Method and device for prompting tumor patient test data upload based on image recognition

Abstract

The present invention discloses a method and device for uploading prompts of laboratory data of tumor patients based on image recognition, comprising: obtaining a pre-learned tumor laboratory text set, a tumor laboratory text learning sample, and a risk warning degree label corresponding to the tumor laboratory text learning sample; autonomously pre-learning a risk information mining operator based on the tumor laboratory text in the pre-learned tumor laboratory text set to obtain a pre-learned risk information mining operator; using the pre-learned risk information mining operator and a risk warning classifier to classify the tumor laboratory text learning sample to obtain the inference risk warning degree; based on the inference risk warning degree and the risk warning degree label, obtaining the algorithm training cost; optimizing the algorithm configuration parameters of the pre-learned risk information mining operator and / or the risk warning classifier according to the algorithm training cost. The present invention can improve the accuracy of the risk warning detection algorithm when performing risk warning classification.

Description

Technical Field

[0001] The present invention relates to the field of text processing, and in particular to a method and device for prompting tumor patient test data upload based on image recognition. Background Art

[0002] With the advancement of medical technology and the diversification of tumor treatment methods, the survival of tumor patients has been significantly extended, and tumor diseases have gradually shown a trend of chronicity. However, this change has also posed new challenges to patient management and medical services. Especially for tumor patients receiving chemotherapy, they need to regularly monitor blood indicators between chemotherapy cycles and pay attention to various systemic symptoms to evaluate the side effects and efficacy of chemotherapy. For doctors, accurately understanding the patient's condition after the previous cycle of chemotherapy is crucial to formulating the treatment plan for the next cycle. However, in actual work, many patients are unable to provide complete and accurate information due to various reasons (such as forgetting to bring test reports, forgetting or describing inaccurate symptoms, etc.), resulting in doctors' inaccurate grasp of the patient's adverse reactions to treatment, which may affect the adjustment of treatment plans. In addition, with the diversification of tumor treatment methods and the extension of patient survival, patients' medical records and examination results are increasing. The loss and difficulty of sorting out paper reports have become a common problem faced by doctors and patients. This not only increases the difficulty and cost of medical services, but also affects the individualization and refinement of patient management. With the development of medical informatization and the widespread use of electronic medical record cloud devices and laboratory information cloud devices, a large number of tumor patients' laboratory data are digitally stored. How to quickly and accurately extract key risk information from massive amounts of tumor test text data and provide guidance (non-diagnostic guidance) to patients based on this information is a major challenge facing the current medical information technology field. Summary of the invention

[0003] The object of the present invention is to provide a method and device for prompting tumor patients to upload test data based on image recognition. The embodiment of the present invention is implemented as follows:

[0004] In a first aspect, the present invention provides a method for prompting tumor patient laboratory data upload based on image recognition, which is applied to a cloud device, and the cloud device is communicatively connected to a target terminal. The method includes: in response to a data upload instruction of the target terminal, obtaining a tumor laboratory information image uploaded by a user; obtaining a tumor laboratory text contained in the tumor laboratory information image based on OCR recognition; inputting the tumor laboratory text into a pre-learned risk information mining operator in a risk warning detection algorithm to perform risk information mining to obtain a risk characterization vector; the pre-learned risk information mining operator is obtained by autonomously pre-learning the risk information mining operator in the risk warning detection algorithm using a pre-learned tumor laboratory text set; performing risk warning classification on the risk characterization vector according to the risk warning classifier in the risk warning detection algorithm to obtain a risk warning degree corresponding to the tumor laboratory text, and returning the risk warning degree to the target terminal. Among them, the risk warning detection algorithm is trained through the following training steps: obtaining the pre-learned tumor test text set, tumor test text learning samples and risk warning degree labels corresponding to the tumor test text learning samples; autonomously pre-learning the risk information mining operator according to the tumor test text in the pre-learned tumor test text set to obtain a pre-learned risk information mining operator; calling the pre-learned risk information mining operator and the risk warning classifier to perform risk warning classification on the tumor test text learning samples to obtain the inferred risk warning degree; obtaining the algorithm training cost according to the inferred risk warning degree and the risk warning degree label; optimizing the algorithm configuration parameters of at least one of the pre-learned risk information mining operator and the risk warning classifier according to the algorithm training cost.

[0005] In a second aspect, the present invention provides a cloud device comprising: one or more processors; a memory; one or more computer programs; wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method described above is implemented.

[0006] The beneficial effects of the present invention include at least: the present invention uses a pre-learned tumor test text set, a tumor test text learning sample, and a risk warning degree label corresponding to the tumor test text learning sample, and autonomously pre-learns the risk information mining operator according to the tumor test text in the pre-learned tumor test text set to obtain a pre-learned risk information mining operator, and uses the pre-learned risk information mining operator and the risk warning classifier to classify the tumor test text learning sample to obtain the inference risk warning degree, and obtains the algorithm training cost based on the inference risk warning degree and the risk warning degree label, and optimizes the algorithm configuration parameters of at least one of the pre-learned risk information mining operator and the risk warning classifier according to the algorithm training cost. The present invention uses a large number of pre-learned tumor test text sets to autonomously pre-learn the risk information mining operator in the risk warning detection algorithm, and after obtaining the pre-learned risk information mining operator, the risk warning detection algorithm is fine-tuned using the tumor test text learning sample, which can improve the training quality of the risk warning detection algorithm, and the accuracy of the risk warning detection algorithm in classifying the risk warning of the tumor test text is improved, which helps to provide accurate medical guidance. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] Figure 1 The present invention provides a flowchart of a method for prompting a tumor patient to upload laboratory data based on image recognition.

[0008] Figure 2 It is a schematic diagram of the training process of the risk warning detection algorithm provided by an embodiment of the present invention.

[0009] Figure 3 It is a schematic diagram of the composition of a cloud device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0010] In the embodiment of the present invention, the execution subject of the method for uploading laboratory data of tumor patients based on image recognition is a cloud device, including but not limited to a cloud server, a cloud computer, etc. The cloud device is connected to the network and implements the present invention through interactive operations with other cloud devices in the network. The network where the cloud device is located includes but is not limited to the Internet, wide area network, metropolitan area network, local area network, VPN network, etc. The cloud device is connected to the target terminal in communication, and the target terminal can be a computer, tablet computer, smart phone or other electronic device. Figure 1 As shown, the method includes:

[0011] Step S110: In response to the data upload instruction of the target terminal, the tumor test information image uploaded by the user is obtained.

[0012] In step S110, the cloud device monitors and identifies the data upload instruction from the target terminal in real time. When the user triggers the upload operation on the target terminal, such as clicking an "upload" button, the target terminal sends a data upload instruction containing specific parameters to the cloud device. This instruction may include information such as the user's identity (such as user name, account ID, etc.), the uploaded file type (such as image), and the desired operation (such as storage, processing). Once the data upload instruction is received, the cloud device begins to process the uploaded data. In this example, the uploaded data is an image of tumor test information, which may mean that the user has taken a picture of his or her tumor test report in some way (such as a mobile phone camera) and selected the upload option. The cloud device parses the file path or data stream in the upload instruction and extracts the image data from it. The image data can be stored in a binary format, including information such as pixel values, color depth, and image size. For example, a standard RGB (red, green, and blue) image may consist of three color channels, each channel containing a two-dimensional array of the same size as the image, and each element in the array represents the color intensity of the corresponding pixel. For tumor test report images, these pixel values ​​will reflect elements such as text, charts, and lines on the report.

[0013] After acquiring the image data, the cloud device can perform a series of verification and preprocessing operations on the data to ensure the accuracy and efficiency of subsequent processing. For example, format verification, resizing, denoising, grayscale conversion, rotation and cropping, etc. The verified and preprocessed image data will be stored in the storage medium of the cloud device, such as a hard disk, SSD (solid state drive) or cloud storage service. For easy management and retrieval, the computer will assign a unique identifier (such as UUID, universal unique identifier) ​​to each image and associate it with the user's identity. In addition, the cloud device will also record metadata such as the upload time, size, format, etc. of the image for subsequent statistical analysis and optimization. For example, suppose a patient uses a smartphone to take a photo of his or her tumor test report and selects the upload option through a medical health management application. The application will generate a data upload instruction, which includes the patient's account ID, the path of the image file, and the type of upload request (such as "upload tumor test report"). After receiving this instruction, the cloud device will read the image data from the specified file path and perform necessary verification and preprocessing. For example, if the image size is too large, the computer will scale it to a size suitable for subsequent processing; if the image contains noise, the computer will apply a denoising algorithm to process it. After processing, the image data will be stored in a designated directory on the server and associated with the patient's account ID.

[0014] Step S120: Obtain the tumor test text contained in the tumor test information image based on OCR recognition.

[0015] In step S120, the cloud device uses optical character recognition (OCR) technology to extract text content from the tumor test information image uploaded by the user. OCR technology is a technology that converts text in an image into machine-editable text. It analyzes the image, identifies the text area in the image, and converts the pixel values ​​of these areas into corresponding character codes. OCR technology is widely used in document scanning, image processing, automated data entry and other fields, which will not be described here.

[0016] In tumor test information images, OCR technology can identify text information in the image, such as the patient's name, age, gender, tumor type, test results, etc. This information is crucial for subsequent risk information mining and prompt classification. For example, if OCR technology can accurately identify the patient's tumor type and test results, the cloud device can assess the patient's risk level based on this information and give corresponding medical guidance.

[0017] Step S130: input the tumor test text into the pre-learned risk information mining operator in the risk warning detection algorithm to perform risk information mining to obtain a risk characterization vector; the pre-learned risk information mining operator is obtained by autonomously pre-learning the risk information mining operator in the risk warning detection algorithm using a pre-learned tumor test text set.

[0018] In step S130, the cloud device inputs the tumor test text uploaded by the user into the pre-learned risk information mining operator in the risk warning detection algorithm to mine the risk information in the text and generate a risk characterization vector. The pre-learned risk information mining operator is an unsupervised pre-trained machine learning model that can extract key risk information from tumor test texts and convert it into risk characterization vectors. This operator is trained on a large set of pre-learned tumor test texts, which contains a large number of tumor test reports from different patients, different hospitals and different time periods. Through unsupervised pre-training, the operator can learn the potential structure and pattern of risk information in the text, so that when faced with new tumor test texts, it can quickly identify and extract relevant risk information.

[0019] In step S130, the cloud device implements risk information mining through the following steps:

[0020] 1. Text preprocessing: First, the cloud device preprocesses the input tumor test text to remove irrelevant information, reduce noise and extract useful features. The preprocessing steps may include text cleaning (such as removing punctuation, special characters, etc.), word segmentation (dividing the text into single words or phrases), part-of-speech tagging (assigning a part-of-speech tag to each word) and stop word filtering (removing common words that are meaningless to risk assessment).

[0021] 2. Feature extraction: Next, the cloud device uses feature extraction technology to extract key features from the preprocessed text. These features may include word frequency, co-occurrence, semantic similarity, etc. The purpose of feature extraction is to convert text data into numerical data that can be processed by machine learning models. For example, methods such as Bag of Words or TF-IDF (Term Frequency-Inverse Document Frequency) can be used to calculate the importance of each word in the text and use it as a feature value.

[0022] 3. Risk information mining: After feature extraction, the cloud device inputs the extracted features into the pre-learning risk information mining operator for risk information mining. This operator uses deep learning algorithms (such as convolutional neural networks (CNN), recurrent neural networks (RNN), or Transformer, etc.) to process input features and generate risk representation vectors. The risk representation vector is a high-dimensional numerical vector that contains the potential representation of risk information in the text. Through training, the operator can learn how to map input features to the risk representation vector, so that similar texts have similar representations in the vector space.

