Infectious disease screening method based on preference data and direct preference optimization algorithm

Abstract

The invention provides an infectious disease screening method based on preference data and a direct preference optimization algorithm, and the method comprises the steps: obtaining multi-dimensional health text information inputted by a patient, extracting atomic symptom description in the information, and generating a corresponding medical question; determining a consistency reference score of the atomic symptom description and a preset standard description and a non-reference score determined based on medical question sampling, and forming a comprehensive authenticity score through weighted fusion; further combining clinical preset symptom importance indexes to determine authenticity reward values of symptom descriptions, calculating an optimization probability between any two symptom descriptions based on a Bradley-Terry model, and constructing a factual preference data set; and performing model fine tuning on the preference data set through a direct preference optimization algorithm, and outputting an infectious disease symptom screening result in real time. According to the scheme, the screening accuracy of infectious diseases can be improved, the missed diagnosis rate and the misdiagnosis rate are remarkably reduced, and the screening efficiency is improved.

Description

Technical Field

[0001] The present disclosure relates to the technical field of infectious disease screening. Specifically, it relates to an infectious disease screening method based on preference data and direct preference optimization algorithm. Background Art

[0002] In recent years, infectious diseases (such as the COVID-19 pandemic, influenza, etc.) have been highly prevalent and have a fast transmission speed. Traditional infectious disease screening methods mainly rely on doctors' manual inquiries, which are often inefficient when faced with a large number of patients. Moreover, due to limited medical resources, many high-risk patients may not be screened in a timely manner, while low-risk patients occupy valuable diagnosis and treatment time and resources, leading to a more unbalanced allocation of resources. At the same time, the characteristics of infectious diseases are constantly evolving with epidemiological laws. For example, new mutant strains may present new symptom combinations. Existing screening methods lack a mechanism for real-time capture of emerging disease characteristics and model updates, making it difficult to adjust screening strategies in a timely manner, resulting in an ineffective response to rapidly changing disease risks.

[0003] Currently, machine learning methods have been applied in some medical scenarios, such as patient risk scoring models based on logistic regression or gradient boosting decision trees (GBDT), symptom classification methods based on algorithms such as naive Bayes, etc. However, these methods often have the following deficiencies: limited ability to process multi-dimensional and complex symptom descriptions, especially when faced with free text input, vague descriptions, or inconsistent information before and after from patients, it is difficult for the model to balance efficiency and accuracy; insufficient dynamic adaptation ability to sudden or unknown symptoms, and many models are unable to incorporate new symptoms and update their weights in a timely manner; in frequent iterations or incremental training, if there is no effective constraint on the deviation of the model distribution, the stability of the screening results may decline. Summary of the Invention

[0004] The embodiments of the present disclosure at least provide an infectious disease screening method based on preference data and direct preference optimization algorithm, which can improve the screening accuracy of infectious diseases, significantly reduce the missed diagnosis and misdiagnosis rates, and improve the screening efficiency.

[0005] The embodiments of the present disclosure provide an infectious disease screening method based on preference data and direct preference optimization algorithm, including:

[0006] Obtain multi-dimensional health text information input by a patient, extract atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions;

[0007] Determine a consistency reference score between the atomic symptom description and a preset standard description, and determine a non-reference score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings based on the medical question;

[0008] Fuse the consistency reference score and the reference-free score to obtain a comprehensive authenticity score, and calculate the authenticity reward value corresponding to the atomic symptom description in combination with a preset symptom importance index; determine the preference probability between any two of the atomic symptom descriptions according to the authenticity reward value, and construct a factual preference dataset;

[0009] Utilize the factual preference dataset to fine-tune the pre-trained screening model through a direct preference optimization algorithm, and output the screening results of infectious disease symptoms in real time.

[0010] In an optional implementation manner, extract the atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions, specifically including:

[0011] Perform natural language processing on the multi-dimensional health text information, and split the multi-dimensional health text information into multiple sub-word phrases;

[0012] According to a preset keyword dictionary, identify the atomic symptom descriptions in the sub-word phrase text that match the keyword dictionary;

[0013] For each atomic symptom description, input the atomic symptom description into a pre-trained medical domain language generation model to generate the medical question containing the atomic symptom description.

[0014] In an optional implementation manner, determine the consistency reference score between the atomic symptom description and a preset standard description, specifically including:

[0015] Split the atomic symptom description into multiple independent inspection items, and each independent inspection item corresponds to a specific symptom information in the atomic symptom description;

[0016] Determine the consistency reference score according to the number of the independent inspection items and an indication function indicating whether the independent inspection items are consistent.

[0017] In an optional implementation manner, determine the reference-free score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings of the medical question, specifically including:

[0018] Perform multiple generation samplings on the medical question, and record the question answers of each sampling;

[0019] Determine the target number of occurrences corresponding to the question answer with the most occurrences;

[0020] Determine the reference-free score according to the target number of occurrences and the number of samplings corresponding to the generation sampling.

[0021] In an alternative embodiment, a comprehensive authenticity score is obtained by fusing the consistency reference score and the reference-free score, and a authenticity reward value corresponding to the atomic symptom description is calculated in combination with a preset symptom importance index, which specifically includes:

[0022] The consistency reference score and the reference-free score are weighted and averaged according to a preset first weighting parameter to determine the comprehensive authenticity score corresponding to the atomic symptom description;

[0023] According to a preset second weighting parameter, the comprehensive authenticity score and the symptom importance index are weighted and averaged to determine the authenticity reward value.

[0024] In an alternative embodiment, the preference probability between any two of the atomic symptom descriptions is determined according to the authenticity reward value, which specifically includes:

[0025] For each candidate description pair composed of two atomic symptom descriptions, the reward difference between the authenticity reward value corresponding to the first atomic symptom description and the authenticity reward value corresponding to the second atomic symptom description in the candidate description pair is determined;

[0026] According to a preset adjustment factor and the reward difference, the preference probability indicating that the first atomic symptom description is superior to the second atomic symptom description is determined, where the adjustment factor is used to amplify or reduce the influence of the reward difference on the calculation of the preference probability.

