Auxiliary inquiry system for oral treatment

Abstract

The invention relates to the technical field of inquiry systems, in particular to an auxiliary inquiry system for oral treatment. Comprising an oral disease knowledge graph construction and inquiry knowledge base establishment module, a patient information acquisition and preprocessing module, an AI guide type inquiry module, a symptom correlation analysis and disease preliminary auxiliary prediction module, an auxiliary inquiry report generation module and a doctor feedback interaction and model iteration optimization module. According to the method, the problems of inaccurate patient description and key symptom omission in traditional inquiry are solved by constructing the oral disease knowledge graph, designing a multi-modal information acquisition and preprocessing process, generating an AI guide type inquiry process, combining the knowledge graph and machine learning to realize preliminary auxiliary disease prediction, and establishing a strict data security mechanism; the integrity and accuracy of symptom information collection are ensured through multi-modal information collection and dynamic guide inquiry.

Description

Technical Field

[0001] This invention relates to the field of consultation system technology, and in particular to an auxiliary consultation system for oral treatment. Background Technology

[0002] In the field of oral healthcare, consultation is the first step in the treatment process, and its accuracy and efficiency directly affect the formulation of subsequent diagnostic and treatment plans. With increasing public awareness of oral health, the demand for oral healthcare continues to grow, while problems such as the uneven distribution of high-quality oral healthcare resources and the inconsistent professional levels of medical staff in primary healthcare institutions are becoming increasingly prominent.

[0003] Traditional oral health consultations mainly rely on face-to-face communication between medical staff and patients, which has many limitations: on the one hand, patients have limited knowledge of oral diseases and often find it difficult to accurately describe the details of symptoms (such as the type of pain, frequency of attacks, and accompanying symptoms), leading to information transmission errors; on the other hand, medical staff need to manually record patient information and sort out the correlation of symptoms during the consultation process, which is not only time-consuming, but may also cause key information to be missed due to differences in subjective experience. Summary of the Invention In view of the limitations of traditional oral consultations, which mainly rely on face-to-face communication between medical staff and patients, this invention provides an auxiliary consultation system for oral treatment.

[0004] The technical solution adopted in this invention is: an auxiliary consultation system for oral treatment, comprising: The module for constructing a knowledge graph of oral diseases and building a knowledge base for consultation is used to establish a knowledge graph of oral diseases and a knowledge base for consultation. The patient information collection and preprocessing module collects patient information, including basic information, symptom descriptions, and medical history data, and performs preprocessing. The AI-guided consultation module generates a personalized guided consultation process based on the patient's initial information and knowledge graph, and proactively uncovers symptom information that the patient may have missed. The symptom association analysis and preliminary disease prediction module, based on the complete symptom set obtained from the consultation, combines knowledge graphs and machine learning models to analyze the association between symptoms and achieve preliminary prediction of oral diseases. The auxiliary consultation report generation module organizes the auxiliary consultation analysis results into a structured auxiliary report and presents it intuitively through visualization. The doctor feedback interaction and model iteration optimization module introduces a doctor feedback mechanism, using the auxiliary diagnostic results as the basis for model optimization, thereby enabling continuous iteration of the auxiliary consultation system. In one embodiment, the oral disease knowledge graph construction and consultation knowledge base building module performs knowledge graph construction and consultation knowledge base building, as follows: Knowledge graph construction: First, data was collected from oral medicine databases (such as the PubMed oral medicine sub-database and the Chinese Stomatological Association's official website treatment guidelines) and clinical case databases of tertiary hospital oral departments using web crawlers; Total number of disease samples collected Symptom sample size Number of samples for etiology , forming the original dataset ,in For disease datasets, For symptom datasets, This is a dataset for etiological data; Then, natural language processing techniques are used to perform entity recognition and relation extraction on the raw data, and the strength of association between entities is calculated. The formula is: ; in, For entities With entity The number of times they co-occur in clinical cases For entities Total number of occurrences in all cases For entities Total frequency of occurrence in all cases; filtered by association strength, retaining... Based on the entity relationships, construct an initial knowledge graph; Building a medical consultation knowledge base: Based on knowledge graphs, a structured medical consultation knowledge base is constructed, which includes a standardized medical consultation process library, a symptom assessment indicator library, and a disease diagnosis rule library.

