A cardiothoracic surgery risk early warning method and system based on artificial intelligence

Abstract

This invention provides an artificial intelligence-based method and system for risk warning in cardiothoracic surgery, belonging to the field of cardiothoracic surgery technology. The system includes: a data input module, which collects basic patient information and reports, real-time monitoring signals from surgical equipment, medical imaging data, and patient physiological index data, generating a time-aligned standardized data stream; a feature processing module, which receives the standardized data stream and generates a fused feature vector based on a unified time axis; and a risk prediction and warning module, which combines patient physiological indicators and fused features, applies a dual-channel neural network pre-trained model to calculate the probability value of complication risks, integrates it into a real-time risk probability table, and triggers a warning by comparing it with a risk threshold table. This invention significantly improves the timeliness and accuracy of identifying and predicting complications in cardiothoracic surgery through real-time standardized alignment of multi-source medical data, independent prediction of complication risks using a dual-channel neural network, and real-time updates of the dynamic probability table.

Description

Technical Field

[0001] This invention relates to the field of cardiothoracic surgery technology, specifically to a method and system for early warning of cardiothoracic surgery risks based on artificial intelligence. Background Technology

[0002] Cardiothoracic surgery relies on high-precision procedures such as cardiopulmonary bypass and vascular anastomosis. During the operation, it is necessary to continuously monitor the patient's electrocardiogram, blood pressure, blood oxygen, and the status of various organs. Since the operation lasts for several hours and there are significant individual differences among patients, it is necessary to establish a real-time risk warning mechanism to analyze and identify the risk of complications and ensure the safety of the operation.

[0003] The core bottlenecks of existing cardiothoracic surgery risk warning technologies lie in three aspects: ① The heterogeneity of multi-source medical data makes integration difficult; the original data formats of different surgical devices differ significantly and lack a unified timeline, forcing reliance on manual timestamp calibration, resulting in long data analysis delays; ② Traditional prediction models couple all physiological indicators into a single computational layer, making it impossible to calculate the corresponding risk probability based on different complications, and thus difficult to achieve independent prediction of the risk probability of multiple complications; ③ Existing systems use a fixed-period static evaluation mechanism, making it difficult to respond in real time to intraoperative emergencies. Therefore, there is an urgent need to construct an intelligent early warning system for cardiothoracic surgery that integrates real-time alignment of multi-source data, accurate prediction of independent complications, and dynamic risk assessment, to achieve an intelligent early warning system from raw data synchronization and feature fusion analysis to real-time early warning triggering. Summary of the Invention

[0004] To address the technical problems mentioned in the background section, this invention proposes an artificial intelligence-based method and system for early warning of risks in cardiothoracic surgery.

[0005] Therefore, the technical solution adopted by the present invention is as follows:

[0006] An artificial intelligence-based risk warning system for cardiothoracic surgery, characterized in that the system includes:

[0007] M1, the data input module, collects raw data including patient basic information and reports, real-time monitoring signals of surgical equipment, medical imaging data, and patient physiological indicators throughout the entire cardiothoracic surgery process; it performs standardized processing on the patient basic information and reports, real-time monitoring signals of surgical equipment, and medical imaging data to generate a standardized data stream, which is then transmitted to the feature processing module and the risk prediction and early warning module.

[0008] M2, the feature processing module, receives the standardized data stream, extracts key features, performs fusion processing based on a unified time axis, generates a fused feature vector, and transmits it to the risk prediction and early warning module.

[0009] The key feature extraction process includes structured data generation, dynamic data generation, and image data generation.

[0010] The unified timeline is defined as follows:

[0011] ;

[0012] in, For the first Each time point The total number of time points. To create a unified timeline set;

[0013] The fused feature vector is represented as follows:

[0014] ;

[0015] in, For the first Structured data at each time point, For the first Dynamic data at each time point. For the first Image data at each time point, For the first The fused feature vector of each time point;

[0016] Finally, the fused feature vector is output. ;

[0017] M3, the risk prediction and early warning module, receives the fused feature vector and the patient's physiological indicators, applies a pre-trained prediction model to calculate the risk probability value of the patient's cardiothoracic surgery complications, and integrates them into a risk probability table; it then compares the results with the risk threshold table to make a judgment and issue an early warning.

