Physical examination triage method and system based on deep learning

Abstract

The invention relates to the technical field of medical information services, in particular to a physical examination triage method and system based on deep learning, and achieves the precision and individuation of physical examination recommendation through the fusion of medical knowledge and artificial intelligence. The system intelligently generates questionnaires based on user portraits, efficiently collects key health information, and converts the key health information into structured features. The deep learning model quantifies various physical examination requirements according to the physical examination requirements, personalized adjustment is carried out for user risk features and item importance through a dynamic threshold mechanism, and accurate screening is achieved. The finally generated physical examination scheme gives consideration to safety and efficiency, significantly improves the early detection rate of serious diseases, optimizes medical resource allocation, and provides efficient and accurate health management services for users.

Description

Technical Field

[0001] This application relates to the field of medical information service technology, specifically a deep learning-based physical examination triage method and system. Background Technology

[0002] The core value of health checkup triage lies in its ability to conduct precise risk screening and needs matching based on an individual's multidimensional characteristics (such as age, gender, family history, and lifestyle habits), thereby directing limited health checkup resources to the most needed examinations. This not only avoids meaningless over-examination and saves personal and social costs, but also significantly improves the detection rate of early-stage serious diseases, achieving true preventive medicine and fundamentally improving user experience and health outcomes.

[0003] However, current technological practices have significant shortcomings in the health checkup triage process. The current mainstream model heavily relies on manual inquiry or simple rule-based judgment (such as selecting a package solely based on age and gender), essentially being experience-driven or a passive "package supermarket" style of selection. This model lacks systematic analysis and quantitative assessment of complex personal health information, cannot handle the interactive effects between multiple features, and lacks the ability to continuously learn and optimize. The result is crude recommendations with weak targeting, the omission of numerous personalized risk signals, and the failure to deliver on the promise of personalized health management. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a deep learning-based method and system for physical examination triage to solve the problems in the background art.

[0005] To achieve the above objectives, this application adopts the following technical solution:

[0006] This application discloses a deep learning-based physical examination triage method, comprising the following steps:

[0007] Obtain the user's basic personal information, which includes gender, age, occupation, and place of residence;

[0008] Based on the personal basic information, target questions are selected from a pre-built question bank and a target questionnaire is generated. The target questionnaire is then sent to the user to obtain questionnaire information, which includes lifestyle information, personal medical history, family medical history, and descriptions of recent symptoms.

[0009] Upon receiving questionnaire information from a user, a target entity is extracted from the questionnaire information, and a user feature vector is generated based on the user's basic information and the target entity, wherein the questionnaire information is generated based on the target questionnaire.

[0010] The user feature vector is input into a pre-built physical examination prediction model to obtain the confidence levels of various physical examination items;

[0011] A dynamic screening threshold is constructed for the user, and target physical examination items with a confidence level greater than or equal to the dynamic screening threshold are selected; a physical examination plan is generated based on the target physical examination items, and the physical examination is sent to the user.

[0012] In one embodiment of this application, the method for constructing the question library includes:

[0013] Obtain medical knowledge materials;

[0014] Knowledge extraction was performed on the aforementioned medical knowledge materials to obtain multiple physical examination suggestions;

[0015] Entity extraction is performed on the multiple physical examination suggestions to obtain physical examination categories and multiple identity information limiting entities, wherein the identity information limiting entities are used to limit the value range of personal basic information;

[0016] A knowledge template is constructed based on the physical examination category and multiple identity information-defined entities;

[0017] Based on the physical examination category in each knowledge template, expand the question text, and construct a structured question template based on the physical examination category, the question text, and multiple identity information to limit the entity, wherein the question text is a guiding question text related to the physical examination category;

[0018] Each structured question template is assigned an importance score and type label. A question library is built based on multiple structured question templates and the importance scores and label types of the question text within multiple structured question templates. Structured question templates with the same physical examination category are merged.

[0019] In one embodiment of this application, the importance score of each structured question template is evaluated, including:

[0020] Obtain the basic importance score for each medical examination category. And score the basic importance of the physical examination categories. As the basis for importance scoring for all corresponding question texts The basic importance score Calibrated by an expert system, For physical examination category index, Index the question text;

[0021] Calculate a specific score for each medical examination category based on multiple identity-based entities for each category. And the specificity score of the physical examination category. As a specific score corresponding to all question texts The mathematical expression for the specificity score is:

[0022]

[0023] In the formula, An index for identity information entities. For the first The theoretical range of values ​​for an identity information entity. For the first The corresponding item of the individual examination The scope of each identity information entity is limited;

[0024] Calculate the timeliness score of the question text corresponding to each physical examination category. The timeliness score The mathematical expression is:

[0025]

[0026] In the formula, For the theoretical maximum score, For the first The time-depletion coefficient of individual test items, Indicates the first The difference between the publication time of the issue text for an individual medical examination item and the current time;

[0027] Based on the aforementioned basic importance score The specificity score and the aforementioned timeliness score Calculate importance score The importance score The mathematical expression is:

[0028]

[0029] In the formula, As the first weight, As the second weight, It is the third weight.

[0030] In one embodiment of this application, selecting target questions from a pre-built question bank and generating a target questionnaire based on the personal basic information includes:

[0031] The personal basic information is compared with the identity information-limited entity of multiple structured question templates in a pre-built question library, and the question text in the structured question template that matches the personal basic information is used as the target question text;

[0032] Multiple target question texts are divided based on type labeling to obtain multiple sets of target question texts of different types;

[0033] Extract the target questions with the highest importance scores from multiple types of target question text sets, so that the target questions of multiple types are distributed according to a preset ratio;

[0034] A target questionnaire is generated based on multiple target questions, wherein the target questions in the target questionnaire are ranked according to their importance score.

