A recall-based kidney cancer prediction method, device, and program product

Abstract

The application relates to the field of intelligent medical treatment, in particular to a kidney cancer prediction method and device based on recall and a program product. The method comprises the following steps: S1, acquiring clinical data; S2, inputting the clinical data into a rag module to obtain a recall vector; the rag module comprises a recall module and a storage module; the recall module comprises a text recall unit and a model recall unit; the clinical data is subjected to similarity calculation with text in the storage module through the text recall unit to obtain similar text data; the similar text data is input into a recall model through the model recall unit to obtain the recall vector; and S3, inputting the recall vector into a large language model to perform kidney cancer prediction and obtain a prediction result. The application integrates vector data and text data to perform kidney cancer prediction, and has good clinical value.

Description

Technical Field

[0001] The present application relates to the field of intelligent medical care, and specifically to a recall-based kidney cancer prediction method, device, program product, and computer-readable storage medium. Background Art

[0002] Kidney cancer is a common malignancy of the urinary system. It often has no obvious symptoms in its early stages, and many patients are not diagnosed until the late stages, when the cancer may have already spread to other parts of the body. There is a growing need to improve early diagnosis, improve patient outcomes, and reduce overscreening. Currently, clinical assessment of kidney cancer prognosis relies heavily on traditional statistical methods and physician experience. However, prognosis is influenced by multiple factors, including pathological type, clinical stage, treatment regimen, and individual patient characteristics. Traditional statistical methods are subject to high subjectivity and lack accuracy. The rapid development of artificial intelligence (AI), particularly breakthroughs in natural language processing and deep learning, has opened up new possibilities for accurate prognosis prediction in kidney cancer. However, existing AI prediction systems are mostly based on a single data source and lack the comprehensive utilization of information from multiple sources. Summary of the Invention

[0003] To address the above issues, the present invention proposes a renal cancer prediction method based on recall, which specifically includes:

[0004] S1. Obtain clinical data;

[0005] S2. Inputting the clinical data into a rag module to obtain a recall vector; the rag module includes a recall module and a storage module, the recall module includes a text recall unit and a model recall unit, the clinical data is similarly calculated by the text recall unit with the text in the storage module to obtain similar text data, and the similar text data is input into the recall model by the model recall unit to obtain a recall vector;

[0006] S3. Inputting the recall vector into a large language model to perform kidney cancer prediction to obtain a prediction result.

[0007] Furthermore, the training process of the recall model is as follows:

[0008] S101, obtaining a first training set and a second training set, wherein the first training set and the second training set respectively include N tasks, where N is a natural number greater than 1;

[0009] S102, initializing the parameters of the recall model to be trained to obtain a first recall model;

[0010] S103, copying the parameters and model structure of the first recall model to obtain a second recall model;

[0011] S104: Input the first training set into the second recall model for training to obtain gradient update data;

[0012] S105 , updating the first recall model based on the gradient update data and obtaining a recall model by training with the second training set.

[0013] Optionally, the gradient update data is gradient data obtained by sequentially training the second recall model on N task data, and the Nth task data training is performed on the basis of the completion of the N-1th task data training.

[0014] Optionally, the gradient update data is obtained by performing N task data training in parallel on the second recall model to obtain N gradient data, and gradient fusion is performed on the N gradient data to obtain the gradient update data.

[0015] Optionally, the gradient fusion includes one or more of the following: weighted summation and addition.

[0016] The step S105 is replaced by: training a second recall model based on the gradient update data and the second training set to obtain second gradient update data, acquiring test data, and training the first recall model updated based on the second gradient update data using the test data to obtain a recall model.

[0017] The rag module also includes a vectorization module, and similar text data is vectorized and converted into text vectors through the vectorization module. The text vectors are input into the recall model through the model recall unit to obtain recall vectors.

[0018] Optionally, the recall module also includes a vector recall unit, the clinical data is converted into a clinical vector through a vectorization module, the clinical vector is calculated by the vector recall unit by performing similarity calculation between the clinical vector and the vector in the storage module to obtain a calculated vector, and the calculated vector is input into the recall model by a model recall unit to obtain a recall vector.

[0019] Optionally, the calculated vector is fused with similar text data and then input into a recall model to obtain a recall vector.

[0020] Optionally, the calculated vector is fused with the text vector and then input into the recall model to obtain the recall vector.

