Automatic classification and management system for postoperative medical orders and follow-up of oral patients

Abstract

The application discloses a postoperative medical order and follow-up automatic classification and management system, relates to the technical field of medical management, and comprises a diagnosis and treatment semantic residual extraction engine, which is used for acquiring multi-source diagnosis and treatment data of a patient from an electronic medical record system of a stomatology department; a multidimensional treatment classification engine, which is used for being based on diagnosis and treatment semantic residual feature vectors; a decision shunting module, which is used for judging whether a preset automatic processing condition is met according to a treatment stage confidence gradient value; an individualized medical order intelligent matching and generation engine, which is used for searching a matched medical order component from a pre-constructed medical order component library according to the classification result; an intelligent follow-up plan arrangement engine, which is used for being based on the classification result and an embedded oral treatment standard path atlas; and a multi-channel intelligent pushing engine, which is used for sending the individualized postoperative medical order content and follow-up reminding to the patient through a selected pushing channel, so that the patient can conveniently understand and execute, the number of times that the patient comes to the hospital is reduced, and the convenience of reexamination is improved.

Description

Technical Field

[0001] This invention relates to the field of medical management system technology, specifically to an automatic classification and management system for postoperative medical orders and follow-up of oral patients. Background Technology

[0002] With increasing public awareness of oral health, the demand for oral healthcare services is growing rapidly. Compared to other medical specialties, the treatment process in dentistry (such as implantology, orthodontics, endodontics, and periodontology) has significant unique characteristics: treatment is often not completed in a single visit, but rather relies heavily on a multi-stage, multi-frequency sequence of treatments (for example, root canal treatment requires multiple stages such as access cavity preparation and filling; implant treatment involves a longer period including primary surgery, secondary surgery, and restoration). Therefore, the patient's postoperative care after each visit and adherence to scheduled follow-up appointments directly affect the success of the overall treatment and its long-term outcomes.

[0003] However, in current oral clinical practice and hospital information management, the following prominent technical deficiencies exist regarding postoperative medical orders and follow-up management: While existing Hospital Information Systems (HIS) and Electronic Medical Records (EMR) systems store a large amount of medical records, images, laboratory data, and billing information, this information is often scattered across different subsystems, forming "data silos." This is particularly true in dental records, which contain a large amount of unstructured free text descriptions, with significant differences in writing habits among different doctors, complex tooth position annotation systems (such as the mixing of FDI, Palmer, and Chinese descriptions), and numerous synonymous medical terms. Traditional information extraction techniques based on simple keyword matching or billing coding rules cannot accurately capture the "residual medical semantics" embedded in the medical record text, resulting in the system's inability to truly and accurately reconstruct the patient's actual treatment status.

[0004] Dental treatments are not only diverse, but the precautions required for different stages of the same disease also vary significantly. Most existing follow-up management systems tend to treat each patient visit as an isolated event, lacking the ability to track the data longitudinally over time. They typically only trigger a uniform template based on a "surface-level charge code" or "primary diagnosis," failing to identify the exact stage of the patient's treatment sequence (e.g., whether it is in the "opening and drainage stage" or the "root canal filling stage"), nor can they comprehensively assess individual risk factors such as the patient's systemic diseases (e.g., diabetes), bad habits (e.g., smoking), and the complexity of the surgery. This can easily lead to mismatched or insufficiently targeted medical advice.

[0005] In clinical practice, patients often undergo multiple dental treatments of different areas and types at the same time (e.g., root canal treatment on the left side and implant surgery on the right). Existing systems typically generate prescriptions and follow-up appointment schedules by simply overlaying standardized templates for each treatment. This mechanical overlay not only generates a large amount of redundant information but also easily leads to logical conflicts in the prescriptions (e.g., post-implantation instructions state "no chewing on the affected side," while other routine treatment templates suggest "normal chewing is allowed"), leaving patients confused. Furthermore, the inability to intelligently combine follow-up appointment times forces patients to make frequent hospital visits, significantly reducing their experience and adherence.