[0035] In one embodiment of this application, extracting target entities from the questionnaire information and generating user feature vectors based on the personal basic information and the target entities includes:

[0036] Based on the medical named entity recognition model, logical decision entities or numerical entities of named entities and command entities are extracted from the questionnaire information, wherein the logical decision entities include selection entities, affirmative entities and negative entities;

[0037] The logical decision entity or numerical entity of the named entity or command entity is concatenated to obtain the target entity; and the target entity and the personal basic information are vectorized and concatenated to obtain the user feature vector.

[0038] In one embodiment of this application, the physical examination prediction model adopts a multilayer perceptron network structure, wherein the method for constructing the physical examination prediction model includes:

[0039] S1, Obtain a physical examination data sample, wherein the physical examination data sample includes a personal information sample, a questionnaire information sample, and physical examination items after desensitization processing;

[0040] S2, extract user feature vector samples from the physical examination data samples, and construct a training dataset by combining the physical examination items as training labels;

[0041] S3, Extract training data from the training dataset and input the training data into the multilayer perceptron network structure to obtain the prediction result;

[0042] S4, calculate the loss between the prediction result and the label based on the pre-constructed loss function, and adjust the internal parameters of the multilayer perceptron network structure based on the backpropagation of the loss combined with the gradient descent method;

[0043] S5. Repeat steps S3-S4 until training is complete and the physical examination prediction model is obtained.

[0044] In one embodiment of this application, the mathematical expression of the loss function is:

[0045]

[0046]

[0047]

[0048] In the formula, Indicates the total loss. Represents cross-entropy loss, The weights represent the cross-entropy loss. Indicates the weighted focus loss. This indicates the weighted focus loss weight. Represents the regularization coefficient. This represents the set of all trainable parameters in the model. Indicates the training batch size. Indicates the number of items included in the physical examination. For sample index, For tag indexing, Indicates the first The first sample The true probability of individual test items Indicates the first The first sample Predicted probability of individual test items Indicates the first The importance weight of individual test items Represents the positive sample balance coefficient. This indicates the focus parameter.

[0049] In one embodiment of this application, constructing the dynamic filtering threshold for the user includes:

[0050] Obtain the validation set of the physical examination prediction model;

[0051] For each physical examination item, a confidence threshold that maximizes the F1 score of the predicted result is searched in the validation set, and a dynamic screening threshold is constructed based on the confidence thresholds of multiple physical examination items.

[0052] In one embodiment of this application, generating a physical examination plan based on the target physical examination items includes:

[0053] The examination details and prices for multiple target health check-up items are sent to the user.

[0054] Upon receiving selection information from the user, a medical examination plan is constructed based on the final medical examination items selected by the user.

[0055] This application also provides a deep learning-based physical examination triage system, including:

[0056] The acquisition module is used to acquire the user's basic personal information, which includes gender, age, occupation, and place of residence.

[0057] An interaction module is used to select target questions from a pre-built question bank based on the personal basic information and generate a target questionnaire, and send the target questionnaire to the user to obtain questionnaire information, wherein the questionnaire information includes lifestyle information, personal medical history, family medical history and description of recent symptoms;

[0058] The feature extraction module is used to extract target entities from the questionnaire information received from the user, and generate user feature vectors based on the personal basic information and the target entities, wherein the questionnaire information is generated based on the target questionnaire.

[0059] The prediction module is used to input the user feature vector into a pre-built physical examination prediction model to obtain the confidence levels of various physical examination items.

[0060] The filtering and solution construction module is used to construct the user's dynamic filtering threshold, and filter out target physical examination items with a confidence level greater than or equal to the dynamic filtering threshold; generate a physical examination plan based on the target physical examination items, and send the physical examination to the user.

[0061] The beneficial effects of this application are as follows: This application presents a deep learning-based physical examination triage method and system, which achieves precise and personalized physical examination recommendations by integrating medical knowledge and artificial intelligence. The system intelligently generates questionnaires based on user profiles, efficiently collects key health information, and transforms it into structured features. The deep learning model quantifies various physical examination needs accordingly, and then uses a dynamic threshold mechanism to personalize adjustments based on user risk characteristics and the importance of each item, achieving accurate screening. The final generated physical examination plan balances safety and efficiency, significantly improves the early detection rate of serious diseases, optimizes the allocation of medical resources, and provides users with efficient and accurate health management services. Attached Figure Description

[0062] The present application will be further described below with reference to the accompanying drawings and embodiments:

[0063] Figure 1 This is a diagram illustrating an application scenario of a deep learning-based physical examination triage system in one embodiment of this application.

[0064] Figure 2 This is a flowchart illustrating a deep learning-based physical examination triage method in one embodiment of this application;

[0065] Figure 3 This is a flowchart illustrating the problem library construction process in one embodiment of this application;

[0066] Figure 4 This is a structural diagram of a deep learning-based physical examination triage system shown in one embodiment of this application. Detailed Implementation

[0067] The following specific examples illustrate the implementation of this application. Those skilled in the art can easily understand other advantages and effects of this application from the content disclosed in this specification. This application can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of this application. It should be noted that, unless otherwise specified, the following embodiments and features in the embodiments can be combined with each other.

