Intelligent gastrointestinal tract postoperative complication early warning system and method thereof

Abstract

The invention relates to the technical field of medical information systems, in particular to an intelligent gastrointestinal tract postoperative complication early warning system and a method thereof.The intelligent gastrointestinal tract postoperative complication early warning system comprises a data acquisition module, a data preprocessing module, a deep learning model module in communication connection with the data acquisition module, and a knowledge distillation module in communication connection with the data preprocessing module; the expert knowledge fusion module is in communication connection with the deep learning model module and is used for fusing expert knowledge into a mixed deep learning model; generating a lightweight prediction model; the dynamic weight adjustment module is in communication connection with the deep learning model module and the knowledge distillation module; the multi-task learning module is in communication connection with the deep learning model module; the early warning output module is in communication connection with the knowledge distillation module and the multi-task learning module and is used for generating complication risk early warning based on an output result of the lightweight prediction model; an early warning result and interpretive analysis are displayed through a visual interface, so that the early warning accuracy is improved, and more comprehensive patient state evaluation is provided for doctors.

Description

Technical Field

[0001] The present invention relates to the field of medical information technology, in particular to an intelligent early warning system and method for postoperative complications of the gastrointestinal tract. Background Art

[0002] In recent years, with the continuous improvement of China's medical and health system, maternal and child health care services have gradually become a social focus. With the continuous progress of medical technology, gastrointestinal surgery has become an important means for treating various digestive system diseases. However, postoperative complications are still the key factors affecting the prognosis of patients. Traditional methods for early warning of postoperative complications mainly rely on doctors' experience judgment and routine monitoring indicators, and such methods often have problems such as strong subjectivity and poor timeliness of early warning. In recent years, with the wide application of artificial intelligence technology in the medical field, postoperative complication early warning systems based on machine learning have gradually become a research hotspot.

[0003] Existing intelligent early warning systems usually adopt a single machine learning model, such as Support Vector Machine (SVM) or Random Forest, etc., to predict the risk of postoperative complications. Although these systems have improved the accuracy of early warning to a certain extent, there are still many limitations. First, a single model is difficult to make full use of multi-source heterogeneous medical data, such as structured vital sign data, unstructured medical images, and medical record texts. Second, existing systems often ignore the importance of temporal information and cannot effectively capture the dynamic changes of the patient's state over time. Third, most systems lack effective integration of expert knowledge and are difficult to organically combine clinical experience with machine learning models. In addition, existing systems usually adopt a fixed model structure and weights, lack the ability to dynamically adjust the importance of different data sources, and are difficult to adapt to the special conditions of individual patients.

[0004] On the other hand, existing early warning systems often treat complication prediction as a single task, ignoring the possible correlations between different types of complications. This method not only limits the learning ability of the model but also cannot provide a comprehensive risk assessment for clinicians. At the same time, most systems lack sufficient interpretability when outputting early warning results, and doctors are difficult to understand and trust the prediction basis of the model, which to a certain extent limits the application of early warning systems in clinical practice.

[0005] In view of the above problems, there is an urgent need to develop an intelligent early warning system for postoperative complications of the gastrointestinal tract that can comprehensively utilize multi-source heterogeneous data, integrate expert knowledge, have dynamic adaptability, simultaneously predict the risks of multiple complications, and provide interpretable early warning results. Summary of the Invention

[0006] The present invention aims to solve the above problems existing in the prior art, and provides an intelligent early warning system and method for postoperative complications of the gastrointestinal tract. By innovatively integrating technologies such as multi-source heterogeneous data processing, deep learning model construction, knowledge distillation, dynamic weight adjustment, and multi-task learning, the system realizes efficient and accurate early warning of postoperative complications of the gastrointestinal tract.

[0007] The present invention proposes an intelligent early warning system for postoperative complications of the gastrointestinal tract, including:

[0008] A data acquisition module, used for:

[0009] Collecting multi-source heterogeneous data of patients, including structured data and unstructured data;

[0010] Obtaining the risk assessment results of experts on patients;

[0011] A data preprocessing module, communicatively connected to the data acquisition module, used for:

[0012] Receiving the multi-source heterogeneous data sent by the data acquisition module;

[0013] Performing standardization processing and feature extraction on the multi-source heterogeneous data;

[0014] A deep learning model module, communicatively connected to the data preprocessing module, used for:

[0015] Based on the data after the standardization processing and feature extraction, constructing a hybrid deep learning model;

[0016] Using a temporal convolutional network (TCN) and a long short-term memory network (LSTM) to process the data;

[0017] A knowledge distillation module, communicatively connected to the deep learning model module, used for:

[0018] Integrating expert knowledge into the hybrid deep learning model;

[0019] Generating a lightweight prediction model;

[0020] A dynamic weight adjustment module, communicatively connected to the deep learning model module and the knowledge distillation module, used for:

[0021] Dynamically evaluating the importance of different data sources;

[0022] Adjusting the weights of each data source in the model according to the evaluation results;

[0023] A multi-task learning module, communicatively connected to the deep learning model module, used for:

[0024] Simultaneously predicting multiple related complication risks;

[0025] Optimize the overall prediction performance of the model;

[0026] An early warning output module, communicatively connected to the knowledge distillation module and the multi-task learning module, for:

[0027] Generate a complication risk warning based on the output result of the lightweight prediction model;

[0028] Display the warning result and the interpretive analysis through a visualization interface.

[0029] Preferably, the data acquisition module includes:

[0030] A structured data acquisition unit, for acquiring the patient's vital sign parameters, laboratory test data, and clinical event records;

[0031] An unstructured data acquisition unit, for acquiring the patient's medical images and electronic medical record data;

[0032] An expert evaluation data acquisition unit, for obtaining the evaluation results of medical experts on the risk of postoperative complications of the patient;

[0033] Among them, the structured data acquisition unit, the unstructured data acquisition unit, and the expert evaluation data acquisition unit are all communicatively connected to the data preprocessing module, for transmitting the acquired data to the data preprocessing module.

