Oral diagnosis and treatment patient service platform based on reinforcement learning

Abstract

The invention discloses an oral diagnosis and treatment patient service platform based on reinforcement learning, and relates to the technical field of diagnosis and treatment services, and the platform comprises a health scoring module which integrates multi-source data, generates personalized oral health scores through a dynamic weight distribution model, and displays high-risk items in combination with 3D visualization; the patient management and marketing module tracks patient behaviors, clusters and classifies the patient behaviors, generates a dynamic follow-up visit path by using multi-target reinforcement learning, and matches a precise marketing verbal skill through a knowledge graph; the in-diagnosis auxiliary module synchronizes health data of the patient to the doctor terminal in real time, detects prescription conflicts and performs early warning, and generates personalized communication verbal skills in combination with historical preferences of the patient and real-time emotion analysis; through multi-dimensional data integration and intelligent processing, the whole oral diagnosis and treatment process is optimized, the diagnosis and treatment efficiency and the patient satisfaction degree are improved, the operation cost is reduced, and remarkable innovativeness and practicability are achieved.

Description

Technical Field

[0001] The present invention relates to the field of diagnosis and treatment data service technology, and specifically to an oral diagnosis and treatment patient service platform based on reinforcement learning. Background Art

[0002] Oral health is an important part of overall health and is closely related to people's quality of life. However, the traditional oral diagnosis and treatment model mainly relies on the doctor's clinical experience and the patient's subjective description, lacking a quantitative assessment of the patient's overall health status and the formulation of personalized treatment plans. This model seems powerless in the face of the growing complexity and diversity of oral diseases. When making a diagnosis, doctors can often only rely on the patient's current clinical manifestations and limited medical history records. It is difficult to fully understand the impact of the patient's living habits, environmental factors, etc. on oral health, which leads to one-sided diagnosis and treatment decisions. In addition, under the traditional diagnosis and treatment model, patients lack awareness of their own oral health status and lack effective self-management tools, making the prevention and early intervention of oral diseases difficult to implement.

[0003] In terms of patient management, most dental institutions currently rely primarily on manual follow-up and appointment reminders, a method that is inefficient and prone to omissions. Furthermore, patient classification management is insufficiently refined, making it impossible to develop personalized follow-up strategies based on individual patient characteristics and needs, resulting in high patient churn rates. Traditional dental marketing relies primarily on advertising and word-of-mouth, lacking in-depth patient data mining and precision marketing methods. This makes it difficult for medical institutions to accurately understand patients' potential needs and consumption preferences, making it impossible to provide targeted marketing content, which in turn impacts marketing effectiveness and patient conversion rates.

[0004] During oral care, the efficiency and quality of communication between doctors and patients directly impact treatment outcomes and patient satisfaction. However, the current treatment environment presents several communication barriers, such as doctors' inability to fully answer patients' questions due to time constraints and patients' difficulty understanding specialized terminology. Furthermore, doctors lack real-time decision-making tools when formulating treatment plans, making it difficult to comprehensively consider a patient's multidimensional information, such as allergy history and treatment preferences. This can lead to irrational or incomplete treatment plans. In these cases, patient compliance may decrease, compromising treatment outcomes and oral health recovery.

[0005] In order to solve the above defects, a technical solution is now provided. Summary of the Invention

[0006] The purpose of this invention is to solve the problems of the traditional oral diagnosis and treatment model, such as the lack of comprehensive quantitative evaluation, inefficient patient management and marketing methods, and obstacles in communication and decision-making during diagnosis, and to propose an oral diagnosis and treatment patient service platform based on reinforcement learning.

[0007] The purpose of the present invention can be achieved through the following technical solutions: The dental diagnosis and treatment patient service platform based on reinforcement learning includes: The health scoring module integrates multi-source data, generates personalized oral health scores through a dynamic weight distribution model, and displays high-risk items in 3D visualization; The patient management and marketing module tracks and clusters patient behavior, uses multi-objective reinforcement learning to generate dynamic follow-up paths, and matches precise marketing strategies through knowledge graphs. The in-diagnosis assistance module synchronizes patient health data to the doctor's terminal in real time, detects prescription conflicts and issues warnings, and generates personalized communication language based on patient historical preferences and real-time sentiment analysis.

[0008] Furthermore, the specific execution process of the health scoring module is as follows: Multi-source data collection: Obtain medical records from the hospital HIS system, collect lifestyle data through questionnaires or wearable devices, extract personality tags through social media analysis, and use generative adversarial networks to enhance virtual patient portraits; Data desensitization and feature extraction: Encrypt sensitive information, use NLP technology to extract features of unstructured text, and generate structured vectors; Dynamic weight distribution model training: Design an adaptive weight distribution network based on the large model and dynamically adjust the scoring weight according to the data type; Health score generation and visualization: Outputs an oral health index of 0-100 points, marks high-risk items, and intuitively displays the location of lesions through a 3D oral model.