[0006] Most existing automated medical recommendation systems adopt a "one-size-fits-all" processing logic, lacking the ability to quantitatively assess the confidence level of machine judgment results. When faced with rare cases, cases with ambiguous borderlines, or complex situations such as changes in treatment midway, if the system blindly executes automated recommendations, it is highly likely to cause medical errors and disputes. If all reviews are done manually, it not only consumes a lot of the energy of medical staff, causing nurses to be overwhelmed with tedious education and follow-up registration, but also defeats the original purpose of information management.

[0007] In summary, current technologies lack a comprehensive management solution that can non-invasively integrate fragmented data from multiple sources, accurately perceive the multi-dimensional characteristics of oral treatment stages, intelligently resolve complex conflicting medical orders, and achieve an optimal balance between automated processing and manual review. Therefore, there is an urgent need for a system that can automatically classify and personalize postoperative medical orders and follow-up management for oral patients, in order to effectively improve the quality of medical services, reduce the burden on medical staff, and enhance patient compliance. Summary of the Invention

[0008] The purpose of this invention is to provide an automatic classification and management system for postoperative medical orders and follow-up of oral patients, so as to solve the problems mentioned in the background art.

[0009] To achieve the above objectives, the present invention provides the following technical solution: an automatic classification and management system for postoperative medical orders and follow-up of oral patients, comprising: The diagnostic semantic residue extraction engine is used to obtain multi-source diagnostic data of patients from the electronic medical record system of dentistry. It performs semantic parsing on the unstructured text in the multi-source diagnostic data through a natural language processing pipeline dedicated to the dental field, extracts diagnostic semantic residue features, and encodes them into structured diagnostic semantic residue feature vectors. The natural language processing pipeline dedicated to the dental field includes a tooth position annotation standardization module, a dental terminology synonym disambiguation module, a treatment operation sequence extraction module, and a latent information reasoning module. A multi-dimensional treatment classification engine is used to perform multi-level classification processing based on the residual feature vector of the diagnosis and treatment semantics, using an evidence fusion hierarchical attention network. The output includes classification results including main treatment type, treatment stage, treatment complexity and patient individualized risk stratification, and calculates the treatment stage confidence gradient value corresponding to each level of classification results. The decision-making and triage module is used to determine whether the preset automatic processing conditions are met based on the confidence gradient value of the treatment stage. If the conditions are met, the classification result is transmitted to the personalized medical order intelligent matching and generation engine. If the conditions are not met, the classification result is marked as pending review and pushed to the manual review channel. A personalized medical order intelligent matching and generation engine is used to retrieve matching medical order components from a pre-built medical order component library based on the classification results, and generate personalized postoperative medical order content through a dynamic assembly mechanism of medical order components. The intelligent follow-up plan orchestration engine is used to generate a personalized follow-up plan containing multiple follow-up visit nodes based on the classification results and the built-in standard oral treatment path map. A multi-channel intelligent push engine is used to send the personalized postoperative medical orders and follow-up reminders to the patient through the selected push channels.

[0010] Furthermore, the evidence fusion hierarchical attention network includes: The evidence source weighted attention layer is used to evaluate the contribution of information from different data sources to the current classification task through learnable attention weights; The temporal decay attention layer is used to apply temporal decay weights to the historical medical records of the same patient, where the decay rate is dynamically adjusted according to the treatment type. The counterfactual reasoning verification layer is used to, after outputting the classification result, assume the classification result is another category and check whether there is evidence in the current feature vector that strongly contradicts the assumption. The confidence of the original classification result is adjusted according to the check result.

[0011] Furthermore, the confidence gradient value of the treatment stage is a binary tuple (C_t, ΔC / Δt), where C_t is the confidence of the system in classifying the patient's treatment stage at the current moment, and ΔC / Δt is the rate of change of confidence over time. The preset automatic processing conditions are: C_t is greater than or equal to the first threshold, and the absolute value of ΔC / Δt is less than or equal to the second threshold.

