process performance curve, so as to obtain the globally optimal hyperparameter combination comprises:

[0045] constructing an initial cell automaton state matrix according to the regularization decision tree;

[0046] setting a state evolution rule, and performing cell automaton state iteration evolution on the initial cell automaton state matrix, so as to obtain an updated cell automaton state matrix and a to-be-explored hyperparameter combination list after the iteration condition is reached, the to-be-explored hyperparameter combination list comprising at least one to-be-explored hyperparameter combination;

[0047] calculating the performance index of each to-be-explored hyperparameter combination in the to-be-explored hyperparameter combination list, and selecting the one with the best performance index as the globally optimal hyperparameter combination.

[0048] The application also provides a machine learning-based intraoperative bleeding risk prediction device for spinal surgery, comprising:

[0049] A to-be-predicted patient information acquisition module is configured to acquire to-be-predicted patient information.

[0050] A model acquisition module is configured to acquire a trained massive bleeding risk prediction model.

[0051] A prediction module is configured to input the to-be-predicted patient information into the trained massive bleeding risk prediction model, thereby obtaining a prediction result.

[0052] The application also provides a network calculator comprising the machine learning-based intraoperative bleeding risk prediction device for spinal surgery.

[0053] The application trains a high-precision prediction model through large-scale patient data (including basic information, surgical parameters, blood indicators, etc.), thereby improving the accuracy and stability of intraoperative SBL risk prediction. The application provides Cell Saver use recommendations based on risk thresholds, optimizes blood resource allocation. The application improves clinical decision-making efficiency through an automated tool, reduces blood transfusion-related complications (such as infection, immune response), and improves patient prognosis. BRIEF DESCRIPTION OF DRAWINGS

[0054] Figure 1 FIG. 1 is a flowchart of a machine learning-based intraoperative bleeding risk prediction method for spinal surgery according to an embodiment of the application.

[0055] Figure 2 FIG. 4 is a schematic diagram of ranger model prediction performance in an embodiment of the application.

[0056] Figure 3 FIG. 6 is a schematic diagram of a network calculator interface in an embodiment of the application. DETAILED DESCRIPTION

[0057] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be described in more detail below with reference to the accompanying drawings. In the drawings, the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The described embodiments are some, but not all, embodiments of this application. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain this application, and should not be construed as limiting this application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application. The embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0058] In the description of this application, it should be understood that the terms "center", "longitudinal", "lateral", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this application 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. Therefore, they should not be construed as limiting the scope of protection of this application.

[0059] Example 1:

[0060] like Figure 1 The machine learning-based methods for predicting intraoperative bleeding risk in spinal surgery, as shown, include:

[0061] Obtain information on patients to be predicted;

[0062] Obtain a trained model for predicting the risk of major bleeding;

[0063] The patient's information is input into a trained major bleeding risk prediction model to obtain the prediction results.

[0064] In this embodiment, the machine learning-based method for predicting intraoperative bleeding risk in spinal surgery further includes:

[0065] The massive bleeding risk prediction model is trained;

[0066] The training of the massive bleeding risk prediction model includes:

[0067] Obtain a historical dataset, which includes historical baseline data for multiple patients;

[0068] Obtain multiple different types of machine learning models;

[0069] According to the historical data set, each machine learning model is trained respectively and 10-fold cross-validation is used to optimize the hyperparameters, so as to obtain each trained machine learning model;

[0070] The SHAP value analysis is used to evaluate the importance of each model feature in the historical basic data, and 12 independent prediction factors are cross-screened from the plurality of models;

[0071] One of the trained machine learning models is selected as a trained massive hemorrhage risk prediction model.

[0072] In the embodiment, the historical basic data includes: age characteristics, gender characteristics, height characteristics, weight characteristics, diabetes characteristics, hypertension characteristics, coronary heart disease characteristics, cerebrovascular disease characteristics, liver and kidney function abnormality characteristics, spinal degenerative disease characteristics, spinal deformity characteristics, spinal infection characteristics, spinal tumor characteristics, surgical method characteristics, fusion vertebrae number characteristics, surgical site characteristics, preoperative coagulation function characteristics, preoperative blood routine characteristics, intraoperative Cell Saver use characteristics, intraoperative and postoperative red blood cell transfusion volume characteristics;

[0073] The different types of machine learning models include Ranger models, cforest models, randomForest models, ridge models, and glmnet models.

[0074] The application will be further described in detail below by way of example, and it can be understood that the example does not constitute any limitation on the application.

[0075] Data acquisition and preprocessing:

[0076] The clinical data of 3944 patients undergoing spinal surgery is obtained from the electronic medical record system of the medical institution, covering:

[0077] Patient basic information: age, gender, height, weight.

[0078] Comorbidities: diabetes, hypertension, coronary heart disease, cerebrovascular disease, liver and kidney function abnormalities.

[0079] Spinal disease diagnosis: degenerative disease, deformity, infection, tumor, etc.

[0080] Surgical parameters: surgical method (fusion surgery, decompression surgery), number of fused vertebrae, surgical site (cervical vertebrae, thoracic vertebrae, lumbar vertebrae).

[0081] Blood indicators: preoperative coagulation function (prothrombin activity, APTT, INR), preoperative blood routine (platelet, red blood cell).

[0082] Transfusion management: intraoperative Cell Saver usage, intraoperative and postoperative red blood cell transfusion volume.

[0083] Data preprocessing includes the following steps:

[0084] Cleaning outliers, imputing missing values (e.g., by mean imputation), encoding categorical variables as dummy variables (e.g., diagnosis type), normalizing numerical variables (e.g., age normalized to 0-1 range), and building a database that meets the requirements of machine learning.

[0085] Risk prediction model construction:

[0086] Dataset division: randomly divide the database into a training set (70%) and a test set (30%).