[0005] In one embodiment, the patient information collection and preprocessing module collects patient information including basic information, symptom description, and medical history data, and performs preprocessing; the specific method is as follows: The modal information acquisition module is designed to support patients in inputting information via text, voice, and images. Text input: Patients fill in basic information and symptom descriptions on the system interface; Voice input: Using speech recognition technology to convert patients' verbal descriptions of symptoms into text; Image input: Patients can upload photos of the oral cavity. The system uses image preprocessing algorithms to perform preliminary processing on the images and extract key visual features. Data preprocessing involves cleaning, standardizing, and extracting features from the collected information. The specific steps are as follows: Data cleaning: Removing invalid data and handling missing data. For missing items in the basic information, a predictive imputation method based on other patient information is used, with the following formula: ; in, The age after filling in the blank. For patients with similar symptoms in the past, their ages were considered. For similar patient numbers, For the current patient and the first Symptom similarity among historical patients; Text standardization: The symptom description text is segmented and stop words are removed, converting the text into word vectors. The Word2Vec algorithm is used to train word vectors, with a dimension of 200. The formula is as follows: ; in, The number of words after word segmentation. For the first Word2Vec vectors of words, The average word vector of the symptom description text; Feature fusion: Integrating text features, speech-to-text features, and image features into a unified patient feature vector. The dimension is 200 + 2 = 202, and the formula is: ; in, For text feature weights, These are the image feature weights.

[0006] In one embodiment, the AI-guided consultation module generates a personalized guided consultation process based on the patient's initial information and knowledge graph, and actively uncovers symptom information that the patient may have missed. The specific method is as follows: Initial symptom matching and diagnosis direction determination: The preprocessed feature vector of the patient Similarity calculations were performed between the symptom entity vectors and the knowledge graph using the cosine similarity formula: ; in, Symptom entities in a knowledge graph vector, For symptoms Similarity to patient characteristics. Screening. The symptoms were used as the initial matching symptoms. And based on the associated diseases of these symptoms in the knowledge graph, the direction of the consultation is determined; Dynamic question generation for medical consultation: Based on the consultation direction and standardized processes in the knowledge base, a reinforcement learning algorithm is used to generate dynamic consultation questions. The consultation process is modeled as a Markov decision process, with states... For "the currently acquired set of symptoms", actions Reward for the next consultation question generated The contribution of the information obtained for this problem to the diagnosis of the disease; The reward function is designed as follows: ; in, For new symptoms obtained through this problem, For the current set of candidate diseases, The average contribution weight of new symptoms to candidate diseases. For indicator functions; The goal of training the diagnostic strategy model using deep reinforcement learning is to maximize the cumulative reward. γ is the deduction factor. After the model is trained, the next question is dynamically generated based on the current consultation status.

[0007] In one embodiment, the symptom association analysis and preliminary disease prediction module analyzes the relationships between symptoms based on the complete symptom set obtained from the consultation, combined with a knowledge graph and a machine learning model, to achieve preliminary prediction of oral diseases. The specific method is as follows: First, the symptom association strength calculation is based on the association weights between symptoms and diseases in the knowledge graph, combined with the patient's actual set of symptoms. Calculate the association score between each symptom and the candidate disease. : ; in, For symptoms For disease Contribution weight, For the patient's symptoms Credibility; Then, calculate the disease. Comprehensive correlation score The formula is: ; in, For disease A typical set of symptoms The penalty coefficient for missing symptoms; Machine learning models assist in disease prediction; Random forest models are used to further assist in disease prediction. Patient feature vector Symptom association score ; During model training, 5-fold cross-validation was used to optimize parameters. The number of decision trees was set to 100, and the maximum tree depth was 10. The model output is the predicted probability for each candidate disease. The formula is: ; in, The number of decision trees in the random forest. For the first The prediction function of a decision tree. For indicator functions; Finally, by combining the symptom association score with the model's predicted probability, the final recommendation priority for the disease is calculated. : ; in, For weighting coefficients, according to Sort the predicted diseases from highest to lowest and output the top 3 candidate diseases and their corresponding confidence scores; In one embodiment, the auxiliary consultation report generation module organizes the auxiliary consultation analysis results into a structured auxiliary report and presents it intuitively through visualization, as follows: Structured assisted consultation report generation: Automatically fill in content using natural language generation technology: Patient basic information: age, gender, underlying diseases, taken from pre-processed patient information; Symptom summary: Organized by "Main Symptoms - Accompanying Symptoms - Symptom Characteristics"; Temporary care recommendations: Generate personalized temporary care instructions based on patient symptoms and disease prediction results; Visual presentation design: The results of the consultation are presented using intuitive visual charts.

[0008] In one embodiment, the doctor feedback interaction and model iteration optimization module introduces a doctor feedback mechanism, using the auxiliary diagnostic results as the basis for model optimization. The specific method for achieving continuous iteration of the assisted consultation system is as follows: After the doctor completes the final diagnosis, the system displays a feedback interface to collect the following key information: Diagnostic result matching degree: The doctor's assessment of the matching degree between the top 3 predicted diseases and the actual diagnostic results; Model Iterative Optimization: Incremental learning algorithm is used for iterative optimization.