[0068] It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of this application. Therefore, the drawings only show the layers related to this application and are not drawn according to the actual number, shape and size of the layers in the actual implementation. In the actual implementation, the form, number and proportion of each layer can be arbitrarily changed, and the layer layout may also be more complex.

[0069] Numerous details are explored in the following description to provide a more thorough explanation of embodiments of this application; however, it will be apparent to those skilled in the art that embodiments of this application may be practiced without these specific details.

[0070] Figure 1 This is a diagram illustrating an application scenario of a deep learning-based physical examination triage system in one embodiment of this application. Figure 1 As shown, the scenario design in this application includes a cloud component and a smart terminal 110. The cloud component includes an interactive server 120, a question database server 130, and a prediction model server 140.

[0071] During the physical examination triage, the user accesses the interactive server 120 through the smart terminal 110 and uploads personal basic information to the interactive server 120 through registration information / information upload, etc. The interactive server 120 extracts question text from the question database server 130 according to a specific algorithm based on the personal information, generates a questionnaire, and returns the questionnaire to the smart terminal 110.

[0072] After filling out the questionnaire on the smart terminal 110, the user sends the questionnaire information to the interactive server 120 through the smart terminal 110.

[0073] The interactive server 120 forwards the personal basic information and questionnaire information to the prediction model server 140 for predicting the physical examination items.

[0074] The predicted physical examination items are returned to the interactive server 120. After a final round of filtering and interaction of the physical examination items, the final physical examination plan is obtained.

[0075] In the above process, multiple rounds of interaction were used to extract user information, make accurate recommendations, and select projects, thereby generating a health checkup plan that matches user characteristics and meets user subjective needs. Specifically...

[0076] Figure 2 This is a flowchart illustrating a deep learning-based physical examination triage method in one embodiment of this application, as shown below. Figure 2 As shown, a deep learning-based physical examination triage method in this application mainly includes the following steps:

[0077] S210, Obtain the user's basic personal information, wherein the basic personal information includes gender, age, occupation and place of residence;

[0078] In this application, basic personal information of users is obtained through user-initiated uploads.

[0079] Basic personal information mainly includes gender, age, occupation, and place of residence. These four types of personal information can serve as clues for generating a questionnaire, thereby enabling targeted questioning and avoiding redundant questionnaires that fail to obtain useful information.

[0080] For example: The user's personal information is: Male, 48 years old, software engineer, residing in Haidian District, Beijing.

[0081] Based on this information, the system intelligently matches and generates the following concise questionnaire about recent symptom descriptions from the question database:

[0082] Occupation-related: "Do you experience neck and shoulder pain, or wrist numbness (such as mouse hand) due to prolonged desk work?"

[0083] Regional related: "During the winter smog in Beijing, have you experienced respiratory discomfort such as dry cough or chest tightness?"

[0084] Age + Gender: "Have you experienced any prostate-related symptoms in the past six months, such as increased nighttime urination or difficulty urinating?"

[0085] In addition, it can expand on other medical history, family medical history and other questions based on basic information.

[0086] S220, based on the personal basic information, select a target question from a pre-built question bank and generate a target questionnaire, and send the target questionnaire to the user to obtain questionnaire information, wherein the questionnaire information includes lifestyle information, personal medical history, family medical history and recent symptom descriptions;

[0087] This application requires accurate prediction of physical examination items based on personal information and questionnaire data. The questionnaire data needs to generate different question texts for different groups of people. This application pre-builds a question database and extracts question texts from the database based on personal information. The process of building the question database is shown below.

[0088] Figure 3 Here is a flowchart of the problem library construction process in one embodiment of this application, such as... Figure 3 As shown, the methods for constructing the question bank include:

[0089] (1) Obtaining medical knowledge materials;

[0090] The medical knowledge materials in this application can be medical guidelines or relevant guidance materials on physical examinations.

[0091] (2) Extract knowledge from the medical knowledge materials to obtain multiple physical examination suggestions;

[0092] Specifically, multiple health checkup recommendations are extracted from medical knowledge materials, for example:

[0093] It is recommended that men over 40 years of age, especially those with a family history of PSA infection, undergo PSA testing and digital rectal examination annually.

[0094] "Men over 40 years old in the Beijing-Tianjin-Hebei region need to pay attention to lower urinary tract symptoms (LUTS)," etc.

[0095] (3) Entity extraction is performed on the multiple physical examination suggestions to obtain physical examination categories and multiple identity information limiting entities, wherein the identity information limiting entities are used to limit the value range of personal basic information;

[0096] Regarding the extracted medical examination recommendations, this application uses entity extraction to extract the medical examination items and identity information-restricted entities. Identity information restrictions not mentioned are considered unrestricted. For example:

[0097] Physical examination items: "PSA test / digital rectal examination";

[0098] Identity information restrictions: "Male", "40 years or older", "any occupation", "any region".

[0099] Physical examination item: "Lower urinary tract symptoms (LUTS)";

[0100] Identity information restrictions: "Male", "40 years or older", "Any occupation", "Beijing-Tianjin-Hebei region".

[0101] (4) Construct a knowledge template based on the physical examination category and multiple identity information to define entities;

[0102] The knowledge template processes the above entities into a structured form, for example:

[0103] Knowledge Template 1 = {"PSA test / digital rectal exam", "male", "40 years and older", "any occupation", "any region"}.

[0104] Knowledge Template 2 = {"Lower urinary tract symptoms (LUTS)", "Male", "40 years and older", "Any occupation", "Beijing-Tianjin-Hebei region"}.

[0105] This knowledge template is displayed in text form, but in practical applications, it needs to be stored in code.

