Bladder cancer postoperative recurrence risk prediction system

Abstract

The invention discloses a bladder cancer postoperative recurrence risk prediction system, and relates to the technical field of cancer recurrence risk prediction. The bladder cancer postoperative recurrence risk prediction system comprises a data acquisition module, a data processing module, a feature engineering module, a risk prediction module, a decision support module and a result display module. According to the method, a bladder cancer postoperative recurrence risk prediction model is optimized and improved, that is, an input layer, a full connection layer, a merging layer and a comprehensive processing layer are newly added on a traditional LSTM model, so that the prediction model can fully utilize time sequence data and static feature data; therefore, the influence of time sequence data such as methylation detection and miRNA expression of the patient and static data such as age and gender of the patient on the recurrence risk of the bladder cancer can be fully considered, data can be analyzed from different angles and levels, a more accurate prediction result is provided, the overfitting risk of the prediction model can be reduced, and the prediction accuracy of the bladder cancer is improved. And the generalization ability of the whole prediction model is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of cancer recurrence risk prediction, and in particular to a system for predicting the risk of recurrence after bladder cancer surgery. Background Art

[0002] The bladder cancer postoperative recurrence risk prediction system is an intelligent medical system based on big data and machine learning technology, which aims to help doctors assess the risk of recurrence after surgery. The system provides personalized recurrence risk assessment by analyzing multi-dimensional information such as clinical data, pathological characteristics, and gene expression of patients, thereby guiding doctors to develop more effective treatment plans and follow-up plans.

[0003] The existing method for predicting the recurrence rate of non-muscle invasive bladder cancer uses a muscle invasion judgment model to determine whether a patient has non-muscle invasive bladder cancer. Then, before surgery, the two nodes and an edge between the two nodes are processed based on a graph neural network model to determine the surgical method, tumor invasion degree, tumor spread degree, tumor activity, and tumor health impact. Then, the first recurrence rate of non-muscle invasive bladder cancer is determined based on the surgical method, tumor invasion degree, tumor spread degree, tumor activity, and tumor health impact. Next, the second recurrence rate determination model is used after surgery to determine the second recurrence rate of non-muscle invasive bladder cancer. Finally, based on the non-muscle invasive The first recurrence rate and the second recurrence rate of bladder cancer are used to determine the recurrence rate. The second recurrence rate determination model is a long-short-term neural network model LSTM, and the bladder MR image sequence and urine cytology examination image sequence of the patient at multiple time points after surgery are used for model training and prediction. However, the long-short-term neural network model LSTM still has problems such as gradient disappearance, low training efficiency and limited generalization ability. Moreover, the model is only trained and predicted through image sequence data, but the time series data such as patient methylation detection and miRNA expression and the static data such as patient age, gender and medical history on the risk of bladder cancer recurrence are not considered, which will also make the model prediction accuracy low.

[0004] Based on the above situation, the present invention proposes a bladder cancer postoperative recurrence risk prediction system with high prediction accuracy. Summary of the invention

[0005] In order to overcome the shortcomings of the existing bladder cancer recurrence rate prediction method using the long short-term neural network model LSTM, which still has the problems of gradient vanishing, low training efficiency and limited generalization ability, and only trains the model and predicts through image sequence data, but does not consider the influence of time series data such as patient methylation detection and miRNA expression and static data such as patient age, gender and medical history on the risk of bladder cancer recurrence, which also makes the model prediction accuracy low, the present invention proposes a bladder cancer postoperative recurrence risk prediction system with high prediction accuracy.