[0009] The beneficial effects of this invention are as follows: Compared with the prior art, this invention solves the problems of inaccurate patient descriptions and omissions of key symptoms in traditional consultations by constructing an oral disease knowledge graph, designing a multimodal information collection and preprocessing process, generating an AI-guided consultation process, combining knowledge graphs and machine learning to achieve preliminary auxiliary prediction of diseases, and establishing a strict data security mechanism. Through multimodal information collection and dynamic guided consultation, the completeness and accuracy of symptom information collection are ensured. Secondly, the consultation efficiency is greatly improved. AI automatically completes information recording, symptom analysis and disease prediction, reducing the manual operation time of medical staff and shortening the average waiting time for patients. Attached Figure Description

[0010] Figure 1 This is a schematic diagram of the system framework structure of the present invention. Detailed Implementation

[0011] In the description of this invention, it should be noted that the terms "front", "up", "down", "left", "right", "vertical", "horizontal", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0012] To address the problems existing in the background art, this application proposes the following technical solution: an auxiliary consultation system for oral treatment, comprising: The module for constructing a knowledge graph of oral diseases and building a knowledge base for consultation is used to establish a knowledge graph of oral diseases and a knowledge base for consultation. The module for constructing an oral disease knowledge graph and building a consultation knowledge base performs knowledge graph construction and consultation knowledge base construction, as detailed below: Knowledge graph construction: Construct a knowledge graph of oral diseases, covering disease entities (such as dental caries, periodontitis, pulpitis, etc.), symptom entities (such as toothache, gingival bleeding, halitosis, etc.), etiological entities (such as bacterial infection, dietary habits, oral hygiene, etc.), treatment plan entities (such as fillings, root canal treatment, periodontal cleaning, etc.), and examination items entities (such as oral X-rays, periodontal probing, etc.). First, data was collected from oral medicine databases (such as the PubMed oral medicine sub-database and the Chinese Stomatological Association's official website treatment guidelines) and clinical case databases of tertiary hospital oral departments using web crawlers; Total number of disease samples collected Symptom sample size Number of samples for etiology , forming the original dataset ,in For disease datasets, For symptom datasets, This is a dataset for etiological data; Then, natural language processing techniques (such as the BERT model) are used to perform entity recognition and relation extraction on the raw data, and the strength of association between entities is calculated. The formula is: ; in, For entities With entity The number of times they co-occur in clinical cases For entities Total number of occurrences in all cases For entities Total number of occurrences across all cases. The value range is [0,1]. The closer the value is to 1, the stronger the relationship between the two entities (e.g., "pulpitis" and "nocturnal spontaneous pain"). (Approximately 0.9). Based on association strength screening, [the following was retained]. The entity relationships are used to construct an initial knowledge graph.

[0013] Finally, at least three oral specialists at the associate chief physician level were invited to manually review and revise the initial knowledge graph, supplementing it with rare disease-symptom associations and disease classification information from the latest treatment guidelines, thus forming the final oral disease knowledge graph. ,in For a collection of entities, It is a set of relationships between entities.

[0014] Building a medical consultation knowledge base: Based on knowledge graphs, a structured medical consultation knowledge base is constructed, which includes a standardized medical consultation process library, a symptom assessment indicator library, and a disease diagnosis rule library. Standardized consultation process library: Consultation processes are designed according to the type of oral disease (such as dental pulp disease, periodontal disease, oral mucosal disease). Each process includes 3-5 core consultation nodes (such as initial symptom screening, symptom detail confirmation, medical history inquiry, and lifestyle habit survey) to ensure that the consultation logic is coherent and no key information is omitted.

[0015] Symptom assessment indicator library: Quantitative assessment indicators are set for each symptom entity. For example, the assessment indicators for "tooth pain" include pain intensity (1-10 point visual analog scale), pain type (persistent / intermittent, spontaneous / irritative), and pain duration (…). (Unit: days), and establish a weighted table for the association between indicators and diseases. ,in Indicates symptoms For disease The contribution weight of diagnosis, satisfying For disease (Corresponding typical symptom set).

[0016] Disease Diagnosis Rule Base: This system transforms diagnostic criteria from oral medicine treatment guidelines into machine-readable rules. For example, if a patient experiences tooth sensitivity to hot or cold stimuli (symptoms)... Pain worsens at night (symptom) And there was no bleeding gums (symptom) If the tooth is missing, the preliminary diagnosis is pulpitis (disease). Rule confidence The value is set by experts based on clinical experience (e.g., 0.85) and used to assess the reliability of subsequent disease predictions.