[0023] 4. Generation of risk representation vector: Finally, the cloud device outputs a risk representation vector from the pre-learned risk information mining operator. This vector can be regarded as a representation of the text in the risk space, which contains all the risk-related information in the text. By comparing the risk representation vectors of different texts, the cloud device can evaluate the risk similarity between them and classify risk warnings accordingly.

[0024] The training process of the pre-learning risk information mining operator is an unsupervised pre-training process, which relies on a large number of pre-learning tumor test text sets. The training process is roughly as follows:

[0025] 1. Data collection and preprocessing: First, the cloud device collects a large amount of tumor test text data and preprocesses the data. The preprocessing step is similar to the text preprocessing step mentioned above, aiming to remove irrelevant information, reduce noise and extract useful features.

[0026] 2. Model initialization: Next, the cloud device initializes the pre-learned risk information mining operator. This operator may be a deep neural network model, and its structure depends on the specific task. During initialization, the computer assigns a random value or a preset value to each parameter of the model.

[0027] 3. Unsupervised pre-training: The cloud device then uses the pre-learned tumor test text set to perform unsupervised pre-training on the operator. The purpose of pre-training is to allow the operator to learn the potential structure and pattern of risk information in the text. During the pre-training process, the cloud device does not rely on any labels or supervisory information, but optimizes the model parameters by reconstructing the input data and maximizing mutual information.

[0028] 4. Model optimization: Finally, the cloud device uses optimization algorithms (such as gradient descent, Adam, etc.) to iteratively update the model parameters to minimize the pre-trained objective function. Through multiple iterative training, the operator can gradually learn how to effectively extract risk information from text and generate accurate risk representation vectors.

[0029] The pre-learning risk information mining operator can provide rich feature information for subsequent machine learning models by converting text data into risk characterization vectors, thereby improving the accuracy and reliability of risk assessment. For example, suppose a user uploads a text image containing lung cancer test results, and the cloud device converts it into text data through OCR technology: "Patient name: Li Si, age: 60 years old, tumor type: lung cancer, test results: positive". In step S130, the cloud device first pre-processes the text, such as removing punctuation marks, word segmentation, etc., to obtain the pre-processed text: "Patient name Li Si age 60 years old tumor type lung cancer test results positive". Then, the cloud device uses feature extraction technology to extract key features from the pre-processed text, such as word frequency, co-occurrence relationship, etc. Then, these features are input into the pre-learning risk information mining operator for risk information mining to generate a high-dimensional risk characterization vector. Finally, the cloud device can evaluate the patient's risk level based on this risk characterization vector and give corresponding medical guidance.

[0030] Step S140: performing risk warning classification on the risk characterization vector according to the risk warning classifier in the risk warning detection algorithm, obtaining the risk warning degree corresponding to the tumor test text, and returning the risk warning degree to the target terminal.

[0031] In step S140, the risk warning classifier is a core component in the risk warning detection algorithm, which is responsible for mapping the risk characterization vector to predefined risk warning degree categories. These categories can be divided according to the level of risk, such as "low risk", "medium risk" and "high risk". The classifier learns how to correctly classify new risk characterization vectors by learning a large number of annotated risk characterization vectors and their corresponding risk warning degree labels.

[0032] In step S140, the cloud device performs the following steps to complete the risk warning classification:

[0033] 1. Input risk characterization vector: First, the cloud device passes the risk characterization vector generated in step S130 as input to the risk warning classifier. This vector is a high-dimensional numerical array that contains all risk-related information extracted from the tumor test text.

[0034] 2. Classifier processing: After receiving the risk representation vector, the classifier will process it using the internally learned classification rules or model parameters. This may involve complex mathematical operations and logical judgments, aiming to map the input vector to the most appropriate risk warning level category.

[0035] For example, suppose the risk characterization vector is an array of length 100, where each element represents the value of a feature. The classifier may calculate a score or probability distribution based on the combination and weight of these feature values ​​through linear combination, kernel function transformation, etc., to indicate the possibility that the vector belongs to each category.

[0036] 3. Determine the risk warning level: After the classifier processes the risk characterization vector, it will output one or more scores / probability values ​​corresponding to different risk warning level categories. The cloud device will select the category with the highest score / probability as the final risk warning level.

[0037] For example, if the classifier outputs a probability distribution such as [0.1, 0.6, 0.3], which correspond to "low risk", "medium risk" and "high risk" respectively, the computer will use "medium risk" as the final risk warning level.

[0038] 4. Return the risk level: Finally, the cloud device encodes the determined risk level into an easy-to-understand form (such as text messages, icons, etc.) and returns it to the target terminal through the previously established communication connection. In this way, patients or medical professionals can immediately understand the risk level of tumor test results and make corresponding decisions or actions accordingly.

[0039] To ensure the accuracy and reliability of the risk warning classifier, the cloud device is fully trained and optimized before deployment. This can include the following steps:

[0040] 1. Data collection and annotation: Collect a large amount of tumor test texts and their corresponding risk warning labels to form a training data set. This data should cover various types of tumors, test results, and risk levels to ensure the generalization ability of the classifier.

[0041] 2. Feature selection and extraction: According to the characteristics of tumor test text and the needs of risk assessment, appropriate features are selected for extraction. These features may include word frequency, semantic similarity, text structure, etc., which together form the basis of the risk characterization vector.

[0042] 3. Model selection and training: Select a suitable machine learning model (such as support vector machine SVM, random forest, neural network, etc.) as the classifier and train it using the training data set. During the training process, the model will continuously adjust its internal parameters to minimize the classification error rate or maximize the classification accuracy.

[0043] 4. Model evaluation and optimization: Use the validation data set to evaluate the trained model and check its performance on unseen data. If the performance is poor, you can optimize it by adjusting model parameters, adding training data, changing feature selection strategies, etc.

[0044] Please refer to Figure 2 In the embodiment of the present invention, the risk warning detection algorithm is trained through the following training steps:

[0045] Step S11: obtaining a pre-learned tumor test text set, tumor test text learning samples, and risk warning degree labels corresponding to the tumor test text learning samples.

[0046] The pre-learned tumor test text set is a collection of a large amount of unlabeled tumor test text data. These data can come from multiple hospitals, clinics or research institutions, covering test reports of different patients, different time periods and different tumor types. It is understandable that the acquisition of these data is permitted by laws and regulations. The main purpose of the pre-learned text set is to provide sufficient training data for the risk information mining operator in the risk warning detection algorithm to learn the potential structure and pattern of risk information in the text.

[0047] For example, the pre-learned tumor test text set may contain thousands or even tens of thousands of tumor test reports, each of which contains detailed information such as the patient's name, age, gender, tumor type, test results, etc. These reports exist in text form and may contain various professional terms, abbreviations, and formats, providing rich training materials for risk information mining operators.

[0048] Unlike the pre-learning tumor test text set, the tumor test text learning samples are carefully selected and annotated data sets used to directly train the classifier in the risk warning detection algorithm. These learning samples can contain two parts: the tumor test text itself and the corresponding risk warning degree label. The text part is similar to the data in the pre-learning text set, but the number is smaller and more representative; the label part is a clear annotation of the risk level of the text, which is used to guide the learning process of the classifier. For example, suppose there is a learning sample set containing 1,000 tumor test texts. Each text corresponds to a risk warning degree label, such as "low risk", "medium risk" or "high risk". These labels are based on the evaluation of professional doctors or statistical analysis of historical data, ensuring the accuracy and reliability of the labels.

[0049] The risk warning degree label directly reflects the risk information contained in the text. The generation of labels can rely on professional medical knowledge, historical data and statistical analysis methods. In actual operation, medical experts may need to manually annotate some texts to establish an initial set of labels. Then, these annotated data can be used to train a preliminary classification model to automatically annotate the remaining text. Finally, the accuracy and consistency of all labels are ensured through manual review and correction. For example, when generating risk warning degree labels, a variety of methods can be used. For example, a comprehensive consideration can be made based on factors such as tumor type, severity of test results, age and gender of the patient. For certain specific tumor types or test indicators, it may also be necessary to refer to professional medical guidelines or literature to determine the risk level. The final generated labels will be used as part of the training samples to guide the learning process of the risk warning detection algorithm.

[0050] Step S12: autonomously pre-learning the risk information mining operator according to the tumor test texts in the pre-learned tumor test text set to obtain a pre-learned risk information mining operator.

[0051] Autonomous pre-learning is an unsupervised learning method that does not require labels or supervisory information to guide the learning process. In the training of risk warning detection algorithms, the purpose of autonomous pre-learning is to enable the risk information mining operator to automatically learn useful feature representations and potential risk information structures from a large amount of unlabeled tumor test texts. Through this process, the operator can capture the risk patterns implicit in the text and provide rich feature information for subsequent risk assessment.

[0052] The pre-learned tumor test text set is the key input data in the autonomous pre-learning process. This dataset contains a large amount of unlabeled tumor test text, covering test reports of different patients, different time periods and different tumor types. These text data provide rich training materials for the risk information mining operator, enabling it to learn the diversity and complexity of risk information in the text.

[0053] For example, the pre-learned tumor test text set may contain thousands or even tens of thousands of tumor test reports, each of which contains detailed information such as the patient's name, age, gender, tumor type, test results, etc. These text data may come from multiple hospitals, clinics, or research institutions, and are highly diverse and representative.

[0054] The risk information mining operator is a machine learning model that is responsible for extracting key risk information from tumor test text. This operator may be implemented using deep learning algorithms (such as convolutional neural networks (CNN), recurrent neural networks (RNN), or Transformer, etc.), and has powerful feature extraction and representation learning capabilities. During the autonomous pre-learning process, the operator will continuously adjust its internal parameters and structure to better capture the risk information in the text.

[0055] In step S12, the cloud device performs the following steps to complete autonomous pre-learning:

[0056] 1. Data preprocessing: First, the pre-learning tumor test text set is preprocessed, including text cleaning, word segmentation, part-of-speech tagging, stop word filtering, etc. These preprocessing operations help remove irrelevant information, reduce noise, and extract useful features.

[0057] 2. Feature extraction: Next, feature extraction technology is used to extract key features from the preprocessed text. These features may include word frequency, co-occurrence relationship, semantic similarity, etc., which together constitute the feature representation of the text.

[0058] 3. Model initialization: Initialize the risk information mining operator and assign random parameters or preset parameters to it. These parameters will be continuously adjusted and optimized in the subsequent pre-learning process.

[0059] 4. Unsupervised pre-training: The pre-processed text data is input into the risk information mining operator for unsupervised pre-training. During the pre-training process, the operator will try to learn the potential structure and pattern of risk information in the text, and optimize the feature representation and model performance by continuously adjusting its internal parameters.

[0060] For example, during the pre-training process, the operator may use unsupervised learning algorithms such as autoencoders or generative adversarial networks (GANs). Autoencoders reconstruct input text through the structure of encoders and decoders to learn effective feature representations of text; generative adversarial networks learn the potential distribution of text data through adversarial training of generators and discriminators.

[0061] 5. Model evaluation and adjustment: During the pre-training process, the performance of the model is evaluated regularly, such as by using indicators such as reconstruction error and mutual information to measure the learning effect of the model on text feature representation. Based on the evaluation results, the model parameters and structure are adjusted to optimize the model performance.

[0062] 6. Generation of pre-learning risk information mining operator: After multiple rounds of unsupervised pre-training, the pre-learning risk information mining operator is obtained. This operator can extract key risk information from the text and generate effective feature representation.

[0063] The pre-learning risk information mining operator plays an important role in the risk warning detection algorithm. It can convert the input tumor test text into a high-dimensional risk representation vector, providing rich feature information for subsequent risk assessment. By comparing the risk representation vectors of different texts, the cloud device can evaluate the risk similarity between them and classify the risk warning accordingly.