[0018] The basic architecture of the prediction model is as follows:

[0019] The input to the prediction model is the patient's physiological indicators;

[0020] Before being input into the prediction model, the patient's physiological indicators need to be synchronized with the fused feature vector over time to generate... ;

[0021] The prediction model adopts a dual-channel neural network structure, which is connected to the fused feature vector and independently calculates the risk probability value of each type of cardiothoracic surgery complication.

[0022] The dual-channel neural network structure has three layers:

[0023] The first layer serves as the input layer, receiving time points. The patient's physiological indicators The patient's physiological indicators are expressed as follows: ,in, For the physiological indicators of patient type c. This represents the total number of the patient's physiological indicators.

[0024] The second layer is the distribution layer, which includes the left channel and the right channel.

[0025] The left channel employs a self-attention model to calculate the correlation matrix among all the patient's physiological indicators. The elements of the correlation matrix serve as the probability of influence between the physiological indicators, and the formula is as follows:

[0026] ;

[0027] in, , The weight matrix is ​​trainable. For feature dimensions; It is an incidence matrix; The first of the correlation matrix Line 1 The elements of the column.

[0028] Indicates the physiological indicators right The probability of its influence;

[0029] The right channel employs a routing mechanism to extract strongly correlated physiological indicators corresponding to specific complications from the patient's physiological indicators and directly allocate them to the corresponding complication clinics in the third layer.

[0030] The allocation of the strongly correlated physiological indicators is accomplished using a predefined medical routing matrix R. This is a binary routing flag, representing the c-th row and the first position of the medical routing matrix R. Column elements;

[0031] The allocation decision is made based on the following routing rules:

[0032] ;

[0033] when When, it indicates the first Patient physiological indicators were assigned to the first A clinic for various complications;

[0034] The third layer is the diagnostic layer, where Z independent complication clinics operate simultaneously. Each clinic studies only the specified complication and independently calculates the risk probability value for each different complication.

[0035] For each different complication, first calculate the weighted outcome value for that complication, for time point p. The formula for calculating the weighted outcome value of each complication is as follows:

[0036] ;

[0037] in, For the first The weighted outcome values ​​of the various complications, For the medical routing matrix List, These are the curing parameters;

[0038] Input the weighted result into the sigmoid function to obtain the risk probability value of the corresponding complication:

[0039] ;

[0040] in, For the first Risk probability value of the complication It is a natural constant.

[0041] Furthermore, the standardization process includes unifying the original data format and correcting the acquisition time point;

[0042] The patient's basic information and reports and the medical imaging data are pre-stored before the cardiothoracic surgery, and the surgical equipment's real-time monitoring signals and the patient's physiological indicators are updated in real time during the cardiothoracic surgery.

[0043] The processed data generates a standardized data stream with consistent time points. The standardized data stream includes standardized and time-aligned basic patient information and reports, real-time monitoring signals from surgical equipment, and medical imaging data.

[0044] Furthermore, the structured data generation process is as follows:

[0045] Based on the standardized data stream, the patient's basic information and the text content in the report are identified, and the text content is set as a string set. ,in, For the first The text content,

[0046] The string set is matched with related medical terms to determine whether the string contains a medical term; the medical term is defined as a set. ,in, For the first A medical term,

[0047] The criteria for judgment are as follows:

[0048] ;

[0049] like Include ,but Conversely, ;

[0050] After a successful match, the relevant medical terms are used to extract associated numerical values, using the following formula:

[0051] ;

[0052] in, For the first Related numerical values;

[0053] Finally, structured data is generated based on the medical terms and associated numerical values. , .

[0054] Furthermore, the dynamic data generation process is as follows:

[0055] Based on the standardized data stream, the real-time monitoring signal of the surgical device is set as... ,in, For the first Real-time monitoring signals from the surgical equipment This refers to the total number of sampling points in the real-time monitoring signal of the surgical equipment.

[0056] The total number of sampling points is divided into time windows based on fixed time intervals, and the formula is as follows:

[0057] ;

[0058] in, The duration of the fixed-time segmented window, Sampling frequency, Number of time windows;

[0059] After obtaining the number of time windows, the signal window is calculated;

[0060] Calculate the range, rate of change, and window mean for each time window.

[0061] The formula for calculating the range is:

[0062] ;

[0063] in, For the first One signal window, For the first One range value;

[0064] The formula for calculating the rate of change is:

[0065] ;

[0066] in, For the first The q-th sampling point of a signal window For the first Rate of change; The number of points contained in a single time window, by Calculated;

[0067] The formula for calculating the window mean is:

[0068] ;

[0069] in, The mean of the window;

[0070] Finally, dynamic data is generated based on the range, rate of change, and fluctuation amplitude. , .