[0027] In an alternative embodiment, the screening model is trained based on the following steps:

[0028] A sample preference data set constructed from a plurality of sample atomic symptom descriptions is obtained, and based on the ranking of the sample atoms symptoms according to the corresponding preference probability in the sample preference data set, preferred candidate symptom descriptions and sub-optimal candidate symptom descriptions are selected;

[0029] A contrast objective function is constructed such that the generation probability of the preferred candidate symptom description is higher than that of the sub-optimal candidate symptom description;

[0030] According to the contrast objective function, and using a distance function that measures the difference between the model output distribution and the reference benchmark distribution with the KL divergence regularization term as the regularization term, the total loss function is determined;

[0031] The screening model is trained according to the total loss function until the screening model has the ability to screen out infectious disease symptoms based on the atomic symptom description.

[0032] The embodiments of the present disclosure also provide an infectious disease screening device based on preference data and the direct preference optimization algorithm. The device includes:

[0033] An information acquisition module, configured to acquire multi-dimensional health text information input by a patient, extract atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions;

[0034] An authenticity scoring module, configured to determine a consistency reference score between the atomic symptom description and a preset standard description, and determine a non-reference score corresponding to the atomic symptom description according to the occurrence times of each question answer after multiple generation samplings based on the medical question;

[0035] A preference data analysis module, configured to fuse the consistency reference score and the non-reference score to obtain a comprehensive authenticity score, and calculate a authenticity reward value corresponding to the atomic symptom description in combination with a preset symptom importance index; determine a preference probability between any two of the atomic symptom descriptions according to the authenticity reward value, and construct a factual preference data set;

[0036] A symptom screening module, configured to utilize the factual preference data set to fine-tune a pre-trained screening model through a direct preference optimization algorithm, and output an infectious disease symptom screening result in real time.

[0037] The embodiment of the present disclosure further provides an electronic device, including: a processor, a memory, and a bus, where the memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are executed by the processor, the above-mentioned infectious disease screening method, or the steps in any possible implementation manner of the above-mentioned infectious disease screening method are executed.

[0038] The embodiment of the present disclosure further provides a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, the above-mentioned infectious disease screening method based on preference data and direct preference optimization algorithm, or the steps in any possible implementation manner of the infectious disease screening method based on preference data and direct preference optimization algorithm are executed.

[0039] The embodiment of the present disclosure further provides a computer program product, including a computer program / instructions. When the computer program and instructions are executed by a processor, the above-mentioned infectious disease screening method based on preference data and direct preference optimization algorithm, or the steps in any possible implementation manner of the above-mentioned infectious disease screening method based on preference data and direct preference optimization algorithm are implemented.

[0040] An infectious disease screening method based on preference data and direct preference optimization algorithm provided by an embodiment of the present disclosure obtains multi-dimensional health text information input by a patient, extracts atomic symptom descriptions from the multi-dimensional health text information, and generates medical questions corresponding to the atomic symptom descriptions; determines a consistency reference score between the atomic symptom descriptions and preset standard descriptions, and determines a non-reference score corresponding to the atomic symptom descriptions according to the occurrence times of each question answer after multiple generation samplings based on the medical questions; fuses the consistency reference score and the non-reference score to obtain a comprehensive authenticity score, and calculates a authenticity reward value corresponding to the atomic symptom descriptions in combination with a preset symptom importance index; determines a preference probability between any two of the atomic symptom descriptions according to the authenticity reward value, and constructs a factual preference data set; uses the factual preference data set to fine-tune a pre-trained screening model through the direct preference optimization algorithm, and outputs an infectious disease symptom screening result in real time. It can improve the screening accuracy of infectious diseases, significantly reduce the missed diagnosis and misdiagnosis rates, and improve the screening efficiency.

[0041] To make the above objects, features, and advantages of the present disclosure more obvious and understandable, the following specifically enumerates preferred embodiments and, in conjunction with the accompanying drawings, provides a detailed description as follows. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] To more clearly illustrate the technical solutions of the embodiments of the present disclosure, the following briefly introduces the accompanying drawings required for use in the embodiments. The accompanying drawings herein are incorporated into the specification and constitute a part of the specification. These drawings show embodiments that conform to the present disclosure and, together with the specification, are used to illustrate the technical solutions of the present disclosure. It should be understood that the following drawings only show certain embodiments of the present disclosure and should not be regarded as limiting the scope. For those of ordinary skill in the art, other related drawings can be obtained based on these drawings without creative efforts.

[0043] Figure 1 Shows a flowchart of an infectious disease screening method based on preference data and direct preference optimization algorithm provided by an embodiment of the present disclosure;

[0044] Figure 2 Shows a flowchart of a training method for an infectious disease screening model provided by an embodiment of the present disclosure;

[0045] Figure 3 Shows a schematic diagram of an infectious disease screening device based on preference data and direct preference optimization algorithm provided by an embodiment of the present disclosure;

[0046] Figure 4 Shows a schematic diagram of an electronic device provided by an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0047] To make the objectives, technical solutions, and advantages of the embodiments of the present disclosure clearer, the technical solutions in the embodiments of the present disclosure will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present disclosure. Apparently, the described embodiments are only some, rather than all, of the embodiments of the present disclosure. Components of the embodiments of the present disclosure generally described and illustrated in the figures herein may be arranged and designed in a variety of different configurations. Therefore, the detailed description of the embodiments of the present disclosure provided herein is not intended to limit the scope of the claimed present disclosure, but merely represents selected embodiments of the present disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of the present disclosure without creative efforts fall within the scope of the present disclosure.

[0048] It should be noted that like reference numerals and letters denote like items in the following figures, and thus, once an item is defined in one figure, it need not be further defined and explained in subsequent figures.

[0049] The term "and / or" in this document merely describes an association relationship and indicates that three relationships may exist. For example, A and / or B may represent three cases: A exists alone, both A and B exist simultaneously, and B exists alone. Additionally, the term "at least one" in this document means any one of a plurality or any combination of at least two of a plurality. For example, including at least one of A, B, and C may mean including any one or more elements selected from the set composed of A, B, and C.