[0017] The patient information collection and preprocessing module collects patient information, including basic information, symptom descriptions, and medical history data, and performs preprocessing. The patient information collection and preprocessing module collects patient information including basic information, symptom descriptions, and medical history data, and performs preprocessing; the specific method is as follows: The modal information acquisition module is designed to support patients in inputting information via text, voice, and images. Text input: Patients fill in basic information (age) on the system interface. ,gender Do you have underlying medical conditions such as high blood pressure or diabetes? ), Symptom description (e.g., "Pain in the lower left teeth for 3 days, which is worse when eating hot or cold foods").

[0018] Voice input: Speech recognition technology (such as a Transformer-based speech recognition model) is used to convert the patient's spoken symptoms into text, achieving a high accuracy rate. The accuracy rate needs to reach over 95%. If the recognition result is ambiguous (e.g., "pain" may correspond to "toothache" or "gingival pain"), the system will automatically generate 2-3 candidate options for the patient to confirm.

[0019] Image Input: Patients can upload photos of specific areas of their oral cavity (such as swollen gums or cavities). The system uses image preprocessing algorithms (such as grayscale conversion and edge detection) to perform preliminary processing on the images and extract key visual features (such as the area of ​​the swollen region). Diameter of the cavity This provides visual evidence for symptom assessment.

[0020] Data preprocessing involves cleaning, standardizing, and extracting features from the collected information. The specific steps are as follows: Data cleaning: Remove invalid data (such as blank text descriptions, blurred and unrecognizable pictures), and handle missing data. For missing items in basic information (such as age not filled), use the prediction filling method based on other patient information. The formula is: ; where is the filled age, is the age of historical patients with similar symptoms (similarity ), is the number of similar patients, is the symptom similarity between the current patient and the th historical patient (calculated by text similarity algorithm, such as cosine similarity).

[0021] Improve the accuracy of missing age filling by weighted averaging the ages of similar patients.

[0022] Text standardization: Segment the symptom description text (using a professional segmentation dictionary for oral medicine, including professional terms such as "cavity" and "periodontal pocket"), remove stop words (such as "of" and "了"), and convert the text into word vectors , and use the Word2Vec algorithm to train the word vectors with the dimension set to 200. The formula is: ; where is the number of words after segmentation, is the Word2Vec vector of the th word, is the average word vector of the symptom description text, which is used for subsequent symptom matching and analysis. 3. Feature fusion: Integrate text features (word vectors ), text features converted from speech (processed in the same way as text features), and picture features (such as redness area , cavity diameter , etc.) into a unified patient feature vector with a dimension of 200 + 2 = 202. The formula is: ; where is the text feature weight (0.7), is the picture feature weight (0.3), and the weights are determined through cross-validation to ensure that the contributions of different modality features to subsequent analysis are reasonable.

[0023] AI-guided interrogation module, based on the initial patient information and knowledge graph, generates a personalized guided interrogation process to actively mine symptom information that patients may have missed; In the AI-guided consultation module, a personalized guided consultation process is generated based on the patient's initial information and knowledge graph. The specific methods for proactively uncovering symptom information that the patient may have missed are as follows: Initial symptom matching and diagnosis direction determination: The preprocessed feature vector of the patient Similarity calculations were performed between the symptom entity vectors and the knowledge graph using the cosine similarity formula: ; in, Symptom entities in a knowledge graph The vector (generated during the knowledge graph construction process), For symptoms Screening based on similarity to patient characteristics The symptoms were used as the initial matching symptoms. Based on the associated diseases of these symptoms in the knowledge graph, the direction of the consultation is determined (e.g., if the initial matched symptom is "pain in teeth due to hot or cold stimuli", the consultation direction focuses on symptoms related to dental pulp diseases).

[0024] Dynamic question generation for medical consultation: Based on the consultation direction and standardized processes in the knowledge base, a reinforcement learning algorithm is used to generate dynamic consultation questions. The consultation process is modeled as follows: Markov Decision Process (MDP), State For "the currently acquired set of symptoms", actions Rewards are given for "the next consultation question generated". This refers to "the contribution of the information obtained from this question to the diagnosis of the disease".

[0025] The reward function is designed as follows: ; in, For new symptoms obtained through this problem, The current set of candidate diseases (generated from the initial matching symptom associations), The average contribution weight of new symptoms to candidate diseases (taken from the symptom assessment index library). For indicator functions (if) Not in the previous symptom set In the middle, then ,otherwise ).

[0026] The generated questions can acquire new, diagnostically valuable information. The diagnostic strategy model is trained using deep reinforcement learning (DRL), with the goal of maximizing cumulative reward. (γ is the deduction factor, set to 0.9). After the model is trained, the next question can be dynamically generated according to the current consultation status until the obtained symptom set can cover more than 80% of the typical symptoms of the candidate disease, or the consultation steps reach the preset limit (such as 10 steps), and then the consultation stops.