[0071] Furthermore, the image data generation process is as follows:

[0072] Based on the standardized data stream, the medical image data is set as a two-dimensional matrix. The formula for identifying organ outlines in the medical image data and calculating the number of pixels within the organ outlines is as follows:

[0073] ;

[0074] in, For image width, For image height, For pixel coordinates, Number of pixels;

[0075] The physical area and average grayscale value of the outline are calculated based on the number of pixels.

[0076] The formula for calculating its physical area is:

[0077] ;

[0078] in, The width of a single pixel. The height of a single pixel. For the number of pixels, Physical area;

[0079] The formula for calculating its average gray value is:

[0080] ;

[0081] in, For pixel coordinates, coordinates grayscale value at that location The region within the organ's outline. This represents the average grayscale value.

[0082] Finally, by combining the physical area and average gray level, the image data is obtained. , .

[0083] Furthermore, based on the risk probability values ​​of the corresponding complications, a risk probability table with time nodes is compiled.

[0084] Based on the risk probability table and the preset risk threshold table, the risk probability values ​​of all complications are determined sequentially. If a specific complication is higher than the corresponding risk threshold, an early warning is triggered.

[0085] Compared with the prior art, the advantages of the present invention are as follows:

[0086] 1. This invention solves the problem of heterogeneity of raw data by standardizing the processing of the raw data format of surgical equipment and correcting the time point, generating a time-aligned standardized data stream, realizing the real-time synchronous flow of raw data, and avoiding conflicts caused by time delay.

[0087] 2. This invention uses a dual-channel neural network structure to calculate the probability value of complication risk. The left channel analyzes the correlation between physiological indicators, while the right channel realizes the diversion of physiological indicators and allocates them to the corresponding complication clinics. The corresponding complication clinics calculate and output the risk probability value, realizing independent prediction of the probability of multiple complication risks and avoiding the error interference of traditional prediction models.

[0088] 3. This invention integrates prediction results in real time to generate a risk probability table with time nodes. Based on the real-time updated risk probability table, it compares it with the risk threshold table. When the risk probability value exceeds the standard, an early warning scheme is immediately triggered, realizing real-time response to intraoperative emergencies and improving early warning response capabilities. Attached Figure Description

[0089] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0090] Figure 1This is a schematic diagram of the overall workflow of the present invention;

[0091] Figure 2 This is a schematic diagram of the feature processing module of the present invention.

[0092] Figure 3 This is a schematic diagram of the risk prediction and early warning module of the present invention.

[0093] Figure 4 This is a medical image data diagram of the present invention;

[0094] Figure 5 This is a diagram of the real-time detection signals of the surgical device of the present invention. Detailed Implementation

[0095] To achieve the above objectives, the present invention provides an artificial intelligence-based risk warning system for cardiothoracic surgery, comprising:

[0096] M1, the data input module, collects raw data throughout the entire cardiothoracic surgery process, including patient basic information and reports, real-time monitoring signals from surgical equipment, medical imaging data, and patient physiological indicators. It then standardizes the patient basic information and reports, real-time monitoring signals from surgical equipment, and medical imaging data to generate a standardized data stream, which is then transmitted to the feature processing module and the risk prediction and early warning module. (See...) Figure 1 .

[0097] The standardization process includes unifying the original data format and correcting the acquisition time point;

[0098] The patient's basic information, reports, and medical imaging data are pre-stored before the cardiothoracic surgery, and the surgical equipment's real-time monitoring signals and the patient's physiological indicators are updated in real time during the cardiothoracic surgery.

[0099] In this embodiment, the patient's physiological indicators include Blood pressure = 120, blood oxygen = 98%, urine output = 50 ml / h, ... ;

[0100] The processed data generates a standardized data stream with consistent time points. The standardized data stream includes standardized and time-aligned basic patient information and reports, real-time monitoring signals from surgical equipment, and medical imaging data.

[0101] The standardized data stream serves as the initial output and is directly transmitted to the feature processing module and the risk prediction and early warning module.

[0102] M2, the feature processing module, receives the standardized data stream, extracts key features, performs fusion processing based on a unified time axis, generates a fused feature vector, and transmits it to the risk prediction and early warning module; see [link / details]. Figure 2 .