[0012] Furthermore, the dynamic assembly mechanism of the medical order component includes: Based on the treatment type and treatment stage in the classification results, retrieve all matching medical order components from the medical order component library; Based on the individualized risk stratification of patients in the classification results, the personalized adjustment rules within the medical order component are triggered to modify or supplement the component content. Perform conflict detection and dependency sorting between components, eliminate conflicting content, and establish a presentation order; Filter the component subsets that match the timeliness based on the current postoperative time window; The components are prioritized based on the principle of optimizing the utility density of medical order information. Render the assembled component set into patient-readable text.

[0013] Furthermore, it also includes: The parallel treatment conflict resolution module is used to scan the conflict relationship between all the medical order components to be assembled and select them according to preset rules when multiple parallel treatments are detected for the same patient at the same time. It merges the duplicate general guidance components in multiple treatments, generates zonal guidance content for parallel treatments involving different oral regions, and merges the follow-up timelines of multiple treatments into a unified follow-up calendar. The continuous learning module is used to collect the review results from the manual review channel as labeled data and to incrementally train the classification model in the multi-dimensional treatment classification engine.

[0014] This application also discloses a method for the automatic classification and management of postoperative medical orders and follow-up for oral patients, including the following steps: S1: Obtain multi-source diagnosis and treatment data of the patient from the electronic medical record system of the Department of Stomatology, and retrieve the patient's historical medical records; S2: Semantic parsing of unstructured text in the multi-source diagnostic and treatment data is performed through a dedicated natural language processing pipeline for the oral field, extracting diagnostic and treatment semantic residual features and encoding them into structured diagnostic and treatment semantic residual feature vectors; wherein the semantic parsing includes tooth position annotation standardization processing, oral terminology synonym disambiguation processing, treatment operation sequence extraction processing, and implicit information reasoning processing; S3: Based on the residual feature vector of the diagnosis and treatment semantics, a multi-level classification process is performed using an evidence fusion hierarchical attention network. The output includes classification results including main treatment type, treatment stage, treatment complexity and patient individualized risk stratification. The confidence gradient value of the treatment stage corresponding to each level classification result is calculated. S4: Determine whether the preset automatic processing conditions are met based on the confidence gradient value of the treatment stage. If the conditions are met, proceed to step S5. If the conditions are not met, mark the classification result as pending review and push it to the manual review channel. Proceed to step S5 after manual confirmation. S5: Based on the classification results, retrieve matching medical order components from the pre-built medical order component library, and generate personalized postoperative medical order content through the dynamic assembly mechanism of medical order components; S6: Based on the classification results and the built-in standard oral treatment pathway map, generate a personalized follow-up plan that includes multiple follow-up visit nodes; S7: Send the personalized postoperative medical instructions and first follow-up visit reminder to the patient through the selected push channel.

[0015] Furthermore, the multi-level classification process described in step S3 includes: The first level of primary treatment type classification is used to determine the types of primary treatment items; The second level of treatment stage classification is used to determine the current stage in the overall treatment process; The third level of treatment complexity classification is used to determine the level of treatment complexity. The fourth level of patient-specific risk stratification is used to identify individualized risk factors for each patient. The fifth layer of treatment progress deviation detection is used to detect the deviation between the actual treatment progress and the standard treatment path.

[0016] Furthermore, in step S4, the confidence gradient value for the treatment phase is a binary tuple (C_t, ΔC / Δt), and the judgment logic for the preset automatic processing conditions includes: When C_t is greater than or equal to the first threshold and the absolute value of ΔC / Δt is less than or equal to the second threshold, it is determined to be a high-confidence stable state, which meets the automatic processing conditions. When C_t is greater than or equal to the first threshold but ΔC / Δt is significantly negative, it is judged as a wavering state, automatic processing is suspended and manual review is triggered; When C_t is less than the first threshold, it is determined to be a low confidence state and is transferred to the manual review channel.

[0017] Furthermore, the dynamic assembly mechanism of the medical order component described in step S5 includes: Retrieve matching medical order components based on the treatment type and treatment stage in the classification results; Personalized adjustment rules within the component are triggered based on the patient's individualized risk stratification in the classification results; Perform conflict detection and dependency ordering between components; Components with suitable timeliness are selected based on the current postoperative time window; Prioritize the medical order information utility density based on the optimization principle; Render the assembled component set into patient-readable text.