[0064] Step S13: Retrieve the pre-learned risk information mining operator and risk warning classifier to perform risk warning classification on the tumor test text learning samples to obtain the inferred risk warning degree.

[0065] The pre-learning risk information mining operator is an unsupervised pre-trained machine learning model that can extract key risk information from tumor test texts and convert it into a high-dimensional risk representation vector. This vector contains all risk-related information in the text and is the basis for subsequent risk warning classification. In step S13, the cloud device uses this operator to extract features from tumor test text learning samples and obtains a risk representation vector for each text.

[0066] The risk warning classifier is a supervised learning machine learning model that can classify the input risk characterization vector to determine the risk warning level corresponding to the text. During the training process, the classifier has learned the mapping relationship between different risk characterization vectors and the corresponding risk warning levels. Therefore, in step S13, the cloud device can use the classifier to classify the risk characterization vectors generated by the pre-learning risk information mining operator to obtain the inference risk warning level of each tumor test text learning sample.

[0067] In step S13, the cloud device performs the following steps to generate the inference risk warning level:

[0068] 1. Input tumor test text learning samples: The cloud device first selects a batch of tumor test text learning samples from the training data set as input. These samples contain text content and corresponding risk warning level labels, but in this step, the label information is only used for subsequent performance evaluation and does not directly participate in the reasoning process.

[0069] 2. Feature extraction: For each input sample, the cloud device uses the pre-learned risk information mining operator to extract features. The operator converts the text content into a high-dimensional risk representation vector, which contains all risk-related information in the text.

[0070] For example, suppose the content of a tumor test text learning sample is "Patient Zhang San, lung cancer, elevated tumor markers". The pre-learning risk information mining operator may convert this text into a vector of length 100, where each element represents the value of a feature (such as word frequency, semantic similarity, etc.).

[0071] 3. Classification reasoning: After extracting the risk representation vector, the cloud device inputs it into the risk warning classifier for classification reasoning. The classifier will calculate the probability distribution of one or more categories based on the feature values ​​in the vector and the classification rules or model parameters learned internally.

[0072] For example, suppose the output of the risk warning classifier is a probability distribution containing three categories (low risk, medium risk, high risk), such as [0.1, 0.7, 0.2]. This distribution means that the probability that the text belongs to "low risk" is 10%, the probability that it belongs to "medium risk" is 70%, and the probability that it belongs to "high risk" is 20%.

[0073] 4. Determine the degree of inference risk warning: Finally, the cloud device selects the category with the highest probability as the degree of inference risk warning for the text based on the probability distribution output by the classifier.

[0074] For example, in the example above, the computer would select "medium risk" as the inference risk warning level for the text.

[0075] After obtaining the inferred risk warning level, the cloud device also needs to compare it with the actual risk warning level label corresponding to the text to evaluate the performance of the risk warning detection algorithm. By calculating indicators such as accuracy, recall, and F1 score, the cloud device can understand the performance of the algorithm at different risk levels and optimize the algorithm accordingly. For example, the performance of the algorithm can be improved by adjusting the parameters of the pre-learning risk information mining operator, improving the structure of the risk warning classifier, or increasing the training data.

[0076] Step S14: Obtain the algorithm training cost according to the inference risk warning degree and the risk warning degree label.

[0077] In step S13, the cloud device has used the pre-learned risk information mining operator and risk warning classifier to infer the tumor test text learning samples and obtain the inference risk warning level of each text. These inference results are the algorithm's risk assessment of the current training data, which reflects the algorithm's understanding and classification capabilities of text risk information.

[0078] At the same time, each tumor test text learning sample also comes with a real risk warning level label, which is manually annotated by professional doctors or medical experts based on the patient's actual condition and test results, representing the real risk level of the text. These labels are supervisory information during the algorithm training process, used to guide the algorithm's learning and parameter adjustment.

[0079] The algorithm training cost is an important indicator to measure the performance of the algorithm. It indicates the degree of difference between the prediction results of the algorithm on the given training data and the actual label. In the training of the risk warning detection algorithm, the training cost can be calculated by the loss function.

[0080] The loss function is a mathematical expression that takes the algorithm's prediction results (i.e., the degree of inference risk) and the true label as input, and outputs a scalar value that reflects the difference between the prediction results and the true label. The design of the loss function can be based on specific optimization goals and assumptions, such as minimizing prediction error, maximizing classification accuracy, etc.

[0081] In the training of risk warning detection algorithms, commonly used loss functions include cross-entropy loss, mean squared error loss, etc. These loss functions can calculate the overall loss of the algorithm on the current training data according to different risk warning level categories and predicted probability distributions.

[0082] In step S14, the cloud device calculates the loss value for each tumor test text learning example, and then accumulates or averages these loss values ​​to obtain the average loss value of the entire training set, which is the training cost of the algorithm. This training cost reflects the overall performance of the algorithm on the current training data.

[0083] Step S15: Optimizing algorithm configuration parameters of at least one of the pre-learning risk information mining operator and the risk warning classifier according to the algorithm training cost.

[0084] Algorithm configuration parameters are key components of machine learning models, which determine the structure and behavior of the model. In the risk warning detection algorithm, both the pre-learning risk information mining operator and the risk warning classifier contain a large number of configuration parameters, such as the weights, biases, learning rates, batch sizes, etc. of the neural network. The values ​​of these parameters directly affect the prediction performance and training efficiency of the model.

[0085] During the training process, the cloud device evaluates the performance of the algorithm under the current parameter configuration by calculating the training cost. The training cost is a quantitative indicator that reflects the degree of difference between the algorithm's prediction results and the true label. When the training cost is high, it means that the prediction performance of the algorithm under the current parameter configuration is poor and the parameters need to be optimized.

[0086] The goal of parameter optimization is to find an optimal set of parameter configurations that minimizes the prediction error of the algorithm on the training data. This can be achieved by iteratively adjusting the parameter values ​​and observing the changes in the training cost. In each iteration, the cloud device updates the parameter values ​​based on the current training cost and gradient information in order to reduce the training cost in the next iteration.

[0087] In step S15, the cloud device performs the following steps to optimize the algorithm configuration parameters:

[0088] 1. Gradient calculation: First, the cloud device calculates the gradient of the training cost with respect to each parameter. The gradient is a vector that points to the direction in which the training cost decreases the fastest. In computer vision and deep learning, this can be achieved through the backpropagation algorithm.

[0089] 2. Parameter update: Next, the cloud device updates the parameter value based on the gradient information. There are many ways to update the parameters, such as stochastic gradient descent (SGD), mini-batch gradient descent (Mini-batch SGD), Adam optimizer, etc. These optimization algorithms adjust the parameter values ​​through different strategies in order to minimize the training cost while maintaining the training efficiency.

[0090] 3. Training cost monitoring: During the parameter update process, the cloud device will continuously monitor the changes in training cost. If the training cost continues to decrease, it means that the parameter optimization is effective; if the training cost stops decreasing or fluctuates, it may be necessary to adjust the hyperparameters of the optimization algorithm (such as learning rate, momentum, etc.) or reconsider the model structure.

[0091] 4. Iteration and convergence: Parameter optimization is an iterative process. The cloud device needs to perform gradient calculation and parameter update steps multiple times until the training cost converges to a stable level or reaches a preset number of iterations. During the iterative process, the cloud device can also stop training early based on the performance on the validation set to avoid overfitting.

[0092] Assume there is a simple neural network model, which consists of two fully connected layers and one output layer. During the training process, the cross entropy loss function is used to calculate the training cost, and the Adam optimizer is used to update the network parameters. In each iteration, the cloud device calculates the gradient of each parameter according to the back propagation algorithm. For example, for the weight matrix W1 of the first fully connected layer, its gradient It can be calculated using the chain rule and the derivative of the loss function with respect to the output.

[0093] After obtaining the gradient, the cloud device uses the Adam optimizer to update the weight matrix W1. The Adam optimizer calculates the update amount of each parameter based on the historical gradient information and uses the learning rate to control the update step size. After each iteration, the computer calculates and records the current training cost. If the training cost continues to decrease, it means that the parameter optimization is effective; if the training cost stops decreasing or fluctuates, the learning rate or other hyperparameters may need to be adjusted. Repeat the above steps until the training cost converges to a stable level or reaches the preset number of iterations. During the iteration process, the validation set can also be used to monitor the generalization ability of the model and stop training in advance if necessary.

[0094] Step S15 minimizes the training cost by iteratively adjusting the algorithm configuration parameters, thereby improving the prediction accuracy and generalization ability of the algorithm. During the parameter optimization process, the cloud device needs to calculate the gradient, update the parameters, monitor the training cost, and perform iteration and convergence judgment.

[0095] In one implementation, step S12, autonomously pre-learning the risk information mining operator according to the tumor test text in the pre-learned tumor test text set to obtain the pre-learned risk information mining operator, includes:

[0096] Step S121: obtaining original tumor test texts, positive tumor test texts, and negative tumor test texts in the pre-learned tumor test text set; wherein the original tumor test texts refer to tumor test texts belonging to risk prompts, the positive tumor test texts refer to tumor test texts obtained after risk information is injected into the tumor test texts, and other tumor test texts refer to tumor test texts other than the original tumor test texts and the positive tumor test texts in the pre-learned tumor test text set;

[0097] Step S122: autonomously pre-learning the risk information mining operator according to the original tumor test text, the positive tumor test text and the negative tumor test text to obtain a pre-learned risk information mining operator.

[0098] In step S121, the cloud device obtains three different types of text data from the pre-learned tumor test text set: original tumor test text, positive tumor test text, and negative tumor test text. These three types of text data will play different roles in the subsequent autonomous pre-learning process and jointly promote the learning of the risk information mining operator.

[0099] Original tumor test texts: These texts represent tumor test texts that are risk warnings, that is, they contain risk information that doctors or experts consider important. These texts are the main source for risk information mining operators to learn risk features. For example, an original tumor test text may contain keywords such as "elevated tumor markers" and "abnormal cell proliferation", which are crucial for identifying risk information.

[0100] Positive tumor test texts: These texts are obtained by injecting risk information into the original tumor test texts. Risk information injection refers to artificially adding risk-related information to the text to enhance the risk characteristics of the text. For example, "normal tumor markers" can be modified to "abnormally elevated tumor markers" to generate positive tumor test texts. These texts help operators learn more diverse risk feature representations during autonomous pre-learning.

[0101] Negative tumor test texts: These texts are other texts in the pre-learned tumor test text set except the original tumor test texts and positive tumor test texts. These texts may contain no or only less risk information, so they may be regarded as noise data in the autonomous pre-learning process. However, through reasonable data processing and model design, negative tumor test texts can also provide valuable information for the learning of risk information mining operators.

[0102] In step S122, the cloud device uses the different types of text obtained in step S121 to autonomously pre-learn the risk information mining operator. This process is intended to enable the operator to automatically learn the potential structure and pattern of risk information from text data through unsupervised learning.

[0103] Before autonomous pre-learning, the cloud device first pre-processes the text data. This includes steps such as text cleaning (such as removing punctuation marks, special characters, etc.), word segmentation (dividing the text into individual words or phrases), part-of-speech tagging (assigning a part-of-speech tag to each word), and stop word filtering (removing common words that are meaningless to risk features). These pre-processing operations help reduce data noise, improve data quality, and lay the foundation for subsequent feature extraction and model training.

[0104] After preprocessing, the cloud device uses feature extraction technology to extract key risk features from the text. These features may include word frequency, co-occurrence, semantic similarity, etc. The purpose of feature extraction is to convert text data into numerical data that can be processed by machine learning models. For example, methods such as Bag of Words or TF-IDF (Term Frequency-Inverse Document Frequency) can be used to calculate the importance of each word in the text and use it as a feature value.