[0103] The key feature extraction process is as follows:

[0104] Based on the standardized data stream, the patient's basic information and the text content in the report are identified, and the text content is set as a string set. ,in, For the first The text content,

[0105] The string set is matched with related medical terms to determine whether the string contains a medical term; the medical term is defined as a set. ,in, For the first A medical term,

[0106] The criteria for judgment are as follows:

[0107] ;

[0108] like Include ,but Conversely, ;

[0109] After a successful match, the relevant medical terms are used to extract associated numerical values, using the following formula:

[0110] ;

[0111] in, For the first Related numerical values;

[0112] For example, the The text content is The corresponding number The medical term is At this time, the first Related values ;

[0113] Finally, structured data is generated based on the medical terms and associated numerical values. , .

[0114] Based on the standardized data stream, the real-time monitoring signal of the surgical device is set as... ,in, For the first Real-time monitoring signals from the surgical equipment This refers to the total number of sampling points in the real-time monitoring signal of the surgical equipment.

[0115] The total number of sampling points is divided into time windows based on fixed time intervals, and the formula is as follows:

[0116] ;

[0117] in, The duration of the fixed-time segmented window, Sampling frequency, Number of time windows;

[0118] After obtaining the number of time windows, the signal window is calculated;

[0119] Calculate the range, rate of change, and window mean for each time window.

[0120] The range value is used to reflect the fluctuation amplitude of the real-time monitoring signal of the surgical device within a time window, and the calculation formula is as follows:

[0121] ;

[0122] in, For the first One signal window, For the first One range value;

[0123] The rate of change is used to reflect the average change of adjacent sampling points within the time window, and its calculation formula is as follows:

[0124] ;

[0125] in, For the first The q-th sampling point of a signal window For the first Rate of change; The number of points contained in a single time window, by Calculated;

[0126] The window mean reflects the stability of the real-time monitoring signal of the surgical device, and its calculation formula is as follows:

[0127] ;

[0128] in, The mean of the window;

[0129] Finally, dynamic data is generated based on the range, rate of change, and fluctuation amplitude. , .

[0130] Based on the standardized data stream, the medical image data is set as a two-dimensional matrix. The formula for identifying organ outlines in the medical image data and calculating the number of pixels within the organ outlines is as follows:

[0131] ;

[0132] in, For image width, For image height, For pixel coordinates, Number of pixels;

[0133] The physical area and average grayscale value of the outline are calculated based on the number of pixels.

[0134] The formula for calculating its physical area is:

[0135] ;

[0136] in, The width of a single pixel. The height of a single pixel. For the number of pixels, Physical area;

[0137] The formula for calculating its average gray value is:

[0138] ;

[0139] in, For pixel coordinates, coordinates grayscale value at that location The region within the organ's outline. This represents the average grayscale value.

[0140] Finally, by combining the physical area and average gray level, the image data is obtained. ,

[0141]

[0142] The unified timeline is defined as follows:

[0143] ;

[0144] in, For the first Each time point The total number of time points. To create a unified timeline set;

[0145] The fused feature vector is represented as follows:

[0146] ;

[0147] in, For the first Structured data at each time point, For the first Dynamic data at each time point. For the first Image data at each time point, For the first The fused feature vector of each time point;

[0148] Finally, the fused feature vector is output. .

[0149] It should be noted that if the patient's basic information report, surgical equipment real-time monitoring signal, or medical imaging data at time point p is used, the corresponding data most recent to time point p will be used to fill in the gaps.

[0150] The fused feature vector is used to represent the dynamic changes in the patient's state throughout the entire cardiothoracic surgery process, and is transmitted to the risk prediction and early warning module as the input basis for calculating the risk probability value.

[0151] M3, the risk prediction and early warning module, receives the fused feature vector and the patient's physiological indicators, applies a pre-trained prediction model to calculate the risk probability value of the patient developing complications during cardiothoracic surgery, and integrates them into a risk probability table; it then compares the result with a risk threshold table to make a judgment and issue an early warning; see Figure 3 .

[0152] The basic architecture of the prediction model is as follows:

[0153] The input to the prediction model is the patient's physiological indicators;

[0154] Before being input into the prediction model, the patient's physiological indicators need to be synchronized with the fused feature vector over time to generate... .

