Malignant tumor AI auxiliary diagnosis system based on Internet hospital

Abstract

The invention discloses a malignant tumor AI auxiliary diagnosis system based on an internet hospital, relates to the technical field of medical informatization and artificial intelligence, and improves the diagnosis and treatment efficiency through individualized early warning and closed-loop intervention. And closed-loop management from early warning to intervention is realized through the doctor-patient interaction module. When high-risk early warning occurs, the system immediately notifies a doctor and a patient to take actions, so that the time link from problem discovery to problem processing is shortened.

Description

Technical Field

[0001] The present invention relates to the fields of medical informatization and artificial intelligence technologies, and particularly to an AI-assisted diagnosis system for malignant tumors based on an Internet hospital. Background Art

[0002] Malignant tumors have a high incidence rate and a high recurrence rate, and have become one of the major chronic diseases threatening human health. After undergoing initial treatments such as surgery, radiotherapy, and chemotherapy, tumor patients often need long-term follow-up monitoring to promptly detect the progression or recurrence of tumors. However, traditional tumor follow-up mainly relies on regular outpatient reviews or patients reporting symptoms on their own. This mode has deficiencies in terms of efficiency and accuracy: on the one hand, if patients do not have obvious discomfort during the follow-up interval, they may neglect the review due to busy work and life, resulting in the failure to detect early changes in the condition; on the other hand, it is difficult for doctors to promptly grasp the condition of patients outside the hospital, and the best intervention time may be missed. With the development of Internet hospitals, more and more patients obtain medical services through online consultations and telemedicine. However, the current Internet follow-up for tumor patients lacks intelligent means and cannot fully utilize a large amount of medical data for risk prediction and early warning.

[0003] Some existing technical solutions attempt to apply artificial intelligence to tumor diagnosis and treatment decision-making or follow-up management. For example, the invention with the patent application number CN109166105B discloses an AI-assisted diagnosis system for tumor malignant risk stratification of artificial intelligence medical images, including data collection, preprocessing, model establishment, verification, and stratification diagnosis modules, etc. This system uses AI to distinguish between benign and malignant and perform risk stratification on medical images, and plays a certain role in improving the accuracy of initial tumor diagnosis. However, this technology mainly focuses on the diagnosis link, lacks support for long-term monitoring and early warning after patients are discharged from the hospital, and fails to cover the needs of the entire tumor treatment process management. In addition, another report shows that a certain tumor specialty hospital has constructed an in-hospital intelligent monitoring and early warning system for critically ill patients. Through multi-dimensional data collection and intelligent scoring, it can update vital signs every 5 minutes and trigger a risk warning within 30 seconds, increasing the in-hospital treatment response speed by 50%. This indicates that there is great potential in using artificial intelligence for medical early warning. However, the above system mainly focuses on the monitoring of acute events during hospitalization and has not solved the pain points of out-of-hospital follow-up early warning for tumor patients.

[0004] In summary, the following deficiencies currently exist in the field of tumor follow-up and early warning: data in each medical link is relatively isolated, lacking a unified platform to integrate the complete medical records of patients in scenarios such as outpatient clinics, hospitalizations, and online consultations; there is a lack of an intelligent analysis model for tumor recurrence risk, and follow-up mainly relies on manual experience; the means of out-of-hospital monitoring for patients are limited, and it is impossible to achieve real-time perception and automatic early warning of changes in symptoms and test results; at the same time, the interaction between patients and doctors is insufficient, lacking closed-loop management, and often showing a lag in response when abnormalities occur. Therefore, there is an urgent need for a new system that integrates Internet hospitals and artificial intelligence technology to achieve the full-cycle management of malignant tumor patients, including the integration of in-hospital and out-of-hospital data, intelligent assisted diagnosis, and individualized monitoring and early warning, so as to improve the follow-up efficiency and the early detection rate of tumor recurrence. Summary of the Invention

[0005] The purpose of the present invention is to provide an AI-assisted diagnosis system for malignant tumors based on Internet hospitals to solve the above technical problems.