[0021] The recall model includes one or more of the following: convolutional neural network, hole convolutional neural network, residual network, and Transformer.

[0022] Optionally, the recall model includes an input layer, a hidden layer, a self-attention layer, a feedforward layer, a normalization layer, and an output layer. The input data passes through the input layer and the hidden layer in sequence to obtain a hidden vector, and the hidden vector is input to the self-attention layer and the feedforward layer respectively. The output vector of the self-attention layer, the output vector of the feedforward layer, and the hidden vector are fused and input to the normalization layer to obtain a recall vector, which is then output by the output layer.

[0023] Sort the text similarity to get the top L similar texts, sort the vector similarity to get the top L calculated vectors, fuse the top L similar texts with the top L calculated vectors and input them into the recall model to get the recall vector, where L is a natural number greater than 1;

[0024] Optionally, the text similarities are sorted to obtain the first three similar texts, the vector similarities are sorted to obtain the first three calculated vectors, and the first three similar texts and the first three calculated vectors are fused and input into the recall model to obtain the recall vector.

[0025] The large language model includes one or more of the following: BERT, XLNet, GPT-2, ChatGPT, GPT-3, Baidu Wenxin large model, Alibaba Tongyi large model, Huawei Pangu large model, and Tencent Hunyuan large model.

[0026] Optionally, the method further includes data preprocessing, including: data deduplication, data special character processing, and data standardization.

[0027] Optionally, the storage module includes a text storage module and a vector storage module, wherein the data in the text storage module is converted into vectors through a vector model in a vectorization module and stored to obtain a vector storage module.

[0028] The object of the present invention is to provide a computer program product having a computer program or instructions thereon, comprising: the computer program or instructions being executed by a processor to implement the above-mentioned recall-based kidney cancer prediction method.

[0029] The object of the present invention is to provide a computer device comprising a memory, a processor, and a computer program or instructions stored in the memory, wherein the computer program or instructions are executed by the processor to implement the above-mentioned recall-based kidney cancer prediction method.

[0030] An object of the present invention is to provide a computer-readable storage medium having a computer program or instructions stored thereon, which is executed by a processor to implement the above-mentioned recall-based kidney cancer prediction method.

[0031] Advantages of the present invention:

[0032] 1. The present invention processes data features of different types of clinical data by constructing text recall, vector recall, and model recall structures, thereby improving the ability to capture clinical data features, improving the model's learning of important features, and improving the prediction accuracy and credibility of the overall model.

[0033] 2. The recall model of the present invention adopts a training strategy different from the existing one based on the difference in data types. It extracts features from different data independently and updates the recall model only after the features are integrated, so as to avoid the noise impact of different data on the model, improve the feature extraction ability of the model, and obtain high-quality feature data.

[0034] 3. The present invention performs similarity screening on texts before feature extraction, which speeds up the model's learning of important data, saves model training time, and has practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0036] Figure 1 A schematic flow chart of a method for predicting kidney cancer based on recall provided in an embodiment of the present invention;

[0037] Figure 2 A schematic diagram of a renal cancer prediction system based on recall provided by an embodiment of the present invention;

[0038] Figure 3 A schematic diagram of a renal cancer prediction device based on recall provided in an embodiment of the present invention;

[0039] Figure 4 A flowchart of a recall execution provided by an embodiment of the present invention;

[0040] Figure 5 This is a structural diagram of the recall model provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0041] In order to enable those skilled in the art to better understand the solutions of 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.

[0042] In some of the processes described in the specification and claims of the present invention and the above-mentioned figures, multiple operations that appear in a specific order are included, but it should be clearly understood that these operations may not be executed in the order in which they appear in this article or may be executed in parallel. The serial numbers of the operations, such as S101, S102, etc., are only used to distinguish between different operations, and the serial numbers themselves do not represent any execution order. In addition, these processes may include more or fewer operations, and these operations may be executed in sequence or in parallel. It should be noted that the descriptions of "first", "second", etc. in this article are used to distinguish different messages, devices, modules, etc., and do not represent the order of precedence, nor do they limit "first" and "second" to be different types.

[0043] Figure 1 A schematic diagram of a method for predicting kidney cancer based on recall provided by an embodiment of the present invention specifically includes:

[0044] S1. Obtain clinical data;

[0045] In one embodiment, the clinical data is multi-source medical data such as clinical information, pathology reports, and treatment records of kidney cancer patients.