[0050] Through research, it has been found that currently, machine learning methods have been applied in some medical scenarios, such as patient risk scoring models based on logistic regression or gradient boosting decision trees (GBDT), symptom classification methods based on algorithms such as naive Bayes, etc. However, these methods often have the following deficiencies: limited processing ability for multi-dimensional and complex symptom descriptions, especially when faced with free text input, vague descriptions, or inconsistent information before and after by patients, it is difficult for the model to balance efficiency and accuracy; insufficient dynamic adaptation ability to sudden or unknown symptoms, and many models cannot timely incorporate new symptoms and update their weights; in frequent iterations or incremental training, if there is no effective constraint on the deviation of the model distribution, the screening results may show a decline in stability.

[0051] Based on the above research, the present disclosure provides an infectious disease screening method, apparatus, electronic device and storage medium based on automatically constructing factual preference data and the Direct Preference Optimization (DPO) algorithm. It obtains multi-dimensional health text information input by a patient, extracts atomic symptom descriptions from the multi-dimensional health text information, and generates medical questions corresponding to the atomic symptom descriptions; determines a consistency reference score between the atomic symptom descriptions and preset standard descriptions and a non-reference score determined by sampling based on the medical questions, and forms a comprehensive authenticity score through weighted fusion; further combines clinically preset symptom importance indicators to determine the authenticity reward value of the symptom descriptions, and calculates the preference probability between any two symptom descriptions based on the Bradley-Terry model to construct a factual preference data set; fine-tunes the model with the preference data set through the direct preference optimization algorithm, where a KL divergence regularization term is introduced to ensure the stability of the model, and the infectious disease symptom screening result is output in real time. This solution can improve the screening accuracy of infectious diseases, significantly reduce the missed diagnosis and misdiagnosis rates, and improve the screening efficiency.

[0052] To facilitate the understanding of this embodiment, first, a detailed introduction is given to an infectious disease screening method based on preference data and the direct preference optimization algorithm disclosed in the embodiments of the present disclosure. The execution subject of the infectious disease screening method provided in the embodiments of the present disclosure is generally a computer device with certain computing capabilities. Such computer devices include, for example: terminal devices or servers or other processing devices. The terminal device may be a user equipment (UE), a mobile device, a user terminal, a terminal, a cellular phone, a cordless phone, a personal digital assistant (PDA), a handheld device, a computing device, a vehicle-mounted device, a wearable device, etc. In some possible implementation manners, the infectious disease screening method may be implemented by a processor calling computer-readable instructions stored in a memory.

[0053] See Figure 1 As shown, it is a flowchart of an infectious disease screening method based on preference data and the direct preference optimization algorithm provided in the embodiments of the present disclosure. The method includes steps S101 to S104, where:

[0054] S101. Obtain multi-dimensional health text information input by a patient, extract atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions.

[0055] In a specific implementation, multi-dimensional health text information input by patients is collected in real-time and comprehensively. The multi-dimensional health text information includes at least symptoms, physical signs, travel history, contact history, etc. To prevent data inconsistencies caused by patients' misunderstanding or input errors, after the patient submits a symptom description, subsequent questions will be automatically generated for confirmation, such as: "Did you mention a persistent high fever in your input?", "Has the high fever lasted for more than 48 hours?", etc.

[0056] Here, the multi-dimensional health text information can be collected through a self-help consultation application installed on a mobile device (such as a smartphone, tablet, etc.). The self-help consultation application can include a patient-side application and a doctor-terminal application. The patient inputs a description of the condition through the self-help consultation application. The application is designed with multiple functional modules, including but not limited to self-reporting of symptoms, recording of basic physical signs, collection of contact history and travel history information, etc. The patient can complete the initial consultation by selecting preset symptom items, filling in free-text descriptions, answering customized questions, etc.

[0057] Among them, for the patient-side application, the patient can input a free-text symptom description, upload physiological data such as body temperature, heart rate, blood oxygen, etc., and fill in information such as travel history, contact history, and past medical history through the self-help consultation application on the smartphone or tablet.

[0058] Among them, for the doctor-terminal application, the doctor-side system can display the symptom data, physical sign records, and historical medical history input by the patient. The doctor verifies and supplements the symptoms based on the information input by the patient to ensure the accuracy of all symptom descriptions. If the information provided by the patient is insufficient or ambiguous, the doctor can further confirm or improve the symptom data through online communication with the patient. For the physical sign data (such as body temperature, blood pressure, etc.) provided by the patient, the doctor can check to ensure the validity and reliability of the data. If the patient does not provide a certain piece of data, the doctor can ask the patient to measure or record it. The doctor can directly view information such as the patient's past medical history, medication history, and allergy history through the system to further supplement the patient's disease background.

[0059] As a possible implementation method, the self-help consultation application can provide multiple preset symptom options (such as fever, cough, headache, shortness of breath, etc.). The patient completes symptom input in text / voice form. For atypical symptoms, the condition can also be described in free text to ensure more complete symptom information is collected.

[0060] In a specific implementation, the self-help consultation application can first prompt the user to enter basic information, and then guide the patient to enter physical sign data such as body temperature, heart rate, and blood pressure (if the patient can self-measure or already has relevant data). If the relevant data cannot be self-measured, the sensor (such as a thermometer, a heart rate monitor) can be automatically called to collect the physical sign data. After that, data formatting and standardization processing are automatically performed. The patient can also be allowed to record through a selection box whether there has been recent contact with an infectious disease patient and whether they have ever been to an area with a high incidence of the epidemic.

[0061] As another possible implementation method, the data collection interface of the self-help consultation application is designed with a preset format template to ensure that the input data meets the standards. For example, beside the body temperature input box, it is marked "Unit: °C", prompting the user to enter reasonable data within the range of 30°C to 45°C. The background uses a data filtering algorithm to perform real-time verification on abnormal data (such as body temperature input below 30°C or above 45°C) and prompts the user to review or re-enter.