[0155] The prediction model adopts a dual-channel neural network structure, which is connected to the fused feature vector and independently calculates the risk probability value corresponding to each type of cardiothoracic surgery complication, ensuring that the output value falls within the risk probability value range of 0 to 1.

[0156] The dual-channel neural network structure has three layers:

[0157] The first layer serves as the input layer, receiving the patient's physiological indicators at time point p. The patient's physiological indicators are expressed as follows: ;in, For the physiological indicators of patient type c. This represents the total number of the patient's physiological indicators.

[0158] The second layer is the distribution layer, which includes the left channel and the right channel.

[0159] The left channel employs a self-attention model to calculate the correlation matrix among all patient physiological indicators. This correlation matrix reflects underlying patterns, and its elements represent the probability of influence between physiological indicators. The formula is as follows:

[0160] ;

[0161] in, , The weight matrix is ​​trainable. The feature dimension is 64 in this embodiment; It is an incidence matrix; The first of the correlation matrix Line 1 The elements of the column.

[0162] Indicates the physiological indicators right The probability of influence;

[0163] For example, This indicates the blood pressure value. To indicate urine volume, when This indicates that when blood pressure drops, there is an 87% probability that urine output will decrease simultaneously.

[0164] The right channel employs a routing mechanism to extract strongly correlated physiological indicators corresponding to specific complications from the patient's physiological indicators and directly allocate them to the corresponding complication clinics in the third layer.

[0165] The allocation of the strongly correlated physiological indicators is accomplished using a predefined medical routing matrix R. This is a binary routing flag, representing the c-th row and the first position of the medical routing matrix R. Column elements;

[0166] The allocation decision is made based on the following routing rules:

[0167] ;

[0168] when When, it indicates the first Patient physiological indicators were assigned to the first A clinic for various complications;

[0169] For example, blood pressure and urine output data are assigned to the kidney injury clinic, while blood oxygen and respiratory rate data are assigned to the respiratory clinic.

[0170] The third layer is the diagnostic layer, where Z independent complication clinics operate simultaneously. Each clinic studies only the specified complication and independently calculates the risk probability value for each different complication.

[0171] For each different complication, first calculate the weighted outcome value for that complication, for time point p. The formula for calculating the weighted outcome value of each complication is as follows:

[0172] ;

[0173] in, For the first The weighted outcome values ​​of the complications, For the medical routing matrix List, These are the curing parameters;

[0174] Input the weighted result into the sigmoid function to obtain the risk probability value of the corresponding complication:

[0175] ;

[0176] in, For the first Risk probability value of the complication It is a natural constant.

[0177] The training process of the prediction model includes a dataset construction phase, a learning phase, and a validation phase. The training process is as follows:

[0178] The process of the dataset construction phase is as follows:

[0179] We collected real-life cardiothoracic surgery data from the past 5 years, including preoperative examination records, intraoperative monitoring records, postoperative medication records, and postoperative 72-hour complication records. From these data, we extracted the patients' basic information and reports, real-time monitoring signals from surgical equipment, medical imaging data, and patients' physiological indicators.

[0180] Before extraction, invalid data caused by disconnection of surgical equipment was removed. Occasionally missing data was filled by averaging of nearby time points to ensure data integrity > 98%.

[0181] For example, if blood pressure data is temporarily missing, the average of the blood pressure data from the preceding and following 5 minutes can be used to fill the gap.

[0182] A dataset of at least 10,000 samples was created by cardiothoracic surgeons who marked the occurrence time of all complications and combined with the historical cardiothoracic surgical data. The sample labels in the training set were the risk probability values ​​of the complications, which were frequency values ​​obtained by statistically analyzing real historical cardiothoracic surgical data.

[0183] The dataset was split into training and testing sets in an 8:2 ratio.

[0184] During the learning phase, the prediction model is trained using the training set and tested using the test set. When the prediction model achieves an accuracy rate of ≥85% for three consecutive days, it proceeds to the clinical validation phase.

[0185] If the test accuracy does not meet the target, the learning phase is repeated and the prediction model parameters are adjusted.

[0186] The clinical validation phase includes static testing, dynamic testing, and manual testing.

[0187] The static test was conducted using data from 500 new cardiothoracic surgeries in a blind test. When the complication prediction accuracy was >85%, the static test was passed and the dynamic test was initiated.