[0046] In one embodiment, the method further includes data preprocessing, including: data deduplication, data special character processing, and data standardization.

[0047] In one specific embodiment, multi-source medical data, including clinical information, pathology reports, and treatment records, is collected from patients with renal cancer. The collected data is cleaned and standardized to ensure consistency and accuracy. This data processing includes:

[0048] Deduplication: For the acquired data, search for the corresponding result data according to the user's patient ID and perform deduplication processing on the found result data;

[0049] Standardization: standardize key information such as medical terms, time, address, abbreviations, etc. in user information and map them to standardized expressions;

[0050] Special character processing: filtering out special characters that have no real meaning, such as emoticons, contained in the original user information obtained.

[0051] In a specific embodiment, the application process of the present invention is as follows Figure 4As shown in the figure, after the user chooses to initiate a prognosis prediction request, the instruction is returned to the backend through the http interface. After receiving the incoming API request, the backend starts to call the system to perform prognosis prediction process processing. The processing flow is as follows: data acquisition and data preprocessing, feature extraction through the rag module, and prediction through the LLM model to obtain the prediction results. After that, the prediction results are returned to the frontend through the http interface to display the prognosis prediction results of the current data, making it convenient for doctors and patients to view the prediction results.

[0052] Specifically, the acquired clinical data is preprocessed to obtain preprocessed data, which is then input into the rag module. The preprocessed data is similarly calculated between the text recall unit and the text in the text storage module in the storage module to obtain similar text data. The similar text data is input into the trained recall model in the model recall unit for feature extraction to obtain a recall vector, which is then input into the large language model for prediction to obtain a prediction result.

[0053] S2. Inputting the clinical data into a rag module to obtain a recall vector; the rag module includes a recall module and a storage module, the recall module includes a text recall unit and a model recall unit, the clinical data is similarly calculated by the text recall unit with the text in the storage module to obtain similar text data, and the similar text data is input into the recall model by the model recall unit to obtain a recall vector;

[0054] In one embodiment, the training process of the recall model is:

[0055] S101, obtaining a first training set and a second training set, wherein the first training set and the second training set respectively include N tasks, where N is a natural number greater than 1;

[0056] S102, initializing the parameters of the recall model to be trained to obtain a first recall model;

[0057] S103, copying the parameters and model structure of the first recall model to obtain a second recall model;

[0058] S104: Input the first training set into the second recall model for training to obtain gradient update data;

[0059] S105, updating the first recall model based on the gradient update data and training the recall model through the second training set;

[0060] In one embodiment, the gradient update data is gradient data obtained by sequentially training the second recall model with N task data, and the Nth task data training is performed on the basis of the completion of the N-1th task data training.

[0061] In one embodiment, the gradient update data is obtained by performing N task data training in parallel on the second recall model to obtain N gradient data, and gradient fusion is performed on the N gradient data to obtain the gradient update data.

[0062] In one embodiment, the gradient fusion includes one or more of the following: weighted summation and addition.

[0063] In one embodiment, S105 is replaced by: training a second recall model based on the gradient update data and the second training set to obtain second gradient update data, acquiring test data and training the first recall model updated based on the second gradient update data using the test data to obtain a recall model.

[0064] In one embodiment, the rag module further includes a vectorization module, and similar text data is vectorized and converted by the vectorization module to obtain a text vector, and the text vector is input to the recall model through the model recall unit to obtain a recall vector.

[0065] In one embodiment, the recall module also includes a vector recall unit, the clinical data is converted into a clinical vector through a vectorization module, the clinical vector is calculated by the vector recall unit by performing similarity calculation between the clinical vector and the vector in the storage module to obtain a calculated vector, and the calculated vector is input into the recall model through the model recall unit to obtain a recall vector.

[0066] In one embodiment, the calculated vector is fused with similar text data and then input into a recall model to obtain a recall vector.

[0067] In one embodiment, the calculation vector is fused with the text vector and then input into the recall model to obtain the recall vector.

[0068] In one embodiment, the recall model includes one or more of the following: convolutional neural network, dilated convolutional neural network, residual network, and Transformer.

[0069] In one embodiment, the recall model includes an input layer, a hidden layer, a self-attention layer, a feedforward layer, a normalization layer, and an output layer. The input data passes through the input layer and the hidden layer in sequence to obtain a hidden vector, which is input to the self-attention layer and the feedforward layer respectively. The output vector of the self-attention layer, the output vector of the feedforward layer, and the hidden vector are fused and input to the normalization layer to obtain a recall vector, which is then output by the output layer.