[0006] To solve the above technical problems, the present invention adopts the following technical solutions: An AI-assisted diagnosis system for malignant tumors based on Internet hospitals, comprising: Health record module: Connect to multi-source data, and construct an electronic health record through a natural language processing module and image feature extraction technology; AI-assisted diagnosis module: Integrate structured and unstructured features, and use a machine learning model to evaluate the patient's status; Intelligent early warning module: Based on the dynamic changes of the patient's symptom questionnaire, test results, and imaging reports, perform risk prediction, and then monitor and trigger hierarchical early warning through threshold rules and multivariate risk models; Doctor-patient interaction module: Set up closed-loop follow-up management and feedback new data to the electronic health record.

[0007] Preferably, the multi-source data is specifically structured data and unstructured data; before the natural language processing module, the unstructured data needs to be preprocessed. The natural language processing module uses word segmentation, named entity recognition, and attention mechanism technologies to extract information from structured data and unstructured data, and the image feature extraction technology extracts multi-dimensional features of image texture and morphology.

[0008] Preferably, the unstructured features specifically include text features and image features, and the fusion formula of the text features and image features is: , , , where: represents the text feature vector at the t-th time step, represents the length of the text sequence, represents for all The global text obtained after performing average pooling, represents the size of the image feature map, represents for all of the global image after performing average pooling, represents the convolutional output vector at the (i,j) position on the image feature map, represents and are concatenated at the dimension for calculating the fusion weight, represents the trainable parameter of the first layer of gating, which is used to map the concatenated vector to a scalar, and the gating factor is obtained through sigmoid, , the fusion weight of the text and the image "attention", represents the Sigmoid activation function, which maps the result of the linear transformation to between (0,1) as the fusion weight, represents the fused unstructured overall feature vector; the formula for the fusion of the unstructured feature and the structured one is , , where: represents the structured feature vector, , which is represented as the concatenation of the structured and unstructured vectors and serves as the input to the second layer of gating, represents the trainable parameter of the second layer of gating, , represents the fusion weight, approaching 1 means more dependence on the structured feature, and approaching 0 means more dependence on the unstructured feature, represents the obtained multi-modal diagnostic input vector.

[0009] Preferably, the specific steps for the machine learning model to evaluate the patient's state are as follows: Adopt the sigmoid output of the last layer of the neural network, and its formula is: , where: represents the input vector after multi-modal fusion, , represents the trainable parameter, , represents the sigmoid function, , represents the predicted probability of the patient's recurrence risk; Optimize the parameters by minimizing the binary cross-entropy loss, and its formula is: , where: represents the overall loss value, regarded as the parameter set , represents the total number of samples participating in the training, represents the true label of the i-th sample, represents the probability output of the model predicting the i-th sample as "recurrence", The evaluation formula is: , where Set the risk threshold according to expert experience.

[0010] Preferably, the dynamic change is processed by the exponential smoothing time series coding method, and the formula for exponential smoothing time series coding is: , where represents the multi-dimensional original feature vector collected at time, is the smoothed coding representation at time t, represents the previous follow-up time, represents the smoothing coefficient, with a value between 0 and 1, represents the follow-up interval.

[0011] Preferably, the specific formula for the risk prediction is: where is the exponential smoothing time series coding vector at time t, is the coding compared with the previous follow-up time, is the weight vector obtained by training and learning, <00> is the incremental weight coefficient, c is the bias term, is for the coding vector the measure of uncertainty for each dimension, is the base probability based on sigmoid, is to limit within the interval , is , used as the lower bound, is the upper bound, , X(t) is the predicted probability of recurrence of the patient at time t by the model.

[0012] Preferably, the threshold rule includes dynamic threshold learning, and the specific formula for the dynamic threshold learning is: , where is the q-th quantile dynamic threshold maintained at time t, is the same quantile threshold obtained at the previous update at time , used as the starting point for this gradient update, is the update interval, learning rate, is to calculate the difference between the actual distribution ratio at the old threshold and the target quantile q, , where is the empirical cumulative distribution function, <000> is the risk probability output of the i-th patient at time t, q is the target quantile value, equal to the calculated by the risk prediction formula.

[0013] Preferably, the closed-loop follow-up management includes the following steps: Follow-up reminder: Regularly push notifications for filling out questionnaires, uploading reports, or making appointment reminders; AI assistant Q&A: Patients can ask questions to the AI health assistant at any time; Doctor's remote intervention: When patients seek help, doctors can view the latest data on the platform for guidance.

