Digestive tract operation consumable consumption prediction system based on big data analysis

Abstract

The invention provides an alimentary canal operation consumable consumption prediction system based on big data analysis, and aims to accurately predict the consumable consumption in a data driving mode. According to the system, patient data, operation data, consumable data and hospital data in a hospital EMR, an HIS, a supply chain management system and a hospital management system are collected in a multi-source mode, and after format unification, cleaning, feature extraction and normalization are carried out, a multi-model prediction system is constructed; a logic regression model is adopted to evaluate complication risks, an SVR model is utilized to predict the amount of fault-tolerant consumables when complications exist, and finally a total prediction result is dynamically output in combination with the probability of the complication risks and pushed to an electronic medical record system to assist doctors in preoperative preparation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical data analysis, in particular to a digestive tract surgery consumable consumption prediction system based on big data analysis. BACKGROUND

[0002] In hospital digestive tract surgery, accurate prediction of consumable consumption is the key to optimizing surgical preparation and reducing resource waste or shortage. Traditional consumable preparation relies on experience, which is prone to prediction deviation due to patient individual differences, surgical complexity, and doctor habits, affecting surgical efficiency. With the advancement of hospital informatization construction, EMR, HIS and other systems have accumulated massive multi-dimensional data, providing a data basis for scientific prediction of consumable consumption through big data analysis.

[0003] Existing systems have introduced data collection tools, but mostly stay at the data recording level, lack deep integration of multi-source data, and have not combined machine learning technology to achieve intelligent prediction, resulting in low prediction accuracy, poor adaptability, and inability to cope with complex surgical scenarios.

[0004] To solve the above problems, the present application proposes a digestive tract surgery consumable consumption prediction system based on big data analysis, which collects patient data, surgery data, consumable data and hospital data from hospital EMR, HIS, supply chain management system and hospital management system, etc. After format unification, correlation integration, cleaning, feature extraction and normalization, a prediction system including complication risk assessment, basic consumable prediction and fault-tolerant consumable prediction is constructed. Finally, the dynamic prediction result is output combined with the complication risk probability and pushed to the hospital electronic medical record system to assist the doctor in preoperative preparation, realizing accurate and intelligent prediction of consumable consumption. To solve the technical problems mentioned in the background art, the present application proposes a digestive tract surgery consumable consumption prediction system based on big data analysis.

[0005] To this end, the technical solution adopted by the present application is as follows: The digestive tract surgery consumable consumption prediction system based on big data analysis comprises a data collection module and a surgery consumable consumption prediction module. The data collection module comprises collecting patient data from the hospital electronic medical record system, collecting surgery data from the hospital information system, collecting consumable data from the hospital supply chain management system and obtaining hospital data from the hospital management system as first data, and pre-processing the first data into second data. The surgical consumable consumption prediction module is configured to input the second data into an XGBoost model and an ARIMA model to predict basic consumable consumption, and then input the second data into a logistic regression model to determine the probability of occurrence of a complication, and when the determined probability of occurrence of a complication is greater than an empirical threshold, input the second data into a support vector regression model to predict fault-tolerant consumable consumption.