[0106] (5) Expand the question text based on the physical examination category in each knowledge template, and construct a structured question template based on the physical examination category, the question text and multiple identity information to limit the entity, wherein the question text is a guiding question text related to the physical examination category;

[0107] In this application, an LLM model is used to expand the question text for each knowledge template in the physical examination category. The expanded question text consists of four types of questions, including: habit information, personal medical history, family medical history, and description of recent symptoms.

[0108] Here is an example of guiding an LLM model:

[0109] "[Character] You are a senior general practitioner and questionnaire design expert."

[0110] [Task] Generate user-friendly screening questions for physical examination items.

[0111]

Input

[0112] [Generation Requirements] Questions should be generated strictly according to four categories (1-2 questions per category):

[0113] Habit information: Focus on modifiable risk factors (sedentary lifestyle, diet, exercise, etc.)

[0114] Personal medical history: Previous urological / prostate-related diagnoses or surgeries

[0115] Family medical history: Medical history of first-degree male relatives (limited to first-degree relatives)

[0116] Recent symptoms: Subjective feelings such as pain during urination that have appeared in the past 3 months.

[0117] Based on the above guiding text, the LLM model outputs the following text:

[0118] (a) Habitual information:

[0119] Does your work or daily life require you to sit for long periods of time (e.g., more than 4 hours continuously) more than 3 days a week?

[0120] Do you typically drink less than 1.5 liters of water per day (including plain water, tea, soup, etc.) (about 3 bottles of regular mineral water)?

[0121] (b) Personal medical history:

[0122] Have you ever been diagnosed with prostatitis, benign prostatic hyperplasia (BPH), or undergone prostate-related surgery?

[0123] Do you have diabetes or high blood pressure and are currently taking medication?

[0124] (c) Family medical history:

[0125] Has any of your fathers or brothers been diagnosed with prostate cancer before the age of 65?

[0126] (d) Recent symptoms:

[0127] In the past 3 months, have you frequently experienced urinary frequency (significantly increased number of urinations), urinary urgency (sudden urge to urinate that is difficult to hold), or nighttime urination more than twice?

[0128] In the past three months, have you noticed a significantly weaker urine stream, difficulty urinating, or a feeling of incomplete urination?

[0129] In addition, the question text can be configured with options to allow users to choose directly, thus avoiding situations where users are less willing to give a subjective answer due to the need to input text.

[0130] Finally, the question text generated by the LLM model needs to be submitted to experts for review, and only the approved question texts should be retained. After generating the question text, it should be stored in a knowledge template to form a structured question template. For example:

[0131] Structured Question Template 2 = {"Lower urinary tract symptoms (LUTS)", "Male", "Over 40 years old", "Any occupation", "Beijing-Tianjin-Hebei region", Personal habits (Q1, Q2), Personal medical history (Q1, Q2), Family medical history (Q1, Q2), Recent symptoms (Q1, Q2)}.

[0132] (6) Assess the importance and type of each structured question template, and build a question library based on the importance scores and annotation types of the question texts within multiple structured question templates. Structured question templates with the same physical examination category are merged.

[0133] In this application, the question text needs to be updated over time. Furthermore, when outputting questions, different questions need to be sorted to output the key questions. Therefore, it is also necessary to score the importance of the question text for each structured question template. The importance scoring process includes:

[0134] (6-1) Obtain the basic importance score for each physical examination category. And score the basic importance of the physical examination categories. As the basis for importance scoring for all corresponding question texts The basic importance score Calibrated by an expert system, For physical examination category index, Index the question text;

[0135] First, a baseline importance score is assigned by experts for each category of physical examination, for example, the baseline importance score for prostate examination. (Maximum score 1.0)

[0136] (6-2) Calculate the specificity score for each physical examination category based on multiple identity information of entities for each physical examination category. And the specificity score of the physical examination category. As a specific score corresponding to all question texts The mathematical expression for the specificity score is:

[0137]

[0138] In the formula, An index for identity information entities. For the first The theoretical range of values ​​for an identity information entity. For the first The corresponding item of the individual examination The scope of each identity information entity is limited;

[0139] The specificity score measures the accuracy of the target population for health check recommendations. The stricter the limiting conditions, the more targeted the recommendations, and the higher the score.

[0140] The theoretical range of values ​​for identity information entities is based on their characteristics, for example:

[0141] gender: =2 (male, female);

[0142] age: =100 (0-99 years old, 100 possible values);

[0143] Profession: =50 (assuming there are 50 types of occupations);

[0144] area: (34 provincial-level administrative regions);

[0145] Regarding the aforementioned scope without limitation, the first The range of values ​​for an identity information entity Its range of values ​​is the theoretical range, that is... .

[0146] (6-3) Calculate the timeliness score of the question text corresponding to each physical examination category. The timeliness score The mathematical expression is:

[0147]

[0148] In the formula, For the theoretical maximum score, For the first The time-depletion coefficient of individual test items, Indicates the first The difference between the publication time of the issue text for an individual medical examination item and the current time;

[0149] The timeliness score ensures that the latest medical evidence receives higher weight, and the introduction of the timeliness score can also eliminate outdated knowledge, ensuring that the question bank is always up-to-date.

[0150] In terms of timeliness rating In the calculation formula, the aging attenuation coefficient is used. Limit the decay of different types of physical examinations over time. For example, the decay coefficient of influenza tests is relatively large and they need to be updated frequently.