[0070] In one embodiment, the text similarities are sorted to obtain the first L similar texts, the vector similarities are sorted to obtain the first L calculated vectors, the first L similar texts and the first L calculated vectors are fused and input into the recall model to obtain a recall vector, where L is a natural number greater than 1.

[0071] In one embodiment, the text similarities are sorted to obtain the first three similar texts, the vector similarities are sorted to obtain the first three calculated vectors, and the first three similar texts and the first three calculated vectors are fused and input into the recall model to obtain the recall vector.

[0072] In one embodiment, the storage module includes a text storage module and a vector storage module, wherein the data in the text storage module is converted into vectors through a vector model in a vectorization module and stored to obtain a vector storage module.

[0073] In one embodiment, the model training process of the RAG module obtains a renal cancer clinical dataset and labels and performs data preprocessing on the data in the dataset to obtain training data. The clinical dataset includes multi-source medical data such as clinical information, pathology reports, and treatment records of renal cancer patients.

[0074] In a specific embodiment, the Rag model construction includes vectorization module construction, storage module construction, and recall module construction. The vectorization module construction converts data into vectors through a vector model. The structure of the vector model is:

[0075] L1 embedding layer, through the embedding weight matrix and the ID of the input data mapping, the embedding word vector is obtained as the embedding matrix representation of the input data, and the vector dimension is 768;

[0076] The L2 multi-head attention mechanism layer performs matrix feature extraction on the 768-dimensional feature vector output by the embedding layer through three linear layers. After matrix multiplication, a 768-dimensional vector representation of each input data that incorporates attention information is obtained;

[0077] The L3 forward computation layer passes each input of the multi-head attention mechanism layer through two linear layers, and after activation by the activation layer, outputs the final 768-dimensional vector representation of each data.

[0078] Among them, the loss function used is crossentropy:

[0079] H(p,q)=-∑ x p(x)logq(x)

[0080] Among them, p(x) is the true label of the current input, q(x) is the model's predicted value for each label. Assuming a three-classification task, the correct label of a sample is the first class, then p = [1, 0, 0], and the model prediction value is assumed to be [0.5, 0.4,

[0081] 0.1], then the cross entropy is calculated as follows:

[0082] H(p=[1,0,0],q=[0.5,0.4,0.1)=-(1*log0.5+0*log0.4+0*log0.1)≈0.3

[0083] The present invention stores the clinical data after data preprocessing in the text storage module of the storage module, converts the clinical data after data preprocessing into vectors through the vector model of the vectorization module, and stores the vectors in the vector storage module of the storage module. The design storage structure of the vector storage module is as follows:

[0084] Define the Vectors table to store vector data. This table contains the following fields:

[0085] id (INT, PRIMARY KEY): Unique identifier of the vector.

[0086] vector_type (VARCHAR): The type of the vector (such as "2D", "3D", etc.).

[0087] vector(FLOAT): vector matrix.

[0088] description(TEXT): Description of the vector or related text information.

[0089] image_path (VARCHAR): The path where the vector visualization image is stored.

[0090] The vector data stored in the vector library is as follows: [[-0.029963023960590363,0.03870072960853577,0.06795354932546616,-0.03098524920642376,-0.0545266792178154,.......,0.051449041813611 984],......,[-0.029963023960590363,0.03870072960853577,0.06795354932546616,-0.03098524920642376,-0.0545266792178154,.......,0.051449041813611984]].

[0091] The dimension of the stored vector is: [k, 512], where k is the amount of knowledge and 512 is the dimension of the vector obtained by the knowledge text through the vector representation model.

[0092] Definition of the text storage module: Create a table called TextImages to store text data and related image information. This table contains the following fields:

[0093] id (INT, PRIMARY KEY): Unique identifier of the text.

[0094] text(TEXT): The text content to be stored.

[0095] image_path (VARCHAR): The path where the image containing the text is stored.

[0096] description(TEXT): description of the text or image.

[0097] Construction of model recall unit:

[0098] The structure of the recall model is as follows Figure 5 As shown, after passing through the input layer, it first passes through the hidden layer. The calculation logic of the hidden layer is:

[0099] x=t*W T

[0100] Where t is the input training data, W is the weight matrix of the hidden layer, and T represents the transpose operation of the matrix. The final output x is the vector representation output obtained after passing through the hidden layer.