[0006] The bladder cancer recurrence risk prediction system after surgery includes: The data collection module is used to collect basic information, clinical data, pathological data, imaging data, surgical record data and genetic testing data of bladder cancer patients from the hospital's electronic medical record system, laboratory information system and imaging system to obtain the input total data set ; Data processing module, used to collect the total input data set Preprocess the data in to obtain a valid input data set , according to whether the data changes dynamically over time, the effective input data set Divide the data that changes dynamically over time into a dynamic input data set , and data that does not change dynamically over time constitutes a static input data set ; Feature engineering module for dynamic input data sets and static input datasets Perform feature extraction, feature selection and feature fusion on the data to obtain a dynamic input feature set and static input feature set , the dynamic input feature set The features in are arranged in chronological order to obtain time series data ; The risk prediction module is used to build a bladder cancer postoperative recurrence risk prediction model based on the LSTM model and improve and optimize it, and train the bladder cancer postoperative recurrence risk prediction model. After the training is completed, the time series data and static input feature set Input the bladder cancer postoperative recurrence risk prediction model to evaluate the recurrence risk and survival duration of bladder cancer patients and generate prediction results; A decision support module is used to provide personalized treatment recommendations and follow-up plans based on the prediction results output by the bladder cancer postoperative recurrence risk prediction model, assisting doctors in optimizing the treatment and management strategies for bladder cancer patients; The result display module is used to provide a visual interface for doctors and researchers. After entering the patient information, they can view the predicted results, view the recommended treatment suggestions and follow-up plans, and generate corresponding reports.

[0007] As a preferred aspect of the invention, the bladder cancer postoperative recurrence risk prediction model includes two input layers, an optimized LSTM layer, a fully connected layer, a merging layer, a comprehensive processing layer, two task-specific layers and two output layers, wherein the two input layers are: a time series input layer for receiving time series data ; Static feature input layer, used to receive static input feature set ; The optimized LSTM layer is used to process time series data , capturing long-term dependencies in time series; the fully connected layer is used to process static input feature sets , and further extract static features; the merging layer is used to merge the outputs of the optimized LSTM layer and the fully connected layer to form a new feature vector; the comprehensive processing layer is used to process the merged feature vector and further extract features; the two task-specific layers are: the recurrence probability prediction layer, which is used to map the output of the comprehensive processing layer to the target value of the recurrence probability prediction task; the survival period prediction layer, which is used to map the output of the comprehensive processing layer to the target value of the survival period prediction task; the two output layers are: the recurrence probability output layer, which is used to output a probability value between 0 and 1, indicating the probability of postoperative recurrence of the patient, using the sigmoid activation function and optimized by the binary cross entropy loss function; the survival period output layer, which is used to output a continuous value, indicating the expected survival time, using the linear activation function and optimized by the mean square error loss function; the optimized LSTM layer, the fully connected layer, the merging layer and the comprehensive processing layer are all connected with a Dropout layer to prevent overfitting, and randomly discard 20% of the neuron outputs.

[0008] As a preferred aspect of the invention, the optimized LSTM layer introduces residual connections on the basis of the traditional LSTM model. The structural change of the LSTM model is to add a residual connection to the output of the LSTM, that is, the input is directly added to the output. The optimized LSTM layer also introduces an attention mechanism on the basis of the traditional LSTM model. The LSTM model dynamically assigns weights to different time steps through the attention mechanism to capture time dependencies.

[0009] As a preferred aspect of the invention, the input total data set The data that change dynamically over time include bladder MR images, urine cytology images, DNA methylation levels, expression levels of some miRNAs, and gene mutation status of bladder cancer patients at different time points after surgery, while the data that do not change dynamically over time include the age, gender, and size, location, and number of tumors removed after surgery of bladder cancer patients.

[0010] As a preferred aspect of the invention, the time series data It is three-dimensional data, and the specific format is ,in is the number of samples, indicating a dynamic input data set The number of independent samples contained in , is the time step, which indicates the length of the time series in each sample, that is, the number of time point data contained in each sample. is the number of features, indicating the number of features contained at each time point.