[0006] Further, the surgical consumable consumption prediction process is as follows: Step 1-1 Combine the patient features in the no-complication data set in the second data , surgical features , hospital features , and basic NLP features into an input feature vector and input into an XGBoost model; = wherein is the total number of no-complication samples; corresponding no-complication surgical consumable use labels are represented as: = wherein represents the consumption of consumables in the i th no-complication surgery; Step 1-2 Model training is performed using an objective function containing a loss function and a regularization term, and the objective function is: wherein is the loss function, and are regularization coefficients, is the leaf node score of the tree; represents the predicted consumption of consumables in the i th no-complication surgery by the XGBoost model; Step 1-3 The trained XGBoost model is used to predict the basic consumable consumption of no-complication surgeries, and the formula is: wherein represents the predicted basic consumable consumption of the i th no-complication surgery, K is the total number of trees in the model, represents the learning rate, represents the single-tree prediction value of the k th tree for the i th surgery; represents the input features of the i th surgery; Step 1-4 Extract the historical consumable usage records of the past 36 months in the second data complication-free dataset to form a historical consumable usage time series input into the ARIMA model: wherein, represents the total number of the past 36 months complication-free dataset, represents the usage of consumables in the i-th complication-free surgery; Step 1-5 Train the ARIMA model using the minimum mean square error as the objective function, and the objective function is: wherein, represents the actual usage of consumables in the t-th month of complication-free surgery, is the predicted usage of consumables in the t-th month of complication-free surgery by the ARIMA model; Step 1-6 Predict the time series usage of future basic consumables using the trained ARIMA model, and the formula is: wherein, is an autoregressive polynomial; is a moving average polynomial; is a constant term, is a random error term; T represents a lag operator; represents a d-th order difference operation; represents the predicted usage of complication-free surgery consumables; Step 1-7 Weighted fusion of the XGBoost model and the ARIMA model predicted basic consumable usage of complication-free surgery and the time series usage of future basic consumables, and the formula is: = XGBoost prediction value + (1- ) ARIMA prediction value wherein, is the model weight; represents the final predicted basic consumable usage; After the prediction of the basic consumable usage, the probability of complications occurring during surgery is determined by the results of the complication risk assessment.

[0007] Further, the complication risk assessment process for predicting the probability of complications occurring during surgery is: Step 2-1 Extract the patient characteristics , surgical characteristics , hospital features , and basic NLP features As input features, form a feature matrix , input into a logistic regression model; = wherein, is expressed as a feature matrix dimension; Step 2-2 defines the target variable , is a binary label, indicating whether a complication occurs, 1 indicating a complication, and 0 indicating no complication; Step 2-3 uses the maximum likelihood estimation method to train the logistic regression model to obtain the probability of complications during surgery, and the model expression is: wherein, , and represent the model parameters, represents the probability of a risk event occurring when given input features ; If the predicted probability of complications during surgery is less than the empirical threshold, the basic consumable usage and complication risk assessment is pushed to the hospital's electronic medical record system, otherwise the fault-tolerant consumable prediction is performed.

[0008] Further, characterized in that: after obtaining the probability of complications during surgery predicted by the logistic regression model greater than the empirical threshold, the fault-tolerant consumable prediction is performed to predict the usage of surgical equipment when complications occur during surgery, the specific process being: Step 3-1 combines the patient features , surgical features , hospital features , and complication NLP features in the second data set with complications as input features, into an input feature vector input into a support vector regression model; Step 3-2 defines the target variable , indicating the total consumable usage of the surgery with complications; Step 3-3 trains the model by minimizing the sum of the square norm of the weight vector and the penalty term of the slack variable as the objective function, and the objective function formula is: wherein, represents the weight vector, penalty coefficient, total number of samples with complications, relaxation variable; Step 3-4 is mapped to a high-dimensional space by a radial basis function (RBF) to fit the total consumable usage under the complication scenario The formula of the radial basis function is: wherein, kernel coefficient; input feature vector of the i-th sample; input feature vector of the j-th sample; Euclidean distance square between error vectors; wherein, total consumable usage under the complication scenario; Lagrange multiplier; Step 3-5 is the formula for calculating the fault-tolerant consumable usage: Before calculating the fault-tolerant consumable usage, the second data set with complications is used to perform a basic consumable prediction to obtain the basic consumable usage, and then perform a final calculation of the fault-tolerant consumable usage; Fault-tolerant consumable usage = SVR predicted total consumable usage - basic consumable usage obtained by the basic consumable prediction model; Step 3-6 pushes the basic consumable usage, complication risk assessment, fault-tolerant consumable usage, and final comprehensive prediction result to the electronic medical record system of the hospital.