[0018] Furthermore, it also includes step S8: Automatically send follow-up reminders to patients before each scheduled visit; After a patient's follow-up visit, the system automatically reads the new medical records and returns to step S2, updates the treatment stage classification, and re-triggers steps S5 and S6 to generate medical orders and update follow-up plans, thus forming a closed-loop management system. Collect the review results from the manual review channel as labeled data, and incrementally train the classification model in step S3.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention utilizes a diagnostic semantic residue extraction engine to acquire multi-source diagnostic and treatment data from a dental electronic medical record system, and performs semantic parsing of unstructured text using a dental-specific natural language processing pipeline. This pipeline comprises four core modules: a tooth position annotation standardization module, a dental terminology synonym disambiguation module, a treatment operation sequence extraction module, and a latent information inference module. It systematically extracts diagnostic and treatment semantic residue features scattered throughout the free text of electronic medical records. Compared to existing technologies that rely solely on diagnostic or billing codes for coarse-grained classification, this invention can identify multiple expressions of the same treatment, extract complete treatment operation sequences, and infer implicit clinical information, thereby achieving accurate identification of treatment type and stage.

[0020] 2. The multi-dimensional treatment classification engine of this invention employs an evidence-fusion hierarchical attention network, simultaneously performing five levels of classification processing, including primary treatment type classification, treatment stage classification, treatment complexity grading, patient-specific risk stratification, and treatment progress deviation detection. This multi-dimensional hierarchical classification approach can comprehensively depict the patient's treatment context, not only determining what treatment the patient has received, but also identifying the current stage of treatment, the complexity of the treatment, the patient's individualized risk factors, and whether there are any deviations in treatment progress. Compared to existing technologies that only perform single-dimensional classification, the multi-dimensional classification results provided by this invention can provide a more comprehensive and accurate decision-making basis for subsequent medical order generation and follow-up plan arrangement.

[0021] 3. This invention creatively proposes a treatment stage confidence gradient value as the basis for decision triage. This value is a binary tuple containing the current confidence level and the rate of change of confidence over time, which not only measures the accuracy of classification but also the system's ability to assess the reliability of its own judgment. The decision triage module determines whether preset automatic processing conditions are met based on the treatment stage confidence gradient value. Cases with high confidence and stable status enter the fully automatic processing flow, while cases with low confidence or fluctuating judgment are transferred to the manual review channel. This design enables the system to have metacognitive capabilities regarding the reliability of its own judgment, achieving full automation to improve efficiency in cases with high certainty, and proactively seeking human support to ensure safety in cases with high uncertainty, thus finding the optimal balance between efficiency and safety.

[0022] 4. This invention employs a dynamic assembly mechanism for medical order components to replace the traditional template matching method. Each component in the medical order component library is a structured object containing rich attributes, including a list of applicable treatment types, a list of applicable treatment stages, an applicable postoperative time window, priority, a list of conflicting components, a list of dependent components, and a set of personalized adjustment rules. The system retrieves matching medical order components based on classification results, triggers personalized adjustment rules to modify or supplement component content, performs conflict detection and dependency sorting to eliminate contradictory content, filters timeliness-matched components based on the postoperative time window, prioritizes components according to the principle of optimizing the utility density of medical order information, and finally renders them into patient-readable text. Compared to the problem of fixed templates in existing technologies failing to adapt to individual differences, this invention can generate highly personalized medical order content based on the patient's treatment type, stage, complexity, and risk factors, ensuring that the medical orders received by each patient match their actual situation.