[0105] After extracting the risk features, the cloud device uses these features as input to train the risk information mining operator. The training process can use unsupervised learning algorithms, such as autoencoders, generative adversarial networks (GANs), etc. These algorithms can automatically learn the potential structure and pattern of the data from the feature data and store the learned knowledge in the model parameters.

[0106] During the training process of the autoencoder, the input layer receives the risk feature vector, the encoding layer compresses the input vector into a low-dimensional latent representation (i.e., encoding), and the decoding layer reconstructs the latent representation into the original input vector. By minimizing the reconstruction error (such as mean square error MSE), the autoencoder can learn an effective feature representation. In the training of the risk information mining operator, the feature vectors of the original tumor test text and the positive tumor test text can be used as input, and the latent representation of the risk information can be learned through the training of the autoencoder.

[0107] During the training process of the generative adversarial network, the generator tries to generate fake data similar to real data, while the discriminator tries to distinguish between real data and fake data. Through adversarial training between the generator and the discriminator, the generative adversarial network can learn the real distribution of the data. In the training of the risk information mining operator, the feature vectors of the original tumor test text and the positive tumor test text can be used as real data, and the feature vectors of the negative tumor test text can be used as noise data. Through the training of the generative adversarial network, a more robust risk feature representation can be learned.

[0108] During the model training process, the cloud device needs to regularly evaluate the performance of the model and adjust the model based on the evaluation results. Evaluation indicators may include reconstruction error, mutual information, clustering effect, etc. If the model performance is poor, the model can be optimized by adjusting the model structure, increasing training data, changing feature extraction methods, etc.

[0109] After multiple rounds of unsupervised pre-training, the cloud device obtained a pre-learned risk information mining operator. This operator is able to extract effective risk feature representations from text data and can play an important role in subsequent risk warning detection tasks.

[0110] By acquiring different categories of tumor test texts and using these texts for autonomous pre-learning, the cloud device can enable the risk information mining operator to learn the potential structure and pattern of risk information. This process not only improves the operator's feature extraction ability, but also provides strong support for subsequent risk warning classification tasks. Through reasonable data preprocessing, feature extraction, model training, and evaluation adjustment steps, the cloud device can generate high-quality pre-learning risk information mining operators, laying the foundation for the overall performance improvement of the risk warning detection algorithm.

[0111] In one implementation, step S122, autonomously pre-learning the risk information mining operator according to the original tumor test text, the positive tumor test text and the negative tumor test text to obtain the pre-learned risk information mining operator, includes:

[0112] Step S1221: Obtaining enhanced original tumor test text corresponding to the original tumor test text and enhanced positive tumor test text corresponding to the positive tumor test text; enhancing the original tumor test text means that at least one of the risk warning and the warning level in the original tumor test text is modified without changing the text semantics to obtain the tumor test text; enhancing the positive tumor test text means that the text semantics in the positive tumor test text is modified without changing the risk warning and the warning level to obtain the tumor test text;

[0113] Step S1222: performing risk information mining on the original tumor test text, the enhanced original tumor test text, and the negative tumor test text according to the risk information mining operator, respectively, to obtain an original risk characterization vector corresponding to the original tumor test text, an enhanced original risk characterization vector corresponding to the enhanced original tumor test text, and a negative risk characterization vector corresponding to the negative tumor test text; performing calculations based on the original risk characterization vector, the enhanced original risk characterization vector, and the negative risk characterization vector to obtain a first training generation value;

[0114] Step S1223: performing risk information mining on the positive tumor test text and the enhanced positive tumor test text respectively according to the risk information mining operator to obtain a positive risk characterization vector corresponding to the positive tumor test text and an enhanced positive risk characterization vector corresponding to the enhanced positive tumor test text; performing calculations on the positive risk characterization vector, the enhanced positive risk characterization vector and the negative risk characterization vector to obtain a second training generation value;

[0115] Step S1224: Optimize the operator configuration parameters of the risk information mining operator according to the first training cost value and the second training cost value to obtain a pre-learned risk information mining operator.

[0116] In step S1221, the cloud device first obtains the enhanced original tumor test text corresponding to the original tumor test text and the enhanced positive tumor test text corresponding to the positive tumor test text. These enhanced texts are obtained by making specific changes to the original texts, aiming to enhance the sensitivity and robustness of the risk information mining operator to risk information.

[0117] Enhanced original tumor test text: This is a text obtained by modifying the risk warning and warning level in the original tumor test text, but the semantics of the text remains unchanged. For example, if the original text mentions "slightly elevated tumor markers", the enhanced text may change it to "significantly elevated tumor markers" to highlight the risk information. This modification helps the risk information mining operator learn how to represent different risk levels.

[0118] Enhanced positive tumor test text: This is a text obtained by modifying the text semantics in the positive tumor test text, but the risk warning and warning level remain unchanged. For example, if the positive text mentions "tumor markers are normal", the enhanced text may change it to a sentence with similar semantics but different expression, such as "all test indicators are within the normal range." This modification helps the risk information mining operator learn how to represent the same risk information under different semantic expressions.

[0119] In step S1222, the cloud device uses a risk information mining operator to mine risk information on the original tumor test text, the enhanced original tumor test text, and the negative tumor test text, and obtains their corresponding risk characterization vectors. Then, the cloud device calculates the first training generation value by comparing the distance or similarity between these vectors.

[0120] The risk information mining operator maps text data into a high-dimensional space to form risk representation vectors. These vectors contain the potential representation of risk information in the text. For example, a text containing "elevated tumor markers" may get a vector with a high score on the "risk" dimension.

[0121] The cloud device can calculate the distance between the original risk representation vector and the enhanced original risk representation vector (such as Euclidean distance, cosine similarity, etc.), as well as the distance between them and the negative risk representation vector. These distances reflect the positional relationship of different texts in the risk representation space.

[0122] The cloud device calculates the first training generation value based on these distances. One possible calculation method is to use triplet loss, which encourages the distance between the original risk representation vector and the enhanced original risk representation vector to be smaller than the distance between them and the negative risk representation vector. The specific formula is as follows: L triplet =max(0, margin+D(a, p) D(a, n)), where a is the anchor point (original risk representation vector), p is the positive sample (reinforced original risk representation vector), n is the negative sample (negative risk representation vector), D(·,·) is the distance function, and margin is a hyperparameter used to control the distance between positive and negative samples.

[0123] In step S1223, the cloud device performs similar processing on the positive tumor test text and the enhanced positive tumor test text, and calculates the second training generation value. This step is intended to further enhance the sensitivity of the risk information mining operator to risk information.

[0124] Risk information mining and vector distance calculation: The cloud device performs risk information mining on the positive tumor test text and the enhanced positive tumor test text, obtains their corresponding risk representation vectors, and calculates the distance between these vectors.

[0125] Second training generation value calculation: Similar to the first training generation value, the cloud device uses triple loss or other loss functions to calculate the second training generation value. However, the positive sample at this time is a reinforced positive risk representation vector, and the negative sample is still a negative risk representation vector. This step helps the risk information mining operator learn the risk information representation under different semantic expressions while keeping the risk warning and warning level unchanged.

[0126] In step S1224, the cloud device optimizes the operator configuration parameters of the risk information mining operator according to the first training cost value and the second training cost value. This step is the core of autonomous pre-learning, which aims to minimize the training cost value by iteratively adjusting the parameters, thereby improving the performance of the operator. The cloud device may use optimization algorithms such as stochastic gradient descent (SGD) and Adam to update the parameters of the operator. These algorithms adjust the parameter values ​​according to the gradient information of the training cost value, in order to reduce the training cost value in the next round of iteration. The optimization process is an iterative process, and the cloud device needs to execute steps S1221 to S1223 multiple times and adjust the parameters according to the changes in the training cost value. When the training cost value converges to a stable level or reaches a preset number of iterations, the optimization process ends, and the risk information mining operator obtained at this time is the pre-learning risk information mining operator.

[0127] Step S122 optimizes the risk information mining operator by introducing enhanced text and calculating training cost values. The generation of enhanced text enhances the sensitivity and robustness of the operator to risk information, while the calculation of training cost values ​​provides a clear goal for parameter optimization. By iteratively adjusting parameters and minimizing training cost values, the cloud device finally obtains a pre-learned risk information mining operator with better performance, which provides strong support for subsequent risk warning detection tasks.

[0128] In one embodiment, the risk information mining operator includes a first risk information mining sub-operator and a second risk information mining sub-operator, and the first risk information mining sub-operator and the second risk information mining sub-operator are twin operators containing different operator configuration parameters; step S1222, according to the risk information mining operator, the original tumor test text, the enhanced original tumor test text and the negative tumor test text are respectively subjected to risk information mining, and the original risk characterization vector corresponding to the original tumor test text, the enhanced original risk characterization vector corresponding to the enhanced original tumor test text and the negative risk characterization vector corresponding to the negative tumor test text are obtained, including:

[0129] Step S12221: Perform risk information mining on the original tumor test text according to the first risk information mining sub-operator to obtain the original risk characterization vector; perform risk information mining on the enhanced original tumor test text according to the second risk information mining sub-operator to obtain the enhanced original risk characterization vector; perform risk information mining on the negative tumor test text according to the first risk information mining sub-operator or the second risk information mining sub-operator to obtain the negative risk characterization vector.

[0130] In step S1222, the cloud device uses the risk information mining operator to mine risk information on different types of tumor test texts to generate corresponding risk characterization vectors. In this step, the risk information mining operator is designed to consist of two twin operators: a first risk information mining sub-operator and a second risk information mining sub-operator. The two operators contain different configuration parameters but have similar structures and functions for extracting risk information from texts. In step S12221, the cloud device first selects the first risk information mining sub-operator to process the original tumor test text. The original tumor test text is a text containing clear risk prompts, which directly reflects the patient's health status and potential risks. The first risk information mining sub-operator performs a deep analysis of the original tumor test text and extracts key risk features therein, such as abnormal test indicators, pathological descriptions, etc. These features are then converted into numerical data to form the original risk characterization vector. The original risk characterization vector is a high-dimensional array in which each element represents the value of a feature. These values ​​may represent the importance, frequency of occurrence, or other statistics of the feature for subsequent risk assessment and classification. For the enhanced original tumor test text, the cloud device selects the second risk information mining sub-operator for processing. The enhanced original tumor test text is obtained by modifying the risk prompt or prompt level in the original text, aiming to test the sensitivity of the operator to changes in risk information. The second risk information mining sub-operator also performs in-depth analysis on the enhanced original tumor test text, extracts risk features, and generates an enhanced original risk characterization vector. The enhanced original risk characterization vector is similar to the original risk characterization vector in structure, but due to subtle differences in the text content, the two may differ in value. This difference reflects the operator's ability to respond to changes in risk information. For negative tumor test text (i.e., text that does not contain clear risk prompts), the cloud device can select either the first risk information mining sub-operator or the second risk information mining sub-operator for processing. Since negative tumor test text may not contain useful risk information, the generated negative risk characterization vector may appear as a relatively sparse or uniform numerical distribution, lacking a significant feature pattern. One of the purposes of processing negative tumor test text is to evaluate the performance of the operator on noisy data, to ensure that it can maintain stable output when facing non-risk-related text, and avoid false positives or false negatives. Although the first risk information mining sub-operator and the second risk information mining sub-operator have similar structures, they may generate different risk characterization vectors when processing the same text due to different configuration parameters. These differences may be reflected in feature selection, weight assignment, activation function selection, etc. Different operators may select different feature subsets for subsequent processing. For example, one operator may pay more attention to the changes in laboratory indicators, while another operator may pay more attention to keywords in pathological descriptions. When generating risk characterization vectors, different operators may assign different weights to different features. This reflects the different understandings of the importance of different features by the operators.The choice of activation function will also affect the generation of risk characterization vectors. Different activation functions may result in different distributions and ranges of vectors after nonlinear transformations. The generated risk characterization vectors are the basis for subsequent risk assessment and classification. These vectors contain key risk information in the text and can be input into classifiers, regression models, or other machine learning algorithms for further processing. In risk warning classification tasks, risk characterization vectors can be used as input to the classifier to determine the risk category (such as low risk, medium risk, high risk) to which the text belongs. In risk assessment tasks, risk characterization vectors can be used to predict the patient's specific risk level or the possibility of disease progression. In text mining tasks, risk characterization vectors can also be used to cluster similar texts to better understand the intrinsic structure and distribution of text data.