[0062] Optionally, for the multi-dimensional health text information, to facilitate subsequent processing, this process includes word segmentation and stop word removal, removing irrelevant auxiliary words (such as "de", "le", "ne"), and decomposing the text content; correcting spelling mistakes and unifying different expressions (such as "stomachache" → "abdominal pain"); semantic parsing to analyze the sentence structure and identify the core expressions related to health.

[0063] Furthermore, for the multi-dimensional health text information, natural language processing is performed to split the multi-dimensional health text information into multiple sub-word phrases; according to the preset keyword dictionary, the atomic symptom descriptions that match the keyword dictionary in the sub-word phrase text are identified; for each atomic symptom description, the atomic symptom description is input into a pre-trained language generation model in the medical field to generate a medical question containing the atomic symptom description.

[0064] In a specific implementation, natural language processing technology can be used to perform word segmentation on the long text symptom description input by the patient. For example, for the text "I have had a high fever for 3 consecutive days, accompanied by severe dry cough and obvious shortness of breath", the system decomposes it into multiple phrases or expressions such as "3 consecutive days", "high fever", "severe dry cough", and "obvious shortness of breath". Combining with the pre-constructed professional dictionary of the infectious disease department, the key symptom entities involved in the text, such as "high fever", "dry cough", and "shortness of breath", are identified and part-of-speech tagged for subsequent analysis. In addition, the descriptions of time, quantity, etc. that appear are identified. For example, "3 days" will be extracted and recorded as a parameter of the symptom duration.

[0065] Here, the atomic symptom description refers to the smallest unit that can independently represent a certain symptom feature. For example: "cough", "headache", "shortness of breath", "body temperature 38°C", etc. The process of extracting atomic symptom descriptions can perform named entity recognition through natural language processing (NLP) based on a medical dictionary or a deep learning model to identify medical-related words in the text, such as symptoms, signs, disease names, etc. Use a medical terminology library (such as SNOMED CT, ICD-10) to match keywords in the patient's text, extract symptom descriptions, and then parse the sentence structure to determine which descriptive words belong to the same symptom unit.

[0066] Exemplarily, for the multi-dimensional health text information "I've had a sore throat recently and a bit of a fever." input by the patient, the corresponding parsed atomic symptom 1: "sore throat" (pharyngeal pain); atomic symptom 2: "fever" (elevated body temperature).

[0067] After that, redundant information is removed. For example, in the sentence "I think it might be a cough caused by the flu.", "flu" is not an atomic symptom, and only "cough" is extracted. In the sentence "I felt a bit dizzy last night and much better today.", "much better today" does not belong to the symptom description and needs to be removed, and only "dizzy" is retained. Use a medical vocabulary standard library (such as UMLS, MeSH) to normalize different expressions, such as: "sore throat" → "pharyngalgia", "difficult to catch one's breath" → "shortness of breath", which can improve the accuracy of subsequent matching and analysis.

[0068] Furthermore, once the atomic symptom descriptions are extracted, corresponding medical questions can be automatically generated for each symptom to further collect more detailed information. Specifically, a pre-trained language generation model in the medical field can be used to automatically convert each atomic symptom description into a targeted medical question. First, the atomic symptom description is fed into the language generation model as input, and the model automatically generates a corresponding medical question based on medical domain knowledge. For example, convert "high fever" to "What are the possible causes of a patient having a continuous high fever?"; convert "dry cough" to "What clinical symptoms might severe dry cough imply?" etc. Second, this process is automatically carried out for each atomic symptom description one by one to generate medical questions that facilitate subsequent model multiple sampling and evaluation of answer consistency.

[0069] Optionally, the generation method of medical questions can include rule-based template generation and pre-trained medical language models (such as ChatGPT, T5). For the rule-based template generation method, a set of templates is predefined and symptom-related variables are filled in. For example, "How long has your {symptom} lasted?" and "Under what circumstances does your {symptom} worsen?". In this way, when "headache" is input, it will automatically generate: "How long has your headache lasted?" and "Under what circumstances does your headache worsen?". For the generation method based on pre-trained medical language models, a medical NLP model is trained to be able to automatically generate reasonable questions according to the input symptom description. For example: Input: "chest pain", the model generates: "Is your chest pain stabbing or a feeling of compression?" and "Does the chest pain spread to the arm or back?".

[0070] In this way, through the above processing, the long text description of the multi-dimensional health text information input by the patient is converted into a set of structured data, where each data unit is an atomic symptom description. For example, the text can be converted to: s1: continuous high fever (3 days), s2: severe dry cough, s3: obvious shortness of breath.

[0071] S102. Determine the consistency reference score between the atomic symptom description and the preset standard description, and determine the non-reference score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings based on the medical questions.

[0072] In a specific implementation, the atomic symptom description together with corresponding parameters (such as duration, severity) is used as the input for authenticity evaluation and preference data construction. First, without manual annotation, multi-dimensional authenticity evaluation of the atomic symptom description is performed, and a factual preference data set is constructed. The consistency reference score is used to measure the matching degree between the atomic symptom description and the standardized medical description, that is, the consistency degree of the text. The non-reference score is used to calculate the stability of different answers based on the generation sampling of medical questions, so as to evaluate the credibility of the atomic symptom description.

[0073] Here, the atomic symptom description is split into multiple independent inspection items, and each independent inspection item corresponds to a specific symptom information in the atomic symptom description; according to the number of independent inspection items and the indicator function indicating whether the independent inspection items are consistent, the consistency reference score is determined.

[0074] Specifically, first, by pre-accessing authoritative medical databases or clinical guideline data, such as the clinical materials released by the National Center for Disease Control, as the preset standard description, the atomic symptom description is further split into multiple independent examination items, and each examination item corresponds to a specific piece of information in the description (for example, "high fever" as one item, "lasting for 3 days" as another item). Through the pre-trained language judgment module, each examination item is compared with the authoritative data to determine whether the information of this item is consistent with the authoritative data.