[0027] The symptom association analysis and preliminary disease prediction module, based on the complete symptom set obtained from the consultation, combines knowledge graphs and machine learning models to analyze the association between symptoms and achieve preliminary prediction of oral diseases. In the symptom association analysis and preliminary disease prediction module, based on the complete symptom set obtained from the consultation, and combined with knowledge graphs and machine learning models, the association relationships between symptoms are analyzed to achieve preliminary prediction of oral diseases. The specific methods are as follows: First, the symptom association strength calculation is based on the association weights between symptoms and diseases in the knowledge graph, combined with the patient's actual set of symptoms. Calculate the association score between each symptom and the candidate disease. : ; in, For symptoms For disease The contribution weight (taken from the symptom assessment index library). For the patient's symptoms The credibility of the evidence (such as symptoms clearly described by the patient and supported by pictures). Symptoms only vaguely mentioned ).

[0028] Then, calculate the disease. Comprehensive correlation score The formula is: ; in, For disease A typical set of symptoms A penalty coefficient for missing symptoms (set to 0.3) is used. The contribution of typical symptoms already present in the patient is considered, and key symptoms that are not present are also penalized to ensure that the score accurately reflects the likelihood of the disease (e.g., if the typical symptom of pulpitis, "nighttime pain," is not present, then...). (It will decrease accordingly).

[0029] Machine learning models assist in disease prediction by employing a random forest (RF) model to further aid in disease prediction and improve accuracy. This involves using patient feature vectors... Symptom association score (For all candidate diseases) concatenate them into the model input vector , dimension ( (Number of candidate diseases).

[0030] During model training, 5-fold cross-validation was used to optimize parameters. The number of decision trees was set to 100, and the maximum tree depth was 10. The model output is the predicted probability for each candidate disease. The formula is: ; in, The number of decision trees in the random forest. For the first The prediction function of a decision tree. For indicator functions (if the first) The trees were predicted to be diseased. ,but ,otherwise Finally, by combining the symptom association score with the model's predicted probability, the final recommendation priority for the disease is calculated. : ; in, The weighting coefficient (set to 0.4) balances the results of expert knowledge (symptom association scores) and data-driven approaches (model prediction probabilities). (By...) Sort the candidates from highest to lowest confidence level and output the top 3 candidate diseases and their corresponding confidence scores. (Normalized values), such as "1. Pulpitis (confidence 0.82); 2. Dental caries (confidence 0.65); 3. Periodontitis (confidence 0.31)".

[0031] The auxiliary consultation report generation module organizes the auxiliary consultation analysis results into a structured auxiliary report and presents it intuitively through visualization. In the auxiliary consultation report generation module, the specific method for organizing the auxiliary consultation analysis results into a structured auxiliary report and presenting it intuitively through visualization is as follows: Structured assisted consultation report generation: The report includes the following modules, which are automatically populated with content using Natural Language Generation (NLG) technology: Patient basic information: age, gender, underlying diseases, etc., taken from pre-processed patient information.

[0032] Symptom summary: Organized by "Main Symptoms - Accompanying Symptoms - Symptom Characteristics", such as "Main Symptoms: Pain in the lower left teeth (lasting for 3 days, visual analog scale score of 7); Accompanying Symptoms: Pain worsens with hot and cold stimuli; Symptom Characteristics: No gingival bleeding, no halitosis".

[0033] Disease prediction results: The top 3 candidate diseases are listed, each including a disease description (taken from a knowledge graph) and a comparison of matching symptoms. The corresponding examination items are taken from the diseases in the knowledge graph. The corresponding set of examination items; such as "Pulpitis: Disease introduction: an inflammatory lesion of the dental pulp tissue, often developed from tooth decay; matching symptom ratio: 80% (4 / 5 typical symptoms match); recommended examination items: oral X-ray, pulp vitality test".

[0034] Temporary nursing recommendations: Based on the patient's symptoms and disease prediction results, generate personalized temporary nursing guidance, such as "When the pain is obvious, you can take warm salt water to relieve it, avoid eating food that is too cold or too hot, and avoid vigorous chewing within 24 hours". The nursing recommendations must comply with the oral medicine nursing standards and be marked "temporary recommendations, subject to the doctor's diagnosis".

[0035] Visual presentation design: The consultation results are presented using intuitive visual charts, lowering the barrier to information comprehension. Symptom Association Heatmap: Displays the strength of the association between patient symptoms and candidate diseases in matrix form. The horizontal axis represents candidate diseases, and the vertical axis represents patient symptoms. Cell colors correspond to different shades. (Symptom-Disease Association Score), the darker the color, the stronger the association, helping medical staff quickly identify the correspondence between core symptoms and diseases.