[0087] Model training: train 26 machine learning models on the training set, including Ranger, cforest, randomForest, ridge, glmnet, etc., and use 10-fold cross-validation to optimize hyperparameters (e.g., tree depth, learning rate).

[0088] Feature selection: use SHAP value analysis to evaluate the importance of each model feature, and cross-select 12 independent predictors from multiple models: age, height, weight, preoperative platelets, preoperative red blood cells, spinal lesion type (deformity, degeneration), spinal surgery type (thoracic, lumbar, cervical; vertebral resection, osteotomy), and the number of fused vertebrae.

[0089] Model evaluation: evaluate performance on the test set using ROC curve, AUC, and Brier score; confirm robustness on the external validation set.

[0090] Risk stratification: determine the risk prediction threshold as 0.53 based on decision curve analysis, with patients above 0.53 as high-risk patients and below 0.53 as low-risk patients.

[0091] Web calculator generation (see Figure 3 ):

[0092] Model serialization: save the trained Ranger model in a portable format (e.g., `.rds` file in R language).

[0093] Development framework: develop the web calculator based on the R Shiny framework and deploy it on the shinyapps.io platform.

[0094] Front-end design: the user interface includes text boxes (input age, platelets, etc.) and option selection controls (select diagnosis type, surgery site), with data format verification through JavaScript or R.

[0095] Background processing: Pre-process input data (e.g. dummy variable conversion, standardization), call serialized model to calculate SBL risk probability, generate Cell Saver treatment benefit report (e.g. reduce red blood cell transfusion units, improve postoperative hemoglobin level).

[0096] User experience: Support PC, tablet and mobile phone access, interface display risk probability (e.g. "Predicted SBL risk: 33.1%") and decision recommendation.

[0097] Risk calculation and treatment benefit evaluation:

[0098] Clinicians input patient data through the network calculator, and the system outputs in real time:

[0099] SBL risk probability (in percentage form).

[0100] Treatment benefit report: Quantify the effect of Cell Saver use (e.g. reduce intraoperative blood transfusion volume).

[0101] Decision recommendation: If the probability is greater than 0.53, recommend using Cell Saver; otherwise, suggest avoiding.

[0102] Output in text and graphical form (e.g. risk probability bar chart) for easy doctor decision-making.

[0103] Example 2:

[0104] In this embodiment, different models and different training methods are used by the present application, as follows:

[0105] In this embodiment, the historical basic data includes original clinical data, which includes age characteristics, gender characteristics, height characteristics, weight characteristics, BMI characteristics, diabetes characteristics, hypertension characteristics, coronary heart disease characteristics, cerebrovascular disease characteristics, liver and kidney function abnormalities characteristics, surgical method characteristics, fusion vertebrae number characteristics, surgical site characteristics, surgical estimated duration characteristics, intraoperative Cell Saver use characteristics; preoperative coagulation function characteristics, preoperative blood routine characteristics, preoperative liver and kidney function characteristics, preoperative electrolyte characteristics; intraoperative and postoperative red blood cell transfusion volume characteristics;

[0106] The preoperative coagulation function time series data includes prothrombin time characteristics, activated partial thromboplastin time characteristics, fibrinogen characteristics, D-dimer characteristics, international normalized ratio characteristics;

[0107] The three-dimensional anatomical parameters of the surgical site include the radius of curvature of the vertebral body, the diameter of the spinal canal, the height of the intervertebral space, the bone density, the course of the vertebral artery, the density of the blood vessels around the vertebral body, the position of the intercostal artery or the lumbar artery, the density of the epidural venous plexus, the thickness of the interspinal ligament, the thickness of the yellow ligament, the thickness of the epidural fat, the angle of the surgical approach, the exposure range, and the depth of the instrument implantation.

[0108] The different types of machine learning models include a fractal adaptive Ranger model, a spatio-temporal convolution network, and a causal graph neural network.

[0109] In this embodiment, training the fractal adaptive Ranger model includes:

[0110] An engineering feature matrix is generated according to a historical data set, and the engineering feature matrix is divided into an engineering feature matrix training set, an engineering feature matrix validation set, and an engineering feature matrix test set.

[0111] In this embodiment, generating the engineering feature matrix according to the historical data set includes:

[0112] Temporal features, frequency domain features, and dynamic trend features are extracted from the time series data (e.g., preoperative coagulation function time series data), thereby obtaining the temporal features, the frequency domain features, and the dynamic trend features.

[0113] For the three-dimensional anatomical parameters of the surgical site, anatomical feature calculation is performed to obtain anatomical features, including geometric features, topological features, and tissue features. For example, the geometric features can include a spinal canal stenosis rate = (theoretical diameter - actual diameter) / theoretical diameter, a vertebral body inclination angle (a three-dimensional space angle), and a bone spur volume (mm³). The topological features can include an intervertebral foramen porosity (porosity volume / total volume) and a nerve root canal tortuosity (using Hausdorff distance). The tissue features can include a bone density heterogeneity index (CT value standard deviation) and an epidural fat proportion (based on Hounsfield unit threshold segmentation).

[0114] The physiological features (e.g., gender features, height features, weight features, BMI features, diabetes features, and hypertension features) are converted into vectors.

[0115] The temporal features, the frequency domain features, the dynamic trend features, the geometric features, and the various physiological features are fused to obtain an engineering feature matrix.

[0116] The fractal adaptive feature weighting is performed according to the engineering feature matrix training set to obtain a fractal adaptive weight matrix.

[0117] In the embodiment, the fractal adaptive weight matrix is obtained by fractal adaptive feature weighting according to the engineering feature matrix training set, which comprises:

[0118] The fractal dimension of each dimension in the engineering feature matrix is calculated to obtain a fractal dimension vector.

[0119] Each dimension in the fractal dimension vector is assigned a weight by fractal dimension-driven weight adjustment to generate an adaptive weight matrix.

[0120] The adaptive weight matrix is subjected to a surgical type enhancement operation to obtain a fractal adaptive weight matrix.