[0131] Step S12221 describes in detail how the cloud device uses the first risk information mining sub-operator and the second risk information mining sub-operator to process different types of tumor test texts and generate corresponding risk characterization vectors. This process not only reflects the sensitivity and robustness of the operator in processing risk information, but also provides important data support for subsequent risk assessment and classification tasks. By adjusting the configuration parameters and training strategies of the operators, the performance of the operators can be further optimized and the accuracy and reliability of risk assessment can be improved.

[0132] In one embodiment, in step S12221, risk information mining is performed on the original tumor test text according to the first risk information mining sub-operator to obtain the original risk characterization vector; risk information mining is performed on the enhanced original tumor test text according to the second risk information mining sub-operator to obtain the enhanced original risk characterization vector, including:

[0133] Step S122211: extract local features of the original tumor test text according to the first N-gram model in the first risk information mining sub-operator to obtain corresponding original tumor test text features; perform risk information mining on the original tumor test text features according to the first risk information mining module in the first risk information mining sub-operator to obtain an original risk representation vector;

[0134] Step S122212: Perform local feature extraction on the enhanced original tumor test text according to the second N-gram model in the second risk information mining sub-operator to obtain corresponding enhanced original tumor test text features; perform risk information mining on the enhanced original tumor test text features according to the second risk information mining module in the second risk information mining sub-operator to obtain an enhanced original risk representation vector.

[0135] In step S122211, the cloud device first uses the first risk information mining sub-operator to process the original tumor test text. This process can be divided into two main stages: local feature extraction and risk information mining.

[0136] Local feature extraction is the first step of risk information mining, and its purpose is to identify keywords, phrases or patterns from the text that are valuable for risk assessment. In the first risk information mining sub-operator, this task is undertaken by the first N-gram model.

[0137] The N-gram model is a text analysis method based on a statistical language model. It captures local features in a text by segmenting the text into a sequence of N consecutive words (or characters), namely N-grams. For example, in a bigram, N=2, the model focuses on the frequency of two consecutive words in the text; in a trigram, N=3, the model focuses on the frequency of three consecutive words. In the first risk information mining sub-operator, the first N-gram model is used to extract local features from the original tumor test text. For example, if the original text contains the expression "elevated tumor markers", the first N-gram model may identify N-gram features such as "tumor markers" and "elevated markers".

[0138] After extracting the local features, the cloud device inputs these features into the first risk information mining module for further processing to generate an original risk characterization vector.

[0139] The first risk information mining module is responsible for analyzing N-gram features and generating a high-dimensional risk representation vector based on these features. This vector contains the potential representation of risk information in the text and can be used for subsequent risk assessment and classification tasks.

[0140] Generation of risk representation vectors: The first risk information mining module may use a variety of technologies to generate risk representation vectors, such as word embedding, convolutional neural network (CNN) or recurrent neural network (RNN), etc. These technologies can convert discrete text features into continuous numerical vectors, which is convenient for subsequent calculation and analysis.

[0141] Assume that the original tumor test text is "Patient A, lung cancer, elevated tumor marker AFP". The first N-gram model may extract features such as "lung cancer" and "elevated tumor marker AFP". Then, the first risk information mining module uses these features to generate a risk representation vector, such as [0.8, 0.6, 0.2, ...], where each element represents the weight or score of a feature in the risk space.

[0142] In step S122212, the cloud device processes the enhanced original tumor test text in a similar manner, but this time a second risk information mining sub-operator is used.

[0143] Similar to step S122211, the second N-gram model in the second risk information mining sub-operator is responsible for extracting local features from the enhanced original tumor test text. Since the enhanced original tumor test text is obtained by modifying the risk warning or warning level in the original text, the second N-gram model may extract N-gram features that are slightly different from the original text. For example, if "elevated tumor markers" in the original text is enhanced to "significantly elevated tumor markers", the second N-gram model may identify the new feature of "significantly elevated tumor markers". After extracting the local features, the cloud device inputs these features into the second risk information mining module for processing to generate an enhanced original risk characterization vector.

[0144] Similar to the first risk information mining module, the second risk information mining module is also responsible for analyzing N-gram features and generating risk characterization vectors. However, since the second risk information mining sub-operator has different configuration parameters, the generated enhanced original risk characterization vector may be different from the original risk characterization vector. Since the enhanced original tumor test text and the original tumor test text differ in risk prompts or prompt levels, their risk characterization vectors may also show differences in some dimensions. These differences reflect the sensitivity of the risk information mining operator to changes in risk information.

[0145] Assume that the enhanced original tumor test text is "Patient A, lung cancer, tumor markers significantly increased". The second N-gram model may extract features such as "tumor markers significantly increased". Then, the second risk information mining module uses these features to generate an enhanced original risk representation vector, such as [0.9, 0.5, 0.3, ...], which may show higher scores in some dimensions compared to the original risk representation vector.

[0146] In one embodiment, in step S1222, performing calculations based on the original risk characterization vector, the enhanced original risk characterization vector, and the negative risk characterization vector to obtain a first training cost value includes:

[0147] Step S12222: Obtain a first training cost value based on a first vector distance between an original risk characterization vector and an enhanced original risk characterization vector, and a second vector distance between an original risk characterization vector and a negative risk characterization vector.

[0148] In step S1222, the cloud device performs a series of operations using the original risk characterization vector, the enhanced original risk characterization vector, and the negative risk characterization vector to obtain the first training generation value. This step is a key link in the autonomous pre-learning process of the risk information mining operator. It aims to evaluate the performance of the operator by comparing the distances between different risk characterization vectors and provide guidance for subsequent parameter optimization. The following is a detailed explanation of step S12222.

[0149] In step S12222, the cloud device calculates the first vector distance between the original risk characterization vector and the enhanced original risk characterization vector. This distance reflects the similarity or difference between the two vectors in the risk characterization space. Since the enhanced original tumor test text is obtained by modifying the risk prompt or prompt level in the original text, the positions of the two vectors in the risk characterization space should be relatively close. Vector distances can be calculated using a variety of methods, such as Euclidean distance, cosine similarity, or Manhattan distance. These methods have their own advantages and disadvantages and are suitable for different application scenarios. During the training process of the risk information mining operator, it may be necessary to select a suitable distance measurement method according to the specific situation.

[0150] Next, the cloud device calculates the second vector distance between the original risk characterization vector and the negative risk characterization vector. This distance reflects the difference between the original text and the text that does not contain a clear risk prompt in the risk characterization space. Since the negative tumor test text may not contain useful risk information, the positions of these two vectors in the risk characterization space should be relatively far apart.

[0151] Similarly, the second vector distance may be calculated using a variety of methods, and the specific selection depends on the application scenario and the applicability of the distance measurement method.

[0152] Finally, the cloud device obtains a first training generation value based on the first vector distance and the second vector distance. This value can reflect the performance of the risk information mining operator in distinguishing between texts containing risk warnings and texts that do not contain risk warnings. Specifically, if the first vector distance is small and the second vector distance is large, it means that the operator can accurately identify the risk information in the text and distinguish it from the text that does not contain risk information.

[0153] The calculation of the first training cost value can adopt a variety of strategies, such as directly taking the difference, ratio or weighted sum of two distances, etc. These strategies have their own advantages and disadvantages, and need to be selected according to the specific situation.

[0154] The first training cost is one of the important indicators for evaluating the performance of risk information mining operators. It reflects the ability of the operator to distinguish between texts containing risk warnings and texts that do not contain risk warnings. By minimizing the first training cost, the cloud device can optimize the operator's parameter configuration and improve the accuracy and robustness of its risk information mining.

[0155] In addition, the first training generation value can also be used to guide subsequent model selection and training strategies. For example, when comparing the performance of different risk information mining operators, the first training generation value can be used as one of the evaluation criteria; when adjusting model parameters or optimizing the training process, the optimization effect can also be judged based on the changes in the first training generation value. By calculating the distance between the original risk characterization vector, the enhanced original risk characterization vector and the negative risk characterization vector, and obtaining the first training generation value based on this, the cloud device can evaluate the performance of the operator and provide guidance for subsequent parameter optimization. This process not only helps to improve the accuracy and robustness of risk information mining, but also lays an important foundation for the training and optimization of the entire risk warning detection algorithm.

[0156] In one embodiment, step S1223, risk information mining is performed on the positive tumor test text and the enhanced positive tumor test text according to the risk information mining operator to obtain a positive risk characterization vector corresponding to the positive tumor test text and an enhanced positive risk characterization vector corresponding to the enhanced positive tumor test text, including:

[0157] Step S12231: Perform risk information mining on the positive tumor test text according to the first risk information mining sub-operator to obtain a positive risk characterization vector; perform risk information mining on the enhanced positive tumor test text according to the second risk information mining sub-operator to obtain an enhanced positive risk characterization vector.

[0158] In step S12231, the cloud device processes the positive tumor test text and the enhanced positive tumor test text using the first risk information mining sub-operator and the second risk information mining sub-operator respectively.

[0159] Positive tumor test text is obtained by injecting risk information into the original tumor test text. It contains more significant risk warnings or higher-level risk information than the original text. When processing this type of text, the cloud device selects the first risk information mining sub-operator. The first risk information mining sub-operator is a carefully designed machine learning model that can extract key risk features from the text and generate corresponding risk representation vectors. When processing positive tumor test text, the operator will focus on key information such as risk warning words and abnormal indicator descriptions in the text, and try to convert this information into numerical data.

[0160] Generation of positive risk characterization vector: Through the processing of the first risk information mining sub-operator, the positive tumor test text is converted into a high-dimensional positive risk characterization vector. This vector contains all risk-related information in the text and can be used for subsequent risk assessment and classification tasks.

[0161] Assume that the positive tumor test text is "Patient B, liver cancer, abnormally elevated AFP index". The first risk information mining sub-operator may extract key features such as "liver cancer" and "abnormally elevated AFP index" and generate a positive risk representation vector, such as [0.9, 0.8, 0.4, ...], where each element represents the weight or score of a feature in the risk space.

[0162] Enhanced positive tumor test text is the text obtained by changing the text semantics in the positive tumor test text, but the risk warning and warning level remain unchanged. This type of text is intended to test the sensitivity of the risk information mining operator to text semantic changes while keeping the risk warning unchanged.

[0163] Application of the second risk information mining sub-operator: Similar to the first risk information mining sub-operator, the second risk information mining sub-operator is also a machine learning model that can extract risk features from text and generate risk representation vectors. However, since the two operators have different configuration parameters and internal structures, they may produce different outputs when processing the same text.

[0164] Generation of enhanced positive risk characterization vector: Through the processing of the second risk information mining sub-operator, the enhanced positive tumor test text is converted into a high-dimensional enhanced positive risk characterization vector. This vector is similar to the positive risk characterization vector in structure, but due to subtle differences in text semantics, the two may differ in value.

[0165] Assume that the enhanced positive tumor test text is "Patient B suffers from liver cancer, and his AFP level is significantly elevated". The second risk information mining sub-operator may extract features such as "suffering from liver cancer" and "AFP level is significantly elevated" and generate an enhanced positive risk representation vector, such as [0.8, 0.9, 0.3, ...]. Compared with the positive risk representation vector, this vector may show higher scores or different weight distributions in some dimensions.