[0009] Further, it is characterized in that: The patient data includes patient basic information, age, gender, BMI, ID number, past medical history, number of past operations, ASA score, and preoperative examination results; The surgery includes surgery number, surgery type, surgery duration, and surgery record; The consumable data includes historical consumable usage records and instructions; The hospital data includes hospital level, department resources, and doctor habits; According to the complication markers in the surgery record, the data is strictly divided into two categories: a complication-free surgery data set and a complication surgery data set .

[0010] Further, the preprocessing includes 1) missing value processing For numerical data, a multivariate regression imputation method is adopted, and the specific process is as follows: For the variable with missing values, a regression model is established using the remaining non-missing values of the variable, and the model is represented as: For the variable with missing values , using the remaining non-missing values of the variable , , , ,..., a regression model is established, and the model is represented as: wherein, represents the prediction of the missing value; , ,..., is the regression coefficient; is the error term; n is the number of remaining non-missing value variables; For categorical data, the mode filling is adopted, that is, the highest frequency category value in each field is used to fill in the missing items, and the categorical data represents a data type without mathematical addition and subtraction operations; 2) Feature extraction Patient features, surgical features, hospital features and consumable features are extracted; 3) NLP processing For the text data of the medical history record, the surgical record, the postoperative summary and the consumable instruction manual in the electronic medical record, first, the medical entity in the text data is extracted by using the BiLSTM-CRF model, including the disease name, the surgical complication, the consumable name and the surgical burst; Then, the BERT sentiment analysis model is used to output the surgical complexity score, and the surgical complexity range is [0, 1]; Finally, the extracted medical entity name and surgical complexity are taken as the basic NLP features , the complication type and the complication severity score are taken as the complication NLP features ; 4) Normalization The first data processed above is scaled to a uniform range to obtain the second data x, and the uniform range is [0, 1].

[0011] Compared with the prior art, the advantages of the present application are: 1. The application integrates multi-source data of hospital EMR, HIS, supply chain management system and hospital management system, covers patient data, operation data, consumable data and hospital data, and strictly divides into two data sets of no complications and complications, forms complete data relationship through ID number and operation number association, solves the problem of single data source and low correlation of traditional technology.

[0012] 2. The application processes text data in electronic medical records through NLP technology, uses BiLSTM-CRF model to identify medical entities, BERT model to output operation complexity score, and converts unstructured information into structured features, providing more abundant basis for prediction, overcoming the limitations of traditional technology relying only on structured data.

[0013] 3. The application constructs a hierarchical prediction system, assesses the risk of complications through a logistic regression model, predicts the basic consumable consumption when there is no complication through an XGBoost and ARIMA fusion model, and predicts the additional fault-tolerant consumable consumption when there is a complication through an SVR model, and dynamically adjusts the prediction results according to the complication risk probability, realizing accurate differentiation between regular and special scenarios.

[0014] 4. The prediction results include basic consumption, complication risk, fault-tolerant consumption and final comprehensive value, which can be directly pushed to the hospital electronic medical record system to assist doctors in preoperative preparation, effectively optimize surgical resource allocation, reduce surgical efficiency problems caused by consumable shortage or excess, and are more scientific and operable than traditional experience-dependent preparation methods. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 is a system overview diagram of the present application; Figure 2 is a system flowchart of the present application; Figure 3 is a surgical consumable consumption prediction module of the present application, wherein each symbol in the figure represents as follows: : complication operation data set, : complication operation data set, : patient characteristics, : operation characteristics, : hospital characteristics, : basic NLP features, : complication NLP features. DETAILED DESCRIPTION