[0121] For example, the fractal dimension of each dimension in the engineering feature matrix is calculated to obtain a fractal dimension vector, which comprises:

[0122] For each feature (such as "coagulation function fluctuation rate"), the following operations are performed:

[0123] Scaling data: scaling the feature value sequence by a scaling factor s=0.5 to generate a scaled data subset.

[0124] Calculate the number of data points:

[0125] N1: the number of original data points (such as 1000 samples).

[0126] N2: the number of scaled data points (such as 500 samples).

[0127] The fractal dimension is calculated by the fractal dimension formula, which in the embodiment is as follows:

[0128] Fractal dimension=log(N1 / N2) / log(1 / s);

[0129] Example: if N1=1000, N2=500, and s=0.5, then the fractal dimension=log(2) / log(2)=1.0.

[0130] The fractal dimension is linearly scaled to the interval [1.0, 2.0] to avoid the influence of extreme values on subsequent weight adjustment.

[0131] In the embodiment, each dimension in the fractal dimension vector is assigned a weight by fractal dimension-driven weight adjustment to generate an adaptive weight matrix, which comprises:

[0132] Obtain the initial weight, which can be understood as being artificially set, for example, the coagulation function index weight=1.0 and the BMI weight=0.8.

[0133] The maximum weight is obtained by the following formula:

[0134] Maximum weight = base weight × (1 + fractal dimension × 0.2).

[0135] For example, if the fractal dimension of a feature is 1.8 and the base weight is 1.0, the adjusted weight is 1.0 × (1 + 1.8 × 0.2) = 1.36.

[0136] If the fractal dimension is 1.2 and the base weight is 0.8, the adjusted weight is 0.8 × (1 + 1.2 × 0.2) = 0.8 × 1.24 = 0.99.

[0137] The adjusted weights are normalized according to the feature type (continuous / categorical) to ensure that the sum of the weights of the same type of feature is 1.

[0138] In this embodiment, the operation of the adaptive weight matrix is enhanced by the type of surgery, so as to obtain a fractal adaptive weight matrix, which comprises:

[0139] According to the type of surgery label, a list of key features is extracted from an external knowledge base (such as an ICD coding mapping table).

[0140] For example, if it is a fusion surgery, the key features include BMI, coagulation function indicators, and surgery duration. If it is a tumor surgery, the key features include tumor volume, platelet count, and hemoglobin level.

[0141] The weights of the key features are enhanced, for example, if it is a fusion surgery, the weight of BMI is increased by 20% (such as from 0.99 to 1.19). The weight of coagulation function indicators is increased by 30% (such as from 1.36 to 1.77).

[0142] If it is a tumor surgery, the weight of tumor volume is increased by 50% (such as from 0.8 to 1.2). The weight of platelet count is increased by 25% (such as from 0.7 to 0.875).

[0143] The weights of non-key features are reduced (such as 5%), to avoid imbalance of weight distribution.

[0144] Through the fractal adaptive feature weighting described above, the model can automatically identify features with high complexity and high clinical value, and assign higher weights, providing accurate feature input for subsequent decision tree construction.

[0145] According to the fractal adaptive weight matrix, a regularization tree is constructed based on game theory, so as to obtain a regularization decision tree, a hyperparameter combination, and a performance curve of an iteration process.

[0146] In the present embodiment, a game theory driven regularization tree construction is performed according to the fractal adaptive weight matrix, so as to obtain a regularization decision tree, an initial hyperparameter combination and an iteration process performance curve, which comprises:

[0147] Step 101: Obtain a strategy probability distribution table; specifically, combine a learning rate and a tree depth into a hyperparameter pair (for example, learning rate = 0.05, tree depth = 5), generate an initial combination pool (3 learning rates x 3 tree depths = 9 combinations). Perform strategy probability initialization: assign an initial strategy probability (uniform distribution) to each hyperparameter combination. For example, the probability of the combination of learning rate = 0.05 and tree depth = 5 is 1 / 9 ≈ 11.1%. Embed the fractal adaptive weight matrix into the decision tree construction module.

[0148] Step 102: Randomly select a hyperparameter combination according to the strategy probability distribution table, so as to obtain the hyperparameter combination of the current round; for example, if the probability of the combination of learning rate = 0.08 and tree depth = 6 is 25%, then the combination is selected with a probability of 25%.

[0149] Step 103: Calculate the adjusted importance of the fractal adaptive weight matrix, so as to obtain an adjusted feature importance vector; specifically, for each feature, calculate the adjusted importance: wherein the feature importance = the base importance x the fractal weight, and the base importance is determined by the feature type and can be preset in advance, for example, the base importance of a continuous feature = 1.0.

[0150] For example, the fractal weight of the feature “coagulation function fluctuation rate” = 1.36, and the feature importance = 1.0 x 1.36 = 1.36.

[0151] Step 104: Generate a regularization coefficient according to the learning rate; for example, dynamically calculate λ according to the learning rate: λ = learning rate x 1.5.

[0152] Step 105: Calculate the split criterion of the adjusted feature importance vector according to the regularization coefficient, so as to obtain the split criterion value of each feature in the feature importance vector; for example, calculate the split criterion for each feature: split criterion = Gini impurity - λ x (feature importance)²; for example, the feature importance = 1.36 and λ = 0.12, then the split criterion = Gini impurity - 0.12 x 1.36² ≈ Gini impurity - 0.22.

[0153] Step 106: Construct a regularization decision tree according to the feature importance vector and the split criterion value of each feature; the present application uses a recursive splitting method for construction, and the specific process is as follows:

[0154] Initialize the root node: input the entire training set data into the root node, and set the current tree depth = 0.

[0155] Recursive splitting procedure (loop until stop condition is met):

[0156] Step 1061: Select the optimal splitting feature:

[0157] Inputs: data subset of the current node, feature importance vector, splitting criterion value.