[0011] As a preferred aspect of the invention, the data preprocessing step specifically includes: Missing value processing: processing empty values ​​or missing values ​​in the data set, filling them with linear interpolation or polynomial interpolation, or deleting the entire row or column containing missing values; Outlier processing: using the Z-Score method or IQR method to identify extreme values ​​in the data set that do not conform to the expected pattern, and processing outliers by deleting or replacing them with the median; Data standardization: standardize data of different dimensions through minimum-maximum standardization, where the calculation formula of minimum-maximum standardization is: ; in Represents the original data value, represents the minimum value in the data set. Represents the maximum value in the data set. Represents the normalized data value.

[0012] As a preferred aspect of the invention, the specific operations of feature extraction, feature selection and feature fusion of data are: Feature extraction: automatically identifying and extracting useful feature representations from raw data, that is, extracting data features contained in image data through image processing technology or pre-trained CNN networks. For structured data and self-reported data, the corresponding data features can be directly read; Feature selection: based on existing medical knowledge and the statistical correlation between the features themselves and the predicted target variable, the extracted features are screened to select features related to recurrence risk and survival duration; Feature fusion combines features from different sources by weighted summation to form new comprehensive features, where the weights are optimized through cross-validation. For data features from different modalities, the multimodal fusion model is used to fuse the data features of various modalities to generate new comprehensive features.

[0013] As a preferred aspect of the invention, the training steps of the bladder cancer postoperative recurrence risk prediction model are specifically as follows: The dynamic and static data sets of previous bladder cancer patients were collected, and the collected data were preprocessed. After the preprocessing, the data were subjected to feature extraction, feature selection, and feature fusion to obtain training static feature sets and training time series data; The training static feature set and training time series data obtained in the above steps are divided into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%, respectively. The training set data is used to train the bladder cancer postoperative recurrence risk prediction model, and the appropriate batch size and number of iterations are set. During the training process, the validation set data is used to monitor the performance of the bladder cancer postoperative recurrence risk prediction model to prevent overfitting. After the training is completed, the test set data is used to evaluate the performance of the bladder cancer postoperative recurrence risk prediction model; According to the performance of the bladder cancer postoperative recurrence risk prediction model, the hyperparameters, namely the number of LSTM units, Dropout ratio, batch size and learning rate, were adjusted, and the performance of the bladder cancer postoperative recurrence risk prediction model was gradually optimized until the performance of the bladder cancer postoperative recurrence risk prediction model reached the set standard.

[0014] As a preferred aspect of the invention, the prediction results include the probability of postoperative recurrence of bladder cancer and the survival duration. The decision support module can comprehensively evaluate the probability of postoperative recurrence and the survival duration of bladder cancer patients and classify them according to preset thresholds, and then provide corresponding treatment recommendations and follow-up plans based on different classifications.

[0015] The present invention has the following advantages: 1. The present invention optimizes and improves the prediction model for the risk of recurrence of bladder cancer after surgery, that is, adds an input layer, a fully connected layer, a merging layer, and a comprehensive processing layer to the traditional LSTM model, so that the prediction model can make full use of time series data and static feature data, thereby fully considering the impact of time series data such as patient methylation detection and miRNA expression and static data such as patient age, gender, and medical history on the risk of recurrence of bladder cancer, and can analyze data from different angles and levels to provide more comprehensive and accurate prediction results, and can also reduce the overfitting risk of the prediction model and improve the generalization ability of the overall prediction model.

[0016] 2. The present invention adds a residual connection to the output of the LSTM layer, that is, directly adds the input to the output, which can effectively alleviate the gradient vanishing problem, thereby maintaining the stability of the gradient, and directly passes the input information to the subsequent layers, so that the model can learn useful features faster, thereby improving the training efficiency of the model, and the model can still maintain good training performance in the deep network; and by introducing the attention mechanism in the LSTM layer, the model will be more focused on the key parts of the input sequence, thereby improving the accuracy of the prediction, and when processing long sequence data, the introduced attention mechanism can also effectively capture long-distance dependencies, and dynamically adjust the attention and weights of different time steps, so that the model can better process input sequences of different lengths, thereby not only improving the training efficiency and prediction performance of the model, but also improving the flexibility and expressiveness of the model.