[0036] Disease Confidence Pie Chart: A pie chart showing the confidence percentage of the top 3 candidate diseases, with the area of ​​each sector corresponding to the normalized confidence level. The disease name and specific confidence level value are indicated, such as "pulpitis (45%), dental caries (35%), periodontitis". This allows patients and medical staff to intuitively understand the distribution of disease probabilities.

[0037] Consultation process timeline: Records the time and key nodes of each stage of the consultation (information collection, dynamic consultation, symptom analysis) in the form of a timeline, such as "0-2 minutes: information collection completed; 2-5 minutes: dynamic consultation (3 questions); 5-6 minutes: disease prediction completed", to help medical staff evaluate consultation efficiency and optimize subsequent processes.

[0038] The doctor feedback interaction and model iteration optimization module introduces a doctor feedback mechanism, using the auxiliary diagnostic results as the basis for model optimization, thereby enabling continuous iteration of the auxiliary consultation system. In the doctor feedback interaction and model iteration optimization module, a doctor feedback mechanism is introduced, using the auxiliary diagnostic results as the basis for model optimization. The specific method for achieving continuous iteration of the auxiliary consultation system is as follows: After the doctor completes the final diagnosis, the system displays a feedback interface to collect the following key information: Diagnostic result matching degree: Doctors judge the matching degree between the top 3 candidate diseases predicted by AI and the actual diagnosis results, which are divided into "complete match (actual diagnosis disease is in the top 1 of AI)", "partial match (actual diagnosis disease is in the top 2-3 of AI)" and "no match (actual diagnosis disease is not in the top 3 of AI)" and label the matching degree with Match (1=complete match, 0.5=partial match, 0=no match).

[0039] Symptom importance adjustment: Doctors can adjust the contribution weight of each symptom in the AI ​​analysis. For example, if the AI ​​assigns a weight of 0.3 to "tooth pain" in relation to pulpitis, and the doctor adjusts it to 0.4 based on clinical experience, the system records the adjusted weight. .

[0040] Consultation process optimization suggestions: Doctors can fill in their suggestions for optimizing the dynamic consultation process, such as "Add a question about whether you have ever woken up at night due to pain" or "Reduce redundant questions related to 'dietary habits'". The system will categorize and store the suggestions in the "process optimization suggestion library".

[0041] The model iterative optimization algorithm is based on doctor feedback data and uses an incremental learning algorithm to iteratively optimize the AI ​​model. The specific steps are as follows: Weight update: Based on the symptom weights adjusted by the doctor. Update the original weights in the symptom assessment indicator library. The updated formula is: ; in, For the updated weights, The original weights are set before the update, and β is the weight retention coefficient (set to 0.8) to ensure smooth weight updates and avoid drastic weight fluctuations caused by the subjective judgment of a single doctor.

[0042] Disease prediction model optimization: incorporating the "patient feature vector" feedback from doctors ; Actual diagnosis of disease The "Match" data was used to construct an incremental training dataset. The random forest incremental learning algorithm is used to update the model parameters.

[0043] First, calculate the similarity between the incremental data and the original training data. The formula is: ; in, The original training dataset, This represents the number of samples with similar features between the incremental and original data. If... If the distribution is ≥0.6 (the incremental data is consistent with the original data distribution), then only the node parameters of some decision trees in the random forest are updated; if If the value is less than 0.6 (the incremental data contains new disease feature patterns), then add 10-20 decision trees to the model to ensure that the model can adapt to the new clinical data features.

[0044] Consultation process optimization: Regularly (e.g., monthly) perform statistical analysis on suggestions in the "Process Optimization Suggestion Library" and calculate the mention frequency F of each suggestion. suggest If a certain suggestion F suggest If the suggestion is mentioned ≥5 times (by 5 or more doctors), it will be incorporated into the consultation process optimization plan. For example, if the frequency of mentioning "adding the consultation question about 'nighttime awakening pain'" is 8, then this question will be added to the dynamic consultation process related to dental endodontics, and the reinforcement learning consultation strategy model will be retrained and the priority of question generation will be adjusted.

[0045] The optimization effect is evaluated using the following metrics after each model iteration: Disease prediction accuracy The calculation formula is: ; in, The number of cases where the AI ​​prediction results match the doctor's actual diagnosis (full match + partial match). To assess the total number of cases within the period.

[0046] The goal is to make Each iteration improves efficiency by 1%-2%, consistently remaining above 90%. - Consultation efficiency improvement rate Comparison of average consultation time before and after optimization and The calculation formula is: ; The goal is to make This means that each optimization reduces consultation time by more than 5%, improving the time efficiency for both patients and doctors.