[0075] Here, the consistency reference score between the atomic symptom description and the preset standard description can be calculated by the following formula:

[0076]

[0077] Where r i represents the consistency reference score, that is, the authenticity score obtained by evaluating through external authoritative data; N i represents the number of independent examination items; c ij represents an independent examination item. χ(c ij ) is an indicator function, which judges the jth examination item c ij . When this examination item is consistent with the preset standard description, it takes the value of 1, otherwise it takes the value of 0.

[0078] Furthermore, through the pre-trained model, the medical question is sampled and generated multiple times (for example, 20 times), and the answers to the questions for each sampling are recorded. Using a simple matching algorithm or heuristic rules, the generated answers are classified and counted to determine which answer appears most frequently.

[0079] Specifically, for the medical question, it is sampled and generated multiple times, and the answers to the questions for each sampling are recorded; the target number of occurrences corresponding to the question answer that appears most frequently is determined; according to the target number of occurrences and the number of samplings corresponding to the generated sampling, the non-reference score is determined.

[0080] Here, the non-reference score corresponding to the atomic symptom description can be calculated by the following formula:

[0081]

[0082] Where c i represents the non-reference score of the atomic symptom description (that is, the internal confidence score), and this score reflects the consistency of the model in generating answers to the questions corresponding to this description; n max represents the number of times of the question answer that appears most frequently after multiple samplings of the medical question; n represents the total number of samplings.

[0083] S103. Integrate the consistency reference score and the reference-free score to obtain a comprehensive authenticity score, and calculate the authenticity reward value corresponding to the atomic symptom description in combination with a preset symptom importance index; determine the preference probability between any two atomic symptom descriptions according to the authenticity reward value, and construct a factual preference dataset.

[0084] In specific implementation, the authenticity reward value is a comprehensive index for measuring the credibility of the atomic symptom description, which is calculated based on the consistency reference score, the reference-free score, and the symptom importance index. The consistency reference score measures the matching degree between the symptom description input by the user and the standard medical description. The reference-free score evaluates the reliability of the symptom description based on the sampling stability of the medical question.

[0085] Specifically, perform a weighted average of the consistency reference score and the reference-free score according to a preset first weighting parameter to determine the comprehensive authenticity score corresponding to the atomic symptom description; according to a preset second weighting parameter, perform a weighted average of the comprehensive authenticity score and the symptom importance index to determine the authenticity reward value.

[0086] Here, the comprehensive authenticity score can be calculated by the following formula:

[0087] S i =α·r i +(1-α)·c i , 0≤α≤1

[0088] Wherein, S i represents the comprehensive authenticity score of the atomic symptom description; α represents the first weighting parameter; r i represents the consistency reference score; c i represents the reference-free score; the first weighting parameter is used to balance the influence of external reference data and internal confidence assessment, and its value ranges from 0 to 1: for example, when α is larger, the influence of external data is greater.

[0089] Here, fuse the comprehensive authenticity score with a preset symptom importance index formulated in advance by infectious disease experts, and the authenticity reward value can be calculated by the following formula:

[0090] R(s i )=γ·I(s i )+(1-γ)·S i

[0091] Wherein, R(s i ) represents the authenticity reward value corresponding to the atomic symptom description; I(s i) represents a preset symptom importance index, which can be preset by infectious disease experts and reflects the relative importance of symptoms in clinical diagnosis (for example, "continuous high fever" may take a value of 0.9); γ represents the second weighting parameter, which is used to adjust the weights of the symptom importance index and the comprehensive authenticity score in the final reward value, and its value ranges from 0 to 1.

[0092] Furthermore, for each candidate description pair composed of two atomic symptom descriptions, determine the reward difference between the authenticity reward value corresponding to the first atomic symptom description and the authenticity reward value corresponding to the second atomic symptom description in the candidate description pair; according to the preset adjustment factor and the reward difference, determine the preference probability for indicating that the first atomic symptom description is superior to the second atomic symptom description, where the adjustment factor is used to amplify or reduce the influence of the reward difference on the preference probability calculation.

[0093] In a specific implementation, based on the Bradley-Terry model, pairwise comparisons are made between different candidate symptom descriptions under the input of the same patient. For any two candidate descriptions, calculate the difference in their authenticity reward values, and use the following formula to calculate the preference probability:

[0094]

[0095] where P(s i superior to s j ) represents the preference probability that the candidate description s i is considered superior to s j in the pairwise comparison between the first atomic symptom description s i and the second atomic symptom description s j ; R(s i ) and R(s j ) respectively represent the authenticity reward values of the first atomic symptom description s i and the second atomic symptom description s j ; ω represents the adjustment factor, which is used to amplify or reduce the influence of the reward value difference on the probability calculation, and its value is determined by specific application experience.

[0096] After that, screen the atomic symptom descriptions corresponding to the preference probability greater than the preset threshold to construct a preference data set. In the previous stage, the preference probability between any two atomic symptom descriptions has been calculated to measure the reliability of different symptom descriptions. In this step, reliable data is screened out through the set threshold as the preference data set. The preference data set can reduce false symptoms, improve the accuracy of the screening model, and help the model more accurately distinguish between infectious disease-related symptoms and irrelevant symptoms.

[0097] S104. Screen the atomic symptom descriptions corresponding to the preferred probability greater than the preset threshold to construct a preference dataset, and output the infectious disease symptom screening result after processing the preference dataset through a pre-trained screening model.

[0098] In a specific implementation, the screening model uses the screened preference data as input. The screening model can be a classification model based on deep learning or machine learning, such as: Transformer or BERT model (suitable for symptom analysis of text input); Random Forest, XGBoost (suitable for structured data screening); Multimodal fusion model (combining various data such as text and images).

[0099] Among them, the screening model can output a screening evaluation report. After medical staff view the generated screening evaluation report, they can provide real-time feedback on symptom ranking, risk assessment, and overall screening results through a dedicated interface. The feedback includes scoring, comments, and suggestions, etc.

[0100] It should be noted that all data are protected by AES-256 encryption technology and TLS encryption protocol during the processes of collection, transmission, and storage; the data is anonymized, and multi-level access control is adopted to ensure that only authorized personnel can access sensitive data to meet the requirements of relevant regulations (such as GDPR, PIPL).