[0017] 3. The present invention optimizes and improves the prediction model for the risk of recurrence after bladder cancer surgery, that is, two task-specific layers and an output layer are added to the model, so that the prediction model can perform multi-task learning and prediction, that is, while predicting the risk of recurrence after surgery for bladder cancer patients, it can also predict the survival time of bladder cancer patients, so that doctors can consider the patient's recurrence risk and expected survival time at the same time, thereby formulating a more personalized treatment plan for each patient, and hospitals can also reasonably allocate limited medical resources according to the patient's prognosis, thereby improving the efficiency of medical services, and the knowledge sharing mechanism contained in the multi-task learning framework also helps to improve the generalization ability and accuracy of the overall model. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the structure of a system for predicting the risk of postoperative recurrence of bladder cancer used in an embodiment of the present invention. DETAILED DESCRIPTION

[0019] In order to enable persons skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention.

[0020] Example 1, a system for predicting the risk of recurrence after bladder cancer surgery, such as Figure 1 As shown, including: The data collection module is used to collect basic information, clinical data, pathological data, imaging data, surgical record data and genetic testing data of bladder cancer patients from the hospital's electronic medical record system, laboratory information system and imaging system to obtain the input total data set ; Data processing module, used to collect the total input data set Preprocess the data in to obtain a valid input data set , according to whether the data changes dynamically over time, the effective input data set Divide the data that changes dynamically over time into a dynamic input data set , and data that does not change dynamically over time constitutes a static input data set ; Feature engineering module for dynamic input data sets and static input datasets Perform feature extraction, feature selection and feature fusion on the data to obtain a dynamic input feature set and static input feature set , the dynamic input feature set The features in are arranged in chronological order to obtain time series data ; The risk prediction module is used to build a bladder cancer postoperative recurrence risk prediction model based on the LSTM model and improve and optimize it, and train the bladder cancer postoperative recurrence risk prediction model. After the training is completed, the time series data and static input feature set Input the bladder cancer postoperative recurrence risk prediction model to evaluate the recurrence risk and survival duration of bladder cancer patients and generate prediction results; A decision support module is used to provide personalized treatment recommendations and follow-up plans based on the prediction results output by the bladder cancer postoperative recurrence risk prediction model, assisting doctors in optimizing the treatment and management strategies for bladder cancer patients; The result display module is used to provide a visual interface for doctors and researchers. After entering the patient information, they can view the predicted results, view the recommended treatment suggestions and follow-up plans, and generate corresponding reports.

[0021] The total input data set The data that change dynamically over time include, but are not limited to, bladder MR images, urine cytology images, DNA methylation levels, expression levels of some miRNAs, and gene mutation status of bladder cancer patients at different time points after surgery, while the data that do not change dynamically over time include, but are not limited to, the age, gender, and size, location, and number of tumors removed after surgery of bladder cancer patients.

[0022] The time series data It is three-dimensional data, and the specific format is ,in is the number of samples, indicating a dynamic input data set The number of independent samples contained in , is the time step, which indicates the length of the time series in each sample, that is, the number of time point data contained in each sample. is the number of features, indicating the number of features contained at each time point.

[0023] The steps of data preprocessing specifically include: Missing value processing: processing the null values ​​or missing values ​​in the data set. When there are few missing values ​​and a small amount of data, linear interpolation or polynomial interpolation is used to fill them. When there are few missing values ​​and a large amount of data, the entire row containing the missing values ​​is deleted. When the proportion of missing values ​​in a column is very high, the entire column containing the missing values ​​is deleted. Outlier processing: using the Z-Score method or IQR method to identify extreme values ​​in the data set that do not conform to the expected pattern, and processing outliers by deleting or replacing them with the median; Data standardization: standardize data of different dimensions through minimum-maximum standardization, where the calculation formula of minimum-maximum standardization is: ; in Represents the original data value, represents the minimum value in the data set. Represents the maximum value in the data set. Represents the normalized data value.