[0034] Preferably, the data preprocessing module includes:

[0035] A data standardization unit, for uniformly formatting the data from different sources;

[0036] A feature extraction unit, for:

[0037] Automatically learn hidden features from structured data using an LSTM neural network;

[0038] Use a ResNet model to identify and extract lesion features from medical images;

[0039] Adopt the ICD-10 coding system to standardize the coding of electronic medical record texts;

[0040] A data fusion unit, for integrating the multi-source data after standardization and feature extraction to generate a unified feature vector.

[0041] Preferably, the deep learning model module includes:

[0042] A TCN sub-module, for capturing the long-term dependencies of time series data;

[0043] An LSTM sub-module, for processing variable-length sequence data;

[0044] An attention mechanism unit, configured to:

[0045] Calculate the importance weights of different features;

[0046] Fuse the outputs of the TCN sub-module and the LSTM sub-module;

[0047] Wherein, the TCN sub-module and the LSTM sub-module process the input data in parallel, and their outputs are fused by the attention mechanism unit to form the final deep learning feature representation.

[0048] Preferably, the knowledge distillation module includes:

[0049] A teacher model unit, configured to make predictions based on a complete deep learning model;

[0050] A student model unit, configured to build a lightweight prediction model;

[0051] A knowledge transfer unit, configured to:

[0052] Calculate the difference between the outputs of the teacher model and the student model;

[0053] Transfer the knowledge of the teacher model to the student model by minimizing this difference;

[0054] Wherein, knowledge transfer adopts the temperature-scaled soft label method to retain the uncertainty information of the teacher model's decision.

[0055] Preferably, the dynamic weight adjustment module includes:

[0056] A weight initialization unit, configured to assign initial weights to each data source or feature;

[0057] A performance evaluation unit, configured to evaluate the model performance on the training and test sets;

[0058] A weight update unit, configured to:

[0059] Calculate new weights based on the ratio of the mean square error and the cross-entropy loss;

[0060] Gradually adjust the weights using a dynamic learning rate to avoid drastic fluctuations;

[0061] Wherein, weight update adopts a batch update strategy and is updated once after processing a certain number of samples.

[0062] Preferably, the multi-task learning module includes:

[0063] A task decomposition unit, configured to decompose the complication prediction problem into multiple related sub-tasks;

[0064] A shared representation learning unit for learning the underlying feature representations shared by multiple tasks;

[0065] Task-specific layer units for constructing dedicated output layers for each subtask;

[0066] A multi-task loss calculation unit for:

[0067] Calculating the loss functions for each subtask;

[0068] Performing a weighted sum of the losses of multiple tasks through learnable weight coefficients;

[0069] Among them, the shared representation learning unit uses a hard parameter sharing mechanism, while the task-specific layer unit adopts a soft parameter sharing method.

[0070] Preferably, the warning output module includes:

[0071] A risk quantification unit for converting the model output into specific risk probability values;

[0072] A warning generation unit for:

[0073] Generating warning signals based on a preset risk threshold;

[0074] Adopting different warning strategies for different levels of risk;

[0075] A visualization unit for:

[0076] Generating a graphical display of risk warnings;

[0077] Providing an interpretability analysis of the model decision-making process;

[0078] A warning push unit for sending warning information to relevant medical staff in real time.

[0079] Preferably, the intelligent early warning system for postoperative complications of the gastrointestinal tract further includes:

[0080] A preoperative risk assessment module, communicatively connected to the data acquisition module and the deep learning model module, for:

[0081] Performing a preliminary risk assessment based on the various physiological indicators of the patient before surgery;

[0082] Generating suggestions for preoperative medication and examinations;

[0083] A model update module, communicatively connected to the deep learning model module and the knowledge distillation module, for:

[0084] Regularly retraining the model using newly added clinical data;

[0085] Evaluate the performance change of the model and decide whether to update the deployed prediction model;

[0086] The system monitoring module, communicatively connected to all other modules, is used for:

[0087] Monitor the running status of each component of the system in real time;

[0088] Trigger the alarm and recovery mechanism in case of an anomaly.

[0089] The intelligent gastrointestinal postoperative complication early warning method based on the system includes the following steps:

[0090] S1, Collect multi-source heterogeneous data of patients and expert risk assessment results through the data collection module;

[0091] S2, Use the data preprocessing module to perform standardization processing and feature extraction on the collected data;

[0092] S3, Build and train a hybrid deep learning model based on TCN and LSTM in the deep learning model module;

[0093] S4, Use the knowledge distillation module to integrate expert knowledge into the deep learning model and generate a lightweight prediction model;

[0094] S5, Evaluate the importance of different data sources through the dynamic weight adjustment module and dynamically adjust the weights of each data source in the model;

[0095] S6, Simultaneously predict multiple related complication risks in the multi-task learning module to optimize the overall prediction performance of the model;

[0096] S7, Generate a complication risk warning based on the warning output module and display the warning result and interpretive analysis through a visualization interface;

[0097] Among them, steps S3 to S6 are iteratively executed to continuously optimize the model performance; after step S7, the model is updated regularly to ensure that the system adapts to the latest clinical data characteristics.

[0098] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0099] First of all, the system of the present invention effectively integrates structured vital sign data, unstructured medical images and electronic medical record texts, as well as expert assessment results through multi-source heterogeneous data fusion technology. This comprehensive data integration not only improves the accuracy of early warning, but also provides doctors with a more comprehensive assessment of the patient's status. For example, the system can simultaneously consider the blood pressure change trend of the patient, the postoperative CT image characteristics, and the medication records in the medical record to more accurately identify potential complication risks.

[0100] Secondly, the present invention adopts a hybrid deep learning model based on Temporal Convolutional Network (TCN) and Long Short-Term Memory Network (LSTM). This design fully considers the temporal characteristics of medical data. TCN can effectively capture local temporal dependencies, while LSTM is good at dealing with long-term dependencies. The combination of the two enables the system to better understand the dynamic change process of the patient's state. This is crucial for early detection of the potential risks of postoperative complications, because the early signs of many complications are often reflected in the subtle change trends of certain indicators.