[0009] Furthermore, the specific steps of using the generative adversarial network to enhance the virtual patient portrait in the health scoring module are as follows: Real data feature modeling: Analyze the features of existing multi-source data and extract common distribution patterns; A variational autoencoder is used to model the latent space of sparse features and generate low-dimensional continuous vector representations; Virtual patient data generation: Design a conditional generative adversarial network, using the patient's basic attributes as conditional input to generate a complete portrait of the virtual patient: Structured data generation: simulate medical records and living habits; Unstructured data generation: synthesizing social media behavior labels; Data credibility verification and screening: Build a discriminator network to distinguish the distribution of real data and synthetic data, and only retain virtual data that passes the KS test; Introducing a rule base of dental experts to filter out synthetic samples with logical conflicts; Data fusion: After mixing virtual data with real data, differential privacy technology is used to add noise to the fused data set to ensure individual non-traceability; Under the federated learning framework, cross-institutional data joint training is achieved through encrypted parameter exchange; Dynamic feedback and model iteration: Monitor the performance of the health scoring model on virtual data, trigger retraining of the data generator, and modify the generation rules based on actual clinical feedback.

[0010] Furthermore, the specific operation steps for training the dynamic weight distribution model in the health scoring module are as follows: Basic framework selection, including: input layer receiving multimodal data feature vectors; The weight distribution layer introduces a dynamic weight gating unit into the Transformer's attention mechanism, with each feature dimension corresponding to a learnable weight coefficient; The output layer generates a comprehensive health score and weight distribution report; Data types include: clinical data: diagnosis and treatment records, imaging reports, laboratory indicators; behavioral tags: living habits, social activity, consumption preferences; environmental data: water quality and air quality in the place of residence; The initialization rule sets the initial weights based on medical prior knowledge and allows the model to be dynamically adjusted during training; Attention gating mechanism: In each layer of the Transformer, an attention score is calculated for each feature dimension. The calculation logic is: ,in is the query vector, are the key vector and value vector of the i-th feature, is the dimension scaling factor; weight gating parameters are optimized through back propagation so that the model automatically learns the importance of features; Introducing an external medical rule library to adjust the corresponding feature weights when preset conditions are triggered; The loss function design includes: the main loss is the health score prediction error, using mean square error, MSE; The auxiliary loss is the sparsity constraint of the weight distribution, i.e. L1 regularization; The total loss is ,in is a hyperparameter, is the weight parameter in the model; During the pre-training phase, the GPT-4 base model was adjusted on a public oral health dataset to learn general feature representations. The training steps are as follows: Weight distribution network training: freeze the basic model parameters and only train the weight gating units to optimize dynamic distribution capabilities; Joint fine-tuning: Unfreeze some underlying parameters and jointly optimize the feature extraction and weight distribution modules; Validation indicators: Accuracy of health score: compared with the results of independent evaluation by doctors; Weight interpretability: Analyze the rationality of weights through SHAP values; The weight assignment model is updated in real time based on new patient data, allowing doctors to manually correct abnormal weights.

[0011] Furthermore, the specific operation steps of the patient management and marketing module are as follows: Patient behavior tracking and clustering: Monitor post-diagnosis behavior, including follow-up appointments and health course click-through rates, and classify patients into categories using clustering algorithms. The process is as follows: Data sources include: review appointment records and treatment completion status obtained from HIS and electronic medical records; APP or mini-program behavior logs, including health course click-through rate, article reading time, and online consultation frequency; Define key behavioral indicators, including: Compliance indicators: review completion rate on time, completion rate of health tasks; Loss risk indicators: No APP login for 30 consecutive days, and number of appointment cancellations > 2 times; The acquired data were preprocessed and feature extracted, and a density clustering and K-means hybrid model were used through clustering algorithms. Density clustering was used to identify abnormal patients; K-means was used to classify the patient group, and the number of clusters was determined by the elbow rule. Feature input: After standardization, the feature matrix is input, including compliance score, churn risk score, and consumption preference; the number of K-means clusters is optimized using the silhouette coefficient;

[0012] Clustering result analysis includes: High compliance: review completion rate > 90% and health course click rate > 70%; Potential churn type: No login days > 15 days and number of appointment cancellations ≥ 1; Wait-and-see type: low-frequency interaction but not reaching the churn threshold; Abnormal: Complaints or fake orders identified by the density clustering algorithm; and collaborate with dentists to define clinically significant categories; Dynamic Path Generation: Using the Q-learning algorithm, we set a reward function based on patient categories to generate the optimal follow-up path. We also introduce a multi-objective optimization framework into the Q-learning algorithm, while balancing the three core indicators of return visit rate, patient satisfaction, and intervention cost. Marketing content matching: Build a marketing knowledge graph, associate patient tags with product libraries, and generate personalized recommendation scripts; Effect feedback and model iteration: record the patient conversion rate after strategy execution, optimize the decision model through A / B testing, and update the path rule library.