[0101] The hidden layer calculation result is divided into three parts: 1) calculated by the self-att layer; 2) calculated by the feedforward layer; 3) added with the calculation results of 1) and 2). These three parts can be expressed by the following formula:

[0102] y=LayerNorm(x+MLP(x)+Attention(x))

[0103] Here, Attention represents the self-attack layer, MLP refers to the FeedForward layer in the figure, and x represents the output vector of the hidden layer. LayerNorm represents the normalization layer, which adds the results of the three parts and then inputs them into the normalization layer to obtain the final output vector. Furthermore, the calculation of the self-attack layer is specifically as follows:

[0104] q=x*W q T

[0105] k=x*W k T

[0106] v=x*W v T

[0107] Among them, W q , W k , W v Represent three different weight matrices, T represents transpose, q, k, v represent the output results of the three linear calculation layers, and then matrix operations are performed on q, k, v:

[0108]

[0109] Among them, d is a hyperparameter, set to 64, and att_output is the output result calculated by the self-att layer.

[0110] The calculation process of FeedForward is:

[0111] f=x*W f T

[0112] Among them, W f is the weight matrix of the FeedForward layer, and f is the output of the current calculation.

[0113]

[0114] Among them, e is the natural constant e, and feed_output is the output result of the FeedForward layer.

[0115] The calculation process of LayerNorm is:

[0116]

[0117] Among them, μ is the mean, σ is the variance, g i is the weight matrix of the LayerNorm layer, is the final result of the calculation.

[0118] In a specific embodiment, the training strategy of the recall model is as follows:

[0119] The data set is divided into 10 parts: data1, data2, data3, data4...data10; each data set corresponds to a task and has a total of 30 data items, of which 10 are used as the first training set, 10 are used as the validation set, and 10 are used as the second training set. First, the weight matrix of the initialized model network is recorded as the first recall model, and the network structure and the initialized weight matrix are copied to obtain the second recall model. Subsequently, the 10 first data sets are used in the network structure and weight matrix of the second recall model for separate training. During training, the difference between the present invention and the general network training process is that only 10 first data sets are used for separate training, and the gradients of the 10 data sets are accumulated, but not updated to the weight matrix of the second recall model. Instead, the accumulated gradients recorded in this process are recorded. (The second recall model executes the 10 data sets in parallel to obtain 10 gradients, and the 10 gradients are weighted summed or added to obtain the fusion weight).

[0120] Afterwards, based on the weight matrix of the second recall model, the present invention uses the following method to obtain the parameters obtained by further training to update the weight matrix:

[0121] W1=W-lr*t

[0122] Where W is the previously initialized weight matrix, lr is the learning rate, which is a hyperparameter, t is the cumulative gradient obtained after the 10 tasks are trained in the previous step, and W1 is the updated weight matrix;

[0123] Afterwards, the second recall model of the present invention uses the weight matrix W1 as the basis and uses 10 second training sets to train on the basis of W1. After this step of training is completed, we obtain a final version of the weight matrix W2, where:

[0124]

[0125] Where lr is the learning rate, t k is the gradient value of each task.

[0126] For new tasks with new small sample data volumes, the first recall model of the present invention directly uses the weight W2 and undergoes small sample fine-tuning training, ultimately obtaining a model trained using the self-training paradigm.

[0127] In one embodiment, during the training of the second recall model, the gradient can also be replaced by executing 10 data sets in sequence. The training of each data set is based on the training of the previous data set to obtain the gradient of the training of the 10th data set, and the weight matrix W1 is obtained based on the gradient at this time.

[0128] In one embodiment, the training process of the recall model may also be:

[0129] The weight matrix of the initialized model network is recorded as the first recall model, and the network structure and the initialized weight matrix are copied to obtain the second recall model;

[0130] Subsequently, the 10 first data sets are used in the network structure and weight matrix of the second recall model for separate training. During training, the difference between the present invention and the general network training process is that only 10 first data sets are used for separate training, and the gradients of the 10 data sets are accumulated at the same time, but are not updated to the weight matrix of the second recall model. Instead, the accumulated gradients recorded in this process are recorded.