[0014] The beneficial effects of the present invention are as follows: 1. The present invention integrates data comprehensively: This system integrates the online consultation data of the Internet hospital with the outpatient and inpatient data within the hospital, constructs a complete digital health record for patients, and improves the accuracy of auxiliary diagnosis and risk assessment.

[0015] 2. The present invention conducts intelligent analysis and early warning to make up for the deficiencies of manual follow-up: Through the AI-assisted diagnosis and intelligent early warning module, the system can conduct real-time machine learning analysis on the changes in patients' conditions, automatically identify the early signs of tumor recurrence, and classify and give early warnings.

[0016] 3. The present invention improves the diagnosis and treatment efficiency through individualized early warning and closed-loop intervention: The present invention triggers different levels of early warnings according to the individual conditions of patients, and realizes closed-loop management from early warning to intervention through the doctor-patient interaction module. When a high-risk early warning appears, the system immediately notifies doctors and patients to take actions, shortening the time link from problem discovery to handling.

[0017] 4. The present invention optimizes the medical service experience by enhancing patient participation: This system encourages patients to actively participate in their own health management through the Internet hospital platform. With the help of tools such as AI Q&A, patients can conveniently obtain medical consultations and medication guidance outside the hospital, enhancing their understanding of their own conditions and their cooperation with treatment. Brief Description of the Drawings

[0018] Figure 1 It is a flowchart of an AI-assisted diagnosis system for malignant tumors based on the Internet hospital. Detailed Embodiments

[0019] In order to make the technical means, creative features, achieved purposes, and functions of the present invention easy to understand, the present invention will be further described below in conjunction with specific embodiments. However, the following embodiments are only the preferred embodiments of the present invention, not all of them. Based on the embodiments in the implementation manners, other embodiments obtained by those skilled in the art without creative efforts all belong to the protection scope of the present invention. The experimental methods in the following embodiments are all conventional methods unless otherwise specified. The materials, reagents, etc. used in the following embodiments can all be obtained from commercial channels unless otherwise specified.