[0188] The dynamic test connects to an external cardiothoracic surgery simulation system to verify the predictive model's response capability in emergencies. When an early warning is triggered within 5 minutes, the test proceeds to manual testing.

[0189] The manual test involves inviting cardiothoracic surgeons to compare predictions from 100 high-risk cases. The test is considered passed when the prediction model's results match the clinical judgment of the cardiothoracic surgeons by more than 90%.

[0190] Based on the risk probability values ​​of the corresponding complications, a risk probability table with time points is compiled, as shown in the table below:

[0191]

[0192] The risk probability table serves as a real-time updated risk probability value to capture the evolution of the disease.

[0193] Based on the risk probability table and the preset risk threshold table, the risk probability values ​​of all complications are determined sequentially. If a specific complication is higher than the corresponding risk threshold, an early warning is triggered.

[0194] An example of the risk threshold table is as follows:

[0195]

[0196] Based on the above risk probability table and risk threshold table, it can be seen that if the risk probability value of acute kidney injury is higher than the corresponding risk threshold, the corresponding early warning scheme will be triggered.

[0197] The early warning system includes visual alerts and acoustic prompts.

[0198] The visual alerts will display corresponding red warning boxes on the surgical equipment based on the different types of complications.

[0199] For example, [WARNING]! High risk of acute kidney injury (85%) exceeds the threshold (60%);

[0200] The acoustic prompt emits a 3-second prompt sound once every 2 minutes.

[0201] A method for early warning of cardiothoracic surgery based on artificial intelligence, the method includes the following steps:

[0202] S1. Data Input: Collect basic patient information and reports, real-time monitoring signals of surgical equipment, medical imaging data, and patient physiological index data throughout the entire cardiothoracic surgery process; standardize the patient information and reports, real-time monitoring signals of surgical equipment, and medical imaging data to generate a standardized data stream;

[0203] S2. Feature processing: Receive the standardized data stream, extract key features, perform fusion processing based on a unified time axis, and generate a fused feature vector.

[0204] S3. Risk prediction and early warning: Receive the fused feature vector and the patient's physiological indicators, apply a pre-trained prediction model to calculate the risk probability value of the patient developing various cardiothoracic surgical complications, and integrate them into a risk probability table; determine the risk based on the risk threshold table.

[0205] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A risk warning system for cardiothoracic surgery based on artificial intelligence, characterized in that, include: M1, the data input module, collects raw data including patient basic information and reports, real-time monitoring signals of surgical equipment, medical imaging data, and patient physiological indicators throughout the entire cardiothoracic surgery process; it performs standardized processing on the patient basic information and reports, real-time monitoring signals of surgical equipment, and medical imaging data to generate a standardized data stream, which is then transmitted to the feature processing module and the risk prediction and early warning module. M2, the feature processing module, receives the standardized data stream, extracts key features, performs fusion processing based on a unified time axis, generates a fused feature vector, and transmits it to the risk prediction and early warning module. The key feature extraction process includes structured data generation, dynamic data generation, and image data generation. The unified timeline is defined as follows: ； in, For the first Each time point The total number of time points. To create a unified timeline set; The fused feature vector is represented as follows: ； in, For the first Structured data at each time point, For the first Dynamic data at each time point. For the first Image data at each time point, For the first The fused feature vector of each time point; Finally, the fused feature vector is output. ; M3, the risk prediction and early warning module, receives the fused feature vector and the patient's physiological indicators, applies a pre-trained prediction model to calculate the risk probability value of the patient's cardiothoracic surgery complications, and integrates them into a risk probability table; it then compares the results with the risk threshold table to make a judgment and issue an early warning. The basic architecture of the prediction model is as follows: The input to the prediction model is the patient's physiological indicators; Before being input into the prediction model, the patient's physiological indicators need to be synchronized with the fused feature vector over time to generate... ; The prediction model adopts a dual-channel neural network structure, which is connected to the fused feature vector and independently calculates the risk probability value of each type of cardiothoracic surgery complication. The dual-channel neural network structure has three layers: The first layer serves as the input layer, receiving the patient's physiological indicators at time point p. The patient's physiological indicators are expressed as follows: ;in, For the physiological indicators of patient type c. This represents the total number of the patient's physiological indicators. The second layer is the distribution layer, which includes the left channel and the right channel. The left channel employs a self-attention model to calculate the correlation matrix among all patient physiological indicators. The elements of the correlation matrix serve as the probability of influence between physiological indicators, and the formula is as follows: ； in, , The weight matrix is ​​trainable. For feature dimensions; It is an incidence matrix; The first of the correlation matrix Line number Column elements; Indicates the physiological indicators right The probability of its influence; The right channel employs a routing mechanism to extract strongly correlated physiological indicators corresponding to specific complications from the patient's physiological indicators and directly allocate them to the corresponding complication clinics in the third layer. The allocation of the strongly correlated physiological indicators is accomplished using a predefined medical routing matrix R. This is a binary routing flag, representing the c-th row and the first position of the medical routing matrix R. Column elements; The allocation decision is made based on the following routing rules: ； when When, it indicates the first Patient physiological indicators were assigned to the first A clinic for various complications; The third layer is the diagnostic layer, where Z independent complication clinics operate simultaneously. Each clinic studies only the specified complication and independently calculates the risk probability value for each different complication. For each different complication, first calculate the weighted outcome value for that complication, for time point p. The formula for calculating the weighted outcome value of each complication is as follows: ； in, For the first The weighted outcome values ​​of the various complications, For the medical routing matrix List, These are the curing parameters; Input the weighted result into the sigmoid function to obtain the risk probability value of the corresponding complication: ； in, For the first Risk probability value of various complications It is a natural constant.