[0131] Afterwards, based on the weight matrix of the first recall model, the present invention uses the following method to obtain the parameters obtained by further training to update the weight matrix:

[0132] W1=W-lr*t

[0133] Where W is the previously initialized weight matrix, lr is the learning rate, which is a hyperparameter, t is the cumulative gradient obtained after the 10 tasks are trained in the previous step, and W1 is the updated weight matrix;

[0134] Afterwards, the first recall model of the present invention uses the weight matrix W1 as the basis, and is trained using the 10 second training sets based on W1. After this step of training is completed, we obtain a final version of the weight matrix W2, where:

[0135]

[0136] Where lr is the learning rate, t k is the gradient value of each task.

[0137] In a specific embodiment, the text recall unit calculates the relevance of the text query input by the user with each knowledge segment si in the knowledge text segment set TextImages stored in the text storage module. The calculation steps are as follows: the sentence query is first segmented to generate a word list [wt]. For the sentence si to be compared with the sentence query, the relevance score of each word wt with si is calculated. Finally, the relevance scores of wt relative to si are weighted and summed. The calculation method is as follows:

[0138]

[0139] Among them, idf(wt) is the idf value of word wt, fi is the frequency of word wt in sentence s2, k1 and b are adjustment factors, usually set to 2 and 0.75 respectively, len(si) is the length of sentence si, and avgsl is the average length of all sentences.

[0140] In a specific embodiment, the vector recall unit converts the query input by the user into a vector representation through a vector representation model to obtain a 512-dimensional representation vector a. Then, the representation vector matrix b of all knowledge texts in the vector library is loaded. Furthermore, a matrix calculation is performed using the n-dimensional representation vector a corresponding to the current query and the representation vector matrix b of all knowledge texts in the vector library to obtain the similarity score between the query and all knowledge base texts. The calculation formula is:

[0141]

[0142] Where a is the representation vector obtained by converting the query input by the user into a vector representation through the vector representation model, b is the representation vector matrix of all knowledge texts, and the value range of cosθ is [-1, 1]. It is then normalized to a value in the interval [0, 1]. The normalization formula is:

[0143] d=0.5*(1+cosθ)

[0144] Among them, d represents the similarity between the final user input query and the knowledge base text.

[0145] S3. Inputting the recall vector into a large language model to perform kidney cancer prediction to obtain a prediction result.

[0146] In one embodiment, the large language model includes one or more of the following: BERT, XLNet, GPT-2,

[0147] ChatGPT, GPT-3, Baidu Wenxin Big Model, Alibaba Tongyi Big Model, Huawei Pangu Big Model, Tencent Hunyuan Big Model.

[0148] In a specific embodiment, the output of the rag module is then fed into a large language model for prediction, which also includes prompt word embedding. The output of the rag module is embedded into the prompt word to obtain an embedding result, which is then fed into the large language model for prediction. Prompt word embedding includes: defining a prompt word, which is divided into three parts:

[0149] pre_prompt,cen_prompt,end_prompt,where pre_prompt="Analyze the detailed information of the following kidney cancer patient in depth, paying special attention to the impact of factors such as pathological type, stage, and genetic mutation on prognosis. Please predict the patient's future in X years:

[0150] 1. The likelihood of recurrence, and factors that may increase the risk of recurrence;

[0151] 2. Overall survival rate, including the specific value and confidence interval of the X-year survival rate;

[0152] 3. Is it necessary to adjust the current treatment plan to improve the prognosis? If so, please provide specific suggestions.

[0153] cen_prompt = the output of the rag module; end_prompt = "Please integrate all the above information and provide a prognosis prediction and explanation with a scientific and rigorous attitude." Concatenate the contents of pre_prompt, cen_prompt, and end_prompt, denoted as merge_prompt, and input merge_prompt into the LLM model to obtain the final prognosis prediction output by the model.

[0154] In one embodiment, the execution process of the present invention can also be: performing data preprocessing on the acquired clinical data to obtain preprocessed data, then inputting the preprocessed data into the rag module, performing similarity calculation between the preprocessed data and the text of the text storage module in the storage module through the text recall unit to obtain similar text data, then converting the similar text data into a text vector through the vectorization module, inputting the text vector into the trained recall model in the model recall unit for feature extraction to obtain a recall vector, and inputting the recall vector into the large language model for prediction to obtain a prediction result.