[0024] The specific operations of feature extraction, feature selection and feature fusion of data are as follows: Feature extraction: automatically identifying and extracting useful feature representations from raw data, that is, extracting data features contained in image data through image processing technology or pre-trained CNN networks. For structured data and self-reported data, the corresponding data features can be directly read; Feature selection: based on existing medical knowledge and the statistical correlation between the features themselves and the predicted target variable, the extracted features are screened to select the features that are highly correlated with the risk of recurrence and the duration of survival from the large amount of data features mentioned above; Feature fusion combines features from different sources by weighted summation to form new comprehensive features, where the weights can be optimized through cross-validation. For data features from different modalities (i.e. different types, such as images and texts), the multimodal fusion model is used to fuse the data features of various modalities to generate new comprehensive features.

[0025] The bladder cancer postoperative recurrence risk prediction model includes two input layers, an optimized LSTM layer, a fully connected layer, a merging layer, a comprehensive processing layer, two task-specific layers and two output layers, wherein the two input layers are: a time series input layer for receiving time series data , such as multiple examination results of bladder cancer patients; static feature input layer, used to receive static input feature set , such as the age, gender, and tumor size of bladder cancer patients; the optimized LSTM layer is used to process time series data , capturing long-term dependencies in time series; the fully connected layer is used to process static input feature sets , extracting high-level representations of static features; the merging layer is used to merge the outputs of the optimized LSTM layer and the fully connected layer to form a new feature vector; the comprehensive processing layer is used to process the merged feature vector and extract higher-level feature representations; the two task-specific layers are: the recurrence probability prediction layer, which is used to map the output of the comprehensive processing layer to the target value of the recurrence probability prediction task; the survival period prediction layer, which is used to map the output of the comprehensive processing layer to the target value of the survival period prediction task; the two output layers are: the recurrence probability output layer, which is used to output a probability value between 0 and 1, indicating the probability of postoperative recurrence of the patient, using the sigmoid activation function and optimized by the binary cross entropy loss function; the survival period output layer, which is used to output a continuous value, indicating the expected survival time, using the linear activation function and optimized by the mean square error loss function; the optimized LSTM layer, the fully connected layer, the merging layer and the comprehensive processing layer are all connected with a Dropout layer to prevent overfitting, and randomly discard 20% of the neuron outputs.

[0026] Compared with the traditional LSTM model, the above prediction model adds an input layer, a fully connected layer, a merging layer, and a comprehensive processing layer, which enables the prediction model to make full use of time series data and static feature data, so as to fully consider the impact of time series data such as patient methylation detection and miRNA expression and static data such as patient age, gender and medical history on the risk of bladder cancer recurrence, and can analyze data from different angles and levels to provide more comprehensive and accurate prediction results, and can also reduce the overfitting risk of the prediction model and improve the generalization ability of the overall prediction model; Moreover, compared with the traditional LSTM model, the above prediction model has added two task-specific layers and one output layer, which enables this prediction model to perform multi-task learning and prediction, that is, it can predict the risk of recurrence of bladder cancer patients after surgery while also predicting the survival time of bladder cancer patients, so that doctors can consider the patient's recurrence risk and expected survival time at the same time, so as to formulate a more personalized treatment plan for each patient, and hospitals can also reasonably allocate limited medical resources according to the patient's prognosis, thereby improving the efficiency of medical services. The knowledge sharing mechanism contained in the multi-task learning framework also helps to improve the generalization ability and accuracy of the overall model.

[0027] The optimized LSTM layer introduces a residual connection on the basis of the traditional LSTM model. The structural change of the LSTM model is to add a residual connection to the output of the LSTM, that is, to add the input directly to the output. This structure can effectively alleviate the gradient vanishing problem, thereby maintaining the stability of the gradient, and directly pass the input information to the subsequent layers through the residual connection, which will enable the model to learn useful features faster, thereby improving the training efficiency of the model and enabling the model to maintain good training performance in the deep network. In addition, the optimized LSTM layer also introduces an attention mechanism on the basis of the traditional LSTM model. The introduced attention mechanism will enable the model to focus more on the key parts of the input sequence, thereby improving the accuracy of the prediction, and when processing long sequence data, the introduced attention mechanism can also effectively capture long-distance dependencies, and dynamically adjust the attention and weights of different time steps, so that the model can better process input sequences of different lengths, thereby not only improving the training efficiency and prediction performance of the model, but also improving the flexibility and expressiveness of the model.