[0158] Process: For each feature, calculate the optimal split point according to the splitting criterion value, and select the feature with the largest splitting criterion value as the splitting feature of the current node.

[0159] Output: Optimal splitting feature and split point of the current node (e.g., "coagulation function fluctuation rate > 1.2").

[0160] Step 1062: Check tree depth limit:

[0161] Inputs: current tree depth, maximum depth parameter (tree depth = 6).

[0162] Process: If the current tree depth ≥ maximum depth, stop splitting and mark as a leaf node.

[0163] Otherwise, go to Step 1063.

[0164] Step 1063: Recursively build sub-tree:

[0165] Inputs: data subset of the current node, optimal splitting feature and split point, current tree depth + 1.

[0166] Process: Divide the data into left and right sub-trees according to the optimal splitting feature and split point.

[0167] Repeat the above steps for each sub-tree until the stop condition is met.

[0168] In this embodiment, the stop conditions are as follows:

[0169] Tree depth limit: current tree depth ≥ maximum depth (e.g., tree depth = 6).

[0170] Purity gain threshold: reduction in Gini impurity after splitting < 0.01 (threshold adjustable).

[0171] Step 107: Validate the validation set performance indicators of the decision tree model by the validation set; for example, input the validation set data into the decision tree model to predict the bleeding risk level. Calculate the mean square error (MSE) and classification accuracy (AUC) on the validation set:

[0172] Step 108: Obtain the benefit value of the current hyperparameter combination according to the validation set performance indicator; for example, calculate the benefit according to the performance indicator: benefit = -MSE + 0.1 x AUC (for example, MSE = 280 ml², AUC = 0.92, then benefit = -280 + 0.1 x 0.92 ≈ -279.91).

[0173] Step 109: Update the strategy probability according to the benefit value of the current hyperparameter combination; adjust the strategy probability according to the benefit value: for example, if the benefit > -280 (high benefit), the combination probability rises by 20%. If the benefit ≤ -280 (low benefit), the combination probability decreases by 15%.

[0174] Step 110: Repeat the steps 101 to 109 until the maximum number of iterations is reached or the convergence condition is reached, and obtain the iteration process performance curve and the hyperparameter combination after the maximum number of iterations is reached or the convergence condition is reached.

[0175] In this embodiment, the convergence condition can be as follows:

[0176] Calculate the strategy probability change rate of the adjacent two iterations: change rate = |current probability - last probability| / last probability x 100%, if the change rate of all combinations < 5%, determine that the strategy probability is stable.

[0177] Performance saturation check: check the MSE reduction of the last 3 iterations: MSE reduction = last MSE - current MSE; if the reduction of the last 3 iterations < 10 ml², determine that the performance is saturated.

[0178] Iteration number check: if the maximum number of iterations (20 times) has been reached, forcibly terminate the iteration.

[0179] According to the hyperparameter combination and the iteration process performance curve, the hyperparameter optimization of the cellular automaton is performed, so as to obtain the globally optimal hyperparameter combination.

[0180] In this embodiment, according to the hyperparameter combination and the iteration process performance curve, the hyperparameter optimization of the cellular automaton is performed, so as to obtain the globally optimal hyperparameter combination, which includes:

[0181] According to the regularized decision tree, an initial cellular automaton state matrix is constructed;

[0182] Set the state evolution rule, and perform cellular automaton state iteration evolution on the initial cellular automaton state matrix, so as to obtain the updated cellular automaton state matrix and the list of hyperparameter combinations to be explored after the iteration condition is reached, and the list of hyperparameter combinations to be explored includes at least one hyperparameter combination to be explored (the feature combination corresponding to the state 2 cell).

[0183] In this embodiment, the state evolution itself is also an iterative process, with the following examples:

[0184] Operations at each iteration:

[0185] Step 2.1: State matrix initialization

[0186] Divide the hyperparameter space (learning rate, tree depth, λ) into a three-dimensional grid, with each cell representing a hyperparameter combination.

[0187] Initialize cell states: unexplored (state 0), low yield (state 1), high yield (state 2).

[0188] Step 2.2: Neighbor state analysis

[0189] For each cell, check the state distribution of its molar neighbors (8-neighborhood).

[0190] Example: If there are 3 state 2 cells in the neighborhood, the current cell may upgrade to state 2.

[0191] Step 2.3: State update rules

[0192] Rule 1: If there is a state 2 cell in the neighborhood, the current cell state is upgraded to 2 (prefer to explore high-yield areas).

[0193] Rule 2: If the proportion of state 1 cells in the neighborhood is ≥ 50%, the current cell state is downgraded to 1 (avoid repeated exploration of low-yield areas).

[0194] Rule 3: Unexplored cells (state 0) prefer to explore the highest-yield combination in the neighborhood.

[0195] Step 2.4: Hyperparameter combination generation

[0196] According to the updated state matrix, select the hyperparameter combination corresponding to the state 2 cell (e.g. learning rate = 0.07, tree depth = 5, λ = 0.15).

[0197] Iteration termination conditions:

[0198] Maximum number of iterations (30) is reached.

[0199] The number of state 2 cells in the state matrix changes by less than 5% (for two consecutive iterations).

[0200] The MSE reduction of the last 3 explorations is less than 10ml² (performance saturation).

[0201] Determine whether the iterated cellular automaton state matrix (state 0 / 1 / 2 distribution) converges or reaches the number of iterations, and after reaching, calculate the performance index of each set of to-be-explored hyperparameter combination in the to-be-explored hyperparameter combination list, and select the set with the best performance index as the global optimal hyperparameter combination.