[0047] While some existing oral healthcare apps attempt to provide simple consultation assistance functions, most only focus on symptom keyword matching, lacking in-depth analysis of symptom correlations and failing to provide personalized diagnoses by combining patients' historical health data and oral disease pathological characteristics. Furthermore, these tools do not fully utilize the advantages of artificial intelligence technologies such as natural language processing, machine learning, and knowledge graphs, making it difficult to achieve intelligent guidance during the consultation process, accurate extraction of symptom information, and preliminary prediction of potential diseases. They cannot provide effective decision support for medical staff, nor can they effectively alleviate problems such as long waiting times and poor consultation experiences for patients.

[0048] In this embodiment, by constructing an oral disease knowledge graph, designing a multimodal information collection and preprocessing process, generating an AI-guided consultation process, combining the knowledge graph and machine learning to achieve preliminary auxiliary disease prediction, generating structured reports and visualizations, introducing doctor feedback to iteratively optimize the model, and establishing a strict data security mechanism, the problems of inaccurate patient descriptions and omissions of key symptoms in traditional consultations are solved. Through multimodal information collection and dynamic guided consultations, the completeness and accuracy of symptom information collection are ensured. Secondly, consultation efficiency is greatly improved. AI automatically completes information recording, symptom analysis, and disease prediction, reducing the manual operation time of medical staff and shortening the average waiting time for patients. It provides reliable diagnostic references for medical staff in primary healthcare institutions, reduces the differences in consultation quality between different institutions, and improves patients' awareness of the disease and their medical experience through visualized reports and personalized nursing suggestions.

[0049] Although embodiments of the invention have been shown and described, the scope of the invention will be defined by the appended claims and their equivalents by those skilled in the art.

Claims

1. An auxiliary consultation system for oral treatment, characterized in that, include: The module for constructing a knowledge graph of oral diseases and building a knowledge base for consultation is used to establish a knowledge graph of oral diseases and a knowledge base for consultation. The patient information collection and preprocessing module collects patient information, including basic information, symptom descriptions, and medical history data, and performs preprocessing. The AI-guided consultation module generates a personalized guided consultation process based on the patient's initial information and knowledge graph, and proactively uncovers symptom information that the patient may have missed. The symptom association analysis and preliminary disease prediction module, based on the complete symptom set obtained from the consultation, combines knowledge graphs and machine learning models to analyze the association between symptoms and achieve preliminary prediction of oral diseases. The auxiliary consultation report generation module organizes the auxiliary consultation analysis results into a structured auxiliary report and presents it intuitively through visualization. The doctor feedback interaction and model iteration optimization module introduces a doctor feedback mechanism, using the auxiliary diagnostic results as the basis for model optimization, thereby enabling continuous iteration of the auxiliary consultation system.

2. The auxiliary consultation system for oral treatment according to claim 1, characterized in that, The module for constructing an oral disease knowledge graph and building a consultation knowledge base performs knowledge graph construction and consultation knowledge base construction, as detailed below: Knowledge graph construction: First, data was collected from oral medicine databases and clinical case databases of oral departments in tertiary hospitals using web crawlers; Total number of disease samples collected Symptom sample size Number of samples for etiology , forming the original dataset ,in For disease datasets, For symptom datasets, This is a dataset for etiological data; Then, natural language processing techniques are used to perform entity recognition and relation extraction on the raw data, and the strength of association between entities is calculated. The formula is: ； in, For entities With entity The frequency of co-occurrence in clinical cases For entities Total number of occurrences in all cases For entities Total number of occurrences in all cases; filtered by association strength, retaining... Based on the entity relationships, construct an initial knowledge graph; Building a medical consultation knowledge base: Based on knowledge graphs, a structured medical consultation knowledge base is constructed, which includes a standardized medical consultation process library, a symptom assessment indicator library, and a disease diagnosis rule library.