[0101] As a possible implementation, refer to Figure 2 As shown, it is a flowchart of a training method for an infectious disease screening model provided by an embodiment of the present disclosure. The method includes steps S201 to S204, where:

[0102] S201. Obtain a sample preference dataset constructed from multiple sample atomic symptom descriptions, and select preferred candidate symptom descriptions and sub-optimal candidate symptom descriptions according to the ranking of the sample atoms in the sample preference dataset corresponding to the preferred probability.

[0103] S202. Construct a contrast objective function that makes the generation probability of the preferred candidate symptom description higher than that of the sub-optimal candidate symptom description.

[0104] S203. According to the contrast objective function, and using a distance function for measuring the difference between the model output distribution and the reference benchmark distribution as a regularization term, determine the total loss function.

[0105] S204. Train the screening model according to the total loss function until the screening model has the ability to screen out infectious disease symptoms based on the atomic symptom description.

[0106] In a specific implementation, the screening model constructs factual preference data pairs by collecting sample atomic symptom descriptions for model training and optimization, so that when generating screening results, candidate outputs with high authenticity and high clinical importance are preferentially selected.

[0107] Here, the sample preference data set is used as the input data corresponding to the screening model. The model outputs candidate screening results, and the preferred candidate symptom descriptions and sub-optimal candidate symptom descriptions are selected from the candidate outputs according to the previously constructed sample preference data set.

[0108] Specifically, the optimization goal of the screening model is to make the generation probability of the preferred candidate significantly higher than that of the sub-optimal candidate under a given input.

[0109] Here, the optimization goal can be expressed by the following formula:

[0110]

[0111] where L 差异 represents the contrast optimization objective function, aiming to make the generation probability of the preferred candidate y A significantly higher than that of the sub-optimal candidate y B under a given input x. P(y|x) represents the probability that the model generates the candidate output y given the input x. κ is a scaling factor used to adjust the influence of the logarithmic probability difference, and its specific value can be determined through experiments. y A and y B represent the better candidate description and the less optimal candidate description respectively, and their sorting is based on the previously constructed preference data set.

[0112] Furthermore, to prevent the model from over-optimizing the contrast objective during training and deviating from the original data distribution, and to further ensure the stability and rationality of the model output, a regularization constraint is introduced into the objective function. This regularization term helps to keep the model from deviating too much from the original data distribution during optimization, ensuring the stability and rationality of the generated results. The regularization term can be a KL divergence regularization term, which is used to control the stability of the model output distribution.

[0113] Here, the total loss function includes the contrast optimization objective and the regularization term, and can be expressed by the following formula:

[0114] L 总 = L 差异 + λ·D(P θ (·|x), P 基 (·|x))

[0115] where L 总 represents the total loss function, which is a weighted combination of the contrast optimization objective L 差异 and the regularization term. D(Pθ (·|x), P 基 (·|x)) represents a distance function for measuring the difference between the model output distribution P θ (·|x) and the reference benchmark distribution P 基 (·|x). The distance metric can include cross or other probability distribution distance metrics. λ represents the regularization weight parameter, which is used to control the proportion of the regularization term in the total loss.

[0116] Preferably, the regularization weight parameter can be dynamically adjusted by the following formula:

[0117]

[0118] where λ represents the regularization weight parameter; λ0 represents the initial regularization weight; t represents the current training step; T represents the preset total number of training steps.

[0119] As a possible implementation, newly emerged case data and medical staff feedback data are regularly integrated. The feedback data is automatically stored in the system and used for subsequent online incremental training and parameter fine-tuning of the model. And an online incremental training is performed on the screening model using a mini-batch update strategy (for example, the batch size is 32), so as to ensure that the model can always adapt to the latest clinical data and the epidemic situation.

[0120] An infectious disease screening method based on preference data and direct preference optimization algorithm provided by an embodiment of the present disclosure obtains multi-dimensional health text information input by a patient, extracts atomic symptom descriptions in the multi-dimensional health text information, and generates medical questions corresponding to the atomic symptom descriptions; determines a consistency reference score between the atomic symptom descriptions and preset standard descriptions, and determines a non-reference score corresponding to each atomic symptom description according to the number of occurrences of the answers corresponding to each question after multiple generation samplings based on the medical questions; fuses the consistency reference score and the non-reference score to obtain a comprehensive authenticity score, and calculates a authenticity reward value corresponding to the atomic symptom description in combination with a preset symptom importance index; determines a preference probability between any two of the atomic symptom descriptions according to the authenticity reward value, and constructs a factual preference data set; uses the factual preference data set to fine-tune a pre-trained screening model through a direct preference optimization algorithm, and outputs an infectious disease symptom screening result in real time. It can improve the screening accuracy of infectious diseases, significantly reduce the missed diagnosis and misdiagnosis rates, and improve the screening efficiency.

[0121] Next, the above solution will be described in conjunction with specific embodiments.

[0122] Taking COVID-19 screening as an example, the key symptoms and their corresponding importance indicators can be set as follows: continuous high fever: I(s high fever) = 0.9; dry cough: I(s dry cough) = 0.8; shortness of breath: I(s shortness of breath) = 0.85; loss of taste and smell: I(s loss of smell) = 0.95. For example, a patient enters through a smartphone "I have had a high fever for 3 consecutive days, accompanied by a severe dry cough and obvious shortness of breath". First, preprocess this text, and use the described word segmentation, entity recognition, and question generation technologies to automatically extract the atomic symptom descriptions s1, s2, and s3, which represent "continuous high fever (3 days)", "severe dry cough", and "shortness of breath" respectively. Using the reference data evaluation and reference-free confidence evaluation methods, calculate the reference scores r1, r2, r3 and the internal confidence scores c1, c2, c3 for s1, s2, and s3 respectively. For example, if the support degree of "continuous high fever" in authoritative data is relatively high and the answers are stable after multiple question-and-answer samplings, then its comprehensive authenticity score S1 can reach 0.85; combined with the preset clinical importance I(s1) = 0.9 and the balance coefficient, for example, take γ = 0.7, and calculate the reward value of "continuous high fever": R(s1) = 0.7×0.9 + 0.3×0.85 = 0.63 + 0.255 = 0.885. The system then makes a pairwise comparison between s1 and s2, and calculates their priority probability:

[0123]

[0124] When the probability is greater than 0.5, the system determines that "continuous high fever" is superior to "severe dry cough". All pairwise comparison results form a factual preference data set, which is used to guide subsequent model training.