[0028] The training steps of the bladder cancer postoperative recurrence risk prediction model are specifically as follows: Collect dynamic data sets (such as multiple examination results) and static data sets (such as age, gender, tumor size, etc.) of previous bladder cancer patients, and preprocess the collected data. After the preprocessing is completed, perform feature extraction, feature selection and feature fusion on the data to obtain training static feature sets and training time series data; The training static feature set and training time series data obtained in the above steps are divided into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%, respectively. The training set data is used to train the bladder cancer postoperative recurrence risk prediction model, and the appropriate batch size and number of iterations are set. During the training process, the validation set data is used to monitor the performance of the prediction model to prevent overfitting. After the training is completed, the test set data is used to evaluate the performance of the prediction model. According to the performance of this prediction model, the hyperparameters, namely the number of LSTM units, Dropout ratio, batch size and learning rate, are adjusted, and the performance of this prediction model is gradually optimized until the performance of this prediction model reaches the set standard.

[0029] The prediction results include the probability of postoperative recurrence of bladder cancer and the duration of survival. The decision support module can comprehensively evaluate the probability of postoperative recurrence and the duration of survival of bladder cancer patients and classify them according to preset thresholds, and then provide corresponding treatment recommendations and follow-up plans based on different classifications.

[0030] The specific methods for grading the probability of postoperative recurrence and survival duration of bladder cancer patients and providing corresponding treatment recommendations and follow-up plans according to different gradings are as follows: The bins, among which The probability of recurrence is low. The probability of recurrence is medium. The probability of recurrence is high; Short expected survival, The expected survival period is medium. It is divided into long expected survival period; Treatment recommendations and follow-up plans: For patients with a low probability of recurrence and a long expected survival, regular monitoring and conservative treatment are recommended, i.e., local excision surgery or minimally invasive treatment to avoid side effects caused by overtreatment, urine cytology and imaging examinations (such as CT, MRI) are performed every 3 months for 2 years, followed by long-term follow-up once a year; For patients with a moderate probability of recurrence and a moderate expected survival, a combination of surgery, radiotherapy, chemotherapy and other treatments are used, and individualized according to the patient's specific situation. Immunotherapy or other new therapies can also be considered as adjuvant treatment to reduce the risk of recurrence. A comprehensive physical examination, including blood tests and imaging examinations, is required every 2-3 months for at least 3 years, and then gradually extended to once every six months. For patients with a high probability of recurrence and a short expected survival, more aggressive treatment options are used, such as radical cystectomy, systemic chemotherapy combined with targeted therapy or immunotherapy, while for patients who are not suitable for intensive treatment, the focus is on improving quality of life, controlling symptoms and alleviating pain, and providing appropriate palliative care. At the same time, the condition needs to be evaluated every week or every two weeks, and the treatment plan needs to be adjusted in a timely manner; For patients with a low probability of recurrence but a short expected survival, unnecessary medical interventions should be minimized, and the focus should be on improving the quality of life in the remaining time. The existing symptoms should be alleviated, such as pain management and infection control. The frequency of follow-up visits should be adjusted dynamically according to the patient's condition, without excessively disturbing or ignoring the patient's needs. For patients with an intermediate probability of recurrence but a long expected survival, strategies to prevent recurrence should be strengthened, such as taking specific medications or receiving preventive treatments regularly and having a comprehensive checkup every 6 months to ensure that any potential problems are detected in time; For patients with a high probability of recurrence but a long expected survival period, a multidisciplinary team is organized to discuss the best treatment plan, integrate the opinions of experts from various departments, try personalized precision medicine plans, conduct detailed medical examinations once a month, and closely monitor changes in the disease through more frequent follow-up, and adjust treatment strategies in a timely manner.