[0013] Furthermore, the specific steps for introducing a multi-objective optimization framework into the Q-learning algorithm in the patient management and marketing module are as follows: The definition of reward items includes: Follow-up reward (R1): patients who complete follow-up visits receive +10 points, and those who do not complete follow-up visits receive -20 points (basic target); patient satisfaction reward (R2): based on the satisfaction survey results after follow-up (1-5 points), the reward is converted proportionally (e.g., 5 points → +5, 1 point → -5); intervention cost penalty (R3): points are deducted proportionally based on the cost of implementing the strategy (e.g., the labor cost of telephone follow-up visits, the amount of coupons); Multi-objective fusion calculation logic: ,in It is a dynamic weight coefficient, and the initial value is set according to the business priority; They are follow-up reward, patient satisfaction reward and intervention cost penalty; is the comprehensive reward value; Improve the multi-objective Q-learning algorithm and increase the state dimension, including: Satisfaction history: average satisfaction scores of the last three follow-up visits; Cost accumulation value: intervention cost consumed in the current cycle; Patient value label: categorized by spending power; Pareto optimal action screening: For each state, calculate all actions value; eliminate dominated actions; select the action with the highest comprehensive score from the Pareto frontier; update online through the policy gradient method according to real-time business goals ; Data collection includes satisfaction measurement and cost accounting standardization. Satisfaction measurement is achieved by automatically sending a simplified survey after follow-up. For patients who do not respond, NLP is used to analyze the emotional tendencies in communication records. Cost accounting standardization is achieved by establishing an intervention action cost database. Telephone follow-up visits are charged 5 yuan per visit, SMS messages are charged 0.1 yuan per visit, and coupons are charged 50% of the face value. Cost data is updated in real time through real-time connection with the financial system. Use historical data to train a basic Q-learning model, learn a single-objective strategy, freeze some network layers, and gradually introduce R2 and R3 rewards for adjustment. After deployment, continue to collect new data and update model parameters every 24 hours.

[0014] The policy matching logic includes: High-value patients: prioritize actions with the highest cost and highest satisfaction; Patients at risk of churn: Use the combination with the lowest cost and highest reach rate; Satisfaction-sensitive patients: Insert satisfaction survey actions and adjust subsequent strategies in real time based on the results.

[0015] Furthermore, the specific operation steps of the diagnosis auxiliary module are as follows: Real-time data synchronization: Synchronize patient health scores and historical medication records to the doctor's terminal during the diagnosis and treatment process; Conflict detection and warning: If the current prescription conflicts with the patient's allergy history, a pop-up alert will be automatically triggered and an alternative will be recommended; Communication script generation: Based on the patient's personality tag, a soothing script template is generated. By analyzing the patient's historical medical records, communication preferences, and feedback data, a personalized script knowledge base is constructed. The process is as follows: Access patient interaction data, including: Historical communication records: diagnosis and treatment conversation texts, satisfaction survey results; Behavioral preference data: patient response types to push messages and communication channel choices; Feedback label: patient-initiated labeling preferences; Use NLP technology to extract key interaction features and build a labeling system: communication style, information density preference, and emotional response pattern; To build a personalized speech knowledge base, we first layer the knowledge base. The general speech layer is the basic medical explanation template. The patient portrait layer generates patient-specific labels based on historical data; The dynamic adaptation layer is the real-time session context; Reconstruct the knowledge graph, associate patient tags with speech templates, and establish a feedback chain for speech effects; A summary of patient communication preferences is displayed in the sidebar of the doctor's terminal, and a one-click optimization function is provided: after the doctor enters the original dialogue, it is automatically replaced with the patient-adapted version. After the diagnosis and treatment, the manual modifications of the doctor's dialogue are automatically recorded for optimizing the knowledge base.

[0016] Compared with the prior art, the present invention has the following beneficial effects: (1) The present invention integrates multi-source data, including medical records, living habits, social media tags, etc., through a health scoring module, uses a dynamic weight distribution model to generate personalized oral health scores, and combines them with 3D visualization to display high-risk items; the multi-dimensional data integration and dynamic weight adjustment method make oral health assessment more comprehensive and accurate; doctors can formulate more personalized diagnosis and treatment plans based on detailed health scores and visualization information, thereby improving treatment effects and patient satisfaction; at the same time, the generative adversarial network enhances virtual patient portraits, further enriching the data sources and providing more powerful support for clinical decision-making; (2) In the present invention, the patient management and marketing module tracks patient behavior and performs cluster classification, uses multi-objective reinforcement learning to generate dynamic follow-up paths, and matches precise marketing strategies through knowledge graphs; intelligent patient management methods can effectively reduce patient churn rates, improve patient return visit rates and treatment compliance; through dynamic path generation and strategy matching, medical institutions can develop personalized follow-up and marketing plans for different types of patients, improve resource utilization efficiency, and reduce operating costs; (3) In this invention, the in-diagnosis auxiliary module can synchronize patient health data to the doctor's terminal in real time, detect prescription conflicts and issue warnings, and generate personalized communication language by combining the patient's historical preferences with real-time emotional analysis. This not only improves the efficiency and quality of doctor-patient communication, but also reduces misunderstandings and disputes caused by information asymmetry. Doctors can adjust communication methods and treatment plans based on the patient's real-time emotions and preferences, thereby enhancing the patient's sense of trust and cooperation. In addition, the real-time data synchronization and conflict detection functions can effectively avoid medical errors, improve the safety and reliability of diagnosis and treatment, and further enhance the overall level of medical services. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] In order to facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings; Figure 1 This is the overall system block diagram of the present invention. DETAILED DESCRIPTION

[0018] The following will clearly and completely describe the technical solutions of the present invention in conjunction with the embodiments. Obviously, the embodiments described are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making any creative efforts are within the scope of protection of the present invention.