[0101] Furthermore, the present invention skillfully integrates expert experience into the deep learning model through knowledge distillation technology. This not only improves the prediction accuracy of the model, but also enhances the interpretability and credibility of the model. For example, the system can learn the key indicator combinations that experts focus on when evaluating the risk of complications, so as to provide more clinically practical results when giving early warnings. At the same time, knowledge distillation also realizes the lightweight of the model, enabling the system to be quickly deployed and run in resource-constrained environments, which is particularly important for clinical scenarios that require real-time monitoring.

[0102] In addition, the dynamic weight adjustment mechanism of the present invention enables the system to adaptively adjust the importance of different data sources in the model according to their contributions to the prediction performance. This design greatly improves the adaptability and robustness of the system. For example, for some patients, vital sign data may be more indicative, while for other patients, imaging features may be more critical. The system can dynamically adjust the weights of each data source according to the specific situation, so as to provide more personalized and accurate early warnings.

[0103] The present invention also innovatively adopts a multi-task learning framework to simultaneously predict multiple related complication risks. This method not only improves the overall prediction performance of the model, but also provides doctors with a more comprehensive risk assessment. For example, the system can simultaneously evaluate the risks of multiple complications such as infection, bleeding, and anastomotic leakage in patients, and analyze the potential associations between these risks, providing support for formulating comprehensive prevention and treatment plans.

[0104] Finally, the early warning output module of the present invention transforms the complex model prediction results into clinically decision-making support information that is easy to understand and apply through visualization technology and interpretability analysis. This greatly improves the usability and credibility of the early warning results. For example, the system can display the risk levels of different complications through intuitive charts and use SHAP values to explain the key factors leading to high-risk early warnings, helping doctors quickly understand the basis of the early warnings and formulate corresponding intervention measures.

[0105] In summary, the intelligent early warning system for postoperative complications of the gastrointestinal tract of the present invention realizes efficient, accurate, and interpretable early warning of complication risks through innovative technologies such as multi-source heterogeneous data fusion, combination of deep learning and expert knowledge, dynamic weight adjustment, and multi-task learning. This system can not only improve the accuracy and timeliness of early warning, but also provide comprehensive and personalized decision-making support for medical staff, thereby effectively improving the quality of postoperative management and prognosis of patients undergoing gastrointestinal surgery. This has important clinical significance and social value for reducing the incidence of postoperative complications, reducing waste of medical resources, and improving the quality of life of patients. Brief Description of the Drawings

[0106] Figure 1 It is a logical block diagram of the overall system of the present invention. Detailed Embodiments

[0107] Refer to Figure 1 , the present invention provides an intelligent early warning system and method for postoperative complications of the gastrointestinal tract. The system includes a data acquisition module 1, a data preprocessing module 2, a deep learning model module 3, a knowledge distillation module 4, a dynamic weight adjustment module 5, a multi-task learning module 6, and an early warning output module 7. These modules work together to realize the intelligent early warning of postoperative complications of patients undergoing gastrointestinal surgery.

[0108] Specifically, the data acquisition module 1 is used to acquire multi-source heterogeneous data of patients, including structured data and unstructured data, and obtain the risk assessment results of experts on patients. In a preferred embodiment of the present invention, the structured data may include the vital sign parameters of patients (such as body temperature, blood pressure, heart rate, etc.), laboratory test data (such as blood routine, electrolytes, etc.), and clinical event records. The unstructured data may include medical images (such as CT, MRI, etc.) and electronic medical record texts. The expert risk assessment results are usually scores or classifications given by experienced clinicians based on the overall situation of patients.

[0109] The data preprocessing module 2 is communicatively connected to the data acquisition module 1 and is used to receive the multi-source heterogeneous data sent by the data acquisition module 1 and perform standardization processing and feature extraction on these data. Standardization processing is a key step to ensure that data from different sources and different scales can be uniformly processed. For example, for vital sign data, the Z-score standardization method can be used:

[0110]

[0111] Among them, Z is the standardized value, X is the original value, μ is the average value of this indicator, and σ is the standard deviation. Feature extraction is to extract features meaningful for the prediction task from the original data. For structured data, statistical methods (such as mean, variance, trend, etc.) can be used to extract features; for unstructured data, such as medical images, convolutional neural networks can be used to extract deep features.

[0112] The deep learning model module 3 is communicatively connected to the data preprocessing module 2, and a hybrid deep learning model is constructed based on the data after normalization processing and feature extraction. The present invention innovatively uses a temporal convolutional network (TCN) and a long short-term memory network (LSTM) to process the data. TCN is good at capturing local time dependencies, while LSTM is good at dealing with long-term dependencies. The combination of the two can more comprehensively capture the characteristics of time series data.

[0113] In an embodiment of the present invention, the core formula of TCN can be expressed as:

[0114]

[0115] Among them, F(s) is the output feature, f(i) is the convolutional kernel, x is the input sequence, d is the dilation factor, and k is the convolutional kernel size.

[0116] The knowledge distillation module 4 is communicatively connected to the deep learning model module 3, and is used to integrate expert knowledge into the hybrid deep learning model and generate a lightweight prediction model. The core idea of knowledge distillation is to use a complex "teacher model" to guide the learning of a simple "student model". In the present invention, the teacher model can be a complex deep learning model containing expert knowledge, while the student model is a lightweight model with a simple structure but capable of rapid deployment.

[0117] The loss function of knowledge distillation usually includes two parts:

[0118]

[0119] Among them, KL is the KL divergence, which is used to measure the difference in the output distributions of the teacher model and the student model, CE is the cross-entropy loss, which is used to measure the difference between the output of the student model and the true label, z t and z s are the output logits of the teacher model and the student model respectively, T is the temperature parameter, and α is the balance coefficient.

[0120] The dynamic weight adjustment module 5 is communicatively connected to the deep learning model module 3 and the knowledge distillation module 4, and is used to dynamically evaluate the importance of different data sources and adjust the weights of each data source in the model according to the evaluation results. This dynamic adjustment mechanism can enable the model to better adapt to the individual differences of different patients and changes in data quality.