[0017] To achieve the above object, the present application is realized by the following technical scheme, the present application provides a digestive tract surgery consumable consumption prediction system based on big data analysis, system overview, please refer to Figure 1 , specific process, please refer to Figure 2 , the specific implementation process includes: Step one, data collection and integration Data collection: the collected data includes patient data, operation data, consumable data and hospital data, and according to the complication mark in the operation record, the data is strictly divided into two categories: no complication operation data set And complication operation data set ; Patient data is obtained from the electronic medical record system, referred to as EMR system, including patient basic information, age, gender, BMI, ID number, past medical history, past operation times, ASA score and preoperative examination results; Operation data is obtained from hospital information system (HIS), including operation number, operation type, operation time and operation record; Consumable data is obtained from hospital supply chain management system, including historical consumable use record and instruction manual; Hospital data is obtained from hospital management system, including hospital level, department resources and doctor habits; Data integration: The data of the above collected no complication operation data set And complication operation data set Is uniformly converted into JSON data format which can be recognized by the system, through the two key fields of patient ID number and operation number, the patient data, operation data, consumable data and hospital data are associated and integrated, if the data conflict, the patient data is given priority, a multi-dimensional data relationship covering the whole process of patient diagnosis and treatment is constructed, and the integrated data is uniformly stored in distributed data warehouse, the bottom layer storage adopts Hadoop HDFS architecture, which is used as the first data of the system.

[0018] Step two, data preprocessing Data cleaning: after the first data is imported, systematic data cleaning operation is carried out, including missing value processing, feature extraction, NLP processing and normalization; The method for processing data missing value is as follows: For numerical data, multivariate regression interpolation method is adopted, the specific process is as follows: for the variable with missing value, the remaining non-missing value variable is used to establish a regression model, the model is represented as: For the variable with missing value Use the remaining non-missing values ​​of the variable , , , , ..., Establish a regression model, which is expressed as follows: in, This indicates a prediction of missing values; , , ..., These are the regression coefficients; is the error term; n is the number of other non-missing value variables; Feature extraction is performed on the data after missing values ​​have been processed, including patient features, surgical features, hospital features, and consumable features; NLP processing: In the natural language processing (NLP) stage, text data of medical records, surgical records and postoperative summaries in electronic medical records are collected first, and relevant text information in the instructions for use of consumables is extracted at the same time. For entity recognition tasks, the BiLSTM-CRF model is used to annotate and recognize medical entities in text, including disease names, surgical complications, consumable names, and surgical emergencies. For free text in surgical records describing surgical difficulty and patient condition during surgery, a BERT-based sentiment analysis model is used to output a surgical complexity score, with the surgical complexity ranging from [0,1]. Simultaneously, the identified surgical complications are transformed into structured features, including binary indicator features of complication types. and complication severity score To facilitate subsequent model building, the binary indicator feature is a feature encoding method that represents attributes or states using 0 and 1. Finally, the extracted medical entity names and surgical complexity were used as basic NLP features. , characteristics related to complications and NLP characteristics as complications ; The extracted features are normalized, and the normalized data is used as the second set of data. For continuous features with clearly defined ranges and actual physical meaning, including but not limited to patient age and operation duration, after outlier processing using box plots, Min-Max normalization is applied to map them to the [0,1] interval, as shown in the formula: Among them, among them, The first data is represented as The second data is represented as The maximum and minimum values of the data are represented as The second data is stored in JSON format, contains explicit field definitions, and is stored in a distributed data warehouse to support subsequent model reading.