[0023] 5. This invention includes a parallel treatment conflict resolution module. When multiple parallel treatments are detected for the same patient within the same time period, it automatically scans the conflict relationships between all components of the medical orders to be assembled and selects those that conflict according to preset rules. It merges duplicate common guidance components from multiple treatments, generates zonal guidance content for parallel treatments involving different oral regions, and combines the follow-up timelines of multiple treatments into a unified follow-up calendar. Compared to the problems of information inconsistencies and overload caused by separately pushing medical orders for multiple parallel treatments in existing technologies, this invention can generate integrated and consistent medical order content and follow-up plans, making it easier for patients to understand and execute, reducing the number of patient visits, and improving the convenience of follow-up visits. Attached Figure Description

[0024] Figure 1 This is a diagram illustrating the architecture of the automatic classification and management system for post-oral surgery medical observation and follow-up of the present invention. Figure 2 This is a confidence determination space diagram for the decision-making and diversion module of the present invention; Figure 3 This is a timeline comparison chart of the intelligent follow-up plan for implantation sequence therapy of the present invention; Figure 4 This is a flowchart illustrating the dynamic assembly and screening mechanism of the personalized medical order component of the present invention. Figure 5 This is a diagram showing the three-dimensional layered architecture and component deployment structure of the system of the present invention. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1-5 The automatic classification and management system for postoperative medical orders and follow-up of oral patients provided by this invention adopts a layered modular architecture design. The entire system is deployed between the existing hospital information system and the electronic medical record system. It achieves passive data reading through standardized data interfaces without making any intrusive modifications to the existing system.

[0027] This system mainly consists of six functional modules: a diagnostic semantic residue extraction engine, a multi-dimensional treatment classification engine, a decision triage module, a personalized medical order intelligent matching and generation engine, an intelligent follow-up plan orchestration engine, and a multi-channel intelligent push engine. In addition, it includes a parallel treatment conflict resolution module and a continuous learning module. The modules communicate asynchronously through message queues to ensure high availability and scalability of the system.

[0028] The overall data flow of the system is as follows: After a patient completes a treatment in the dental department, the hospital information system automatically generates a corresponding medical record. Upon detecting a new record, the treatment semantic residue extraction engine automatically acquires all relevant data from the patient's current visit and historical medical data. After processing by a dedicated natural language processing pipeline for the dental field, a structured treatment semantic residue feature vector is extracted. This feature vector is then passed to the multi-dimensional treatment classification engine, which uses an evidence fusion-based hierarchical attention network for multi-level classification processing, outputting the classification results and the corresponding treatment stage confidence gradient values. The decision-making and triage module determines whether the preset automatic processing conditions are met based on the treatment stage confidence gradient values. Cases that meet the conditions enter the fully automatic processing flow, while cases that do not meet the conditions are marked as pending review and pushed to the manual review channel. The classification results that pass the review are passed to the personalized medical order intelligent matching and generation engine and the intelligent follow-up plan orchestration engine, which respectively generate personalized postoperative medical orders and personalized follow-up plans containing multiple follow-up visit nodes. Finally, the multi-channel intelligent push engine sends the medical orders and follow-up reminders to the patient through the selected push channels.

[0029] The implementation of the diagnostic semantic residue extraction engine includes: The diagnostic semantic residual extraction engine is the data entry point for the entire system. Its core function is to obtain multi-source diagnostic data of patients from the electronic medical record system of the Department of Stomatology, and to perform semantic parsing on the unstructured text in the multi-source diagnostic data through a natural language processing pipeline dedicated to the field of stomatology, extract diagnostic semantic residual features and encode them into structured diagnostic semantic residual feature vectors.

[0030] This invention defines fragmented clinical information scattered in unstructured text as "clinical semantic residues." Although these information are not captured by any structured fields, they contain semantic clues that are crucial for treatment classification and medical order matching.

[0031] Regarding data source access, this engine supports obtaining raw data from the following multi-source medical data: free text fields such as outpatient medical records, inpatient medical records, surgical records, and progress notes in electronic medical record systems; nursing operation descriptions and patient response records in nursing information systems; drug names, specifications, usage and dosage, and prescription time in prescription systems; item codes, cost details, and billing timestamps in billing systems; descriptive text of image reports in image archiving and communication systems; test items and result data in laboratory information systems; and follow-up appointment records in appointment systems. Data access adopts the standard HL7 or FHIR interface protocol. For older systems that do not support standard interfaces, data synchronization can also be performed through database views.