[0151] (6-4) Based on the aforementioned basic importance score The specificity score and the aforementioned timeliness score Calculate importance score The importance score The mathematical expression is:

[0152]

[0153] In the formula, As the first weight, As the second weight, It is the third weight.

[0154] Finally, a weighted average is used to merge the three evaluation metrics of the question text into an importance score. This facilitates the subsequent ranking of the matched question texts.

[0155] Based on the question database constructed above, the process of extracting questions and generating questionnaires includes:

[0156] S221, compare the personal basic information with the identity information-limited entities of multiple structured question templates in the pre-built question library, and take the question text in the structured question template that matches the personal basic information as the target question text;

[0157] Specifically, the user's four basic pieces of information—gender, age, occupation, and place of residence—are compared with the "identity information limiting entities" preset for each question template in the question database. Only when the user information meets all the limiting conditions (value range) of a template will the question corresponding to that template be selected.

[0158] Rule-based precision filtering. Its logic is to ensure that the questions asked are strictly correlated with the statistical characteristics of the users.

[0159] S222, Based on type labeling, multiple target question texts are divided to obtain a set of target question texts of multiple types;

[0160] Based on the pre-labeled "type tags", all the questions selected in the previous step are categorized into four independent sets: "lifestyle information", "personal medical history", "family medical history" and "recent symptom description".

[0161] S223, extract the target questions with the highest importance scores from the target question text sets of multiple types respectively, so that the target questions of multiple types are distributed according to a preset ratio;

[0162] Set a preset target percentage for each question type (e.g., habits 30%, personal history 25%, family history 20%, symptoms 25%). Based on the total number of questions planned for the questionnaire, calculate the number of questions to be drawn for each type, and then select the corresponding number of questions with the highest importance score from the question set for that type.

[0163] Quota sampling is combined with the principle of importance priority. The formula can be simplified to: Number of questions selected in each category = Total number of questions × Preset category ratio, and questions are selected in descending order of importance within each category.

[0164] This achieves in-depth optimization and balance of the questionnaire. On the one hand, it ensures the comprehensiveness of the assessment through proportional control, without neglecting any health dimension; on the other hand, it ensures that the most critical and valuable questions are prioritized in each dimension through importance ranking, thereby maximizing the efficiency and quality of information collection within a limited number of questions.

[0165] S224, Generate a target questionnaire based on multiple target questions, wherein the target questions in the target questionnaire are ranked according to their importance score.

[0166] All selected questions are then sorted according to their importance score from highest to lowest, generating an ordered questionnaire list. Further adjustments can be made based on the logical relationships between question types (e.g., placing basic habit questions first and specific symptom questions later). Prioritization and cognitive flow optimization are then implemented. Placing the most important questions first aligns with the general principle of "decreasing user attention."

[0167] The above process significantly improves the reliability of data collection and user experience. The most important information is collected first, reducing the risk of missing core data due to user fatigue or abandonment midway. The logical question flow also makes the answering process smoother and more natural, improving the accuracy of answers and the completion rate of the questionnaire.

[0168] S230, upon receiving questionnaire information from a user, extracting the target entity from the questionnaire information and generating a user feature vector based on the personal basic information and the target entity, wherein the questionnaire information is generated based on the target questionnaire;

[0169] The process of extracting target entities and generating user feature vectors includes:

[0170] S231, Based on the medical named entity recognition model, extract logical judgment entities or numerical entities of named entities and command entities from the questionnaire information, wherein the logical judgment entities include selection entities, affirmative entities and negative entities;

[0171] First, a pre-trained NER model (such as BERT-BiLSTM-CRF) from the medical field is used to identify NER from free text:

[0172] Disease entities (hypertension, diabetes);

[0173] Physical symptoms (chest pain, dizziness);

[0174] Physical body parts (lungs, heart);

[0175] Time-based entities (3 months, 2 years);

[0176] Frequency entities (daily, occasionally);

[0177] It also identifies logical relationships in structured answers based on a rule engine: for example, options "A / B / C" (in conjunction with the multiple-choice questions mentioned above), or affirmative entities ("Yes", "Have", "Confirmed") and negative entities ("No", "None", "Do not have").

[0178] The above process combines the semantic understanding capabilities of deep learning models for unstructured text with the precise parsing capabilities of rule engines for structured data. Named entity recognition captures medical semantics, logical entities handle binary decision-making, and numerical entities quantify risk levels.

[0179] S232, the logical decision entity or numerical entity of the named entity or command entity is concatenated to obtain the target entity; and the target entity and the personal basic information are vectorized and concatenated to obtain the user feature vector.

[0180] After extracting the target entity, a word embedding model is used to perform vector processing on the target entity and personal basic information, thereby obtaining a computer-recognizable user feature vector.

[0181] S240, The user feature vector is input into the pre-built physical examination prediction model to obtain the confidence levels of various physical examination items;

[0182] This application uses deep learning to construct a health check prediction model that represents the relationship between user feature vectors and the confidence levels of health check items. The method for constructing the health check prediction model includes:

[0183] S1, Obtain a physical examination data sample, wherein the physical examination data sample includes a personal information sample, a questionnaire information sample, and physical examination items after desensitization processing;

[0184] Historical data collected from hospitals / health check centers underwent triple desensitization processing:

[0185] (1) Removal of personal identifiers (name, ID number, etc.);

[0186] (2) Generalization of quasi-identifiers (age grouping, occupational category, region to province);

[0187] (3) Disturbance of sensitive information (adding noise to rare diseases and special occupations).