[0121] The weight adjustment can be carried out based on the contribution of each data source to the prediction performance. For example, the weights can be calculated using the following formula:

[0122]

[0123] where w i is the weight of the i-th data source, β i is the contribution measure of this data source to the prediction performance (such as the correlation coefficient or feature importance), and T is the temperature parameter used to control the smoothness of the weight distribution.

[0124] The multi-task learning module 6 is communicatively connected to the deep learning model module 3 and is used to simultaneously predict multiple related complication risks, thereby optimizing the overall prediction performance of the model. Multi-task learning can utilize the commonalities between different tasks and improve the generalization ability of the model. In the early warning of gastrointestinal postoperative complications, related complications may include infection, bleeding, anastomotic leakage, etc.

[0125] The loss function of multi-task learning can be expressed as:

[0126]

[0127] where L i is the loss function of the i-th task, α i is the weight coefficient of this task, and N is the total number of tasks.

[0128] The early warning output module 7 is communicatively connected to the knowledge distillation module 4 and the multi-task learning module 6, and is used to generate a complication risk early warning based on the output result of the lightweight prediction model, and display the early warning result and the interpretive analysis through a visualization interface. The early warning result can include information such as the risk level and specific risk factors, providing decision-making support for medical staff.

[0129] The intelligent gastrointestinal postoperative complication early warning system of the present invention realizes efficient and accurate complication risk early warning by integrating multi-source heterogeneous data, deep learning technology, and expert knowledge. This system can not only improve the accuracy of the early warning, but also provide interpretable early warning results, helping medical staff better understand and apply the early warning information, thereby improving the postoperative management and prognosis of patients.

[0130] In a preferred embodiment of the present invention, the data acquisition module 1 includes a structured data acquisition unit 11, an unstructured data acquisition unit 12, and an expert evaluation data acquisition unit 13. These units work together to ensure that the system can comprehensively collect various relevant data of patients, laying a solid foundation for subsequent early warning analysis.

[0131] The structured data acquisition unit 11 is mainly responsible for collecting the patient's vital sign parameters, laboratory test data, and clinical event records. In an embodiment of the present invention, the vital sign parameters may include body temperature, blood pressure, heart rate, respiratory rate, and blood oxygen saturation, etc. These parameters are usually recorded at regular intervals (such as every 4 hours) to form time series data. The laboratory test data may include indicators such as blood routine, comprehensive biochemical analysis, and coagulation function. The clinical event records may include the time points and detailed information of important clinical events such as medication administration, surgical operations, and occurrence of complications.

[0132] The unstructured data acquisition unit 12 is mainly responsible for collecting the patient's medical images and electronic medical record data. The medical images may include the results of imaging examinations such as CT, MRI, and X-ray before and after surgery. These image data are usually stored in DICOM format, containing a large amount of visual information and metadata. The electronic medical record data may include text information such as doctors' diagnostic records, surgical records, and nursing records. Although these unstructured data are difficult to be directly used for numerical calculations, they often contain rich clinical information and are of great value for accurately predicting the risk of complications.

[0133] The expert evaluation data acquisition unit 13 is mainly responsible for obtaining the evaluation results of medical experts on the risk of postoperative complications in patients. In an embodiment of the present invention, this may be achieved through a standardized evaluation form. For example, experts may need to score the risk of various complications (such as infection, bleeding, anastomotic leakage, etc.) in patients from 1 to 5 points, where 1 point indicates a very low risk and 5 points indicates a very high risk. Such expert evaluation results can convert clinical experience into quantifiable data and provide valuable input for the early warning model of the system.

[0134] Preferably, the structured data acquisition unit 11, the unstructured data acquisition unit 12, and the expert evaluation data acquisition unit 13 are all communicatively connected to the data preprocessing module 2. This design ensures that all the collected data can be timely transmitted to the data preprocessing module 2 for further processing. For example, the structured data may need to perform operations such as missing value filling and outlier processing; the unstructured data may need to perform feature extraction or transformation; the expert evaluation data may need to perform standardization processing. Through this close connection and collaboration, the system of the present invention can efficiently process and integrate various types of data and provide high-quality input for the subsequent deep learning model.

[0135] In another embodiment of the present invention, the data preprocessing module 2 includes a data standardization unit 21, a feature extraction unit 22, and a data fusion unit 23. The design of these units aims to convert heterogeneous data from different sources into a unified format that can be effectively processed by the deep learning model.

[0136] The data normalization unit 21 is mainly responsible for uniformly formatting data from different sources. For numerical vital signs and laboratory test data, methods such as Z-score normalization or Min-Max normalization can be used. For example, for body temperature data, the following Min-Max normalization formula can be used:

[0137]

[0138] where X is the original body temperature value, X min and X max are the minimum and maximum values of all body temperature records of this patient, respectively. This normalization method can map all numerical values to the interval [0,1], which is beneficial to the unified processing of data with different scales.

[0139] The function of the feature extraction unit 22 is more complex, and different feature extraction methods need to be adopted for different types of data. For structured data, the present invention preferably uses an LSTM neural network to automatically learn implicit features. A key advantage of LSTM is its ability to capture long-term dependencies, which is particularly useful for analyzing vital sign data that changes over time. For example, for heart rate data, LSTM can learn that not only the current heart rate value is important, but also the change trend of the heart rate is an important feature.

[0140] For medical image data, the present invention preferably uses a ResNet model for feature extraction. The residual connection design of ResNet enables it to train very deep networks without encountering the problem of gradient disappearance, so as to extract high-level semantic features from complex medical images. In an embodiment of the present invention, a pre-trained ResNet-50 model can be used and then fine-tuned on an image dataset related to gastrointestinal surgery.

[0141] For electronic medical record texts, the present invention adopts the ICD-10 coding system for standardized coding. ICD-10 is an internationally common disease classification system that can convert diagnoses described in natural language into standardized codes. For example, "acute appendicitis" may be coded as "K35". This coding method not only standardizes the text information, but also implies the hierarchical relationship between diseases, which is beneficial for the model to understand the semantic information of diseases.