[0031] It should be understood that those skilled in the art can make improvements or changes based on the above description, and all such improvements and changes should fall within the scope of protection of the appended claims of the present invention. Parts not described in detail in this specification belong to the prior art known to those skilled in the art.

Claims

1. A bladder cancer recurrence risk prediction system after surgery, characterized in that: Included are: The data collection module is used to collect basic information, clinical data, pathological data, imaging data, surgical record data and genetic testing data of bladder cancer patients from the hospital's electronic medical record system, laboratory information system and imaging system to obtain the input total data set ; Data processing module, used to collect the total input data set Preprocess the data in to obtain a valid input data set , according to whether the data changes dynamically over time, the effective input data set Divide the data that changes dynamically over time into a dynamic input data set , and data that does not change dynamically over time constitutes a static input data set ; Feature engineering module for dynamic input data sets and static input datasets Perform feature extraction, feature selection and feature fusion on the data to obtain a dynamic input feature set and static input feature set , the dynamic input feature set The features in are arranged in chronological order to obtain time series data ; The risk prediction module is used to build a bladder cancer postoperative recurrence risk prediction model based on the LSTM model and improve and optimize it, and train the bladder cancer postoperative recurrence risk prediction model. After the training is completed, the time series data and static input feature set Input the bladder cancer postoperative recurrence risk prediction model to evaluate the recurrence risk and survival duration of bladder cancer patients and generate prediction results; A decision support module is used to provide personalized treatment recommendations and follow-up plans based on the prediction results output by the bladder cancer postoperative recurrence risk prediction model, assisting doctors in optimizing the treatment and management strategies for bladder cancer patients; The result display module is used to provide a visual interface for doctors and researchers. After entering the patient information, they can view the predicted results, view the recommended treatment suggestions and follow-up plans, and generate corresponding reports.

2. The bladder cancer postoperative recurrence risk prediction system according to claim 1, characterized in that: The bladder cancer postoperative recurrence risk prediction model includes two input layers, an optimized LSTM layer, a fully connected layer, a merging layer, a comprehensive processing layer, two task-specific layers and two output layers; The two input layers are: time series input layer, which is used to receive time series data ; Static feature input layer, used to receive static input feature set ; The optimized LSTM layer is used to process time series data , capturing long-term dependencies in time series; The fully connected layer is used to process the static input feature set , and further extract static features; The merging layer is used to merge the outputs of the optimized LSTM layer and the fully connected layer to form a new feature vector; The comprehensive processing layer is used to process the merged feature vectors and further extract features; The two task-specific layers are: the recurrence probability prediction layer, which is used to map the output of the comprehensive processing layer to the target value of the recurrence probability prediction task; The life span prediction layer is used to map the output of the comprehensive processing layer to the target value of the life span prediction task; The two output layers are: the recurrence probability output layer, which is used to output a probability value between 0 and 1, indicating the probability of postoperative recurrence of the patient, using the sigmoid activation function and the binary cross entropy loss function for optimization; The survival output layer is used to output a continuous value representing the expected survival time. It uses a linear activation function and is optimized with a mean square error loss function. The optimized LSTM layer, fully connected layer, merging layer, and comprehensive processing layer are all connected with a Dropout layer to prevent overfitting and randomly discard 20% of the neuron outputs.

3. The system for predicting the risk of postoperative recurrence of bladder cancer according to claim 2, characterized in that: The optimized LSTM layer introduces a residual connection on the basis of the traditional LSTM model. The structural change of the LSTM model is to add a residual connection to the output of the LSTM, that is, the input is directly added to the output. The optimized LSTM layer also introduces an attention mechanism on the basis of the traditional LSTM model. The LSTM model dynamically assigns weights to different time steps through the attention mechanism to capture time dependencies.