[0020] Example 1, as Figure 1As shown in the figure, a malignant tumor AI-assisted diagnosis system based on an Internet hospital includes: Health record module: Connect multi-source data, and build an electronic health record through a natural language processing module and image feature extraction technology. The multi-source data specifically refers to structured data and unstructured data; before the natural language processing module, the unstructured data needs to be preprocessed. The natural language processing module uses techniques such as word segmentation, named entity recognition, and attention mechanism to extract information from structured data and unstructured data. The image feature extraction technology extracts multi-dimensional features such as image texture and morphology. The unstructured features specifically include text features and image features. The fusion formula for the text features and image features is: , , , where: represents the text feature vector at the t-th time step, represents the length of the text sequence, represents the global text obtained by averaging all after average pooling, represents the size of the image feature map, represents the global image obtained by average pooling all ; represents the convolutional output vector at the (i,j) position on the image feature map, represents cascaded with in the dimension for calculating the fusion weight, represents the trainable parameter of the first-layer gating, which is used to map the cascaded vector to a scalar, and the gating factor is obtained through sigmoid, , the "attention" fusion weight of text and image, with a value close to 1, depending on text features, and close to 0, depending on image features, represents the Sigmoid activation function, which maps the result of the linear transformation to between (0,1) as the fusion weight, represents the fused unstructured overall feature vector, that is, the text pooling vector and the image pooling vector are mixed proportionally; the fusion formula for unstructured features and structured features is , , where: represents the structured feature vector, which includes numerically represented laboratory test values (such as tumor marker concentrations), vital signs, questionnaire scale scores, etc., and is represented as a unified dimension through preprocessing (such as standardization, missing value imputation), , represents the concatenation of structured and unstructured vectors, which is used as the input of the second-layer gating, Represents the trainable parameters for the second-layer gating, which are used to determine the fusion ratio of structured information and unstructured information. , represents the fusion weight. A value close to 1 indicates a greater dependence on structured features, while a value close to 0 indicates a greater dependence on unstructured features. Represents the obtained multi-modal diagnostic input vector, which will be used as the input to an AI-assisted diagnostic model (such as a deep neural network or a random forest). The specific steps for the machine learning model to evaluate the patient's status are as follows: Adopts the sigmoid output of the last layer of the neural network, and its formula is: , where: Represents the input vector after multi-modal fusion. , represents the trainable parameter. , represents the sigmoid function. , represents the predicted probability of the patient's recurrence risk. Optimizes the parameters by minimizing the binary cross-entropy loss, and its formula is: , where: Represents the overall loss value, regarded as the parameter set , Represents the total number of samples participating in the training. Represents the true label of the i-th sample. 0 indicates "no recurrence", and 1 indicates "recurrence". Represents the probability output of the model predicting "recurrence" for the i-th sample. The evaluation formula is: , where Sets the risk threshold according to expert experience. AI-assisted diagnostic module: Fuses structured and unstructured features and uses a machine learning model to evaluate the patient's status. The specific steps for the machine learning model to evaluate the patient's status are as follows: Adopts the sigmoid output of the last layer of the neural network, and its formula is: , where: Represents the input vector after multi-modal fusion. , represents the trainable parameter. , represents the sigmoid function. , represents the predicted probability of the patient's recurrence risk. Optimizes the parameters by minimizing the binary cross-entropy loss, and its formula is: , where: Represents the overall loss value, regarded as the parameter set , Represents the total number of samples participating in the training. Represents the true label of the i-th sample. 0 indicates "no recurrence", and 1 indicates "recurrence". It represents the probability output of the model predicting "recurrence" for the i-th sample. The evaluation formula is as follows: , where Set the risk threshold according to expert experience; Specifically: Calculate the recurrence probability: The system integrates all the current available medical information of the patient into a unified "fusion feature vector". This vector contains the key information extracted and fused from structured data (such as blood test values, questionnaire scores) and unstructured data (such as the text description of the imaging report, key visual features of the image). The fusion feature vector is input into a pre-trained neural network model. After the neural network model processes this information, a specific function (Sigmoid function) is used in its output layer. The function is to convert the complex calculation results inside the model into a numerical value between 0 and 1, that is, the predicted probability (P_recur) of the patient's current tumor recurrence.

[0021] To put it simply: After the model "digests" various examination reports, symptom descriptions and imaging results of the patient, it outputs a numerical value. The closer this value is to 1, the greater the possibility that the model believes the patient has a recurrence; the closer it is to 0, the smaller the possibility of recurrence. For example, an output of 0.75 means that the model evaluates a 75% recurrence risk.

[0022] Model training and optimization: This neural network model is trained using the real data of a large number of historical patients. During the training process, the system has a clear goal: to make the model's predictions as accurate as possible, especially not to miss the cases that actually have a recurrence. To achieve this goal, the system adopts a strategy called "minimizing binary cross-entropy loss" to adjust the internal parameters of the model (i.e., weights and bias terms).

[0023] The core logic of this strategy is: For each historical patient sample used for training, the system knows its true outcome: either it has actually recurred (labeled as 1) or it has not recurred (labeled as 0).

[0024] The model will calculate a predicted probability (P) based on the historical data of this patient.

[0025] The system will calculate the "gap" or "degree of error" between the model prediction (P) and the true result (0 or 1), and this quantified "degree of error" is the "loss".

[0026] The key point lies in the penalty mechanism: If the patient actually relapses (the true value is 1), but the model predicts a very low probability (for example, P=0.1): the system will consider this a very serious error (missed diagnosis) and give the model a very large penalty (very high loss value), forcing the model to pay more attention to such situations in subsequent adjustments.

[0027] If the patient does not actually have a relapse (the true value is 0), but the model predicts a high probability (for example, P=0.9): the system will also consider this an error (misdiagnosis) and impose a certain penalty, but the severity of this penalty is usually lower than that of a missed diagnosis (this reflects the clinical priority principle of "it is better to have a false alarm than to miss a diagnosis").

[0028] If the model predicts a probability close to the true outcome (e.g., P = 0.85 for a true relapse, or P = 0.15 for no relapse): the system considers the prediction relatively accurate, and the penalty (loss) given is small.