[0142] The data fusion unit 23 is responsible for integrating the multi-source data after normalization and feature extraction to generate a unified feature vector. In a preferred embodiment of the present invention, an attention mechanism is used to achieve the fusion of multi-source data. The attention mechanism can automatically learn the importance weights of different data sources, thereby achieving more intelligent data fusion. Specifically, the following formula can be used to calculate the fused feature vector:

[0143]

[0144] Among them, F is the fused feature vector, and F i is the feature vector of the i-th data source, and α i is the corresponding attention weight, and N is the number of data sources. The attention weight α i can be learned through a small neural network to ensure that the model can dynamically adjust the importance of different data sources according to specific situations.

[0145] The deep learning model module 3 of the present invention includes a TCN sub-module 31, an LSTM sub-module 32, and an attention mechanism unit 33. This design makes full use of the advantages of different types of neural networks and can capture the characteristics of time series data more comprehensively. The TCN sub-module 31 is mainly used to capture the long-term dependencies of time series data. A key advantage of TCN is that its receptive field can grow exponentially with the increase of network depth, which enables it to effectively process long sequence data. In an embodiment of the present invention, the core operation of TCN can be expressed as:

[0146] H l+1 = f(W l *H l + b l ),

[0147] Among them, H l and H l+1 are the feature maps of the l-th layer and the l+1-th layer respectively, W l is the convolution kernel, * represents the dilated convolution operation, b l is the bias term, and f is the activation function (such as ReLU). The dilation factor of the dilated convolution usually grows exponentially with the increase of the number of layers. For example, it can be set to 2 l .

[0148] The LSTM sub-module 32 is mainly used to process variable-length sequence data. The gating mechanism of LSTM enables it to selectively remember or forget information, which is particularly useful for capturing complex time dependencies.

[0149] The attention mechanism unit 33 is used to fuse the outputs of the TCN sub-module and the LSTM sub-module. In a preferred embodiment of the present invention, a multi-head self-attention mechanism is used to implement this fusion process. Multi-head self-attention can simultaneously focus on different aspects of features, thereby providing a richer representation. Specifically, the attention calculation can be expressed as:

[0150]

[0151] Among them, Q, K, and V are the query, key, and value matrices respectively, and d k is the dimension of the key. Multi-head attention repeats this process multiple times and then concatenates the results:

[0152] MultiHead(Q, K, V) = Concat(head 1 ,..., head h )W O ,

[0153] wherein, and W O are learnable parameter matrices.

[0154] With this design, the deep learning model module 3 of the present invention can make full use of the advantages of TCN and LSTM, and achieve intelligent feature fusion through the attention mechanism, thereby improving the prediction performance of the model. This composite network structure enables the system of the present invention to better adapt to the complex clinical problem of early warning of gastrointestinal postoperative complications, and provides more accurate decision-making support for the postoperative management of patients. The knowledge distillation module 4 of the present invention includes a teacher model unit 41, a student model unit 42, and a knowledge transfer unit 43. This design aims to effectively transfer the knowledge in a complex deep learning model to a more lightweight model, thereby improving the deployment efficiency of the model while maintaining the prediction accuracy.

[0155] The teacher model unit 41 mainly makes predictions based on a complete deep learning model. In a preferred embodiment of the present invention, the teacher model can be a complex network structure including multiple TCN and LSTM layers, and expert knowledge is also integrated. For example, some rules based on medical expert experience can be added to the model, such as certain specific combinations of vital signs may indicate a high-risk state. Although this complex model structure has strong expressive power, it may have a large computational overhead and is not suitable for real-time operation in resource-constrained environments.

[0156] The student model unit 42 is responsible for constructing a lightweight prediction model. In the embodiment of the present invention, the student model can be a simplified neural network, such as a feed-forward neural network including only a few fully connected layers. Although this simplified structure has weaker expressive power, it has high computational efficiency and is suitable for rapid deployment in various hardware environments.

[0157] The role of the knowledge transfer unit 43 is to calculate the difference between the outputs of the teacher model and the student model, and transfer the knowledge of the teacher model to the student model by minimizing this difference. In the present invention, temperature-scaled soft labels are used for knowledge transfer, and this method can retain the uncertainty information of the teacher model's decision. Specifically, the loss function of knowledge distillation can be expressed as:

[0158]

[0159] Among them, KL represents the Kullback-Leibler divergence, which is used to measure the difference in the output distributions of the teacher model and the student model; CE represents the cross-entropy loss, which is used to measure the difference between the output of the student model and the true label; z t and z s are the logits outputs of the teacher model and the student model respectively; T is the temperature parameter, usually set to a value greater than 1, which is used to "soften" the probability distribution; α is the balance coefficient, which is used to adjust the ratio of the distillation loss and the true label loss. Preferably, the present invention adopts a dynamically adjusted temperature parameter. For example, the temperature can be adjusted according to the confidence of the teacher model:

[0160] T = max(1, β(1 - max(p t ))),

[0161] where p t is the output probability of the teacher model, and β is an adjustable hyperparameter. This dynamic adjustment strategy can provide a more "soft" target when the teacher model is uncertain, and a more "hard" target when the teacher model is very certain.

[0162] The dynamic weight adjustment module 5 of the present invention includes a weight initialization unit 51, a performance evaluation unit 52, and a weight update unit 53. This design enables the system to dynamically adjust the importance of different data sources in the model according to their contributions to the prediction performance.

[0163] The weight initialization unit 51 is responsible for assigning initial weights to each data source or feature. In an embodiment of the present invention, the initial weights can be set based on expert knowledge. For example, for the early warning of postoperative complications in the gastrointestinal tract, higher initial weights may be given to surgical-related parameters (such as the duration of surgery, blood loss, etc.), and lower initial weights may be given to some routine examination parameters. The initial weights can be represented as a vector:

[0164] W 0 = [w 1 , w 2 ,..., w n ,

[0165] where n is the number of data sources or features, and w i represents the initial weight of the i-th data source or feature.