[0032] After data acquisition, preprocessing is performed, including three steps: data cleaning, format standardization, and data association. Data cleaning mainly removes duplicate records, corrects obvious formatting errors, and handles missing and outlier values. Format standardization unifies the time format, encoding format, and units of measurement from different data sources into the system's internal standard format. Data association establishes a correspondence between the patient's unique identifier and records from each data source, ensuring that all medical data for the same patient can be completely linked.

[0033] The oral medicine-specific natural language processing pipeline includes four core processing modules: tooth position labeling standardization module, oral terminology synonym disambiguation module, treatment operation sequence extraction module, and implicit information reasoning module.

[0034] The tooth position labeling standardization module is used to identify and unify the expressions of different tooth position labeling systems, normalizing them into preset standard tooth position codes. Multiple tooth position labeling systems exist in electronic dental records, and different doctors have significantly different writing habits. Common labeling systems include the FDI standard of the World Dental Federation, Palmer labeling, Universal labeling, and various Chinese description methods. For example, the same tooth, "upper right first molar," may be described as "16," "UR6," "upper right 6," "upper right 6," etc.

[0035] This module's implementation includes the following steps: First, it identifies various tooth position description patterns using a predefined regular expression pattern library. The FDI pattern matches two-digit combinations, the Palmer pattern matches specific symbols plus numbers, the Universal pattern matches numbers from one to thirty-two, the Chinese pattern matches combinations of numbers (up, down, left, right) plus numbers, and the abbreviation pattern matches combinations such as UL, UR, LL, LR plus numbers. Then, it determines the specific meaning based on contextual information. For example, "right upper 6" corresponds to "16" in FDI in the Chinese context, and "UL6" corresponds to "26" in FDI. Finally, it uniformly converts all tooth position descriptions to FDI standard codes and establishes a conversion mapping table to achieve standardized conversion. For generic descriptions such as "full mouth," "maxilla," and "mandible," the module marks them as special types for subsequent processing.

[0036] The oral terminology synonym disambiguation module is used to map synonymous heterogeneous expressions to corresponding standard concept nodes based on a pre-built oral medicine terminology map. Oral medicine terminology exhibits a large number of synonymous heterogeneous expressions; the same concept may have multiple expressions in different contexts. For example, synonyms for "root canal treatment" include "RCT," "root filling," "root canal filling procedure," and "pulp removal"; synonyms for "implant placement" include "implant placement" and "implantation procedure."

[0037] This module performs terminology standardization mapping based on a pre-built oral medicine terminology ontology. This ontology includes commonly used terms and their synonyms from various subspecialties of oral medicine, and establishes hierarchical relationships and attribute associations between terms. During processing, the module first uses a specialized oral medicine word segmenter to segment and tag the text, identifying medical terms. Then, it searches for matching term nodes in the ontology and performs disambiguation based on contextual information. For example, "filling" refers to the completion stage of root canal treatment in "root canal filling," but refers to caries restoration in "composite resin filling." The module determines its accurate meaning through contextual judgment. Finally, the terms are uniformly mapped to standard concept nodes for output.

[0038] The treatment procedure sequence extraction module is used to extract timestamped treatment procedure sequences from chronologically arranged medical records. Many dental treatments require multiple visits, and accurately identifying which stage of the treatment sequence the patient is currently in is crucial for matching medical orders.

[0039] The specific implementation steps of this module are as follows: First, all treatment records for the same patient are sorted by visit time to establish a chronological order. Then, treatment operations in each visit are identified through terminology matching, such as "opening and draining the pulp chamber," "root canal preparation," and "root canal filling." Next, the identified operations are combined into an operation sequence according to the time sequence. For example, the sequence "opening and draining the pulp chamber on January 15, 2024; root canal preparation on January 22, 2024; root canal filling on February 5, 2024" can be extracted from three visit records. Finally, based on a predefined treatment path template, the current treatment stage is identified. The above sequence can be identified as the treatment stage information "three-stage root canal treatment completed." For interrupted treatment sequences, the module will make special markings to indicate the potential risk of treatment interruption.