[0188] S2, extract user feature vector samples from the physical examination data samples, and construct a training dataset by combining the physical examination items as training labels;

[0189] The process for extracting user feature vector samples is the same as described above; please refer to the previous text for understanding, and it will not be repeated here. The dataset adopts a typical ratio: 70% training, 15% validation, and 15% testing.

[0190] S3, Extract training data from the training dataset and input the training data into the multilayer perceptron network structure to obtain the prediction result;

[0191] The model in this application is based on a multilayer perceptron, which consists of an input layer, hidden layers (typically 2-4 layers), and an output layer. This application uses a multilayer perceptron to learn a complex mapping from features to labels through nonlinear transformations.

[0192] S4, calculate the loss between the prediction result and the label based on the pre-constructed loss function, and adjust the internal parameters of the multilayer perceptron network structure based on the backpropagation of the loss combined with the gradient descent method;

[0193] The mathematical expression for the loss function is:

[0194]

[0195]

[0196]

[0197] In the formula, Indicates the total loss. Represents cross-entropy loss, The weights represent the cross-entropy loss. Indicates the weighted focus loss. This indicates the weighted focus loss weight. Represents the regularization coefficient. This represents the set of all trainable parameters in the model. Indicates the training batch size. Indicates the number of items included in the physical examination. For sample index, For tag indexing, Indicates the first The first sample The true probability of individual test items Indicates the first The first sample Predicted probability of individual test items Indicates the first The importance weight of individual test items Represents the positive sample balance coefficient. This indicates the focus parameter.

[0198] In the above loss function, This is a regularization term.

[0199] The fitting problem in this application is to map user feature vectors to confidence levels of various health check items, so a binary cross-term loss function is used as the base loss.

[0200] Considering that this application is applied to the field of predicting health checkup items, which is characterized by extreme class imbalance (positive rates of some items <1%) and asymmetric costs of misclassification (cost of missed diagnosis >> cost of misdiagnosis), a weighted focus loss is introduced. The weighted focus loss is calculated using importance weights. To reflect the differences in importance among different examinations, a positive sample balance coefficient is used. Ensure that rare but important checks are not overlooked, through Differentiated treatment of different risk groups reflects the characteristic of the medical field that "the cost of missed diagnosis is greater than the cost of misdiagnosis".

[0201] S5. Repeat steps S3-S4 until training is complete and the physical examination prediction model is obtained.

[0202] In this application, during the initial training phase Higher level, ensuring stable learning; in the later stages of training Appropriate increases can optimize hard samples. Finally, after the target number of training iterations is reached, or the loss between the predicted results and the actual labels is small and stable, training is complete, resulting in the physical examination prediction model. The physical examination prediction model is deployed on a server to provide prediction services.

[0203] S250, construct a dynamic screening threshold for the user, and filter out target physical examination items with a confidence level greater than or equal to the dynamic screening threshold; generate a physical examination plan based on the target physical examination items, and send the physical examination to the user.

[0204] The dynamic filtering values ​​in this application are obtained using the following method:

[0205] S2501, Obtain the validation set of the physical examination prediction model;

[0206] S2502, for each physical examination item, search in the validation set for a confidence threshold that makes the F1 score of the predicted result reach the highest level, and construct a dynamic screening threshold based on the confidence thresholds of multiple physical examination items.

[0207] In this application, the problem of determining the confidence threshold is transformed into an optimization problem of F1 score search, namely:

[0208]

[0209] In this application, the mathematical expression for the F1 score is:

[0210]

[0211] In the formula, Indicates the confidence threshold Next, the The overall evaluation score of individual test items is considered to be above 0.7, indicating good model performance. Precision rate (the proportion of samples predicted as positive that were actually positive).

[0212] ). Recall rate (the proportion of samples that are actually positive that are correctly predicted as positive).

[0213] The specific search algorithm can be the gradient search algorithm. The gradient search algorithm is existing technology and will not be elaborated on here.

[0214] After identifying the target health checkup items that meet the criteria, the final checkup items are determined by the user, influenced by factors such as budget and actual needs. This approach, while providing accurate recommendations, empowers the user with the decision-making authority. This application determines the final health checkup plan through a single round of interaction, specifically including:

[0215] S2511, sends the examination content and price of multiple target physical examination items to the user;

[0216] S2512, upon receiving selection information from the user, constructs a physical examination plan based on the final physical examination items selected by the user.

[0217] In this application, the user has the final say in selecting the medical examination items, allowing them to choose based on their actual budget and needs. This avoids disputes that might arise from automatic system selection.

[0218] The beneficial effects of this plan are reflected in the organic combination of its three stages:

[0219] Questionnaire interaction phase: Personalized questionnaires are generated through intelligent matching, greatly improving the efficiency and targeting of information collection. Users do not need to answer a large number of irrelevant questions, resulting in a smooth experience, while the system obtains high-quality structured health data.

[0220] In the health checkup prediction stage: deep models can integrate multi-dimensional features to quantitatively assess users' potential needs for various examinations. The prediction results are more accurate and comprehensive than those relying on human experience or fixed rules, laying the foundation for precise prevention.

[0221] In the final screening and interaction stage, dynamic thresholds ensure a medically sound balance between sensitivity and specificity in recommendations, effectively reconciling the conflict between "avoiding missed diagnoses" and "preventing over-testing." The final interaction design, which empowers the user, respects user preferences and budget constraints while enhancing trust and compliance through transparent recommendation reasoning.

[0222] Overall, the solution represents a shift from extensive physical examinations to personalized and precise health management, improving the efficiency of medical resource utilization while optimizing user experience and health benefits.