[0166] The performance evaluation unit 52 is used to evaluate the model performance on the training and test sets. In the present invention, the performance evaluation adopts a combination of multiple metrics to comprehensively measure the performance of the model in different aspects. For example, accuracy, precision, recall, and F1-score can be considered simultaneously. For an imbalanced binary classification problem such as complication warning, the AUC-ROC (area under the receiver operating characteristic curve) metric can be particularly concerned. The comprehensive performance score can be expressed as:

[0167] Score = λ 1 Accuracy + λ 2 Precision + λ 3 Recall + λ 4 F1 + λ 5 AUC,

[0168] where λ i is the weight coefficient of each metric and can be adjusted according to specific clinical needs.

[0169] The weight update unit 53 is responsible for calculating new weights based on the performance evaluation results and gradually adjusting the weights using a dynamic learning rate to avoid drastic fluctuations. In the preferred embodiment of the present invention, the weight update adopts a gradient-based method. Specifically, the weights can be updated using the following formula:

[0170]

[0171] where and are the weights of the i-th data source or feature in the t-th and t + 1-th iterations respectively, η is the learning rate, and Loss is the loss function of the model. To achieve a dynamic learning rate, the Adam optimizer can be adopted, which can adaptively adjust the learning rate of each parameter:

[0172]

[0173] where and are the unbiased estimates of the first moment and the second moment respectively, and ∈ is a small constant for numerical stability.

[0174] The multi-task learning module 6 of the present invention includes a task decomposition unit 61, a shared representation learning unit 62, a task-specific layer unit 63, and a multi-task loss calculation unit 64. This design can simultaneously learn multiple related tasks, thereby improving the generalization ability and prediction accuracy of the model.

[0175] The task decomposition unit 61 is responsible for decomposing the complication prediction problem into multiple related subtasks. In an embodiment of the present invention, different types of complication predictions can be regarded as different subtasks. For example, the risks of multiple complications such as infection, bleeding, and anastomotic leakage can be predicted simultaneously. This task decomposition can not only provide more fine-grained prediction results but also utilize the correlations between different complications to improve the overall prediction performance.

[0176] The shared representation learning unit 62 is used to learn the underlying feature representations shared by multiple tasks. In the present invention, a hard parameter sharing mechanism is adopted, that is, multiple tasks share the same underlying network structure. This sharing mechanism can be expressed as:

[0177] h = f shared (x; θ shared ),

[0178] where x is the input feature, h is the shared representation, and f shared is the shared network, and θ shared are the shared parameters.

[0179] The task-specific layer unit 63 constructs a dedicated output layer for each subtask. These task-specific layers can capture the unique features of each complication prediction task. In a preferred embodiment of the present invention, the task-specific layers adopt a soft parameter sharing method, that is, the parameters of different tasks, although not exactly the same, are constrained to be similar. This can be achieved by adding a parameter similarity regularization term to the loss function:

[0180]

[0181] where θ i and θ j are the parameters of the i-th and j-th tasks respectively.

[0182] The multi-task loss calculation unit 64 is responsible for calculating the loss function of each subtask and weighted summing the losses of multiple tasks through learnable weight coefficients. In the present invention, the multi-task loss function can be expressed as:

[0183]

[0184] where L i is the loss function of the i-th task, w i is the corresponding weight coefficient, N is the total number of tasks, L reg is the parameter similarity regularization term, and λ is the regularization coefficient. The weight coefficient w i can be learned through the gradient descent method to automatically balance the importance of different tasks.

[0185] Through this design of multi-task learning, the system of the present invention can simultaneously predict the risks of multiple postoperative complications, not only improving the comprehensiveness of prediction, but also enhancing the overall prediction accuracy through mutual learning between tasks. This is of great significance for comprehensive risk assessment and personalized treatment plan formulation in clinical practice. The warning output module 7 of the present invention includes a risk quantification unit 71, a warning generation unit 72, a visualization unit 73, and a warning push unit 74. This design aims to convert the complex model prediction results into clinically decision-making support information that is easy to understand and apply.

[0186] The risk quantification unit 71 is responsible for converting the model output into specific risk probability values. In a preferred embodiment of the present invention, the sigmoid function is used to convert the original output (logits) of the model into a probability value between 0 and 1:

[0187]

[0188] where z is the original output of the model, and P(y = 1|x) is the probability of the occurrence of the complication given the input x. To improve the reliability of risk quantification, the present invention also introduces a calibration technique. For example, the temperature scaling method can be used to calibrate the probability:

[0189]

[0190] where T is the temperature parameter, which can be optimized by the cross-entropy loss on the validation set. The warning generation unit 72 generates warning signals based on a preset risk threshold and adopts different warning strategies for different levels of risk. In the embodiment of the present invention, the risk level can be divided into three levels: low, medium, and high. For example:

[0191] Low risk: 0 ≤ P < 0.3;

[0192] Medium risk: 0.3 ≤ P < 0.7;

[0193] High risk: 0.7 ≤ P ≤ 1;

[0194] For different risk levels, the system will generate warning signals of different levels. For example, for high-risk patients, the system may trigger more frequent monitoring and more proactive preventive measures.

[0195] The visualization unit 73 is responsible for generating a graphical display of the risk warning and providing an interpretability analysis of the model decision-making process. In the present invention, a variety of visualization techniques are adopted to enhance the intuitiveness and interpretability of the warning results. For example, a radar chart can be used to simultaneously display the risk levels of multiple complications.

[0196] In addition, the present invention also uses SHAP (SHapley Additive exPlanations) values to explain the prediction results of the model. SHAP values can quantify the contribution of each feature to the prediction result, thereby helping doctors understand which factors lead to high-risk warnings. The calculation of SHAP values is based on cooperative game theory and can be expressed as:

[0197]

[0198] where φ i is the SHAP value of feature i, N is the set of all features, and v is the mapping function from the feature subset to the model output.