[0040] The implicit information reasoning module is used to perform logical reasoning on extracted information based on a clinical knowledge graph to obtain implicit clinical information. Some information in electronic medical records is not explicitly recorded, but can be obtained through logical reasoning using clinical knowledge.

[0041] This module performs reasoning based on a pre-built clinical knowledge graph, which contains various types of reasoning rules. Drug diagnosis association rules can infer possible diagnoses or treatment types based on prescribed medications. For example, a prescription for metronidazole and amoxicillin may indicate that the treatment involves anaerobic infection control. Operational risk association rules can infer postoperative risk points requiring special attention based on descriptions in the surgical record. For example, a description in the surgical record stating "the implant tip is approximately 2 mm from the inferior alveolar nerve canal" suggests the need for additional monitoring for lower lip numbness. Laboratory result risk stratification rules can determine a patient's risk level based on laboratory indicators. For example, an HbA1c level greater than 7% indicates a high-risk level for diabetes. Treatment sequence stage identification rules can infer the current treatment stage based on completed treatment procedures.

[0042] After processing by the above four modules, the system encodes the extracted multi-dimensional information into a structured diagnostic semantic residual feature vector. This feature vector contains the following main dimensions: treatment type feature, recording the type of the main treatment procedure, such as implant treatment, root canal treatment, periodontal treatment, etc.; treatment stage feature, recording the current treatment stage, such as preoperative assessment, intermediate treatment stage, treatment completion, etc.; anatomical site feature, recording the specific tooth location involved, encoded using the FDI standard; surgical complexity feature, recording the complexity of the treatment, such as simple, moderate, complex, etc.; patient systemic factor feature, recording systemic diseases and risk factors affecting healing, such as diabetes, smoking history, allergy history, etc.; temporal stage feature, recording the time interval since the last treatment and its position in the overall treatment sequence; and implicit risk feature, recording potential risk points obtained through inference. The feature vector is stored in a structured format for easy processing by the subsequent classification engine.

[0043] The multi-dimensional treatment classification engine is used to perform multi-level classification processing based on residual feature vectors of diagnosis and treatment semantics, and adopts evidence fusion hierarchical attention network to output classification results including main treatment type, treatment stage, treatment complexity and patient individualized risk stratification, and calculates the confidence gradient value of treatment stage corresponding to each level of classification results.

[0044] Multi-level classification processing includes five levels of classification tasks: The first level is the main treatment type classification, used to determine the main treatment procedures. This system has established a treatment type classification system covering all subspecialties of oral medicine. The primary classification includes ten major categories: implant treatment, endodontic treatment, periodontal treatment, prosthodontics, orthodontics, oral and maxillofacial surgery, pediatric dentistry, oral prevention, treatment of oral mucosal diseases, and treatment of temporomandibular joint disorders. Each major category is further subdivided into several secondary and tertiary categories, forming a complete hierarchical classification structure. For example, implant treatment is subdivided into secondary categories such as preoperative assessment and treatment plan design, simple implant placement, complex implant placement with bone augmentation, immediate implantation, post-implantation surgery, and implant restoration.

[0045] The second layer is the treatment stage classification, used to determine the current stage in the overall treatment process. Different treatment types have different stage division methods. Taking implant treatment as an example, the stages include the preoperative assessment and treatment plan design stage, the completion of the first-stage implant surgery, the osseointegration waiting period, the completion of the second-stage implant surgery, the restoration period, and the long-term maintenance period. Taking root canal treatment as an example, the stages include the preoperative assessment stage, the pulp chamber opening and drainage stage, the root canal preparation stage, the completion of root canal filling, the crown restoration stage, and the treatment completion and maintenance period. Accurate stage classification is crucial for matching stage-specific medical orders and determining the next follow-up appointment.