[0019] It should be understood that the terms “include” and “comprising” used in the specification and claims of the present disclosure indicate the presence of described features, integers, steps, operations, elements and / or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0020] It should also be understood that the terminology used in this disclosure is for the purpose of describing specific embodiments only and is not intended to limit the disclosure. As used in this disclosure and the claims, the singular forms "a," "an," and "the" are intended to include the plural forms unless the context clearly indicates otherwise. It should be further understood that the term "and / or" as used in this disclosure and the claims refers to any and all possible combinations of one or more of the associated listed items, including and including these combinations.

[0021] like Figure 1 As shown in the figure, the oral diagnosis and treatment patient service platform based on reinforcement learning includes a health scoring module, a patient management and marketing module, and an in-diagnosis assistance module; The health scoring module integrates multi-source data (medical records, lifestyle habits, and social media tags) to generate personalized oral health scores through a dynamic weight distribution model (based on a large model). It also combines 3D visualization to display high-risk items, enhancing comprehensiveness of assessments and supporting clinical decision-making. Multi-source data collection: Obtain medical records from the hospital's HIS system, collect lifestyle data through questionnaires / wearable devices, and combine social media analysis to extract personality traits (such as consumption preferences and social activity). Generative adversarial networks are then used to enhance the virtual patient portrait. The process is as follows: Real-world data feature modeling: Analyze existing multi-source data (HIS records, questionnaires, and social media tags) to extract common distribution patterns (e.g., the relationship between age and dental caries rates, and the association between personality traits and treatment compliance). Use a variational autoencoder (VAE) to model sparse features (e.g., social activity) in the latent space, generating a low-dimensional continuous vector representation. Virtual patient data generation: A conditional generative adversarial network (cGAN) is designed to generate a complete portrait of the virtual patient based on basic patient attributes (such as gender and age). Structured data generation: simulated medical records (such as "number of fillings in the past three years = 2") and lifestyle habits (such as "average number of daily tooth brushing times = 1.5"); unstructured data generation: synthesized social media behavior tags (such as "high frequency of browsing beauty content → strong demand for whitening").

[0022] Data credibility verification and screening: Construct a discriminator network to distinguish the distribution of real data and synthetic data, and only retain virtual data that passes the KS test (p>0.05); introduce a rule library of oral medicine experts (such as "there is no dental implant record for patients under 18 years old") to filter synthetic samples with logical conflicts; privacy-preserving data fusion: After mixing virtual data with real data, use differential privacy technology to add noise to the fused data set to ensure individual non-traceability; under the federated learning framework, realize cross-institutional data joint training through encrypted parameter exchange to avoid leakage of original data; dynamic feedback and model iteration: monitor the performance of the health scoring model on virtual data (such as the AUC value drops by >5%), triggering retraining of the data generator; based on actual clinical feedback, modify the generation rules (such as increasing the "weight of the smoking rate of periodontitis patients").

[0023] Data desensitization and feature extraction: Encrypt sensitive information (such as name and ID card), use NLP technology to extract key features of unstructured text (such as chief complaint description), and generate structured vectors; Dynamic weight allocation model training: Based on the large model, an adaptive weight allocation network is designed to dynamically adjust the scoring weights based on the data type (e.g., clinical data accounts for 60% and behavioral labels account for 30%). The specific process is as follows: Basic framework selection: Input layer: Receives multimodal data feature vectors (such as clinical data, behavioral labels, and structured features of environmental data); Weight distribution layer: Introduces dynamic weight gating units into the Transformer's attention mechanism, with each feature dimension corresponding to a learnable weight coefficient; Output layer: Generates a comprehensive health score and weight distribution report (e.g., clinical data weight = 60%, behavioral label weight = 30%);

[0024] Data type classification includes: Clinical data (structured): diagnosis and treatment records, imaging reports, laboratory indicators (initial weight = 60%); behavioral labels (semi-structured): living habits, social activity, consumption preferences (initial weight = 30%); environmental data (unstructured): water quality and air quality in the residential area (initial weight = 10%); Initialization rules: Set initial weights based on medical prior knowledge (e.g., clinical data has a greater impact on diagnosis), and allow the model to dynamically adjust during training. Attention gating mechanism: In each layer of the Transformer, an attention score is calculated for each feature dimension. The calculation logic is: ,in is the query vector, are the key vector and value vector of the i-th feature, The model automatically learns feature importance (e.g., the contribution of behavioral labels to follow-up predictions) by optimizing weight gating parameters through backpropagation. External medical rule bases are introduced (e.g., "Periodontitis patients need to focus on monitoring smoking behavior"), and when specific conditions are triggered, the weight of relevant features is forcibly adjusted (e.g., the weight of the smoking label is increased from 10% to 25%). Loss function design: Main loss: health score prediction error (mean square error, MSE); Auxiliary loss: sparsity constraint of weight distribution (L1 regularization) to prevent weight from being overly concentrated on a single feature; Total loss: ,in is a hyperparameter, is the weight parameter in the model; Pre-training phase: Fine-tune the GPT-4 base model on a public oral health dataset to learn common feature representations. The training steps are as follows:

[0025] Weight allocation network training: freeze the basic model parameters, only train the weight gating unit, and optimize the dynamic allocation capability; joint fine-tuning: unfreeze some underlying parameters, and jointly optimize the feature extraction and weight allocation modules; verification indicators: health score accuracy: compared with the results of independent doctor evaluation (Pearson correlation coefficient>0.85); weight interpretability: analyze the rationality of weights through SHAP value (such as the contribution of clinical data to caries score>50%); update the weight allocation model in real time according to new patient data, allowing doctors to manually correct abnormal weights.

[0026] Health score generation and visualization: Outputs an oral health index of 0-100 points, marks high-risk items (such as caries probability >80%), and intuitively displays the location of lesions through a 3D oral model.

[0027] The patient management and marketing module tracks and clusters patient behavior (reviews, click-through rates), uses multi-objective reinforcement learning (balancing return visit rates, satisfaction, and costs) to generate dynamic follow-up paths, matches precise marketing strategies with knowledge graphs, and continuously optimizes strategies to reduce churn and improve conversion efficiency. Patient behavior tracking and clustering: Monitor post-diagnosis behavior (such as follow-up appointments and health course click-through rates) and use clustering algorithms to classify patients into categories such as high adherence and potential churn. The specific process is as follows: Data sources include: review appointment records and treatment completion status obtained from HIS and electronic medical records (EMR); app / mini program behavior logs (such as health course click-through rate, article reading time, and online consultation frequency); definition of key behavioral indicators, including: compliance indicators: review completion rate on time, health task completion rate (such as daily tooth brushing check-in); churn risk indicators: not logging into the app for 30 consecutive days, and number of appointment cancellations >2; The acquired data was preprocessed and feature extracted. A density-based clustering (DBSCAN) + K-means hybrid model was used for clustering. DBSCAN identified abnormal patients (such as those who were extremely active or completely silent) and avoided noise interference. K-means categorized the mainstream patient population (the number of clusters was determined using the elbow rule). Feature input was standardized and input into a feature matrix, including compliance scores, churn risk scores, and consumer preferences (such as click-through rates for whitening products). The silhouette score was used to optimize the number of K-means clusters (for example, K=4 yielded the highest score). Clustering result analysis includes: high compliance type: review completion rate > 90% + health course click rate > 70% (accounting for 20%); potential churn type: no login days > 15 days + number of cancellations ≥ 1 (accounting for 30%); Wait-and-see type: low-frequency interactions but not reaching the churn threshold (accounting for 40%); abnormal type: special behaviors identified by DBSCAN (such as frequent complaints or fake orders); and clinical significance categories are defined in collaboration with dentists.

[0028] Dynamic path generation: Using the Q-learning algorithm, we set a reward function based on patient categories to generate the optimal follow-up path. We also introduced a multi-objective optimization framework into the Q-learning algorithm, while balancing the three core indicators of return visit rate, patient satisfaction, and intervention cost. The specific steps are as follows: The definition of reward items includes: Follow-up reward (R1): patients who complete follow-up visits receive +10 points, and those who do not complete follow-up visits receive -20 points (basic target); patient satisfaction reward (R2): based on the satisfaction survey results after follow-up (1-5 points), the reward is converted proportionally (e.g., 5 points → +5, 1 point → -5); intervention cost penalty (R3): points are deducted proportionally based on the cost of implementing the strategy (e.g., the labor cost of telephone follow-up visits, the amount of coupons); Multi-objective fusion calculation logic: ,in It is a dynamic weight coefficient, and the initial value is set according to the business priority; They are follow-up reward, patient satisfaction reward and intervention cost penalty; is the comprehensive reward value; the multi-objective Q-learning algorithm is improved and the state dimension is increased, including: Satisfaction history: average satisfaction scores from the last three follow-up visits; cumulative cost: intervention costs consumed in the current cycle; patient value label: patient classification based on spending power (e.g., high-value patients receive increased β weights); Pareto optimal action screening: For each state, calculate all possible actions value; eliminate dominated actions (e.g., the R1, R2, and R3 of an action are all lower than those of another action); select the action with the highest comprehensive score (e.g., the maximum weighted sum) from the Pareto frontier; and update the policy gradient method online based on real-time business goals (e.g., the repeat diagnosis rate during the promotion period). ; Data collection includes satisfaction measurement and cost accounting standardization. Satisfaction measurement is achieved by automatically sending a simplified survey after follow-up. For patients who do not respond, NLP is used to analyze the emotional tendencies in communication records. Cost accounting standardization is achieved by establishing an intervention action cost database, with telephone follow-up = 5 yuan / time, SMS = 0.1 yuan / time, and coupons = face value × 50%. Cost data is updated in real time through the financial system. Use historical data to train a basic Q-learning model, learn a single-objective (return visit rate) strategy, freeze some network layers, and gradually introduce R2 and R3 rewards for fine-tuning. After deployment, continuously collect new data and update model parameters every 24 hours. The strategy matching logic includes: High-value patients: give priority to high-cost, high-satisfaction actions (such as "one-on-one video follow-up with a doctor + customized care plan"); patients at risk of churn: adopt a low-cost, high-reach rate combination (such as "SMS reminders + limited-time coupons"); satisfaction-sensitive patients: mandatory insertion of satisfaction surveys and real-time adjustment of subsequent strategies based on the results.