[0019] Step three, consumable usage prediction Consumable usage prediction includes three modules: basic consumable prediction, complication risk assessment, and fault-tolerant consumable prediction, and the specific implementation process is as shown in Figure 3 ; 1) Basic consumable prediction: The basic consumable prediction adopts a combination of XGBoost model and ARIMA model to predict the basic consumable usage under the condition of no complications; First, train the XGBoost model, the specific process is as follows: The patient characteristics , surgical characteristics , hospital characteristics , and basic NLP characteristics in the complication-free data set in the second data are combined into an input feature vector input into the XGBoost model; = Wherein, is the total number of complication-free samples; The corresponding complication-free surgical consumable usage label is represented as: = Wherein, represents the usage amount of consumables in the i-th complication-free surgery; Second, train the XGBoost model, balance the model prediction accuracy and complexity through the objective function containing the loss function and the regularization term, train the model, and the objective function is: Wherein, is the loss function, and are regularization coefficients, is the leaf node score of the tree; represents the usage amount of consumables in the i-th complication-free surgery predicted by the XGBoost model; When predicting the basic consumable usage of the surgery without complications using the trained XGBoost model, the final value is obtained by weighted summation of the prediction results of multiple trees, and the formula is: wherein, represents the predicted value of the basic consumable usage of the i-th surgery without complications, and K is the total number of trees in the model, represents the learning rate, represents the single-tree prediction value of the k-th tree for the i-th surgery; represents the input features of the i-th surgery; Then, the ARIMA model is trained, and the specific process is as follows: First, for each consumable, the historical consumable usage records of the past 36 months are extracted from the non-complication data set in the second data to form a historical consumable usage time series: wherein, represents the total number of non-complication data sets for the past 36 months, represents the usage of consumables in the i-th non-complication surgery; The fitting effect of the model on the historical time series data is measured by minimizing the mean square error as the objective function, that is, the squared sum of the difference between the predicted value of the model for the past 36 months of non-complication surgery basic consumable usage and the actual usage is minimized: wherein, represents the actual consumable usage of the t-th month of non-complication surgery, represents the consumable usage of the t-th month of non-complication surgery predicted by the ARIMA model; Based on the historical consumable usage time series of non-complication surgery, through the autoregressive term for detecting the influence of historical values on current values, the moving average term for detecting the influence of historical errors on current values, and the difference operation for eliminating the trend of the sequence, the time series usage of future basic consumables is predicted: wherein, is an autoregressive polynomial; is a moving average polynomial; is a constant term, is a random error term; T represents a lag operator; represents a d-order difference operation; represents the predicted basic consumable usage of non-complication surgery; ​​The XGBoost model and the ARIMA model are fused: The prediction results of the two models are superimposed according to the weight to obtain the final basic consumable prediction value : = XGBoost prediction value + (1- ) ARIMA prediction value Wherein, is the model weight; After the basic consumable consumption prediction is completed, the probability of occurrence of complications in the operation is judged through the results of the complication risk assessment; 2) Complication risk assessment: The complication risk assessment adopts a logistic regression model to predict the probability of occurrence of complications in the operation, and the specific process is: First, the patient characteristics , operation characteristics , hospital characteristics and basic NLP characteristics in the second data are taken as input characteristics to form a feature matrix , which is input into the logistic regression model; = Wherein, indicates the dimension of the feature matrix ; The target variable is a binary label, indicating whether a complication occurs, 1 indicating a complication, and 0 indicating no complication; The logistic regression model is trained by using the maximum likelihood estimation method to obtain the probability of occurrence of complications, and the model expression is: Wherein, , and represent model parameters, represents the probability of occurrence of the risk event when the input characteristics are given; If the predicted probability of occurrence of complications in the operation is less than the empirical threshold value, the basic consumable consumption and the complication risk assessment are pushed to the electronic medical record system of the hospital, otherwise the fault-tolerant consumable prediction is performed; 3) Fault-tolerant consumable prediction The fault-tolerant consumable prediction adopts a support vector regression (SVR) model to predict the total consumable consumption under the condition of complications, and the fault-tolerant consumable consumption is calculated by the difference between the total consumable consumption and the basic consumable consumption, and the specific process is: First, using the complication dataset Train the SVR model; The patient features , surgical features , hospital features and complication NLP features are combined as input features into an input feature vector Input into the support vector regression model; Second, define the target variable , representing the total consumable usage of the operation with complications; Train the model by minimizing the sum of the square norm of the weight vector and the penalty term of the slack variable as the objective function, and the formula of the objective function is: Where, represents the weight vector, is the penalty coefficient, is the total number of samples with complications, is the slack variable; Map to high-dimensional space by radial basis kernel function (RBF), fit the total consumable usage in the complication scenario , the formula of the radial basis kernel function is: Where, is the kernel coefficient; represents the input feature vector of the i-th sample; represents the input feature vector of the j-th sample; represents the Euclidean distance square between error vectors; Where, represents the total consumable usage in the complication scenario; is the Lagrange multiplier; Finally, the fault-tolerant consumable usage calculation formula is: Fault-tolerant consumable usage = SVR predicted total consumable usage - basic consumable predicted model obtained basic usage; Final comprehensive prediction result = basic consumable predicted usage + fault-tolerant consumable predicted usage; Step 4, if the probability of complications occurring in the operation predicted by the logistic regression model is less than the empirical threshold, push the basic consumable usage and the complication risk assessment to the hospital's electronic medical record system, otherwise push the basic consumable usage, the complication risk assessment, the fault-tolerant consumable usage and the final comprehensive prediction result to the hospital's electronic medical record system to assist the doctor in preoperative material preparation.