[0223] This application presents a deep learning-based health checkup triage method that integrates medical knowledge and artificial intelligence to achieve precise and personalized health checkup recommendations. The system intelligently generates questionnaires based on user profiles, efficiently collects key health information, and transforms it into structured features. The deep learning model quantifies various health checkup needs based on this information and then uses a dynamic threshold mechanism to personalize adjustments according to user risk characteristics and the importance of each test item, achieving accurate screening. The final generated health checkup plan balances safety and efficiency, significantly improves the early detection rate of serious diseases, optimizes the allocation of medical resources, and provides users with efficient and accurate health management services.

[0224] like Figure 4 As shown, this application also provides a deep learning-based physical examination triage system, including:

[0225] The acquisition module is used to acquire the user's basic personal information, which includes gender, age, occupation, and place of residence.

[0226] An interaction module is used to select target questions from a pre-built question bank based on the personal basic information and generate a target questionnaire, and send the target questionnaire to the user to obtain questionnaire information, wherein the questionnaire information includes lifestyle information, personal medical history, family medical history and description of recent symptoms;

[0227] The feature extraction module is used to extract target entities from the questionnaire information received from the user, and generate user feature vectors based on the personal basic information and the target entities, wherein the questionnaire information is generated based on the target questionnaire.

[0228] The prediction module is used to input the user feature vector into a pre-built physical examination prediction model to obtain the confidence levels of various physical examination items.

[0229] The filtering and solution construction module is used to construct the user's dynamic filtering threshold, and filter out target physical examination items with a confidence level greater than or equal to the dynamic filtering threshold; generate a physical examination plan based on the target physical examination items, and send the physical examination to the user.

[0230] This application presents a deep learning-based health checkup triage system that integrates medical knowledge with artificial intelligence to achieve precise and personalized health checkup recommendations. The system intelligently generates questionnaires based on user profiles, efficiently collects key health information, and transforms it into structured features. The deep learning model then quantifies various health checkup needs based on this information and uses a dynamic threshold mechanism to personalize adjustments according to user risk characteristics and the importance of each test item, achieving accurate screening. The final generated health checkup plan balances safety and efficiency, significantly improves the early detection rate of serious diseases, optimizes the allocation of medical resources, and provides users with efficient and accurate health management services.

[0231] This embodiment also provides an electronic terminal, including: a processor and a memory;

[0232] The memory is used to store computer programs, and the processor is used to execute the computer programs stored in the memory so that the terminal performs any of the methods in this embodiment.

[0233] As will be understood by those skilled in the art, the computer-readable storage medium described in this embodiment allows for the implementation of all or part of the steps in the above method embodiments by computer program-related hardware. The aforementioned computer program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0234] The electronic terminal provided in this embodiment includes a processor, a memory, a transceiver, and a communication interface. The memory and the communication interface are connected to the processor and the transceiver and complete communication between them. The memory is used to store computer programs, the communication interface is used to perform communication, and the processor and the transceiver are used to run the computer programs, so that the electronic terminal performs the steps of the above method.

[0235] In this embodiment, the memory may include random access memory (RAM) and may also include non-volatile memory, such as at least one disk storage device.

[0236] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0237] In the above embodiments, although the present application has been described in conjunction with specific embodiments thereof, many substitutions, modifications, and variations of these embodiments will be apparent to those skilled in the art based on the foregoing description. The embodiments of the present application are intended to cover all such substitutions, modifications, and variations falling within the broad scope of the appended claims.

[0238] The above embodiments are merely illustrative of the principles and effects of this application and are not intended to limit this application. Any person skilled in the art can modify or alter the above embodiments without departing from the spirit and scope of this application. Therefore, all equivalent modifications or alterations made by those skilled in the art without departing from the spirit and technical concept disclosed in this application should still be covered by the claims of this application.

Claims

1. A deep learning-based physical examination triage method, characterized in that, Including the following steps: Obtain the user's basic personal information, which includes gender, age, occupation, and place of residence; Based on the personal basic information, target questions are selected from a pre-built question bank and a target questionnaire is generated. The target questionnaire is then sent to the user to obtain questionnaire information, which includes lifestyle information, personal medical history, family medical history, and descriptions of recent symptoms. Upon receiving questionnaire information from a user, a target entity is extracted from the questionnaire information, and a user feature vector is generated based on the user's basic information and the target entity, wherein the questionnaire information is generated based on the target questionnaire. The user feature vector is input into a pre-built physical examination prediction model to obtain the confidence levels of various physical examination items; A dynamic screening threshold is constructed for the user, and target physical examination items with a confidence level greater than or equal to the dynamic screening threshold are selected; a physical examination plan is generated based on the target physical examination items, and the physical examination is sent to the user.

2. The deep learning-based physical examination triage method according to claim 1, characterized in that, The method for constructing the question database includes: Obtain medical knowledge materials; Knowledge extraction was performed on the aforementioned medical knowledge materials to obtain multiple physical examination suggestions; Entity extraction is performed on the multiple physical examination suggestions to obtain physical examination categories and multiple identity information limiting entities, wherein the identity information limiting entities are used to limit the value range of personal basic information; A knowledge template is constructed based on the physical examination category and multiple identity information-defined entities; Based on the physical examination category in each knowledge template, expand the question text, and construct a structured question template based on the physical examination category, the question text, and multiple identity information to limit the entity, wherein the question text is a guiding question text related to the physical examination category; Each structured question template is assigned an importance score and type label. A question library is built based on multiple structured question templates and the importance scores and label types of the question text within multiple structured question templates. Structured question templates with the same physical examination category are merged.