[0029] Model training involves continuously adjusting its parameters to minimize the total loss calculated across all training samples. This means the model has learned to predict recurrence risk with minimal error (especially significant missed diagnoses) based on historical data.

[0030] Risk grading and decision making based on probability: After the model is trained and put into use, when it calculates the predicted probability of recurrence (P_recur) for a new patient, this probability value itself is not the final conclusion.

[0031] The system compares this probability value with a pre-set risk threshold (θ).

[0032] This threshold is set by medical experts based on clinical experience and understanding of the characteristics of different tumors (such as aggressiveness). It represents a critical point for clinical decision-making.

[0033] The classification rules are very simple: If the calculated recurrence probability is greater than or equal to this threshold (P_recur ≥ θ), the system determines that the patient is currently in a high-risk state.

[0034] If the calculated recurrence probability is less than this threshold (P_recur<θ), the system determines that the patient is currently in a low-risk state.

[0035] For example: Suppose the expert sets the high risk threshold θ to 0.7 (i.e. 70%).

[0036] If the model calculates that patient A’s P_recur = 0.75 → 0.75 ≥ 0.7 → a high-risk warning is triggered.

[0037] If the model calculates that the P_recur of patient B = 0.65 → 0.65 < 0.7 → low risk (continue with regular follow-up).

[0038] Intelligent warning module: Based on the dynamic changes of the patient's symptom questionnaire, test results, and imaging reports, it conducts risk prediction, and then monitors and triggers graded warnings through threshold rules and multivariate risk models. The dynamic changes are processed using the exponential smoothing time series coding method, and the formula for exponential smoothing time series coding is: , where represents the multi-dimensional original feature vector collected at time, is the smoothed coding representation at time t, represents the previous follow-up time, represents the smoothing coefficient, with a value between 0 and 1, represents the follow-up interval; the threshold rules include dynamic threshold learning, and the specific formula for risk prediction is: where is the exponential smoothing time series coding vector at time t, is the coding compared to the previous follow-up time, is the weight vector obtained through training and learning, used to determine the contribution of each dimension feature to the risk score, is the incremental weight coefficient. If w > 0, the recent upward change will further amplify the risk score. If w < 0, it will weaken the false alarm risk brought by short-term noise. c is the bias term to ensure that there is still a baseline score when all features approach zero, is the measurement of the uncertainty of each dimension of the coding vector , reflecting the confidence level of the model in the current state. k is the uncertainty adjustment coefficient, used to amplify or reduce the impact on the final probability, is the base probability based on sigmoid, is to limit to the interval inside, is , used as the lower bound, is the upper bound, , X(t) is the predicted probability of the patient's recurrence at time t by the model. The threshold rules include dynamic threshold learning, and the specific formula for its dynamic threshold learning is: , where is the q-th quantile dynamic threshold maintained at time t. q usually takes values of 0.75 or 0.90 (red warning), is the same quantile threshold obtained at the previous update at time , used as the starting point for this gradient update, The update interval, which is the time difference between two threshold adjustments, can correspond to the patient follow-up cycle or a fixed scheduling frequency. The learning rate (gradient step size) determines the strength of the "bias" correction during each update. When it is larger, the adjustment is faster but may be unstable; when it is smaller, the adjustment is smoother but the convergence is slower. is to calculate the actual distribution ratio at the old threshold The difference from the target quantile q, , where is the empirical cumulative distribution function, which counts the proportion of patients among all N patients at time t whose risk scores do not exceed p. Here, 1{*} is the indicator function, which takes 1 when the condition is met and 0 otherwise. is the risk probability output of the i-th patient at time t, and q is the target quantile value, which is equal to that calculated by the risk prediction formula ; Doctor-patient interaction module: Set up closed-loop follow-up management and feedback new data to the electronic health record. The closed-loop follow-up management includes the following steps: Follow-up reminder: Regularly push notifications for filling out questionnaires, uploading reports, or making appointment for medical treatment; AI assistant answering questions: Patients can ask questions to the AI health assistant at any time; Doctor's remote intervention: When a patient asks for help, the doctor can view the latest data on the platform for guidance.