[0202] In this embodiment, by introducing a fractal adaptive feature encoding, by calculating the feature complexity and assigning a fractal weight, both the key clinical indicators (such as the dynamic trend of the coagulation function time series) are retained, and the three low-information features are removed through weight screening, and the balance between information density and calculation efficiency is achieved in the engineering features; secondly, the cellular automaton hyperparameter optimization breaks through the local optimal trap of traditional grid search, and through the Moore neighbor rule to drive the state evolution, it efficiently locates the global optimal combination in the three-dimensional hyperparameter space (learning rate, tree depth, regularization coefficient), which can reduce the MSE by 18.2% through clinical data verification; the dynamic decision tree provides a performance feedback benchmark for cellular automaton hyperparameter optimization, and the hyperparameter combination optimized by cellular automaton hyperparameter optimization reversely improves the model generalization ability of stage three, this "encoding-optimization-feedback" closed loop design finally improves the AUC of bleeding risk prediction from 0.79 to 0.93, and the clinical application value is significant.

[0203] Embodiment 3:

[0204] A network calculator construction method for intraoperative massive bleeding risk prediction and autologous blood transfusion decision-making in spinal surgery based on machine learning

[0205] S1 data collection: collect the corresponding clinical indicators and operation data of spinal surgery patients from the electronic medical record system of medical institutions.

[0206] S2 massive bleeding prediction model construction: use the collected multiple clinical indicators as prediction variables, construct 26 different machine learning models on the training set, and evaluate the model performance by cross-validation and other methods. In the training set and the test set, respectively, the performance of each model is evaluated by the receiver operating characteristic curve (ROC) and the AUC index. The SHAP value is used for explanatory analysis of the top five models with the best performance, and the intersection of the features selected by each model is taken to simplify and retrain the model. Independent external cohort data is introduced, and the constructed model is externally verified by ROC curve, calibration curve and Brier score. The machine learning model with the best performance is selected for subsequent sensitivity analysis, decision curve analysis, interaction analysis and treatment benefit evaluation.

[0207] S3 Network calculator generates the trained machine learning model into a portable format for loading and calling in the network environment. And pre-process the input variables, ensure that the patient data input by the user is consistent with the format during model training. The model output is the prediction probability of intraoperative significant blood loss and the benefit evaluation of intraoperative autologous blood transfusion.

[0208] S4 Spinal surgery intraoperative massive bleeding probability calculation and intraoperative autologous blood transfusion treatment benefit evaluation: collect the key information data of the surgical patient, and input into the network calculator, the corresponding prediction result can be obtained.

[0209] As a preferred scheme of the present application, in the step S1, the corresponding clinical indicators of the spinal surgery patient include but are not limited to patient basic information (age, gender, height, weight), comorbidities (diabetes, hypertension, coronary heart disease, cerebrovascular disease, liver and kidney function abnormalities), spinal disease diagnosis, spinal surgery related parameters (specific way of operation, number of fused vertebral bodies, operation site, etc.), blood indicators (preoperative coagulation function indicators, preoperative blood routine, etc.), intraoperative blood management measures (application of intraoperative hemostatic drugs and hemostatic measures) and blood transfusion management information (intraoperative autologous recovery and use, intraoperative and postoperative red blood cell transfusion). The collected data is standardized, missing value processed and data cleaned to establish a database that meets the requirements of statistics and machine learning.

[0210] As a preferred scheme of the present application, in the step S2, the 12 independent prediction factors include patient age, height, weight, preoperative platelet, preoperative red blood cell, spinal lesion type (whether involving deformity, degeneration), spinal surgery method (whether the operation site involves cervical spine, thoracic spine, lumbar spine and sacral spine; whether the operation involves vertebral body resection (including osteotomy); the number of vertebral body fusion).

[0211] As a preferred scheme of the present application, in the step S2, according to the performance evaluation of the massive bleeding prediction model, high and low risk patients are distinguished, specifically including: in the test set, the performance of the model is evaluated by the ROC curve, and the best Ranger model is selected, the test set AUC is 0.814 (95% confidence interval: 0.790-0.839), and the AUC of the external validation set is 0.816 (95% confidence interval: 0.763-0.870). Combined with the decision curve analysis, it is determined that the critical value of the prediction probability of significant blood loss risk is 0.53. Patients with prediction probability higher than 0.53 are determined as high risk; those lower than 0.53 are determined as low risk.

[0212] As a preferred scheme of the present application, in the step S4, the network calculator displays the risk prediction result and the benefit evaluation data in text form, generates a detailed report, and intuitively presents the intraoperative massive hemorrhage risk of the patient and the treatment benefit that may be brought by the use of CellSaver. The report provides personalized blood management and transfusion decision support for the clinician, enabling him / her to reasonably select the intraoperative autologous blood transfusion strategy according to the risk status of the patient.

[0213] Compared with the prior art, the present application has the following beneficial effects:

[0214] Significant improvement in prediction accuracy and stability: Compared with the traditional single logistic regression model, the present application can more fully capture the complex relationships in the multi-dimensional clinical data of patients by using multiple machine learning models and 10-fold cross-validation technology, significantly improving the accuracy and stability of massive hemorrhage risk prediction. Through multi-level verification of internal training set, test set and independent external verification set data, the present application ensures the robustness and consistency of the model in different patient groups and subgroups (such as different genders, BMIs, and surgical sites), and has a wide range of applications, providing solid data support for clinical practical application.

[0215] Personalized transfusion management and decision support: Combined with decision curve analysis, the critical value (0.53) of the significant blood loss risk prediction probability is determined, providing a quantitative standard for the classification of high-risk and low-risk patients. Based on the prediction results, the system can output an intraoperative autologous blood transfusion benefit evaluation report (such as reducing the intraoperative total allogeneic red blood cell transfusion volume, reducing the postoperative transfusion demand, and improving the postoperative hemoglobin level), thereby realizing a more targeted personalized transfusion management strategy.

[0216] Real-time online application and data visualization: By constructing a network calculator, real-time calling and risk prediction of the model are realized, and the clinician can quickly obtain the prediction results and treatment recommendations on different terminal devices. The intuitive display of text helps doctors quickly understand the risk level of the patient and the corresponding transfusion management measures, improving the efficiency of clinical decision-making.