[0166] Through the processing of step S12231, the cloud device generates a positive risk characterization vector and an enhanced positive risk characterization vector, which respectively represent the representation of the positive tumor test text and the enhanced positive tumor test text in the risk space. These vectors not only contain the key risk information in the text, but also reflect the risk information mining operator's understanding and processing capabilities of text data. In addition, by comparing the differences between the positive risk characterization vector and the enhanced positive risk characterization vector, the cloud device can evaluate the risk information mining operator's sensitivity to changes in text semantics while keeping the risk prompt unchanged. This evaluation helps to optimize the operator's parameter configuration and improve its robustness, so that it can more accurately identify risk information when processing actual text data.

[0167] In one embodiment, step S12231, performing risk information mining on the positive tumor test text according to the first risk information mining sub-operator to obtain a positive risk characterization vector; performing risk information mining on the enhanced positive tumor test text according to the second risk information mining sub-operator to obtain an enhanced positive risk characterization vector, includes:

[0168] Step S122311: extract local features of the positive tumor test text according to the first N-gram model in the first risk information mining sub-operator to obtain corresponding positive tumor test text features; perform risk information mining on the positive tumor test text features according to the first risk information mining module in the first risk information mining sub-operator to obtain a positive risk representation vector;

[0169] Step S122312: Perform local feature extraction on the enhanced positive tumor test text according to the second N-gram model in the second risk information mining sub-operator to obtain the corresponding enhanced positive tumor test text features; perform risk information mining on the enhanced positive tumor test text features according to the second risk information mining module in the second risk information mining sub-operator to obtain the enhanced original risk representation vector.

[0170] In step S122311, the cloud device processes the positive tumor test text using the first risk information mining sub-operator. This process can be divided into two main stages: local feature extraction and risk information mining.

[0171] Local feature extraction is the first step of risk information mining, and its purpose is to identify keywords, phrases or patterns from the text that are valuable for risk assessment. In the first risk information mining sub-operator, this task is undertaken by the first N-gram model.

[0172] When processing positive tumor test text, the first N-gram model will scan the text content and identify all possible N-gram features. For example, if the text content is "Patient C, gastric cancer, abnormally elevated CEA index", the first N-gram model may extract binary or ternary features such as "Patient C gastric cancer" and "abnormally elevated gastric cancer CEA index".

[0173] After extracting the local features, the cloud device inputs these features into the first risk information mining module for further processing to generate a positive risk characterization vector.

[0174] The first risk information mining module is responsible for analyzing N-gram features and generating a high-dimensional risk representation vector based on these features. This vector contains the potential representation of risk information in the text and can be used for subsequent risk assessment and classification tasks.

[0175] The first risk information mining module may use a variety of technologies to generate risk representation vectors, such as word embedding, convolutional neural network (CNN) or recurrent neural network (RNN), etc. These technologies can convert discrete text features into continuous numerical vectors to facilitate subsequent calculations and analysis.

[0176] Assume that the first risk information mining module uses word embedding technology to process N-gram features. For each N-gram feature, it will find the corresponding vector representation from the pre-trained word embedding matrix, and aggregate these vectors (such as sum, average or weighted sum) to generate the final risk characterization vector. For example, for the above N-gram features, "patient type C gastric cancer" and "abnormal increase in gastric cancer CEA index" may correspond to two vectors [v1, v2, ...] and [w1, w2, ...] respectively, then the positive risk characterization vector can be calculated as [v1+w1, v2+w2, ...].

[0177] In step S122312, the cloud device processes the enhanced positive tumor test text in a similar manner, but this time a second risk information mining sub-operator is used.

[0178] Similar to step S122311, the second N-gram model in the second risk information mining sub-operator is responsible for extracting local features from the enhanced positive tumor test text.

[0179] Since the enhanced positive tumor test text is obtained by modifying the text semantics in the positive tumor test text (but the risk warning and warning level remain unchanged), the second N-gram model may extract a slightly different feature set from the first N-gram model. For example, if the original text "abnormal increase in CEA index of gastric cancer" is rewritten as "patient C is diagnosed with gastric cancer, and his CEA level is significantly increased", the second N-gram model may extract new features such as "patient C is diagnosed with gastric cancer" and "diagnosed with gastric cancer, his CEA level is significantly increased".

[0180] After extracting the local features, the cloud device inputs these features into the second risk information mining module for processing to generate an enhanced positive risk characterization vector.

[0181] Similar to the first risk information mining module, the second risk information mining module is also responsible for analyzing N-gram features and generating risk characterization vectors. However, since the second risk information mining sub-operator has different configuration parameters and internal structures, the generated enhanced positive risk characterization vector may be different from the positive risk characterization vector.

[0182] Since there are differences in textual semantics between the enhanced positive tumor test text and the positive tumor test text (although the risk warning and warning level remain unchanged), their risk representation vectors may also show differences in some dimensions. These differences reflect the sensitivity of the risk information mining operator to changes in textual semantics.

[0183] Assume that the second risk information mining module also uses word embedding technology to process N-gram features. For the new feature set extracted from the enhanced positive tumor test text, the second risk information mining module will find the corresponding vector representation from the pre-trained word embedding matrix and aggregate these vectors to generate an enhanced positive risk representation vector. Due to the difference in feature sets, this vector may differ from the positive risk representation vector in terms of value.

[0184] Steps S122311 and S122312 describe in detail how the cloud device uses the first risk information mining sub-operator and the second risk information mining sub-operator to perform risk information mining on positive tumor test texts and enhanced positive tumor test texts. This process includes two core links: local feature extraction and risk information mining, which are undertaken by the N-gram model and the risk information mining module respectively. Through this process, the cloud device can generate a risk characterization vector containing risk information in the text, providing important data support for subsequent risk assessment and classification tasks. At the same time, since the first risk information mining sub-operator and the second risk information mining sub-operator have different configuration parameters and internal structures, they may produce different outputs when processing the same or similar texts, which helps to enhance the robustness and generalization ability of the risk information mining operator.

[0185] In one implementation, in step S1223, operations are performed based on the positive risk characterization vector, the enhanced positive risk characterization vector and the negative risk characterization vector to obtain a second training generation value, including: step S12232: obtaining the second training generation value based on the third vector distance between the positive risk characterization vector and the enhanced positive risk characterization vector, and the fourth vector distance between the original risk characterization vector and the negative risk characterization vector.

[0186] In step S12232, the cloud device calculates the third vector distance between the positive risk characterization vector and the enhanced positive risk characterization vector. This distance reflects the similarity or difference between the two vectors in the risk characterization space. Since the positive tumor test text and the enhanced positive tumor test text are consistent in risk prompts and prompt levels, but the text semantics are different, the positions of the two vectors in the risk characterization space should be relatively close, but may be slightly different due to semantic differences. Next, the cloud device calculates the fourth vector distance between the original risk characterization vector and the negative risk characterization vector. This distance reflects the difference between the text containing clear risk prompts and the text not containing clear risk prompts in the risk characterization space. Since the original tumor test text contains risk prompts, while the negative tumor test text does not, the positions of the two vectors in the risk characterization space should be relatively far. Similar to the calculation of the third vector distance, the fourth vector distance can also be measured by a variety of methods, and the specific selection depends on the application scenario and the applicability of the distance measurement method. Finally, the cloud device obtains the second training generation value based on the third vector distance and the fourth vector distance. This value can reflect the performance of the risk information mining operator in distinguishing texts that contain risk warnings (whether positive or reinforcing positive) from texts that do not contain risk warnings. Specifically, if the third vector distance is small and the fourth vector distance is large, it means that the operator can accurately identify the risk information in the text and distinguish it from the text that does not contain risk information.

[0187] The calculation of the second training cost value can adopt a variety of strategies, such as directly taking the difference, ratio or weighted sum of the two distances, etc. These strategies have their own advantages and disadvantages, and need to be selected according to the specific situation.

[0188] The second training generation value reflects the operator's ability to distinguish between texts that contain risk warnings (whether positive or reinforcing positive) and texts that do not contain risk warnings. By minimizing the second training generation value, the cloud device can optimize the operator's parameter configuration and improve the accuracy and robustness of its risk information mining. In addition, the second training generation value can also be used to guide subsequent model selection and training strategies. For example, when comparing the performance of different risk information mining operators, the second training generation value can be used as one of the evaluation criteria; when adjusting model parameters or optimizing the training process, the optimization effect can also be judged based on the changes in the second training generation value.

[0189] In one implementation, the pre-learning risk information mining operator includes a first pre-learning risk information mining sub-operator and a second pre-learning risk information mining sub-operator, the first pre-learning risk information mining sub-operator is obtained by calling the first risk information mining sub-operator for autonomous pre-learning, and the second pre-learning risk information mining sub-operator is obtained by calling the second risk information mining sub-operator for autonomous pre-learning; step S13, calling the pre-learning risk information mining operator and the risk warning classifier to classify the tumor test text learning sample for risk warning, and obtaining the inference risk warning degree, including:

[0190] Step S131: inputting the tumor test text learning sample into the first pre-learning risk information mining sub-operator and the second pre-learning risk information mining sub-operator respectively to perform risk information mining, and obtaining a first learning sample risk characterization vector, a second learning sample risk characterization vector and a third learning sample risk characterization vector; the first learning sample risk characterization vector is a vector output by the first pre-learning risk information mining sub-operator and optimized for significance, the second learning sample risk characterization vector is a vector output by the second pre-learning risk information mining sub-operator and optimized for significance, and the third learning sample risk characterization vector is a vector output by the second pre-learning risk information mining sub-operator and not optimized for significance;

[0191] Step S132: classifying the risk warning degree of the third learning sample risk representation vector according to the risk warning classifier to obtain the support coefficient of the third learning sample risk representation vector at each risk warning degree;

[0192] Step S133: Determine the degree of inference risk warning according to the support coefficient.

[0193] In step S131, the cloud device inputs the tumor test text learning samples into the first pre-learning risk information mining sub-operator and the second pre-learning risk information mining sub-operator for risk information mining. These two sub-operators are obtained through autonomous pre-learning based on the first risk information mining sub-operator and the second risk information mining sub-operator, respectively, and have different configuration parameters and internal structures, so they can generate different risk characterization vectors.

[0194] The first pre-learning risk information mining operator performs in-depth analysis on the input text, extracts key risk features, and generates the first learning sample risk representation vector. This vector contains all risk-related information in the text and is an important input for subsequent classification tasks.

[0195] The second pre-learning risk information mining sub-operator is similar to the first operator. It also mines risk information on the input text, but generates the second learning sample risk representation vector. Since the two operators have different configurations, the generated risk representation vectors will also be different.

[0196] Based on the generation of the risk characterization vector of the first learning sample and the risk characterization vector of the second learning sample, the cloud device also performs saliency optimization (weight optimization) on these vectors. Saliency optimization is a feature selection or weighting strategy that aims to highlight the most important features for the classification task while suppressing minor or noise features. Through saliency optimization, the cloud device can further improve the performance of the classifier.

[0197] Assuming that the risk representation vector of the first learning example is [0.5, 0.3, 0.2], after significance optimization, it may become [0.6, 0.1, 0.0], where the weight of the first feature increases, while the weights of the last two features decrease. This indicates that the first feature is more important for the classification task.

[0198] In addition to the first and second learning sample risk characterization vectors, the cloud device also generates a third learning sample risk characterization vector. This vector is generated by the second pre-learning risk information mining sub-operator, but is not significantly optimized. It exists to provide an unoptimized comparison benchmark to evaluate the effect of significant optimization.

[0199] In step S132, the cloud device uses a risk warning classifier to classify the third learning sample risk characterization vector to determine its support coefficient at different risk warning levels. The risk warning classifier is a trained model that can output the probability or confidence that the vector belongs to each category based on the input risk characterization vector.