[0125] Those skilled in the art can understand that in the above method of the specific implementation manner, the writing order of each step does not mean a strict execution order that constitutes any limitation to the implementation process, and the specific execution order of each step should be determined by its function and possible internal logic.

[0126] Based on the same inventive concept, an infectious disease screening device corresponding to the infectious disease screening method is also provided in the embodiments of the present disclosure. Since the principle of solving problems by the device in the embodiments of the present disclosure is similar to the above infectious disease screening method in the embodiments of the present disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0127] Please refer to Figure 3 , Figure 3 which is a schematic diagram of an infectious disease screening device based on preference data and direct preference optimization algorithm provided by the embodiments of the present disclosure. As Figure 3 shown in, the infectious disease screening device 300 based on preference data and direct preference optimization algorithm provided by the embodiments of the present disclosure includes:

[0128] An information acquisition module 310, configured to acquire multi-dimensional health text information input by a patient, extract atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions.

[0129] A truthfulness scoring module 320, configured to determine a consistency reference score between the atomic symptom description and a preset standard description, and determine a non-reference score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings based on the medical question.

[0130] A preference data analysis module 330, configured to fuse the consistency reference score and the non-reference score to obtain a comprehensive truthfulness score, and calculate a truthfulness reward value corresponding to the atomic symptom description in combination with a preset symptom importance index; determine a preference probability between any two of the atomic symptom descriptions according to the truthfulness reward value, and construct a factual preference data set.

[0131] A symptom screening module 340, configured to fine-tune a pre-trained screening model through a direct preference optimization algorithm and output an infectious disease symptom screening result in real time.

[0132] Descriptions of the processing flows of the various modules in the device and the interaction flows between the modules may refer to the relevant descriptions in the above method embodiments and will not be elaborated here.

[0133] An infectious disease screening device based on preference data and a direct preference optimization algorithm provided by an embodiment of the present disclosure acquires multi-dimensional health text information input by a patient, extracts atomic symptom descriptions from the multi-dimensional health text information, and generates medical questions corresponding to the atomic symptom descriptions; determines a consistency reference score between the atomic symptom description and a preset standard description, and determines a non-reference score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings based on the medical question; fuses the consistency reference score and the non-reference score to obtain a comprehensive truthfulness score, and calculates a truthfulness reward value corresponding to the atomic symptom description in combination with a preset symptom importance index; determines a preference probability between any two of the atomic symptom descriptions according to the truthfulness reward value, and constructs a factual preference data set; uses the factual preference data set to fine-tune a pre-trained screening model through a direct preference optimization algorithm and output an infectious disease symptom screening result in real time. It can improve the screening accuracy of infectious diseases, significantly reduce the missed diagnosis and misdiagnosis rates, and improve the screening efficiency.

[0134] Corresponding to Figure 1 the infectious disease screening method based on preference data and a direct preference optimization algorithm in Figure 4 shown, a schematic structural diagram of an electronic device 400 provided by an embodiment of the present disclosure includes:

[0135] a processor 41, a memory 42, and a bus 43; the memory 42 is used to store execution instructions, including an internal memory 421 and an external memory 422; the internal memory 421 here is also called the main memory, which is used to temporarily store the operation data in the processor 41 and the data exchanged with the external memory 422 such as a hard disk. The processor 41 exchanges data with the external memory 422 through the internal memory 421. When the electronic device 400 runs, the processor 41 communicates with the memory 42 through the bus 43, so that the processor 41 executes Figure 1 the steps of the infectious disease screening method based on preference data and direct preference optimization algorithm in

[0136] The embodiments of the present disclosure further provide a computer-readable storage medium, on which a computer program is stored. When the computer program is run by a processor, it executes the steps of the infectious disease screening method based on preference data and direct preference optimization algorithm described in the above method embodiments. Among them, the storage medium may be a volatile or non-volatile computer-readable storage medium.

[0137] The embodiments of the present disclosure further provide a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, they can execute the steps of the infectious disease screening method based on preference data and direct preference optimization algorithm described in the above method embodiments. For details, please refer to the above method embodiments and will not be elaborated here.

[0138] Among them, the above computer program product can be specifically implemented in a way of hardware, software or a combination thereof. In an alternative embodiment, the computer program product is specifically embodied as a computer storage medium. In another alternative embodiment, the computer program product is specifically embodied as a software product, such as a Software Development Kit (SDK), etc.

[0139] Those skilled in the art can clearly understand that for the convenience and conciseness of description, the specific working processes of the devices described above can refer to the corresponding processes in the foregoing method embodiments, and will not be elaborated herein. In several embodiments provided by the present disclosure, it should be understood that the disclosed devices and methods can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of the units is only a logical function division, and there may be other division methods in actual implementation. For another example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the displayed or discussed couplings or direct couplings or communication connections to each other can be through some communication interfaces. The indirect couplings or communication connections of the devices or units can be in electrical, mechanical or other forms.

[0140] The units described as separate components may or may not be physically separated, and the components displayed as units may or may not be physical units, that is, they can be located in one place or distributed to multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0141] In addition, in each embodiment of the present disclosure, the functional units can be integrated into one processing unit, or each unit can exist physically alone, or two or more units can be integrated into one unit.

[0142] If the functions are implemented in the form of software function units and sold or used as independent products, they can be stored in a non-volatile computer-readable storage medium executable by a processor. Based on such an understanding, the technical solution of the present disclosure, in essence, or the part that contributes to the prior art, or a part of this technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in each embodiment of the present disclosure. The foregoing storage medium includes: various media such as USB flash drives, mobile hard disks, read-only memories (ROMs), random access memories (RAMs), magnetic disks, or optical discs that can store program codes.