[0217] The embodiment 3 of the present application is further described in an exemplary manner below, and it can be understood that the example does not constitute any limitation on the present application.

[0218] Data collection and preprocessing: In this step, first, the relevant clinical indicators and surgical data of patients undergoing spinal surgery are collected from the electronic medical record system of the medical institution. The collected data include but are not limited to: patient basic information: age, gender, height, weight; comorbidities: diabetes, hypertension, coronary heart disease, cerebrovascular disease, liver and kidney function abnormalities, etc.; spinal disease diagnosis: diagnosis type (such as degenerative disease, deformity, infection, tumor, etc.); spinal surgery related parameters: surgical method, number of fused vertebral bodies, surgical site (cervical spine, thoracic spine, lumbar spine, sacral spine, etc.); blood indicators: preoperative coagulation function indicators, preoperative blood routine; intraoperative blood management measures: such as the application of hemostatic drugs and hemostatic measures; blood transfusion management information: including the use of autologous blood recovery during surgery and the intraoperative and postoperative red blood cell transfusion situation. The collected data are subjected to data cleaning, missing value completion, standardization and normalization processing, and a database that meets the requirements of statistical analysis and machine learning is established, providing a high-quality data foundation for subsequent model construction.

[0219] Big hemorrhage risk prediction model construction uses the database constructed in S1 to predict the risk of intraoperative massive hemorrhage (defined as intraoperative blood loss ≥500 milliliters) in patients undergoing spinal surgery using machine learning methods. The specific steps are as follows:

[0220] Data set division: all data are randomly divided into training set (70%) and test set (30%).

[0221] Model training: 26 different machine learning models are used for training on the training set, including but not limited to Ranger, cforest, randomForest, ridge, glmnet, etc. During the training process, 10-fold cross-validation is applied to optimize the model hyperparameters, ensuring that the model has good generalization ability. In this embodiment, the Ranger model is preferred as a specific embodiment, with an AUC value of 0.814 (95% confidence interval: 0.790-0.839) in the test set and an AUC value of 0.816 (95% confidence interval: 0.763-0.870) in the independent external validation set; but the present application is not limited to this specific model, and other models with equivalent or better performance are also within the scope of the present application.

[0222] Feature selection: SHAP (SHapley Additive exPlanations) value analysis was used to evaluate the importance of candidate variables in each model, and key features with significant impact on major bleeding risk were selected. After cross-comparison of multiple best-performing models (such as Ranger, cforest, randomForest, ridge, glmnet), 12 independent predictors were finally determined, mainly including patient age, height, weight, preoperative platelet count, preoperative red blood cell count, spinal lesion type (whether involving deformity or degeneration), spinal surgery method (whether involving thoracic, lumbar and cervical spine, whether involving vertebral resection or osteotomy), and the number of vertebral fusion, etc.

[0223] Model evaluation and risk stratification: In the test set, the prediction performance of each model was evaluated by ROC curve, AUC and Brier score, and the critical value of significant bleeding risk prediction probability was determined as 0.53 by decision curve analysis. Patients with prediction probability higher than 0.53 were judged as high risk, and those with prediction probability lower than 0.53 were judged as low risk.

[0224] Web calculator generation: Based on the risk prediction model obtained by training, a web-based calculator was developed through R Shiny framework. The specific steps are as follows:

[0225] Development platform and framework: The web calculator in this embodiment is developed by Shiny framework of R language. R Shiny is a web application development tool based on R language, which can combine the data processing and model calculation capabilities of R language with interactive web pages, making it convenient for clinicians to call risk prediction models in real time through browsers. The web calculator has been deployed on the shinyapps.io platform and runs online using the uniform resource locator https: / / fxzmodel.shinyapps.io / SBLprediction /

[0226] Data entry module: The user interface is built with Shiny's fluidPage, which has multiple input controls, including text boxes for entering patient clinical data and option selection controls for selecting preset options (such as disease type, surgery method, etc.), for entering patient's basic information (such as age, height, weight), spinal disease diagnosis, surgery parameters (surgery method, fusion vertebral number, surgery site, etc.), blood indicators (preoperative blood routine data), and blood transfusion management information.

[0227] Data format verification: On the front end, JavaScript or R language internal input verification logic is used to ensure that the user-entered data conforms to the preset format, preventing model prediction errors due to data format errors.

[0228] Background data preprocessing: The received input data is further processed by the background using R language, including missing value filling, classification variable conversion to dummy variable, numerical variable standardization, etc. to ensure that the input data is completely consistent with the data format during training.

[0229] Model loading and calling: The preprocessed data is then passed to the serialized machine learning model (preferably Ranger model) and the background uses the model to calculate the risk prediction. The prediction output is a continuous significant blood loss risk probability. For example, if the calculation result shows a risk of 33.1%, the system will output this probability as the result.

[0230] Benefit evaluation module: In addition to the risk probability, the background also calls the pre-developed benefit evaluation algorithm to generate data on the benefits of autologous blood transfusion treatment based on the predicted risk level, such as reducing red blood cell transfusion units, adjusting postoperative hemoglobin levels, and other quantitative indicators.

[0231] Decision support suggestions: Based on the pre-set risk threshold (0.53), the system provides specific treatment suggestions for high-risk patients, such as prompting the use of Cell Saver autologous blood transfusion technology to reduce intraoperative and postoperative red blood cell transfusion needs.

[0232] Interface-friendly and real-time response: The network calculator is designed with user experience in mind, with a clean and responsive interface that supports access from multiple mobile terminal devices (PC, tablet, mobile phone, etc.) to facilitate real-time access to decision support data during preoperative or perioperative periods. These mobile terminals should not impose any restrictions on the functionality and scope of use of the embodiments of the present disclosure.