2. The system according to claim 1, characterized in that, The standardization process includes unifying the original data format and correcting the acquisition time point; The patient's basic information and reports and the medical imaging data are pre-stored before the cardiothoracic surgery, and the surgical equipment's real-time monitoring signals and the patient's physiological indicators are updated in real time during the cardiothoracic surgery. The processed data generates a standardized data stream with consistent time points. The standardized data stream includes standardized and time-aligned basic patient information and reports, real-time monitoring signals from surgical equipment, and medical imaging data.

3. The system according to claim 2, characterized in that, The structured data generation process is as follows: Based on the standardized data stream, the patient's basic information and the text content in the report are identified, and the text content is set as a string set. ,in, For the first The text content, The string set is matched with related medical terms to determine whether the string contains a medical term; the medical term is defined as a set. ,in, For the first A medical term, The criteria for judgment are as follows: ； like Include ,but Conversely, ; After a successful match, the relevant medical terms are used to extract associated numerical values, using the following formula: ； in, For the first Related numerical values; Finally, structured data is generated based on the medical terms and associated numerical values. , .

4. The system according to claim 3, characterized in that, The dynamic data generation process is as follows: Based on the standardized data stream, the real-time monitoring signal of the surgical device is set as... ,in, For the first Real-time monitoring signals from the surgical equipment This refers to the total number of sampling points in the real-time monitoring signal of the surgical equipment. The total number of sampling points is divided into time windows based on fixed time intervals, and the formula is as follows: ； in, The duration of the fixed-time segmented window, Sampling frequency, Number of time windows; After obtaining the number of time windows, the signal window is calculated; Calculate the range, rate of change, and window mean for each time window. The formula for calculating the range is: ； in, For the first One signal window, For the first One range value; The formula for calculating the rate of change is: ； in, For the first The first signal window One sampling point, For the first Rate of change; The number of points contained in a single time window, by Calculated; The formula for calculating the window mean is: ； in, The mean of the window; Finally, dynamic data is generated based on the range, rate of change, and fluctuation amplitude. , .

5. The system according to claim 3, characterized in that, The image data generation process is as follows: Based on the standardized data stream, the medical image data is set as a two-dimensional matrix. The formula for identifying organ outlines in the medical image data and calculating the number of pixels within the organ outlines is as follows: ； in, For image width, For image height, For pixel coordinates, Number of pixels; The physical area and average grayscale value of the outline are calculated based on the number of pixels. The formula for calculating its physical area is: ； in, The width of a single pixel. The height of a single pixel. For the number of pixels, Physical area; The formula for calculating its average gray value is: ； in, For pixel coordinates, coordinates grayscale value at that location The region within the organ's outline. This represents the average grayscale value. Finally, by combining the physical area and average gray level, the image data is obtained. , .

6. The system according to claim 1, characterized in that, Based on the risk probability values ​​of the corresponding complications, a risk probability table with time nodes is compiled. Based on the risk probability table and the preset risk threshold table, the risk probability values ​​of all complications are determined sequentially. If a specific complication is higher than the corresponding risk threshold, an early warning is triggered.

Images