[0029] Marketing content matching: Build a marketing knowledge graph, associate patient tags with product libraries, and generate personalized recommendation scripts; effect feedback and model iteration: record the patient conversion rate after strategy execution, optimize the decision model through A / B testing, and update the path rule library.

[0030] The in-diagnosis assistance module synchronizes patient health data to the doctor's terminal in real time, detects prescription conflicts and issues early warnings, and generates personalized communication scripts based on patient historical preferences and real-time sentiment analysis, improving doctor-patient communication efficiency and treatment compliance. Real-time data synchronization: Synchronizes patient health scores and historical medication records to the doctor's terminal during the diagnosis and treatment process; Conflict detection and warning: If the current prescription conflicts with the patient's allergy history (such as cephalosporins), a pop-up alert will automatically trigger and an alternative plan will be recommended; Communication script generation: Based on the patient's personality tag, a soothing script template is generated. By analyzing the patient's historical medical records, communication preferences, and feedback data, a personalized script knowledge base is constructed. The process is as follows: Obtain patient interaction data, including: historical communication records: text of diagnosis and treatment conversations (e.g., doctor's notes, patient questions), satisfaction survey results; behavioral preference data: patient response types to push messages (e.g., preference for graphic explanations > video), communication channel selection (e.g., preference for phone calls > app messages); feedback tags: patient-initiated preferences (e.g., "hope to reduce the use of technical terms" and "prefer step-by-step explanations"); Use NLP technology to extract key interaction features (such as "pain level" and "treatment duration" that frequently appear in patients' questions) and build a labeling system: communication style (direct / euphemistic), information density preference (concise / detailed), and emotional response mode (needs comfort / needs rational explanation); To build a personalized conversational knowledge base, we first layer the knowledge base into three layers: the general conversational layer: basic medical explanation templates (e.g., dental caries treatment steps); the patient profile layer: generating patient-specific tags based on historical data (e.g., "prefers step-by-step explanations - dislikes technical terminology - high anxiety tendency"); and the dynamic adaptation layer: real-time conversation context (e.g., current treatment stage, explained content). Reconstruct the knowledge graph and associate patient tags with speech templates (e.g., "high anxiety tendency" is associated with a library of soothing speech templates); establish a speech effect feedback chain (e.g., if the historical adoption rate of a speech template is greater than 90%, the weight is increased);

[0031] The personalized speech generation logic includes: Context awareness: Real-time analysis of the current diagnosis and treatment stage (such as preoperative preparation and intraoperative operation) and matching of stage-specific speech templates; Tag-driven generation: If the patient's tag is "dislike professional terms", "enamel demineralization" will be automatically converted to "damage to the protective layer of the tooth surface"; For patients who prefer step-by-step explanations, insert progress prompts (e.g., "Next is step 2: clean the carious area, which will take about 2 minutes"); Optimization of historical feedback: If a patient has previously rated a certain type of speech technique low (e.g., “the explanation is too brief”), a more detailed version of the template will be automatically called; A summary of patient communication preferences is displayed in the sidebar of the doctor's terminal, and a one-click optimization function is provided: after the doctor enters the original dialogue, it is automatically replaced with the patient-adapted version. After the diagnosis and treatment, the manual modifications of the doctor's dialogue are automatically recorded for optimizing the knowledge base.

[0032] The preferred embodiments of the present invention disclosed above are intended only to help illustrate the present invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the present invention to specific embodiments. Obviously, many modifications and variations are possible based on the contents of this specification. These embodiments are selected and described in detail in this specification to better explain the principles and practical applications of the present invention, thereby enabling those skilled in the art to better understand and utilize the present invention. The present invention is limited only by the claims and their full scope and equivalents.

Claims

1. A dental diagnosis and treatment patient service platform based on reinforcement learning, characterized by: include: The health scoring module integrates multi-source data, generates personalized oral health scores through a dynamic weight distribution model, and displays high-risk items in 3D visualization. The patient management and marketing module is used to track and cluster patient behavior, generate dynamic follow-up paths using multi-objective reinforcement learning, and match precise marketing strategies through knowledge graphs; The in-diagnosis assistance module is used to synchronize patient health data to the doctor's terminal in real time, detect prescription conflicts and issue warnings, and generate personalized communication language based on patient historical preferences and real-time sentiment analysis.