[0233] In specific implementation, considering that patients undergoing spinal surgery have different degrees of risk of major bleeding during the perioperative period, and due to individual differences (such as surgical method, severity of illness, patient's basic characteristics, etc.), the probability of intraoperative massive bleeding is not the same, therefore, the preoperative data information of the patient can be collected to form a key information data set containing patient basic information (such as age, gender, height, weight), disease diagnosis (such as degenerative disease, deformity, infection, tumor, etc.), surgical parameters (such as surgical method, number of fused vertebral bodies, surgical site, etc.), blood indicators (such as preoperative coagulation function, blood routine data), and related blood management measures. After inputting this data set into the network calculator, a risk prediction model based on machine learning (preferably using a Ranger model) is used to automatically calculate the risk of intraoperative massive bleeding in patients undergoing spinal surgery and output the corresponding risk probability.

[0234] According to the predicted probability of intraoperative massive bleeding risk, the surgical patients can be divided into high-risk and low-risk categories. According to the present embodiment, it is determined by decision curve analysis that the patients with a predicted probability of more than 0.53 are high-risk massive bleeding patients, and more active blood management and intervention measures need to be taken before surgery; the patients with a predicted probability of less than 0.53 are low-risk patients, and part of the unnecessary blood management detection can be exempted, thereby saving the economic cost of the patients. It is worth noting that the risk threshold of 0.53 is only the best value obtained by applying decision curve analysis in the present embodiment, and any changes, modifications or replacements of the present embodiment can lead to changes in the threshold value, but do not affect the core technical idea of the present application.

[0235] In summary, the present application provides a spine surgery massive bleeding risk prediction model based on machine learning and a network calculator construction method thereof, which can accurately predict the intraoperative massive bleeding risk of patients by comprehensive analysis of multiple clinical indicators of the patients before surgery, and provide quantitative risk probability and treatment benefit evaluation. This method not only helps doctors to reasonably stratify and individualize blood management for intraoperative massive bleeding risk, but also provides automatic electronic decision support for optimizing medical resource allocation, reducing patient medical costs, preventing blood transfusion-related adverse reactions, etc., and has significant clinical application value and promotion prospect.

[0236] The present application has the following advantages:

[0237] 1) High-precision risk prediction: Through multi-model comparison and SHAP value feature selection, the Ranger model performs well in the test set and external validation set (AUC is 0.814 and 0.816 respectively, see Figure 2 ), which surpasses the traditional logistic regression method and significantly improves the accuracy and stability of SBL prediction.

[0238] 2) Individualized blood management: The stratification strategy based on the 0.53 risk threshold recommends CellSaver use for high-risk patients and avoids unnecessary intervention for low-risk patients, optimizes resource allocation, and reduces blood transfusion complications (such as infection, immune response).

[0239] 3) Real-time automated tool: The network calculator based on R Shiny supports real-time data input and result output, with an intuitive interface and multi-terminal access, which greatly improves the efficiency of clinical decision-making.

[0240] 4) Explainability and robustness: SHAP value analysis provides feature importance explanation, enhancing model transparency; subgroup analysis (gender, BMI, surgical site, etc., AUC≥0.895) verifies the wide applicability of the model.

[0241] 5) Clinical translation potential: Decision curve analysis showed that the risk-adapted strategy had higher net benefit at all thresholds, and the interaction analysis proved that Cell Saver significantly reduced the transfusion requirement in high-risk patients, providing a new standard for blood management in spine surgery.

[0242] The application also provides a machine learning-based intraoperative bleeding risk prediction device for spine surgery, comprising a patient information to be predicted acquisition module, a model acquisition module, and a prediction module, wherein,

[0243] The patient information to be predicted acquisition module is used to acquire patient information to be predicted;

[0244] The model acquisition module is used to acquire a trained massive bleeding risk prediction model;

[0245] The prediction module is used to input the patient information to be predicted into the trained massive bleeding risk prediction model, thereby obtaining a prediction result.

[0246] The application also provides a network calculator comprising the machine learning-based intraoperative bleeding risk prediction device for spine surgery as described above.

[0247] Finally, it should be pointed out that: the above examples are only used to illustrate the technical solutions of the application, and not to limit them. Although the application has been described in detail with reference to the foregoing examples, those skilled in the art should understand that they can still modify the technical solutions recorded in the foregoing examples, or make equivalent replacements for some technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the application.

Claims

1. A machine learning-based method for predicting intraoperative bleeding risk in spinal surgery, characterized in that, The machine learning-based method for predicting intraoperative bleeding risk in spinal surgery includes: Obtain information on patients to be predicted; Obtain a trained model for predicting the risk of major bleeding; The patient's information is input into a trained major bleeding risk prediction model to obtain the prediction results.

2. The method for predicting intraoperative bleeding risk in spinal surgery based on machine learning as described in claim 1, characterized in that, The machine learning-based method for predicting intraoperative bleeding risk in spinal surgery further includes: The massive bleeding risk prediction model is trained; The training of the massive bleeding risk prediction model includes: Obtain a historical dataset, which includes historical baseline data for multiple patients; Obtain multiple different types of machine learning models; Each machine learning model is trained based on the historical dataset and its hyperparameters are optimized using 10-fold cross-validation to obtain the trained machine learning models. The importance of various model features in historical baseline data was evaluated using SHAP value analysis, and 12 independent predictors were selected from multiple models through cross-selection. One of the trained machine learning models is selected as the trained major bleeding risk prediction model.

3. The method for predicting intraoperative bleeding risk in spinal surgery based on machine learning as described in claim 2, characterized in that, The historical baseline data includes: age characteristics, gender characteristics, height characteristics, weight characteristics, diabetes characteristics, hypertension characteristics, coronary heart disease characteristics, cerebrovascular disease characteristics, abnormal liver and kidney function characteristics, spinal degenerative disease characteristics, spinal deformity characteristics, spinal infection characteristics, spinal tumor characteristics, surgical method characteristics, number of fused vertebrae characteristics, surgical site characteristics, preoperative coagulation function characteristics, preoperative blood routine characteristics, intraoperative Cell Saver usage characteristics, and intraoperative and postoperative red blood cell transfusion volume characteristics. The different types of machine learning models include the Ranger model, cforest model, randomForest model, ridge model, and glmnet model.