[0039] Example 2: When in use, Patient A registers and binds through the Internet hospital APP of the hospital after discharge. Their personal basic information and in-hospital medical record data are automatically obtained by the health record module to establish an electronic health record. The file of Patient A details the diagnosis type, stage, treatment plan (surgical resection range, chemotherapy plan dosage, etc.) of their malignant tumor, as well as the examination results of each postoperative review. Through docking with the hospital database interface, the system continuously updates the new examination reports and test results of this patient. For example, three months after Patient A's discharge, they had a CT examination at a local hospital and uploaded the report to the Internet hospital platform. The health record module automatically stores the text of the imaging report and uses NLP technology to extract the postoperative changes and suspicious abnormal signals described in it as the input for subsequent AI analysis.

[0040] During the follow-up period, Patient A received a reminder of the follow-up task pushed by the doctor-patient interaction module every two weeks and filled out a standardized tumor symptom questionnaire through the APP, including whether there has been recent weight loss, changes in pain level, newly emerged discomfort symptoms, etc. The health record module organized the questionnaire answers together with the previous data and provided them to the AI-assisted diagnosis module for analysis. At this time, the AI-assisted diagnosis module called the pre-trained large tumor follow-up model to evaluate the latest information of Patient A: the model read his recent laboratory tests (it was found that the tumor marker CEA increased by, for example, 30% compared with last month), the results of the symptom questionnaire (reported persistent dull pain in the abdomen recently), and the characteristics of the historical imaging reports. By integrating these multi-source data, the AI model calculated that the current risk score of Patient A for possible tumor recurrence was relatively high, and the determination result was "medium risk". This preliminary diagnosis conclusion was stored in the system and transmitted to the intelligent warning module.

[0041] Based on the risk level given by the AI module and the preset threshold strategy, the intelligent warning module triggered a yellow warning for Patient A (indicating that increased vigilance is required). The system immediately prompted Patient A on the APP side in the form of an information notice: "Your follow-up results show that you need to be focused on. Please contact your doctor in time or have further examinations." At the same time, Doctor B, the oncologist in charge of Patient A's follow-up, also received a warning message push on his doctor-side system. Since this was a yellow warning, Doctor B could choose to conduct a key re-examination during the next appointment or communicate with the patient through the platform first. Doctor B checked Patient A's follow-up data and, considering the changes in his symptoms and indicators, decided to recommend that the patient have a comprehensive examination in advance to rule out recurrence. So Doctor B sent a message to Patient A through the chat function provided by the doctor-patient interaction module, suggesting that he make an appointment for imaging examinations and blood re-examinations.

[0042] The above shows and describes the basic principles, main features and advantages of the present invention. Those skilled in the art of this industry should understand that the present invention is not limited by the above embodiments. The above embodiments and the descriptions in the specification are only preferred examples of the present invention and are not used to limit the present invention. Without departing from the spirit and scope of the present invention, the present invention will have various changes and improvements, and these changes and improvements all fall within the scope of the present invention claimed. The scope of protection claimed by the present invention is defined by the appended claims and their equivalents.

Claims

1. An AI-assisted diagnosis system for malignant tumors based on Internet hospitals, characterized in that, Including: Health record module: Connecting multi-source data, constructing an electronic health record through a natural language processing module and image feature extraction technology; The specific multi-source data is structured data and unstructured data; the unstructured features include text features and image features. The unstructured data is preprocessed before the natural language processing module. The natural language processing module uses techniques such as word segmentation, named entity recognition, and attention mechanism to extract information from structured data and unstructured data. The image feature extraction technology extracts multi-dimensional features of image texture and morphology; AI-assisted diagnosis module: Fusing structured and unstructured features, using a machine learning model to evaluate the patient's condition.

2. The AI-assisted diagnosis system for malignant tumors based on the Internet hospital according to claim 1, wherein: The fusion formula of the text feature and the image feature is: Wherein: represents the text feature vector at the t-th time step, represents the length of the text sequence, represents for all the global text obtained after average pooling, represents the size of the image feature map, represents for all the global image after average pooling, represents the convolution output vector at the (i, j) position on the image feature map, represents cascaded with in the dimension for calculating the fusion weight, represents the trainable parameter of the first layer of gating, which maps the cascaded vector to a scalar, and obtains the gating factor through sigmoid, , the "attention" fusion weight of text and image, represents the Sigmoid activation function, which maps the result of the linear transformation to between (0, 1) as the fusion weight, represents the fused unstructured overall feature vector.