3. The deep learning-based physical examination triage method according to claim 2, characterized in that, Each structured question template's question text is scored for importance, including: Obtain the basic importance score for each medical examination category. And score the basic importance of the physical examination categories. As the basis for importance scoring for all corresponding question texts The basic importance score Calibrated by an expert system, For physical examination category index, Index the question text; Calculate a specific score for each medical examination category based on multiple identity-based entities for each category. And the specificity score of the physical examination category. As a specific score corresponding to all question texts The mathematical expression for the specificity score is: In the formula, An index for identity information entities. For the first The theoretical range of values ​​for an identity information entity. For the first The corresponding item of the individual examination The scope of each identity information entity is limited; Calculate the timeliness score of the question text corresponding to each physical examination category. The timeliness score The mathematical expression is: In the formula, For the theoretical maximum score, For the first The time-depletion coefficient of individual test items, Indicates the first The difference between the publication time of the issue text for an individual medical examination item and the current time; Based on the aforementioned basic importance score The specificity score and the aforementioned timeliness score Calculate importance score The importance score The mathematical expression is: In the formula, As the first weight, As the second weight, It is the third weight.

4. The deep learning-based physical examination triage method according to claim 2, characterized in that, Based on the aforementioned basic personal information, target questions are selected from a pre-built question bank to generate a target survey questionnaire, including: The personal basic information is compared with the identity information-limited entity of multiple structured question templates in a pre-built question library, and the question text in the structured question template that matches the personal basic information is used as the target question text; Multiple target question texts are divided based on type labeling to obtain multiple sets of target question texts of different types; Extract the target questions with the highest importance scores from multiple types of target question text sets, so that the target questions of multiple types are distributed according to a preset ratio; A target questionnaire is generated based on multiple target questions, wherein the target questions in the target questionnaire are ranked according to their importance score.

5. The deep learning-based physical examination triage method according to claim 1, characterized in that, Extracting target entities from the questionnaire information and generating user feature vectors based on the personal basic information and the target entities, including: Based on the medical named entity recognition model, logical decision entities or numerical entities of named entities and command entities are extracted from the questionnaire information, wherein the logical decision entities include selection entities, affirmative entities and negative entities; The logical decision entity or numerical entity of the named entity or command entity is concatenated to obtain the target entity; and the target entity and the personal basic information are vectorized and concatenated to obtain the user feature vector.

6. The deep learning-based physical examination triage method according to claim 1, characterized in that, The physical examination prediction model adopts a multilayer perceptron network structure, wherein the construction method of the physical examination prediction model includes: S1, Obtain a physical examination data sample, wherein the physical examination data sample includes a personal information sample, a questionnaire information sample, and physical examination items after desensitization processing; S2, extract user feature vector samples from the physical examination data samples, and construct a training dataset by combining the physical examination items as training labels; S3, Extract training data from the training dataset and input the training data into the multilayer perceptron network structure to obtain the prediction result; S4, calculate the loss between the prediction result and the label based on the pre-constructed loss function, and adjust the internal parameters of the multilayer perceptron network structure based on the backpropagation of the loss combined with the gradient descent method; S5. Repeat steps S3-S4 until training is complete and the physical examination prediction model is obtained.

7. The deep learning-based physical examination triage method according to claim 6, characterized in that, The mathematical expression for the loss function is: In the formula, Indicates the total loss. Represents cross-entropy loss, The weights represent the cross-entropy loss. Indicates the weighted focus loss. This indicates the weighted focus loss weight. Represents the regularization coefficient. This represents the set of all trainable parameters in the model. Indicates the training batch size. Indicates the number of items included in the physical examination. For sample index, For tag indexing, Indicates the first The first sample The true probability of individual test items Indicates the first The first sample Predicted probability of individual test items Indicates the first The importance weight of individual test items Represents the positive sample balance coefficient. This indicates the focus parameter.

8. The deep learning-based physical examination triage method according to claim 1, characterized in that, Constructing the dynamic filtering threshold for the user includes: Obtain the validation set of the physical examination prediction model; For each physical examination item, a confidence threshold that maximizes the F1 score of the predicted result is searched in the validation set, and a dynamic screening threshold is constructed based on the confidence thresholds of multiple physical examination items.

9. A deep learning-based physical examination triage method according to claim 1, characterized in that, A medical examination plan is generated based on the target medical examination items, including: The examination details and prices for multiple target health check-up items are sent to the user. Upon receiving selection information from the user, a medical examination plan is constructed based on the final medical examination items selected by the user.

10. A deep learning-based physical examination triage system, characterized in that, include: The acquisition module is used to acquire the user's basic personal information, which includes gender, age, occupation, and place of residence. An interaction module is used to select target questions from a pre-built question bank based on the personal basic information and generate a target questionnaire, and send the target questionnaire to the user to obtain questionnaire information, wherein the questionnaire information includes lifestyle information, personal medical history, family medical history and description of recent symptoms; The feature extraction module is used to extract target entities from the questionnaire information received from the user, and generate user feature vectors based on the personal basic information and the target entities, wherein the questionnaire information is generated based on the target questionnaire. The prediction module is used to input the user feature vector into a pre-built physical examination prediction model to obtain the confidence levels of various physical examination items. The filtering and solution construction module is used to construct the user's dynamic filtering threshold, and filter out target physical examination items with a confidence level greater than or equal to the dynamic filtering threshold; generate a physical examination plan based on the target physical examination items, and send the physical examination to the user.

Images