[0200] The risk warning classifier is, for example, a model based on machine learning or deep learning, such as support vector machine (SVM), random forest, neural network, etc. It can learn the mapping relationship between risk characterization vectors and risk warning levels, and classify new risk characterization vectors accordingly. The output of the classifier can be a probability distribution or confidence vector, indicating the possibility that the input vector belongs to each category. These probabilities or confidence values ​​are called support coefficients, which reflect the degree of trust the classifier has in different categories.

[0201] Assume that after the risk warning classifier classifies the risk representation vector of the third learning sample, it outputs a support coefficient vector containing three elements [0.1, 0.7, 0.2], corresponding to the three categories of low risk, medium risk and high risk respectively. This indicates that the classifier believes that the vector is most likely to belong to the medium risk category (70%), followed by high risk (20%), and finally low risk (10%).

[0202] In step S133, the cloud device determines the inference risk warning level according to the support coefficient. This process may involve analysis and interpretation of the support coefficient vector to select the most likely risk warning level as the final result.

[0203] The cloud device will check the support coefficient vector to find the largest element and its corresponding category. This largest element represents the classifier's highest confidence in the category to which the input vector belongs. Based on the analysis results of the support coefficient, the cloud device will select the category corresponding to the largest element as the inference risk prompt level. This level will serve as the cloud device's risk assessment result for the input tumor test text learning sample.

[0204] Continuing with the above example, since the support coefficient vector is [0.1, 0.7, 0.2], where 0.7 is the maximum value, the cloud device will determine medium risk as the inference risk warning level.

[0205] In one embodiment, the tumor test text includes historical fused tumor test text; step S130, inputting the tumor test text into the pre-learned risk information mining operator in the risk prompt detection algorithm to perform risk information mining, and obtaining a risk characterization vector, including:

[0206] Step S1301: Obtaining a separate tumor test text from the historical fused tumor test text;

[0207] Step S1302: inputting the individual tumor test text into the pre-learned risk information mining operator to perform risk information mining, and obtaining a risk representation vector of the individual tumor test text;

[0208] Step S140, classifying the risk characterization vector according to the risk warning classifier in the risk warning detection algorithm to obtain the risk warning degree corresponding to the tumor test text, including:

[0209] Step S1401: classifying the risk warning degree of the individual tumor test text risk characterization vector according to the risk warning classifier to obtain the support coefficient of the individual tumor test text risk characterization vector at each risk warning degree;

[0210] Step S1402: determining the risk warning level of each individual tumor test text corresponding to each individual tumor test text according to the support coefficient;

[0211] Step S1403: Merge the risk warning levels of each individual tumor test text to obtain the risk warning level corresponding to the historical fused tumor test text.

[0212] In step S1301, the cloud device parses the historical fused tumor test text to extract each individual tumor test text contained therein. This step is the basis for subsequent processing, ensuring that the cloud device can perform independent risk information mining for each individual text.

[0213] The cloud device will structure the historical fusion tumor test text, identify and separate each individual text. This may involve analysis of the text format and segmentation of the content. For example, if the historical fusion text is stored in a specific format (such as CSV, JSON or XML), the cloud device will parse the text content according to the definition of the format.

[0214] For example, suppose a user uploads a CSV file containing the tumor test data of three patients, with each row representing the test results of one patient. The cloud device first identifies the column headers of the CSV file, and then divides the data into three separate tumor test files based on the column headers, each containing the complete test information of one patient.

[0215] In step S1302, the cloud device inputs each individual tumor test text extracted in step S1301 into the pre-learned risk information mining operator for risk information mining. This step aims to extract key risk features from each text and convert them into risk representation vectors.

[0216] The pre-learned risk information mining operator is an unsupervised pre-trained machine learning model that can automatically learn the potential structure and pattern of risk information from text data. After receiving a single tumor test text, the operator will perform an in-depth analysis on it, extract risk-related features (such as abnormal indicators, pathological descriptions, etc.), and generate a corresponding risk characterization vector. The risk characterization vector is a high-dimensional array, in which each element represents the weight or score of a feature in the risk space. These weights or scores reflect the importance of the feature to risk assessment. For example, a text containing "tumor markers are significantly elevated" may get a high score on a feature dimension related to tumor risk.

[0217] For example, suppose there is a single tumor test text with the content "Patient A, lung cancer, abnormally elevated CEA index". After receiving this text, the pre-learning risk information mining operator will identify the key risk features (such as "lung cancer", "abnormally elevated CEA index") and generate a risk representation vector, such as [0.8, 0.6, 0.2, ...], where 0.8 may represent the weight of the feature "lung cancer", 0.6 may represent the weight of "abnormally elevated CEA index", and 0.2 may represent the weight of other less important features.

[0218] In step S1401, the cloud device uses a risk warning classifier to classify the risk warning level of each individual tumor test text risk characterization vector generated in step S1302. This step is to determine the risk warning level corresponding to each text.

[0219] The risk warning classifier is a trained supervised learning model that can predict the corresponding risk warning level based on the input risk representation vector. This model may be a multi-classifier that can distinguish different risk levels (such as low risk, medium risk, high risk, etc.). The classifier will output a probability distribution or confidence vector, indicating the probability or confidence that the input vector belongs to each risk level. These probability or confidence values ​​are called support coefficients, which reflect the classifier's confidence in different risk levels.

[0220] For example, continuing the above example, suppose the risk warning classifier classifies the risk representation vector [0.8, 0.6, 0.2, ...] and outputs a support coefficient vector [0.1, 0.8, 0.1], corresponding to low risk, medium risk, and high risk. This indicates that the classifier believes that the probability that this text belongs to the medium risk level is the highest (80%).

[0221] In step S1402, the cloud device determines the risk warning level of each individual tumor test text based on the support coefficient obtained in step S1401. This step may involve analysis and interpretation of the support coefficient to select the most likely risk level as the final result. The cloud device will check the support coefficient vector to find the largest element and its corresponding risk level. This largest element represents the classifier's highest confidence in the risk level to which the input vector belongs. Based on the analysis results of the support coefficient, the cloud device will select the risk level corresponding to the largest element as the risk warning level of the individual tumor test text. For example, continuing with the above example, since the support coefficient vector is [0.1, 0.8, 0.1], where 0.8 is the maximum value, the cloud device will determine medium risk as the risk warning level of this individual tumor test text.

[0222] In step S1403, the cloud device combines the risk warning levels of each individual tumor test text determined in step S1402 to obtain the overall risk warning level corresponding to the historical fused tumor test text. There are many methods for combining the risk warning levels of individual texts, and feasible methods include weighted average, voting mechanism, fuzzy comprehensive evaluation, etc. These methods have their own advantages and disadvantages, and the specific choice depends on the application scenario and needs.

[0223] One feasible merging strategy is weighted averaging. In this method, the cloud device assigns a weight to each individual text based on its importance (such as timestamp, source reliability, etc.), and then performs a weighted summation of the risk warning levels. The text with the greater weight has a greater impact on the final result. For example, suppose there are two separate tumor test texts, and their risk warning levels are medium risk and high risk, respectively, and they are assigned the same weight (0.5). Then, the overall risk warning level of the historical fused tumor test text can be calculated by weighted averaging, that is, .

[0224] Another merging strategy is a voting mechanism. In this method, the cloud device counts the risk warning level of each individual text and selects the risk level with the most occurrences as the final result. If the number of occurrences is the same, further processing measures (such as considering weights, timestamps, etc.) may be taken to determine the final result.

[0225] For example, suppose there are three separate tumor test texts, and their risk indication levels are low risk, medium risk, and high risk. According to the simple majority voting principle, since medium risk and high risk each appear once, and low risk only appears once, the cloud device may select medium risk or high risk with the most occurrences as the final result (depending on the voting rules).

[0226] For more complex situations, cloud devices can also be combined using fuzzy comprehensive evaluation methods. This method can handle fuzzy and uncertain information and obtain comprehensive evaluation results by constructing fuzzy evaluation matrices and weight vectors.

[0227] For example, in fuzzy comprehensive evaluation, the cloud device first constructs a fuzzy evaluation matrix, in which each row represents the risk evaluation vector (composed of multiple fuzzy subsets) of a single tumor test text, and each column represents a risk level. Then, the cloud device assigns a weight vector to each risk level and obtains a comprehensive evaluation result through matrix operation.

[0228] In one implementation, step S1403, merging the risk warning levels of each individual tumor test text to obtain the risk warning level corresponding to the historical fused tumor test text includes:

[0229] Step S14031: dividing the individual tumor test texts to be classified for risk warning to obtain multiple text sets, and determining the individual tumor test text risk warning degree corresponding to the individual tumor test text in each text set;

[0230] Step S14032: performing first-order fusion of the risk warning levels of individual tumor test texts in each text set to obtain the risk warning level of the individual tumor test text group corresponding to each text set;

[0231] Step S14033: Perform a second-order fusion of the risk warning levels of the individual tumor test text groups corresponding to each text set to obtain the risk warning levels corresponding to the historical fused tumor test texts.

[0232] In step S1403, the cloud device needs to merge the risk warning levels of each individual tumor test text to obtain the overall risk warning level corresponding to the historical fused tumor test text. To achieve this goal, the cloud device adopts a phased fusion strategy, including three main steps: text set division, first-order fusion, and second-order fusion.

[0233] In step S14031, the cloud device first divides the individual tumor test texts to be classified for risk warning and organizes them into multiple text sets. The purpose of this step is to classify texts with similar characteristics or sources into one category for more effective risk assessment and fusion in the future.

[0234] The division of text collections can be based on various factors, such as the timestamp of the text, the source hospital, patient information, test items, etc. For example, the cloud device can group texts in a similar time period into one group based on the timestamp of the text, or group texts from the same hospital into one group based on the source hospital.

[0235] For example, suppose there is a set of individual tumor test texts from three different hospitals (Hospital A, Hospital B, Hospital C) over a period of one month. The cloud device can divide these texts into three sets based on the hospital source: Hospital A text set, Hospital B text set, and Hospital C text set.

[0236] After the division is completed, the cloud device needs to determine the individual tumor test texts in each text set and their corresponding risk warning levels. This may involve performing a risk assessment on each text to obtain its individual risk warning level.

[0237] In step S14032, the cloud device performs a first-order fusion on the risk warning degree of the individual tumor test text in each text set. The purpose of the first-order fusion is to integrate multiple risk warning degrees in the same set into a comprehensive risk assessment result, that is, the risk warning degree of the individual tumor test text group. The first-order fusion can adopt a variety of methods, such as weighted average, median calculation, majority voting, etc. Among them, weighted average is a commonly used method, which assigns a weight to each individual text according to its importance (such as timestamp, source reliability, etc.) and performs weighted summation on the risk warning degree. In step S14033, the cloud device performs a second-order fusion on the risk warning degree of the individual tumor test text group corresponding to each text set to obtain the overall risk warning degree corresponding to the historical fusion tumor test text. The purpose of the second-order fusion is to integrate the risk assessment results of different text sets into a global risk assessment. Similar to the first-order fusion, the second-order fusion can also adopt methods such as weighted average, median calculation, majority voting, etc. However, since the second-order fusion is performed at the text set level, it may be necessary to consider the differences and importance between different text sets. If the weighted average method is used for second-order fusion, the calculation formula is similar to that of first-order fusion, but the meanings of weights and risk warning levels are different. In this case, the weight may represent the importance of the text collection (such as the number of texts in the collection, the authority of the source hospital, etc.), and the risk warning level represents the comprehensive risk assessment result of the text collection.

[0238] Step S1403 is one of the key steps in processing the historical fused tumor test text. It obtains the overall risk warning level corresponding to the historical fused tumor test text by fusing the risk warning levels of individual tumor test texts in stages (first-order fusion and second-order fusion). In this process, the cloud device first divides the text into multiple sets according to certain division criteria, and then performs first-order fusion on the text in each set to obtain a risk assessment result at the set level. Finally, the cloud device performs second-order fusion on the risk assessment results of each set to obtain a global risk assessment. Through this strategy, the cloud device can more accurately assess the overall risk level of the historical fused tumor test text and provide patients with more accurate medical guidance.