3. The AI-assisted diagnosis system for malignant tumors based on the Internet hospital according to claim 1, wherein: The fusion formula of the unstructured feature and the structured feature: Wherein: represents a structured feature vector, , which is expressed as a concatenation of a structured and an unstructured vector and serves as the input to the second-layer gating, represents the trainable parameters of the second-layer gating, , represents the fusion weight. A value closer to 1 indicates a greater dependence on the structured features, while a value closer to 0 indicates a greater dependence on the unstructured features, represents the obtained multi-modal diagnostic input vector.

4. The malignant tumor AI-assisted diagnosis system based on the Internet hospital according to claim 1, wherein: The specific steps of using the machine learning model to evaluate the patient's condition are: Adopt the sigmoid output of the last layer of the neural network, and its formula is: , where: represents the input vector after multi-modal fusion, , represents the trainable parameter, , represents the sigmoid function, , represents the predicted probability of the patient's recurrence risk; Optimizing the parameters by minimizing the binary cross-entropy loss, the formula is: , where: represents the overall loss value, regarded as the parameter set , represents the total number of samples participating in the training, represents the true label of the i-th sample, represents the probability output that the model predicts the i-th sample as "recurrence", The formula for evaluating the patient's condition is: , where Set the risk threshold according to expert experience.

5. The AI-assisted diagnosis system for malignant tumors based on the Internet hospital according to claim 1, wherein: It also includes an intelligent warning module: Based on the patient's symptom questionnaire, examining the dynamic changes in test results and imaging reports, performing risk prediction, and then monitoring and triggering hierarchical warnings through threshold rules and multi-variable risk models; The specific formula for risk prediction is as follows: In the formula is the exponential smoothing time series coding vector at time t, is the coding compared to the previous follow-up time, is the weight vector obtained through training and learning, is the incremental weight coefficient, c is the bias term, is for the coding vector the measure of uncertainty for each dimension, is the base probability based on sigmoid, is to limit within the interval inside, is , used as the lower bound, is the upper bound, [[ID=3']], X(t) is the predicted probability of the patient's recurrence at time t by the model.

6. The AI-assisted diagnosis system for malignant tumors based on the Internet hospital according to claim 5, wherein: The dynamic changes are processed by using the exponential smoothing time series coding method, and the formula for exponential smoothing time series coding is as follows: In the formula represents the multi-dimensional original feature vector collected at time is the smoothed coding representation at time t, represents the previous follow-up time, represents the smoothing coefficient, and its value ranges between 0 and 1, represents the follow-up interval.

7. An AI-assisted diagnosis system for malignant tumors based on an Internet hospital according to claim 6, characterized in that: The threshold rule described above includes dynamic threshold learning, and the specific formula for the dynamic threshold learning is as follows: , where The q-th quantile dynamic threshold maintained at time t, is the same quantile threshold obtained in the previous update at time , which is used as the starting point for the current gradient update, is the update interval, is the learning rate, is to calculate the difference between the actual distribution ratio at the old threshold and the target quantile q, , where is the empirical cumulative distribution function, is the risk probability output of the i-th patient at time t, and the q target quantile value is equal to that calculated by the risk prediction formula .

8. An AI-assisted diagnosis system for malignant tumors based on an Internet hospital according to claim 7, characterized in that: It also includes a doctor-patient interaction module: Setting up closed-loop follow-up management and feeding back new data to the electronic health record. The closed-loop follow-up management includes the following steps: Follow-up reminder: Regularly push notifications for filling out questionnaires, uploading reports, or making appointment for medical treatment; AI assistant answering questions: Patients can ask questions to the AI health assistant at any time; Doctor's remote intervention: When a patient asks for help, the doctor can view the latest data on the platform for guidance.

9. A storage medium, the storage medium being a computer-readable storage medium for computer-readable storage, characterized in that: The storage medium stores one or more programs, and the one or more programs can be executed by one or more processors to implement the malignant tumor AI-assisted diagnosis system based on the Internet hospital according to any one of claims 1-8.

Images