3. The auxiliary consultation system for oral treatment according to claim 2, characterized in that, The patient information collection and preprocessing module collects patient information including basic information, symptom descriptions, and medical history data, and performs preprocessing; the specific method is as follows: The modal information acquisition module is designed to support patients in inputting information via text, voice, and images. Text input: Patients fill in basic information and symptom descriptions on the system interface; Voice input: Using speech recognition technology to convert patients' verbal descriptions of symptoms into text; Image input: Patients can upload photos of the oral cavity. The system uses image preprocessing algorithms to perform preliminary processing on the images and extract key visual features. Data preprocessing involves cleaning, standardizing, and extracting features from the collected information. The specific steps are as follows: Data cleaning: removing invalid data and handling missing data; for missing items in basic information, a predictive imputation method based on other patient information is used, the formula is: ； in, The age after filling in the blank. For patients with similar symptoms in the past, their ages were considered. For similar patient numbers, For the current patient and the first Symptom similarity among historical patients; Text standardization: The symptom description text is segmented and stop words are removed, converting the text into word vectors. The Word2Vec algorithm is used to train word vectors, with a dimension of 200. The formula is as follows: ； in, The number of words after word segmentation. For the first Word2Vec vectors of words, The average word vector of the symptom description text; Feature fusion: Integrating text features, speech-to-text features, and image features into a unified patient feature vector. The dimension is 200 + 2 = 202, and the formula is: ； in, For text feature weights, These are the image feature weights.

4. The auxiliary consultation system for oral treatment according to claim 3, characterized in that, In the AI-guided consultation module, a personalized guided consultation process is generated based on the patient's initial information and knowledge graph. The specific methods for proactively uncovering symptom information that the patient may have missed are as follows: Initial symptom matching and diagnosis direction determination: The preprocessed feature vector of the patient Similarity calculations were performed between the symptom entity vectors and the knowledge graph using the cosine similarity formula: ； in, Symptom entities in a knowledge graph vector, For symptoms Similarity to patient characteristics; screening The symptoms were used as the initial matching symptoms. And based on the associated diseases of these symptoms in the knowledge graph, the direction of the consultation is determined; Dynamic question generation for medical consultation: Based on the consultation direction and standardized processes in the knowledge base, a reinforcement learning algorithm is used to generate dynamic consultation questions. The consultation process is modeled as a Markov decision process, with states... For "the currently acquired set of symptoms", the action For the next consultation question generated, a reward will be given. The contribution of the information obtained for this problem to the diagnosis of the disease; The reward function is designed as follows: ； in, For new symptoms obtained through this problem, For the current set of candidate diseases, The average contribution weight of new symptoms to candidate diseases. For indicator functions; The goal of training the diagnostic strategy model using deep reinforcement learning is to maximize the cumulative reward. γ is the deduction factor. After the model is trained, the next question is dynamically generated based on the current consultation status.

5. The auxiliary consultation system for oral treatment according to claim 4, characterized in that, In the symptom association analysis and preliminary disease prediction module, based on the complete symptom set obtained from the consultation, and combined with knowledge graphs and machine learning models, the association relationships between symptoms are analyzed to achieve preliminary prediction of oral diseases. The specific methods are as follows: First, the symptom association strength calculation is based on the association weights between symptoms and diseases in the knowledge graph, combined with the patient's actual set of symptoms. Calculate the association score between each symptom and the candidate disease. : ； in, For symptoms For disease Contribution weight, For the patient's symptoms Credibility; Then, calculate the disease. Comprehensive correlation score The formula is: ； in, For disease A typical set of symptoms The penalty coefficient for missing symptoms; Machine learning models assist in disease prediction; Random forest models are used to further assist in disease prediction. Patient feature vector Symptom association score ; During model training, 5-fold cross-validation was used to optimize parameters. The number of decision trees was set to 100, and the maximum tree depth was set to 10. The model output is the predicted probability for each candidate disease. The formula is: ； in, The number of decision trees in the random forest. For the first The prediction function of a decision tree. For indicator functions; Finally, by combining the symptom association score with the model's predicted probability, the final recommendation priority for the disease is calculated. : ； in, For weighting coefficients, according to Sort the predicted diseases from highest to lowest and output the top 3 candidate diseases and their corresponding confidence scores.

6. The auxiliary consultation system for oral treatment according to claim 5, characterized in that, In the auxiliary consultation report generation module, the specific method for organizing the auxiliary consultation analysis results into a structured auxiliary report and presenting it intuitively through visualization is as follows: Structured assisted consultation report generation: Automatically fill in content using natural language generation technology: Patient basic information: age, gender, underlying diseases, taken from pre-processed patient information; Symptom summary: Organized by "Main Symptoms - Accompanying Symptoms - Symptom Characteristics"; Temporary care recommendations: Generate personalized temporary care instructions based on patient symptoms and disease prediction results; Visual presentation design: The results of the consultation are presented using intuitive visual charts.

7. The auxiliary consultation system for oral treatment according to claim 6, characterized in that, In the doctor feedback interaction and model iteration optimization module, a doctor feedback mechanism is introduced, using the auxiliary diagnostic results as the basis for model optimization. The specific method for achieving continuous iteration of the auxiliary consultation system is as follows: After the doctor completes the final diagnosis, the system displays a feedback interface to collect the following key information: Model Iterative Optimization: Incremental learning algorithm is used for iterative optimization.

Images