2. The oral diagnosis and treatment patient service platform based on reinforcement learning according to claim 1 is characterized in that: The specific execution process of the health scoring module is as follows: Multi-source data collection: Obtain medical records from the hospital HIS system, collect lifestyle data through questionnaires or wearable devices, extract personality tags through social media analysis, and use generative adversarial networks to enhance virtual patient portraits; Data desensitization and feature extraction: Encrypt sensitive information, use NLP technology to extract features of unstructured text, and generate structured vectors; Dynamic weight distribution model training: Design an adaptive weight distribution network based on the large model and dynamically adjust the scoring weight according to the data type; Health score generation and visualization: Outputs an oral health index of 0-100 points, marks high-risk items, and intuitively displays the location of lesions through a 3D oral model.

3. The oral diagnosis and treatment patient service platform based on reinforcement learning according to claim 1 is characterized in that: The specific steps of using the generative adversarial network to enhance the virtual patient portrait in the health scoring module are as follows: Real data feature modeling: Analyze the features of existing multi-source data and extract common distribution patterns; A variational autoencoder is used to model the latent space of sparse features and generate low-dimensional continuous vector representations; Virtual patient data generation: Design a conditional generative adversarial network, using the patient's basic attributes as conditional input to generate a complete portrait of the virtual patient: Structured data generation: simulate medical records and living habits; Unstructured data generation: synthesizing social media behavior labels; Data credibility verification and screening: Build a discriminator network to distinguish the distribution of real data and synthetic data, and only retain virtual data that passes the KS test; Introducing a rule base of dental experts to filter out synthetic samples with logical conflicts; Data fusion: After mixing virtual data with real data, differential privacy technology is used to add noise to the fused data set to ensure individual non-traceability; Under the federated learning framework, cross-institutional data joint training is achieved through encrypted parameter exchange; Dynamic feedback and model iteration: Monitor the performance of the health scoring model on virtual data, trigger retraining of the data generator, and modify the generation rules based on actual clinical feedback.

4. The oral diagnosis and treatment patient service platform based on reinforcement learning according to claim 2 is characterized in that: The specific steps for training the dynamic weight distribution model in the health scoring module are as follows: Basic framework selection, including: input layer receiving multimodal data feature vectors; The weight distribution layer introduces a dynamic weight gating unit into the Transformer's attention mechanism, with each feature dimension corresponding to a learnable weight coefficient; The output layer generates a comprehensive health score and weight distribution report; Data types include: clinical data: diagnosis and treatment records, imaging reports, laboratory indicators; behavioral tags: living habits, social activity, consumption preferences; environmental data: water quality and air quality in the place of residence; The initialization rule sets the initial weights based on medical prior knowledge and allows the model to be dynamically adjusted during training; Attention gating mechanism: In each layer of the Transformer, an attention score is calculated for each feature dimension. The calculation logic is: ,in is the query vector, are the key vector and value vector of the i-th feature, is the dimension scaling factor; weight gating parameters are optimized through back propagation so that the model automatically learns the importance of features; Introducing an external medical rule library to adjust the corresponding feature weights when the preset conditions are triggered; The loss function design includes: the main loss is the health score prediction error, using mean square error, MSE; The auxiliary loss is the sparsity constraint of the weight distribution, i.e. L1 regularization; The total loss is ,in is a hyperparameter, is the weight parameter in the model; During the pre-training phase, the GPT-4 base model was adjusted on a public oral health dataset to learn general feature representations. The training steps are as follows: Weight distribution network training: freeze the basic model parameters and only train the weight gating units to optimize dynamic distribution capabilities; Joint fine-tuning: unfreeze the underlying parameters and jointly optimize the feature extraction and weight distribution modules; Validation indicators: Accuracy of health score: compared with the results of independent evaluation by doctors; Weight interpretability: Analyze the rationality of weights through SHAP values; The weight assignment model is updated in real time based on new patient data, allowing doctors to manually correct abnormal weights.