[0155] In one embodiment, the execution process of the present invention can also be: performing data preprocessing on the acquired clinical data to obtain preprocessed data, then inputting the preprocessed data into the rag module, performing similarity calculation between the preprocessed data and the text of the text storage module in the storage module through the text recall unit to obtain similar text data, and at the same time, converting the preprocessed data into a clinical vector through the vectorization module, performing similarity calculation between the clinical vector and the vector of the vector storage module in the storage module through the vector recall unit to obtain a calculation vector, fusing the similar text data with the calculation vector and inputting them into the trained recall model in the model recall unit for feature extraction to obtain a recall vector, and inputting the recall vector into the large language model for prediction to obtain a prediction result.

[0156] In one embodiment, the execution process of the present invention can also be: performing data preprocessing on the acquired clinical data to obtain preprocessed data, then inputting the preprocessed data into the rag module, performing similarity calculation between the preprocessed data and the text of the text storage module in the storage module through the text recall unit to obtain similar text data, then converting the similar text data into a text vector through the vectorization module, and at the same time, converting the preprocessed data into a clinical vector through the vectorization module, performing similarity calculation between the clinical vector and the vector of the vector storage module in the storage module through the vector recall unit to obtain a calculation vector, fusing the text vector and the calculation vector and inputting them into the trained recall model in the model recall unit for feature extraction to obtain a recall vector, and inputting the recall vector into the large language model for prediction to obtain a prediction result.

[0157] The disclosed embodiments of the present invention further provide a computer program product or system, including a computer program, which implements the steps of the above-mentioned recall-based kidney cancer prediction method when executed by a processor.

[0158] Figure 2 A schematic diagram of a renal cancer prediction system based on recall provided by an embodiment of the present invention specifically includes:

[0159] Acquisition unit: used to obtain clinical data;

[0160] Recall unit: used to input the clinical data into the rag module to obtain a recall vector; the rag module includes a recall module and a storage module, and the recall module includes a text recall unit and a model recall unit. The text recall unit calculates the similarity between the clinical data and the text in the storage module to obtain similar text data, and the model recall unit inputs the similar text data into the recall model to obtain a recall vector;

[0161] Prediction unit: used for inputting the recall vector into the large language model to perform kidney cancer prediction to obtain a prediction result.

[0162] Figure 3 A schematic diagram of a renal cancer prediction device based on recall provided by an embodiment of the present invention specifically includes:

[0163] A memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions, and when the program instructions are executed, any one of the above-mentioned recall-based kidney cancer prediction methods is implemented.

[0164] The disclosed embodiments of the present invention further provide a computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, performs any one of the above-mentioned recall-based kidney cancer prediction methods.

[0165] The validation results of this validation example demonstrate that assigning inherent weights to indications can improve the performance of the present method compared to the default settings. Those skilled in the art will readily appreciate that, for ease of description and brevity, the specific operating processes of the systems, devices, and units described above can be referenced to the corresponding processes in the aforementioned method embodiments and will not be further elaborated upon here. It should be understood that the disclosed systems, devices, and methods can be implemented in other ways within the several embodiments provided herein. For example, the device embodiments described above are merely illustrative. For example, the division of units described is merely a logical functional division. In actual implementation, other divisions may be employed, such as combining or integrating multiple units or components into another system, or omitting or disabling certain features. Furthermore, the coupling, direct coupling, or communication connection shown or discussed may be through interfaces, indirect coupling, or communication connection between devices or units, and may be electrical, mechanical, or other forms. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, i.e., they may be located in one location or distributed across multiple network units. Some or all of the units may be selected to achieve the objectives of the present embodiment as needed. In addition, the functional units in the various embodiments of the present invention may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated units may be implemented in the form of hardware or in the form of software functional units. Those skilled in the art will understand that all or part of the steps in the various methods of the above-mentioned embodiments may be completed by instructing the relevant hardware through a program, and the program may be stored in a computer-readable storage medium, which may include: a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0166] Those skilled in the art will understand that all or part of the steps in the above-mentioned embodiment method can be implemented by instructing the relevant hardware through a program, and the program can be stored in a computer-readable storage medium. The above-mentioned medium storage can be a read-only memory, a disk or an optical disk, etc.

[0167] The above is a detailed introduction to a computer device provided by the present invention. For those skilled in the art, according to the concept of the embodiments of the present invention, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting the present invention.