[0199] The warning push unit 74 is responsible for sending the warning information to relevant medical staff in real time. In an embodiment of the present invention, a multi-level warning push mechanism is adopted. For example, for low-risk warnings, it may just make a mark in the electronic medical record system; for medium-risk warnings, it may notify the attending doctor through the in-hospital message system; and for high-risk warnings, it may notify the doctor and nurse team through multiple methods such as text messages and phone calls at the same time. This multi-level push mechanism can ensure that important warning information can be conveyed in a timely manner while avoiding unnecessary disturbances.

[0200] The present invention also includes a preoperative risk assessment module 8, a model update module 9, and a system monitoring module 10, which further enhance the functionality and reliability of the system.

[0201] The preoperative risk assessment module 8 is communicatively connected to the data acquisition module 1 and the deep learning model module 3, and is used to perform a preliminary risk assessment on the patient before surgery. This module analyzes based on various physiological indicators of the patient before surgery and generates suggestions for preoperative medications and examinations. In a preferred embodiment of the present invention, a simplified neural network model is used for preoperative risk assessment, with the input being the patient's basic information (such as age, gender, BMI, etc.) and the main preoperative examination indicators (such as blood routine, coagulation function, etc.). The output of the model includes a preliminary risk assessment of postoperative complications and specific suggestions, such as:

[0202] Low risk: It is recommended to perform routine preoperative preparations;

[0203] Medium risk: It is recommended to strengthen nutritional support and consider additional imaging examinations;

[0204] High risk: It is recommended to supplement vitamin K before surgery, adjust the anticoagulant drug usage plan, and consider postponing the surgery if necessary;

[0205] This kind of preoperative risk assessment can help doctors better formulate personalized surgical plans and postoperative management plans, thereby reducing the risk of complications.

[0206] The model update module 9 is communicatively connected to the deep learning model module 3 and the knowledge distillation module 4, and is responsible for regularly retraining the model with newly added clinical data, evaluating the change in model performance, and deciding whether to update the deployed prediction model. In the present invention, the model update adopts the incremental learning method, that is, on the basis of the original model, new data is used for fine-tuning. This method can adapt to the change in data distribution while retaining the original knowledge. Specifically, the process of model update can be summarized as follows:

[0207] 1. Collect new clinical data and perform preprocessing;

[0208] 2. Use the new data to fine-tune the existing model;

[0209] 3. Evaluate the performance of the updated model on the validation set;

[0210] 4. If the performance is significantly improved, deploy the updated model; otherwise, keep the original model.

[0211] To avoid catastrophic forgetting, the present invention adopts the Elastic Weight Consolidation (EWC) technology during the fine-tuning process. EWC protects the parameters important for the old tasks by adding a regularization term to the loss function:

[0212]

[0213] where L B (θ) is the loss on the new data, are the parameters of the original model, F i is the Fisher information matrix, which is used to measure the importance of the parameters, and λ is a hyperparameter that weighs the importance of the old and new tasks.

[0214] The system monitoring module 10 is communicatively connected to all other modules, and is used to monitor the running status of each component of the system in real time, and trigger the alarm and recovery mechanism when an anomaly occurs. In the embodiment of the present invention, the system monitoring adopts the Distributed Tracing technology, which can comprehensively monitor each microservice component of the system. Specifically, the system monitoring includes the following aspects:

[0215] 1. Performance monitoring: Trace indicators such as the response time and throughput of each module;

[0216] 2. Error monitoring: Record and analyze the anomalies and errors that occur in the system;

[0217] 3. Resource monitoring: Monitor the CPU, memory, and disk usage of the server;

[0218] 4. Data quality monitoring: Detect outliers, missing values, etc. in the input data;

[0219] When an abnormal situation is monitored, the system will automatically trigger the corresponding processing mechanism. For example, in the case of performance degradation, the system may automatically perform load balancing or resource expansion; for data quality problems, the system may start a data cleaning process or notify the data administrator.

[0220] Finally, the intelligent early warning method for postoperative complications of the gastrointestinal tract of the present invention includes the following steps:

[0221] S1. Collect multi-source heterogeneous data of patients and expert risk assessment results through the data collection module 1;

[0222] S2. Use the data preprocessing module 2 to perform standardization processing and feature extraction on the collected data;

[0223] S3. Construct and train a hybrid deep learning model based on TCN and LSTM in the deep learning model module 3;

[0224] S4. Use the knowledge distillation module 4 to integrate expert knowledge into the deep learning model and generate a lightweight prediction model;

[0225] S5. Evaluate the importance of different data sources through the dynamic weight adjustment module 5 and dynamically adjust the weights of each data source in the model;

[0226] S6. Simultaneously predict multiple related complication risks in the multi-task learning module 6 to optimize the overall prediction performance of the model;

[0227] S7. Generate a complication risk warning based on the warning output module 7 and display the warning result and interpretive analysis through a visualization interface.

[0228] In this method, steps S3 to S6 are a cyclic iterative process, and the system will continuously optimize the model performance. After completing step S7, the system will also regularly perform model updates to ensure that the warning model can adapt to the latest clinical data characteristics.

[0229] Through this comprehensive method, the system of the present invention can comprehensively and accurately evaluate the postoperative complication risks of gastrointestinal surgery patients, provide timely and reliable decision-making support for medical staff, and thus effectively improve the postoperative management quality and prognosis of patients.

[0230] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. Intelligent gastrointestinal postoperative complications early warning system, characterized by ,include: Data acquisition module for: Collect multi-source heterogeneous data of patients, including structured data and unstructured data; Obtain expert risk assessment results for patients; A data preprocessing module is connected to the data acquisition module for: Receiving multi-source heterogeneous data sent by the data acquisition module; Performing standardization processing and feature extraction on the multi-source heterogeneous data; A deep learning model module is connected to the data preprocessing module for: Based on the data after the standardization processing and feature extraction, a hybrid deep learning model is constructed; Use the temporal convolutional network TCN and the long short-term memory network LSTM to process the data; The knowledge distillation module is connected to the deep learning model module for: Incorporating expert knowledge into the hybrid deep learning model; Generate lightweight prediction models; A dynamic weight adjustment module is connected to the deep learning model module and the knowledge distillation module for: Dynamically evaluate the importance of different data sources; Adjust the weight of each data source in the model based on the evaluation results; The multi-task learning module is communicatively connected with the deep learning model module and is used to: Simultaneously predict the risk of multiple related complications; Optimize the overall predictive performance of the model; The early warning output module is connected to the knowledge distillation module and the multi-task learning module for: Generate complication risk warning based on the output of the lightweight prediction model; The warning results and explanatory analysis are displayed through a visual interface.