[0239] The embodiment of the present invention provides a cloud device, such as Figure 3 As shown, the cloud device 100 includes: a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, such as through a bus 102. Optionally, the cloud device 100 may also include a transceiver 104. It should be noted that in actual applications, the transceiver 104 is not limited to one, and the structure of the cloud device 100 does not constitute a limitation on the embodiments of the present invention.

[0240] An embodiment of the present invention provides a cloud device, and the cloud device in the embodiment of the present invention includes: one or more processors; a memory; one or more computer programs, wherein the one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more programs are executed by the processor, the above method is implemented.

Claims

1. A method for prompting tumor patients to upload laboratory data based on image recognition, characterized in that: Applied to a cloud device, the cloud device is communicatively connected with a target terminal, and the method includes: In response to the data upload instruction of the target terminal, acquiring the tumor test information image uploaded by the user; Obtaining tumor test text contained in the tumor test information image based on OCR recognition; Inputting the tumor test text into the pre-learned risk information mining operator in the risk warning detection algorithm to perform risk information mining to obtain a risk representation vector; the pre-learned risk information mining operator is obtained by autonomously pre-learning the risk information mining operator in the risk warning detection algorithm using a pre-learned tumor test text set; Performing risk warning classification on the risk characterization vector according to the risk warning classifier in the risk warning detection algorithm, obtaining the risk warning degree corresponding to the tumor test text, and returning the risk warning degree to the target terminal; The risk warning detection algorithm is trained through the following training steps: Acquire the pre-learned tumor test text set, tumor test text learning samples, and risk warning degree labels corresponding to the tumor test text learning samples; autonomously pre-learning the risk information mining operator according to the tumor test texts in the pre-learned tumor test text set to obtain a pre-learned risk information mining operator; Retrieving the pre-learned risk information mining operator and the risk warning classifier to perform risk warning classification on the tumor test text learning sample to obtain the inferred risk warning degree; Obtaining an algorithm training cost according to the inference risk warning degree and the risk warning degree label; Optimizing algorithm configuration parameters of at least one of the pre-learning risk information mining operator and the risk warning classifier according to the algorithm training cost; The pre-learning risk information mining operator includes a first pre-learning risk information mining sub-operator and a second pre-learning risk information mining sub-operator, wherein the first pre-learning risk information mining sub-operator is obtained by the first risk information mining sub-operator calling the autonomous pre-learning, and the second pre-learning risk information mining sub-operator is obtained by the second risk information mining sub-operator calling the autonomous pre-learning; calling the pre-learning risk information mining operator and the risk warning classifier to perform risk warning classification on the tumor test text learning sample to obtain the inference risk warning degree, including: The tumor test text learning sample is respectively input into the first pre-learning risk information mining sub-operator and the second pre-learning risk information mining sub-operator for risk information mining to obtain a first learning sample risk characterization vector, a second learning sample risk characterization vector and a third learning sample risk characterization vector; the first learning sample risk characterization vector is a vector output by the first pre-learning risk information mining sub-operator and optimized for significance, the second learning sample risk characterization vector is a vector output by the second pre-learning risk information mining sub-operator and optimized for significance, and the third learning sample risk characterization vector is a vector output by the second pre-learning risk information mining sub-operator and not optimized for significance; Classifying the risk warning degree of the third learning sample risk representation vector according to the risk warning classifier to obtain the support coefficient of the third learning sample risk representation vector at each risk warning degree; The degree of the inference risk warning is determined according to the support coefficient.

2. The method according to claim 1, characterized in that The autonomously pre-learning the risk information mining operator according to the tumor test text in the pre-learned tumor test text set to obtain the pre-learned risk information mining operator includes: Obtaining original tumor test texts, positive tumor test texts, and negative tumor test texts in the pre-learned tumor test text set; wherein the original tumor test texts represent tumor test texts belonging to risk prompts, the positive tumor test texts represent tumor test texts obtained after risk information is injected into tumor test texts, and the negative tumor test texts represent tumor test texts other than the original tumor test texts and the positive tumor test texts in the pre-learned tumor test text set; The risk information mining operator is autonomously pre-learned according to the original tumor test text, the positive tumor test text and the negative tumor test text to obtain the pre-learned risk information mining operator.

3. The method according to claim 2, characterized in that The autonomous pre-learning of the risk information mining operator according to the original tumor test text, the positive tumor test text and the negative tumor test text to obtain the pre-learned risk information mining operator includes: Obtaining an enhanced original tumor test text corresponding to the original tumor test text and an enhanced positive tumor test text corresponding to the positive tumor test text; the enhanced original tumor test text represents a tumor test text obtained by changing at least one of the risk warning and the warning level in the original tumor test text without changing the text semantics; the enhanced positive tumor test text represents a tumor test text obtained by changing the text semantics in the positive tumor test text without changing the risk warning and the warning level; According to the risk information mining operator, risk information mining is performed on the original tumor test text, the enhanced original tumor test text and the negative tumor test text respectively to obtain an original risk characterization vector corresponding to the original tumor test text, an enhanced original risk characterization vector corresponding to the enhanced original tumor test text and a negative risk characterization vector corresponding to the negative tumor test text; and calculation is performed according to the original risk characterization vector, the enhanced original risk characterization vector and the negative risk characterization vector to obtain a first training cost value; Perform risk information mining on the positive tumor test text and the enhanced positive tumor test text according to the risk information mining operator to obtain a positive risk characterization vector corresponding to the positive tumor test text and an enhanced positive risk characterization vector corresponding to the enhanced positive tumor test text; perform calculations according to the positive risk characterization vector, the enhanced positive risk characterization vector and the negative risk characterization vector to obtain a second training cost value; The operator configuration parameters of the risk information mining operator are optimized according to the first training cost value and the second training cost value to obtain the pre-learning risk information mining operator.

4. The method according to claim 3, characterized in that The risk information mining operator includes a first risk information mining sub-operator and a second risk information mining sub-operator, wherein the first risk information mining sub-operator and the second risk information mining sub-operator are twin operators including different operator configuration parameters; the risk information mining is performed on the original tumor test text, the enhanced original tumor test text and the negative tumor test text according to the risk information mining operator, respectively, to obtain the original risk characterization vector corresponding to the original tumor test text, the enhanced original risk characterization vector corresponding to the enhanced original tumor test text and the negative risk characterization vector corresponding to the negative tumor test text, including: According to the first risk information mining sub-operator, risk information mining is performed on the original tumor test text to obtain the original risk characterization vector; according to the second risk information mining sub-operator, risk information mining is performed on the enhanced original tumor test text to obtain the enhanced original risk characterization vector; according to the first risk information mining sub-operator or the second risk information mining sub-operator, risk information mining is performed on the negative tumor test text to obtain the negative risk characterization vector.

5. The method according to claim 4, characterized in that performing risk information mining on the original tumor test text according to the first risk information mining sub-operator to obtain the original risk characterization vector; Performing risk information mining on the enhanced original tumor test text according to the second risk information mining sub-operator to obtain the enhanced original risk characterization vector includes: Performing local feature extraction on the original tumor test text according to the first N-gram model in the first risk information mining sub-operator to obtain corresponding original tumor test text features; performing risk information mining on the original tumor test text features according to the first risk information mining module in the first risk information mining sub-operator to obtain the original risk representation vector; According to the second N-gram model in the second risk information mining sub-operator, local features of the enhanced original tumor test text are extracted to obtain corresponding enhanced original tumor test text features; according to the second risk information mining module in the second risk information mining sub-operator, risk information mining is performed on the enhanced original tumor test text features to obtain the enhanced original risk characterization vector.

6. The method according to claim 3, characterized in that: The performing operation according to the original risk characterization vector, the enhanced original risk characterization vector, and the negative risk characterization vector to obtain a first training cost value includes: Obtaining the first training cost value according to a first vector distance between the original risk characterization vector and the enhanced original risk characterization vector, and a second vector distance between the original risk characterization vector and the negative risk characterization vector; The step of performing risk information mining on the positive tumor test text and the enhanced positive tumor test text respectively according to the risk information mining operator to obtain a positive risk characterization vector corresponding to the positive tumor test text and an enhanced positive risk characterization vector corresponding to the enhanced positive tumor test text includes: Performing risk information mining on the positive tumor test text according to the first risk information mining sub-operator to obtain the positive risk characterization vector; performing risk information mining on the enhanced positive tumor test text according to the second risk information mining sub-operator to obtain the enhanced positive risk characterization vector; The performing calculation according to the positive risk characterization vector, the enhanced positive risk characterization vector, and the negative risk characterization vector to obtain a second training cost value includes: The second training cost value is obtained according to a third vector distance between the positive risk characterization vector and the enhanced positive risk characterization vector, and a fourth vector distance between the original risk characterization vector and the negative risk characterization vector.

7. The method according to claim 6, characterized in that performing risk information mining on the positive tumor test text according to the first risk information mining sub-operator to obtain the positive risk characterization vector; Performing risk information mining on the enhanced positive tumor test text according to the second risk information mining sub-operator to obtain the enhanced positive risk characterization vector includes: Performing local feature extraction on the positive tumor test text according to the first N-gram model in the first risk information mining sub-operator to obtain corresponding positive tumor test text features; performing risk information mining on the positive tumor test text features according to the first risk information mining module in the first risk information mining sub-operator to obtain the positive risk representation vector; According to the second N-gram model in the second risk information mining sub-operator, local features of the enhanced positive tumor test text are extracted to obtain corresponding enhanced positive tumor test text features; according to the second risk information mining module in the second risk information mining sub-operator, risk information mining is performed on the enhanced positive tumor test text features to obtain the enhanced original risk characterization vector.

8. The method according to claim 1, characterized in that The tumor test text includes historical fused tumor test text; the step of inputting the tumor test text into the pre-learned risk information mining operator in the risk prompt detection algorithm to perform risk information mining to obtain a risk characterization vector includes: Obtaining a separate tumor test text from the historical fused tumor test text; Inputting the individual tumor test text into the pre-learned risk information mining operator to perform risk information mining to obtain a risk representation vector for the individual tumor test text; The step of performing risk warning classification on the risk characterization vector according to the risk warning classifier in the risk warning detection algorithm to obtain the risk warning degree corresponding to the tumor test text includes: Classifying the risk warning degree of the individual tumor test text risk characterization vector according to the risk warning classifier to obtain the support coefficient of the individual tumor test text risk characterization vector at each risk warning degree; Determine the risk warning level of the individual tumor test text corresponding to each of the individual tumor test texts according to the support coefficient; Merging the risk warning levels of each of the individual tumor test texts to obtain the risk warning level corresponding to the historical fused tumor test text; The step of merging the risk warning levels of the individual tumor test texts to obtain the risk warning level corresponding to the historical fused tumor test texts includes: Dividing the individual tumor test texts to be classified for risk warning to obtain multiple text sets, and determining the individual tumor test text risk warning degree corresponding to the individual tumor test texts in each text set; Performing first-order fusion of the risk warning levels of the individual tumor test texts in each text set to obtain the risk warning level of the individual tumor test text group corresponding to each text set; The risk warning levels of the individual tumor test text groups corresponding to each text set are subjected to second-order fusion to obtain the risk warning levels corresponding to the historical fused tumor test texts.

9. A cloud device, characterized in that: include: one or more processors; Memory; one or more computer programs; The one or more computer programs are stored in the memory and configured to be executed by the one or more processors, and when the one or more computer programs are executed by the processors, the method according to any one of claims 1 to 8 is implemented.

Images