4. The system for predicting the risk of postoperative recurrence of bladder cancer according to claim 1, characterized in that: The total input data set The data that change dynamically over time include bladder MR images, urine cytology images, DNA methylation levels, expression levels of some miRNAs, and gene mutation status of bladder cancer patients at different time points after surgery, while the data that do not change dynamically over time include the age, gender, and size, location, and number of tumors removed after surgery of bladder cancer patients.

5. The system for predicting the risk of recurrence of bladder cancer after surgery according to claim 1, characterized in that: The time series data It is three-dimensional data, and the specific format is ,in is the number of samples, indicating a dynamic input data set The number of independent samples contained in , is the time step, which indicates the length of the time series in each sample, that is, the number of time point data contained in each sample. is the number of features, indicating the number of features contained at each time point.

6. The system for predicting the risk of recurrence of bladder cancer after surgery according to claim 1, characterized in that: The steps of data preprocessing specifically include: Missing value processing: processing empty values ​​or missing values ​​in the data set, filling them with linear interpolation or polynomial interpolation, or deleting the entire row or column containing missing values; Outlier processing: using the Z-Score method or IQR method to identify extreme values ​​in the data set that do not conform to the expected pattern, and processing outliers by deleting or replacing them with the median; Data standardization: standardize data of different dimensions through minimum-maximum standardization, where the calculation formula of minimum-maximum standardization is: ； in Represents the original data value, represents the minimum value in the data set. Represents the maximum value in the data set. Represents the normalized data value.

7. The system for predicting the risk of postoperative recurrence of bladder cancer according to claim 1, characterized in that: The specific operations of feature extraction, feature selection and feature fusion of data are as follows: Feature extraction: automatically identifying and extracting useful feature representations from raw data, that is, extracting data features contained in image data through image processing technology or pre-trained CNN networks. For structured data and self-reported data, the corresponding data features can be directly read; Feature selection: based on existing medical knowledge and the statistical correlation between the features themselves and the predicted target variable, the extracted features are screened to select features related to recurrence risk and survival duration; feature Fusion combines features from different sources by weighted summation to form new comprehensive features, where the weights are optimized through cross-validation. For data features from different modalities, the multimodal fusion model is used to fuse the data features of various modalities to generate new comprehensive features.

8. The system for predicting the risk of postoperative recurrence of bladder cancer according to claim 1, characterized in that: The training steps of the bladder cancer postoperative recurrence risk prediction model are specifically as follows: The dynamic and static data sets of previous bladder cancer patients were collected, and the collected data were preprocessed. After the preprocessing, the data were subjected to feature extraction, feature selection, and feature fusion to obtain training static feature sets and training time series data; The training static feature set and training time series data obtained in the above steps are divided into a training set, a validation set, and a test set according to the ratio of 70%, 15%, and 15%, respectively. The training set data is used to train the bladder cancer postoperative recurrence risk prediction model, and the appropriate batch size and number of iterations are set. During the training process, the validation set data is used to monitor the performance of the bladder cancer postoperative recurrence risk prediction model to prevent overfitting. After the training is completed, the test set data is used to evaluate the performance of the bladder cancer postoperative recurrence risk prediction model; According to the performance of the bladder cancer postoperative recurrence risk prediction model, the hyperparameters, namely the number of LSTM units, Dropout ratio, batch size and learning rate, were adjusted, and the performance of the bladder cancer postoperative recurrence risk prediction model was gradually optimized until the performance of the bladder cancer postoperative recurrence risk prediction model reached the set standard.

9. The system for predicting the risk of postoperative recurrence of bladder cancer according to claim 1, characterized in that: The prediction results include the probability of postoperative recurrence of bladder cancer and the duration of survival. The decision support module can comprehensively evaluate the probability of postoperative recurrence and the duration of survival of bladder cancer patients and classify them according to preset thresholds, and then provide corresponding treatment recommendations and follow-up plans based on different classifications.

Images