2. The intelligent gastrointestinal postoperative complication early warning system according to claim 1 is characterized in that , the data acquisition module includes: Structured data collection unit, used to collect patients' vital signs parameters, laboratory test data and clinical event records; Unstructured data collection unit, used to collect patients’ medical images and electronic medical record data; Expert evaluation data collection unit, used to obtain the evaluation results of medical experts on the risk of postoperative complications of patients; Among them, the structured data acquisition unit, the unstructured data acquisition unit and the expert evaluation data acquisition unit are all communicatively connected to the data preprocessing module, and are used to transmit the collected data to the data preprocessing module.

3. The intelligent gastrointestinal postoperative complication early warning system according to claim 1 is characterized in that , the data preprocessing module includes: Data standardization unit, used to uniformly format data from different sources; Feature extraction unit, used to: Use LSTM neural network to automatically learn implicit features from structured data; Use the ResNet model to identify and extract lesion features from medical images; The ICD-10 coding system is used to standardize the coding of electronic medical record texts; The data fusion unit is used to integrate the multi-source data after standardization and feature extraction to generate a unified feature vector.

4. The intelligent gastrointestinal postoperative complication early warning system according to claim 1 is characterized in that , the deep learning model module includes: TCN submodule, used to capture long-term dependencies of time series data; LSTM submodule, used to process variable-length sequence data; Attention mechanism unit, used to: Calculate the importance weights of different features; Fusion of the outputs of the TCN submodule and the LSTM submodule; Among them, the TCN submodule and LSTM submodule process the input data in parallel, and their outputs are fused by the attention mechanism unit to form the final deep learning feature representation.

5. The intelligent gastrointestinal postoperative complication early warning system according to claim 1 is characterized in that , the knowledge distillation module includes: The teacher model unit, used to make predictions based on the complete deep learning model; Student model unit, used to build lightweight prediction models; Knowledge transfer unit, used to: Calculate the difference between the output of the teacher model and the student model; By minimizing this difference, the knowledge of the teacher model is transferred to the student model; Among them, knowledge transfer adopts a temperature-scaling soft label method to retain the uncertainty information of the teacher model's decision.

6. The intelligent gastrointestinal postoperative complication early warning system according to claim 1 is characterized in that , the dynamic weight adjustment module includes: Weight initialization unit, used to assign initial weights to each data source or feature; Performance evaluation unit, used to evaluate model performance on training and test sets; Weight update unit, used to: Calculate new weights based on the ratio of mean squared error and cross entropy loss; Use dynamic learning rate to gradually adjust weights to avoid drastic fluctuations; Among them, the weight update adopts a batch update strategy, and an update is performed after processing a certain number of samples.

7. The intelligent gastrointestinal postoperative complication early warning system according to claim 1 is characterized in that , the multi-task learning module includes: Task decomposition unit, used to decompose the complication prediction problem into multiple related subtasks; Shared representation learning unit, used to learn the underlying feature representation shared by multiple tasks; Task-specific layer units, used to build specialized output layers for each subtask; Multi-task loss calculation unit, used for: Calculate the loss function for each subtask; The losses of multiple tasks are weighted summed through learnable weight coefficients; Among them, the shared representation learning unit uses a hard parameter sharing mechanism, while the task-specific layer unit adopts a soft parameter sharing method.

8. The intelligent gastrointestinal postoperative complication early warning system according to claim 1 is characterized in that , the warning output module includes: Risk quantification unit, used to convert model output into specific risk probability values; Warning generation unit, used to: Generate early warning signals based on preset risk thresholds; Adopt different early warning strategies for different levels of risks; Visualization unit for: Generate a graphical display of risk warnings; Provide interpretable analysis of the model’s decision-making process; The early warning push unit is used to send early warning information to relevant medical staff in real time.

9. The intelligent gastrointestinal postoperative complication early warning system according to any one of claims 1 to 8, characterized in that , also includes: The preoperative risk assessment module is communicatively connected with the data acquisition module and the deep learning model module, and is used to: Conduct a preliminary risk assessment based on the patient's preoperative physiological indicators; Generate preoperative medication and testing recommendations; A model updating module is connected to the deep learning model module and the knowledge distillation module for: Regularly retrain the model with new clinical data; Evaluate changes in model performance and decide whether to update deployed predictive models; The system monitoring module communicates with all other modules and is used to: Real-time monitoring of the operating status of each component of the system; Trigger alarm and recovery mechanisms when abnormalities occur.

10. An intelligent gastrointestinal postoperative complication warning method based on the system of claim 9, characterized in that , including the following steps: S1, collects patients’ multi-source heterogeneous data and expert risk assessment results through the data collection module; S2, using the data preprocessing module to perform standardization and feature extraction on the collected data; S3, builds and trains a hybrid deep learning model based on TCN and LSTM in the deep learning model module; S4, uses the knowledge distillation module to integrate expert knowledge into the deep learning model and generate a lightweight prediction model; S5, evaluates the importance of different data sources through the dynamic weight adjustment module, and dynamically adjusts the weight of each data source in the model; S6, predict multiple related complication risks simultaneously in the multi-task learning module to optimize the overall prediction performance of the model; S7, generates complication risk warning based on the warning output module, and displays the warning results and explanatory analysis through a visual interface; Among them, steps S3 to S6 are executed iteratively in a loop to continuously optimize the model performance; after step S7, the model is updated regularly to ensure that the system adapts to the latest clinical data characteristics.

Images