Claims

1. A method for predicting kidney cancer based on recall, characterized in that: include: S1. Obtain clinical data; S2. Input the clinical data into the rag module to obtain a recall vector; the rag module includes a recall module, a storage module, and a vectorization module; the recall module includes a text recall unit, a model recall unit, and a vector recall unit; the clinical data is similarly calculated by the text recall unit between the clinical data and the text in the storage module to obtain similar text data; the similar text data is vectorized and converted by the vectorization module to obtain a text vector; the clinical data is converted into a clinical vector by the vectorization module; the clinical vector is similarly calculated by the vector recall unit between the clinical vector and the vector in the storage module to obtain a calculation vector; the calculation vector is fused with the similar text data and then input into the recall model to obtain a recall vector; or the calculation vector is fused with the text vector and then input into the recall model to obtain a recall vector; S3, inputting the recall vector into a large language model to perform kidney cancer prediction to obtain a prediction result; The training process of the recall model is: S101, obtaining a first training set and a second training set, wherein the first training set and the second training set respectively include N tasks, where N is a natural number greater than 1; S102, initializing the parameters of the recall model to be trained to obtain a first recall model; S103, copying the parameters and model structure of the first recall model to obtain a second recall model; S104: Input the first training set into the second recall model for training to obtain gradient update data; S105 , updating the first recall model based on the gradient update data and obtaining a recall model by training with the second training set.

2. The method for predicting kidney cancer based on recall according to claim 1, characterized in that: The gradient update data is gradient data obtained by sequentially training the second recall model on N task data, where the Nth task data training is performed on the basis of the completion of the N-1th task data training; or, the gradient update data is N gradient data obtained by performing gradient fusion on the N gradient data obtained by the second recall model performing parallel training on N task data; The gradient fusion includes one or more of the following: weighted summation and addition.

3. The method for predicting kidney cancer based on recall according to claim 1, characterized in that: The step S105 is replaced by: training a second recall model based on the gradient update data and the second training set to obtain second gradient update data, acquiring test data, and training the first recall model updated based on the second gradient update data using the test data to obtain a recall model.

4. The method for predicting kidney cancer based on recall according to claim 1, wherein: The recall model includes one or more of the following: convolutional neural network, hole convolutional neural network, residual network, and Transformer.

5. The method for predicting kidney cancer based on recall according to claim 1, characterized in that: The recall model includes an input layer, a hidden layer, a self-attention layer, a feedforward layer, a normalization layer, and an output layer. The input data passes through the input layer and the hidden layer in sequence to obtain a hidden vector, which is input to the self-attention layer and the feedforward layer respectively. The output vector of the self-attention layer, the output vector of the feedforward layer, and the hidden vector are fused and input to the normalization layer to obtain a recall vector, which is then output by the output layer.

6. The method for predicting kidney cancer based on recall according to claim 1, characterized in that: Sort the text similarity to get the first L similar texts, sort the vector similarity to get the first L calculated vectors, merge the first L similar texts with the first L calculated vectors and input them into the recall model to get the recall vector, where L is a natural number greater than 1.

7. The method for predicting kidney cancer based on recall according to claim 1, characterized in that: The large language model includes one or more of the following: BERT, XLNet, GPT-2, ChatGPT, GPT-3, Baidu Wenxin large model, Alibaba Tongyi large model, Huawei Pangu large model, and Tencent Hunyuan large model.

8. The method for predicting kidney cancer based on recall according to claim 1, characterized in that: The method also includes data preprocessing, including: data deduplication, data special character processing, and data standardization.

9. The method for predicting kidney cancer based on recall according to claim 1, characterized in that: The storage module includes a text storage module and a vector storage module, wherein the data in the text storage module is converted into vectors through a vector model in a vectorization module and stored to obtain a vector storage module.

10. A computer program product having a computer program or instructions thereon, characterized in that: include: The computer program or instructions are executed by a processor to implement the recall-based kidney cancer prediction method according to any one of claims 1 to 9.

11. A computer device comprising a memory, a processor, and a computer program or instruction stored in the memory, wherein: The computer program or instructions are executed by a processor to implement the recall-based kidney cancer prediction method according to any one of claims 1 to 9.

12. A computer-readable storage medium having a computer program or instruction stored thereon, characterized in that: The computer program or instructions are executed by a processor to implement the recall-based kidney cancer prediction method according to any one of claims 1 to 9.

Images