5. The oral diagnosis and treatment patient service platform based on reinforcement learning according to claim 1 is characterized in that: The specific operation steps of the patient management and marketing module are as follows: Patient behavior tracking and clustering: Monitor post-diagnosis behavior, including follow-up appointments and health course click-through rates, and classify patients into categories using clustering algorithms. The process is as follows: Data sources include: review appointment records and treatment completion status obtained from HIS and electronic medical records; APP or mini-program behavior logs, including health course click-through rate, article reading time, and online consultation frequency; Define key behavioral indicators, including: Compliance indicators: review completion rate on time, completion rate of health tasks; Loss risk indicators: No APP login for 30 consecutive days, and number of appointment cancellations > 2 times; The acquired data were preprocessed and feature extracted, and a density clustering and K-means hybrid model were used through clustering algorithms. Density clustering was used to identify abnormal patients; K-means was used to classify the patient group, and the number of clusters was determined by the elbow rule. Feature input: After standardization, the feature matrix is input, including compliance score, churn risk score, and consumption preference; the number of K-means clusters is optimized using the silhouette coefficient; Clustering result analysis includes: High compliance: review completion rate > 90% and health course click rate > 70%; Potential churn type: No login days > 15 days and number of appointment cancellations ≥ 1; Wait-and-see type: low-frequency interaction but not reaching the churn threshold; Abnormal: Complaints or fake orders identified by the density clustering algorithm; and collaborate with dentists to define clinically significant categories; Dynamic Path Generation: Using the Q-learning algorithm, we set a reward function based on patient categories to generate the optimal follow-up path. We also introduce a multi-objective optimization framework into the Q-learning algorithm, while balancing the three core indicators of return visit rate, patient satisfaction, and intervention cost. Marketing content matching: Build a marketing knowledge graph, associate patient tags with product libraries, and generate personalized recommendation scripts; Effect feedback and model iteration: record the patient conversion rate after strategy execution, optimize the decision model through A / B testing, and update the path rule library.

6. The oral diagnosis and treatment patient service platform based on reinforcement learning according to claim 5 is characterized in that: The specific steps for introducing the multi-objective optimization framework into the Q-learning algorithm in the patient management and marketing module are as follows: The definition of reward items includes: Follow-up reward (R1): patients who complete follow-up visits receive +10 points, and those who do not complete follow-up visits receive -20 points (basic target); patient satisfaction reward (R2): based on the satisfaction survey results after follow-up (1-5 points), the reward is converted proportionally (e.g., 5 points → +5, 1 point → -5); intervention cost penalty (R3): points are deducted proportionally based on the cost of implementing the strategy (e.g., the labor cost of telephone follow-up visits, the amount of coupons); Multi-objective fusion calculation logic: ,in It is a dynamic weight coefficient, and the initial value is set according to the business priority; They are follow-up reward, patient satisfaction reward and intervention cost penalty; is the comprehensive reward value; Improve the multi-objective Q-learning algorithm and increase the state dimension, including: Satisfaction history: average satisfaction scores of the last three follow-up visits; Cost accumulation value: the intervention cost consumed in the current cycle; Patient value label: categorized by spending power; Pareto optimal action screening: For each state, calculate all actions value; eliminate dominated actions; select the action with the highest comprehensive score from the Pareto frontier; update online through the policy gradient method according to real-time business goals ; Data collection includes satisfaction measurement and cost accounting standardization. Satisfaction measurement is achieved by automatically sending a simplified survey after follow-up. For patients who do not respond, NLP is used to analyze the emotional tendencies in communication records. Cost accounting standardization is achieved by establishing an intervention action cost database. Telephone follow-up visits are charged 5 yuan per visit, SMS messages are charged 0.1 yuan per visit, and coupons are charged 50% of the face value. Cost data is updated in real time through real-time connection with the financial system. Use historical data to train a basic Q-learning model, learn a single-objective strategy, freeze some network layers, and gradually introduce R2 and R3 rewards for adjustment. After deployment, continue to collect new data and update model parameters every 24 hours. The policy matching logic includes: High-value patients: prioritize actions with the highest cost and highest satisfaction; Patients at risk of churn: Use the combination of lowest cost and highest reach rate; Satisfaction-sensitive patients: Insert satisfaction survey actions and adjust subsequent strategies in real time based on the results.

7. The oral diagnosis and treatment patient service platform based on reinforcement learning according to claim 1 is characterized in that: The specific operation steps of the diagnosis auxiliary module are as follows: Real-time data synchronization: Synchronize patient health scores and historical medication records to the doctor's terminal during the diagnosis and treatment process; Conflict detection and warning: If the current prescription conflicts with the patient's allergy history, a pop-up alert will be automatically triggered and an alternative will be recommended; Communication script generation: Based on the patient's personality tag, a soothing script template is generated. By analyzing the patient's historical medical records, communication preferences, and feedback data, a personalized script knowledge base is constructed. The process is as follows: Access patient interaction data, including: Historical communication records: diagnosis and treatment conversation texts, satisfaction survey results; Behavioral preference data: patient response types to push messages and communication channel choices; Feedback label: patient-initiated labeling preferences; Use NLP technology to extract key interaction features and build a labeling system: communication style, information density preference, and emotional response pattern; To build a personalized speech knowledge base, we first layer the knowledge base. The general speech layer is the basic medical explanation template. The patient portrait layer generates patient-specific labels based on historical data; The dynamic adaptation layer is the real-time conversation context; Reconstruct the knowledge graph, associate patient tags with speech templates, and establish a feedback chain for speech effects; A summary of patient communication preferences is displayed in the sidebar of the doctor's terminal, and a one-click optimization function is provided: after the doctor enters the original dialogue, it is automatically replaced with the patient-adapted version. After the diagnosis and treatment, the manual modifications of the doctor's dialogue are automatically recorded for optimizing the knowledge base.

Images