[0020] The above description is only specific embodiments of the present application, but the protection scope of the present application is not limited thereto, any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered in the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A system for predicting the amount of consumables used in a digestive tract surgery based on big data analysis, characterized by, The system comprises a data collection module and a surgical consumable consumption prediction module. The data collection module collects patient data from a hospital electronic medical record system, surgical data from a hospital information system, consumable data from a hospital supply chain management system, and hospital data from a hospital management system as first data, and pre-processes the first data into second data. The surgical consumable consumption prediction module inputs the second data into an XGBoost model and an ARIMA model to predict basic consumable consumption, then inputs the second data into a logistic regression model to determine the probability of complications, and when the determined probability of complications is greater than an empirical threshold, inputs the second data into a support vector regression model to predict fault-tolerant consumable consumption.

2. The system of claim 1, wherein, The surgical consumable consumption prediction process is as follows: Step 1-1 Combine patient features from the second data set without complications , procedure features , hospital features , and base NLP features into an input feature vector into an XGBoost model; = wherein, Total number of samples without complications; Corresponding non-complication surgery consumable use label is represented as: = wherein, represents the amount of consumables used in the i-th uncomplicated surgery; Step 1-2: Model training is performed using an objective function comprising a loss function and a regularization term, and the objective function is: wherein, is a loss function, and is a regularization coefficient, is a leaf node score of the tree; represents the amount of consumable used in the i-th complication-free surgery predicted by the XGBoost model. Step 1-3: The trained XGBoost model is used to predict the basic consumable consumption for surgeries without complications, and the formula is: wherein, represents the predicted base consumable usage for the i-th case of complication-free surgery, K is the total number of trees in the model, represents the learning rate, represents the single-tree prediction value of the k-th tree for the i-th case of surgery; represents the input features of the i-th case of surgery; Step 1-4: The historical consumable usage records of the past 36 months in the second data set without complications are extracted to form a historical consumable consumption time series, which is input into an ARIMA model: wherein, represents the total number of data sets without complications for the past 36 months, represents the amount of consumable used in the i-th complication-free surgery; Step 1-5: ARIMA model training is performed using the minimization of mean square error as the objective function, and the objective function is: wherein, represents the actual consumable usage for the tth month without complication surgery, is the ARIMA model predicted consumable usage for the tth month without complication surgery; Step 1-6: The trained ARIMA model is used to predict the time series consumption of future basic consumables, and the formula is: wherein is an autoregressive polynomial; is a moving average polynomial; is the constant term, is the random error term; T denotes the lag operator; denotes the dth order difference operator; denotes the predicted complication-free surgery consumable usage; Step 1-7: The basic consumable consumption predicted by the XGBoost model and the ARIMA model for surgeries without complications and the time series consumption of future basic consumables are weighted and fused, and the formula is: = XGBoost predicted value + (1- ) ARIMA predicted value wherein, are model weights; denotes the final predicted basis consumable usage; After the basic consumable consumption prediction is completed, the probability of complications in surgery is determined by the results of the complication risk assessment.

3. The system of claim 1, wherein, The complication risk assessment process for predicting the probability of complications in surgery is as follows: Step 2-1 patient features in the second data , surgical features , hospital features and basic NLP features are taken as input features to form a feature matrix , which is input into a logistic regression model; = wherein is represented as a feature matrix of dimensionality; Step 2-2 defines the target variable , is a binary label, indicating whether a complication occurs or not, 1 means with complication, 0 means without complication; Step 2-3: The logistic regression model is trained using the maximum likelihood estimation method to obtain the probability of complications in surgery, and the model expression is: wherein, , and denote model parameters, denotes the probability of a risk event occurring given input features x. If the predicted probability of complications in surgery is less than the empirical threshold, the basic consumable consumption and the complication risk assessment are pushed to the hospital's electronic medical record system, otherwise fault-tolerant consumable prediction is performed.