4. The method for predicting intraoperative bleeding risk in spinal surgery based on machine learning as described in claim 2, characterized in that, The historical baseline data includes raw clinical data, which includes characteristics such as age, sex, height, weight, BMI, diabetes, hypertension, coronary heart disease, cerebrovascular disease, abnormal liver and kidney function, surgical method, number of fused vertebrae, surgical site, estimated surgical duration, and intraoperative Cell Saver usage; preoperative coagulation function, preoperative complete blood count, preoperative liver and kidney function, and preoperative electrolytes; and intraoperative and postoperative red blood cell transfusion volume. The preoperative coagulation function time series data includes prothrombin time characteristics, activated partial thromboplastin time characteristics, fibrinogen characteristics, D-dimer characteristics, and international normalized ratio characteristics. Three-dimensional anatomical parameters of the surgical site include vertebral body curvature radius characteristics, spinal canal diameter characteristics, intervertebral space height characteristics, bone density characteristics, vertebral artery course characteristics, vertebral body peripheral vascular density characteristics, intercostal artery or lumbar artery location characteristics, epidural venous plexus density characteristics, interspinous ligament thickness characteristics, ligamentum flavum thickness characteristics, epidural fat thickness characteristics, surgical approach angle characteristics, exposure range characteristics, and instrument implantation depth characteristics. The different types of machine learning models include fractal adaptive Ranger models, spatiotemporal convolutional networks, and causal graph neural networks.

5. The method for predicting intraoperative bleeding risk in spinal surgery based on machine learning as described in claim 4, characterized in that, Training the fractal adaptive Ranger model includes: The engineering feature matrix is ​​generated based on the historical dataset, and then divided into an engineering feature matrix training set, an engineering feature matrix validation set, and an engineering feature matrix test set. Fractal adaptive feature weighting is performed on the training set of the engineering feature matrix to obtain the fractal adaptive weight matrix; Based on the fractal adaptive weight matrix, a game theory-driven regularized tree is constructed to obtain the regularized decision tree, hyperparameter combination, and performance curve of the iterative process. Based on the hyperparameter combination and the performance curve of the iterative process, the hyperparameters of the cellular automata are optimized to obtain the globally optimal hyperparameter combination.

6. The method for predicting intraoperative bleeding risk in spinal surgery based on machine learning as described in claim 5, characterized in that, The step of performing fractal adaptive feature weighting based on the engineering feature matrix training set to obtain the fractal adaptive weight matrix includes: The fractal dimension is calculated for each feature in the engineering feature matrix to obtain the fractal dimension vector; Each feature in each dimension of the fractal dimension vector is assigned a weight through a fractal dimension-driven weight adjustment, thereby generating an adaptive weight matrix. A surgery type enhancement operation is performed on the adaptive weight matrix to obtain a fractal adaptive weight matrix.

7. The method for predicting intraoperative bleeding risk in spinal surgery based on machine learning as described in claim 6, characterized in that, The step of constructing a game-theoretic-driven regularized tree based on the fractal adaptive weight matrix to obtain the regularized decision tree, initial hyperparameter combination, and performance curve of the iterative process includes: Step 101: Obtain the policy probability distribution table; Step 102: Based on the strategy probability distribution table, randomly select a hyperparameter combination to obtain the hyperparameter combination for the current round; Step 103: Calculate the adjusted importance of the fractal adaptive weight matrix to obtain the adjusted feature importance vector; Step 104: Generate regularization coefficients based on the learning rate; Step 105: Calculate the splitting criterion for the adjusted feature importance vector based on the regularization coefficient, thereby obtaining the splitting criterion value for each feature in the feature importance vector; Step 106: Construct a regularized decision tree based on the feature importance vector and the splitting criterion value of each feature; Step 107: Validate the validation set performance metrics of the decision tree model using the validation set; Step 108: Obtain the revenue value of the current hyperparameter combination based on the validation set performance metrics; Step 109: Update the strategy probability based on the payoff value of the current hyperparameter combination; Step 110: Repeat steps 101 to 109 until the maximum number of iterations is reached or the convergence condition is met, and obtain the performance curve of the iteration process and the hyperparameter combination after the maximum number of iterations is reached or the convergence condition is met.

8. The method for predicting intraoperative bleeding risk in spinal surgery based on machine learning as described in claim 7, characterized in that, The step of optimizing the hyperparameters of the cellular automata based on the hyperparameter combination and the performance curve of the iterative process, thereby obtaining the globally optimal hyperparameter combination, includes: Construct the initial cellular automaton state matrix based on the regularized decision tree; Set state evolution rules, perform iterative evolution of the initial cellular automaton state matrix, thereby obtaining the updated cellular automaton state matrix after reaching the iteration conditions and a list of hyperparameter combinations to be explored. The list of hyperparameter combinations to be explored includes at least one set of hyperparameter combinations to be explored. Calculate the performance index of each hyperparameter combination in the list of hyperparameter combinations to be explored, and select the combination with the best performance index as the global optimal hyperparameter combination.

9. A machine learning-based device for predicting intraoperative bleeding risk in spinal surgery, characterized in that, The machine learning-based intraoperative bleeding risk prediction device for spinal surgery includes: A module for acquiring information about patients to be predicted, wherein the module is used to acquire information about patients to be predicted; A model acquisition module, which is used to acquire a trained major bleeding risk prediction model; The prediction module is used to input patient information to be predicted into a trained major bleeding risk prediction model to obtain prediction results.

10. A network calculator, characterized in that, The network calculator includes the machine learning-based intraoperative bleeding risk prediction device for spinal surgery as described in claim 9.

Images