[0143] Finally, it should be noted that the above-described embodiments are only specific embodiments of the present disclosure, used to illustrate the technical solutions of the present disclosure, rather than limiting them. The protection scope of the present disclosure is not limited thereto. Although the present disclosure has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that any person skilled in the art within the technical scope disclosed by the present disclosure can still modify the technical solutions described in the foregoing embodiments or easily think of changes, or perform equivalent replacements on some of the technical features; and these modifications, changes or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present disclosure, and should all be covered within the protection scope of the present disclosure. Therefore, the protection scope of the present disclosure should be subject to the protection scope of the claims.

Claims

1. An infectious disease screening method based on preference data and direct preference optimization algorithm, characterized in that, Including: Obtain multi-dimensional health text information input by a patient, extract atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions; Determine a consistency reference score between the atomic symptom description and a preset standard description, and determine a reference-free score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings based on the medical question; Fuse the consistency reference score and the reference-free score to obtain a comprehensive authenticity score, and calculate a authenticity reward value corresponding to the atomic symptom description in combination with a preset symptom importance index; Determine a preference probability between any two of the atomic symptom descriptions according to the authenticity reward value, and construct a factual preference dataset; Utilize the factual preference dataset to fine-tune a pre-trained screening model through a direct preference optimization algorithm, and real-time output an infectious disease symptom screening result.

2. The method according to claim 1, characterized in that, Extract atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions, specifically including: Perform natural language processing on the multi-dimensional health text information, and split the multi-dimensional health text information into multiple sub-word phrases; According to a preset keyword dictionary, identify the atomic symptom descriptions in the sub-word phrase text that match the keyword dictionary; For each atomic symptom description, input the atomic symptom description into a pre-trained medical domain language generation model to generate the medical question including the atomic symptom description.

3. The method according to claim 1, characterized in that, Determine a consistency reference score between the atomic symptom description and a preset standard description, specifically including: Split the atomic symptom description into multiple independent inspection items, and each independent inspection item corresponds to a specific symptom information in the atomic symptom description; Determine the consistency reference score according to the number of the independent inspection items and an indication function indicating whether the independent inspection items are consistent with each other.

4. The method according to claim 1, characterized in that, Determine a reference-free score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings based on the medical question, specifically including: Perform multiple generation samplings on the medical question, and record the question answers of each sampling; Determine a target number of occurrences corresponding to the question answer with the most occurrences; Determine the reference-free score according to the target number of occurrences and the number of samplings corresponding to the generation sampling.

5. The method according to claim 1, wherein Fuse the consistency reference score and the reference-free score to obtain a comprehensive authenticity score, and calculate a authenticity reward value corresponding to the atomic symptom description in combination with a preset symptom importance index, specifically including: Perform weighted averaging on the consistency reference score and the reference-free score according to a preset first weighting parameter to determine a comprehensive authenticity score corresponding to the atomic symptom description; Perform weighted averaging on the comprehensive authenticity score and the symptom importance index according to a preset second weighting parameter to determine the authenticity reward value.

6. The method according to claim 1, characterized in that, Determine a preference probability between any two of the atomic symptom descriptions according to the authenticity reward value, specifically including: For each candidate description pair composed of every two of the described atomic symptom descriptions, determine the reward difference between the truth reward value corresponding to the first atomic symptom description and the truth reward value corresponding to the second atomic symptom description in the candidate description pair. Based on a preset adjustment factor and the reward difference, determine the preference probability indicating that the first atomic symptom description is superior to the second atomic symptom description, where the adjustment factor is used to amplify or reduce the influence of the reward difference on the preference probability calculation.

7. The method according to claim 1, wherein Train the screening model based on the following steps: Obtain a sample preference data set constructed from multiple sample atomic symptom descriptions, and select a preferred candidate symptom description and a sub-optimal candidate symptom description according to the ranking corresponding to the preference probability of the sample atomic symptoms in the sample preference data set. Construct a contrast objective function that makes the generation probability of the preferred candidate symptom description higher than that of the sub-optimal candidate symptom description. According to the contrast objective function, and using a distance function that measures the difference between the model output distribution and the reference benchmark distribution with a KL divergence regularization term as the regularization term, determine the total loss function. Train the screening model according to the total loss function until the screening model has the ability to screen out infectious disease symptoms based on the atomic symptom descriptions.

8. An infectious disease screening device based on preference data and direct preference optimization algorithm, characterized in that, Includes: An information acquisition module, configured to acquire multi-dimensional health text information input by a patient, extract atomic symptom descriptions from the multi-dimensional health text information, and generate medical questions corresponding to the atomic symptom descriptions. An authenticity scoring module, configured to determine a consistency reference score between the atomic symptom description and a preset standard description, and determine the reference-free score corresponding to the atomic symptom description according to the number of occurrences of each question answer after multiple generation samplings based on the medical question. A preference data analysis module, configured to fuse the consistency reference score and the reference-free score to obtain a comprehensive authenticity score, and calculate the truth reward value corresponding to the atomic symptom description in combination with a preset symptom importance index. Determine the preference probability between any two of the atomic symptom descriptions according to the truth reward value, and construct a factual preference data set. A symptom screening module, configured to use the factual preference data set to fine-tune a pre-trained screening model through a direct preference optimization algorithm, and output an infectious disease symptom screening result in real time.

9. An electronic device, characterized in that, Includes: A processor, a memory, and a bus, where the memory stores machine-readable instructions executable by the processor. When the electronic device runs, the processor communicates with the memory through the bus. When the machine-readable instructions are executed by the processor, the steps of the infectious disease screening method based on preference data and direct preference optimization algorithm according to any one of claims 1 to 7 are executed.

10. A computer-readable storage medium, characterized in that, A computer program is stored on the computer-readable storage medium, and when the computer program is run by a processor, the steps of the infectious disease screening method based on preference data and direct preference optimization algorithm according to any one of claims 1 to 7 are executed.

Images