4. The system of claim 1, wherein: After obtaining the probability of complications in surgery predicted by the logistic regression model greater than the empirical threshold, fault-tolerant consumable prediction is performed to predict the surgical equipment consumption when complications occur, and the specific process is as follows: Step 3-1 combine the second data complication dataset with patient features , procedure features , hospital features , and complication NLP features as input features, combined into an input feature vector into a support vector regression model; Step 3-2 defines the target variable represents the total consumable usage for the surgery with complications; Step 3-3: The model is trained by minimizing the sum of the square norm of the weight vector and the penalty term of the slack variable as the objective function, and the objective function formula is: wherein, denotes a weight vector, denotes a penalty coefficient, denotes the total number of samples with complications, denotes a slack variable; Step 3-4: The total consumable usage under the complication scenario is fitted by mapping to a high-dimensional space through a Radial Basis Function (RBF) The formula of the Radial Basis Function is: wherein, is a kernel coefficient; denotes an input feature vector of the i-th sample; denotes an input feature vector of the j-th sample; denotes the squared Euclidean distance between error vectors; wherein, represents the total consumable usage under the complication scenario; represents the Lagrange multiplier; Step 3-5: The fault-tolerant consumable consumption calculation formula is: prior to calculating the fault tolerant consumable usage, utilizing the second data set of complication data performing a base consumable forecast to obtain a base consumable usage, and then performing a final fault tolerant consumable usage calculation; Fault-tolerant consumable consumption = SVR predicted total consumable consumption - basic consumable consumption obtained by the basic consumable prediction model; Final comprehensive prediction result = basic consumable prediction consumption + fault-tolerant consumable prediction consumption; Step 3-6: The basic consumable consumption, complication risk assessment, fault-tolerant consumable consumption, and final comprehensive prediction result are pushed to the hospital's electronic medical record system.

5. The system of claim 1, wherein: The patient data includes patient basic information, age, gender, BMI, ID number, past medical history, number of previous operations, ASA score and preoperative examination results; The operation includes operation number, operation type, operation duration and operation record; The consumable data includes historical consumable use records and instructions; The hospital data includes hospital level, department resources and doctor habits; The data was strictly divided into two classes according to the complication flag in the surgery records: the no-complication surgery dataset and the complication surgery dataset .

6. The system of claim 1, wherein, The preprocessing includes 1) missing value processing For numerical data, multivariate regression imputation method is adopted, and the specific process is as follows: For variables with missing values, a regression model is established using the remaining non-missing variables, and the model is represented as: For variables with missing values , using the remaining non-missing values of the variable , , , ,... A regression model is built, the model is represented as: wherein, represents the prediction of missing values; , ,..., is a regression coefficient; is an error term; n is the number of remaining non-missing values variables; For categorical data, mode filling is adopted, that is, the highest frequency category value in each field is used to fill in the missing items. The categorical data represents a data type that does not have mathematical addition and subtraction operations; 2) feature extraction Extract patient features, operation features, hospital features and consumable features; 3) NLP processing For the text data of the medical history record, operation record, postoperative summary and consumable use instruction in the electronic medical record, first, the medical entity in the text data is extracted by using the BiLSTM-CRF model, including disease name, operation complication, consumable name and operation emergency; Then, the BERT sentiment analysis model outputs the operation complexity score, and the operation complexity range is [0, 1]; Finally, the extracted medical entity names and the surgical complexity are taken as the basic NLP features The complication type and the complication severity score are taken as the complication NLP features ; 4) normalization The first data processed above is scaled to a unified range to obtain the second data x, and the unified range is [0, 1].

Images