[0046] The third layer is the treatment complexity grading, used to determine the level of complexity of the treatment. The complexity grading is divided into three levels: simple, moderate, and complex. Taking implant treatment as an example, simple implantation refers to a single implant at a standard site; moderately complex implantation refers to multiple implantation or implantation with critical bone volume; and complex implantation refers to implantation accompanied by additional surgeries such as sinus lift, guided bone regeneration, and soft tissue grafting. The complexity level affects the detail of postoperative medical instructions and the frequency of follow-up examinations.

[0047] The fourth layer is individualized risk stratification, used to identify individualized risk factors for each patient. Risk stratification primarily considers the following factors: systemic diseases such as diabetes, hypertension, heart disease, osteoporosis, and immune system disorders; unhealthy lifestyle habits such as smoking and excessive alcohol consumption; special physiological states such as pregnancy and lactation; medication history such as long-term use of anticoagulants, bisphosphonates, and immunosuppressants; and allergy history such as allergies to specific anesthetics or antibiotics. Based on the number and severity of risk factors, patients are categorized into four levels: low risk, low-to-medium risk, medium-to-high risk, and high risk.

[0048] The fifth layer is treatment progress deviation detection, used to detect deviations between the actual treatment progress and the standard treatment path. When the actual number of treatments, intervals, or procedures differ significantly from the standard path, the system will mark the deviation. For example, root canal treatment typically requires two to three visits. If a patient has already visited five times and root canal filling has not been completed, the system will detect this progress deviation and mark it, indicating potential treatment complications or a situation requiring special attention.

[0049] The evidence fusion hierarchical attention network consists of three key components: the evidence source weighted attention layer, the temporal decay attention layer, and the counterfactual reasoning verification layer.

[0050] The evidence source-weighted attention layer is used to evaluate the contribution of information from different data sources to the current classification task through learnable attention weights. Since the information used for classification comes from multiple different data sources, and the contribution of different data sources varies for different classification tasks, it is necessary to differentiate the weighting of information from each data source. The specific implementation of this layer is as follows: Input features are grouped according to data source, including medical record text groups, billing data groups, prescription data groups, laboratory data groups, and imaging data groups, etc. Then, an attention weight is calculated for each data source group using a learnable attention mechanism, assigning higher weights to data sources with higher contributions. For example, for a treatment phase classification task, surgical records and medical history records typically have higher weights than billing codes; while for a patient risk stratification task, test results from the laboratory system have higher weights than medical record text. Finally, the weighted features are fused and output to the next layer for further processing.

[0051] The temporal decay attention layer applies temporal decay weights to the patient's historical treatment records, with the decay rate dynamically adjusted based on the treatment type. Generally, more recent records have higher reference value than older records, but the decay rate varies across different treatment types. The layer is implemented as follows: First, the patient's historical treatment record sequence and its corresponding timestamps are obtained. Then, the decay rate parameter is determined based on the current treatment type. Orthodontic treatment, being a long-term, continuous treatment, has a lower decay rate to ensure slow historical decay, making initial plans from two years ago still highly relevant. Emergency treatment records for acute pulpitis use a higher decay rate because their reference value for judging the current root canal treatment stage decreases over time. Implant treatment, periodontal treatment, etc., use a moderate decay rate. Next, the decay weight for each historical record is calculated based on the time difference and decay rate, with more recent records receiving higher weights and older records receiving lower weights. Finally, the weighted historical features are fused with the current features.

[0052] The counterfactual reasoning validation layer is used to, after outputting the classification result, assume another category and check if there is evidence in the current feature vector that strongly contradicts the assumption. Based on the check results, the confidence level of the original classification result is adjusted. The specific implementation of this layer is as follows: First, the classification result and its confidence level output by the classifier are obtained. Then, counterfactual hypotheses are generated. For each classification task, the top three alternative categories with the highest confidence levels are selected as counterfactual hypotheses. Next, for each counterfactual hypothesis, evidence strongly contradicting the hypothesis is searched in the current feature vector. Conflicting evidence is matched using a predefined rule base. For example, if the assumption is "root canal treatment" but there is clear evidence of "implantation" in the feature vector, a strong contradiction is determined. Finally, the confidence level of the original classification result is adjusted based on the contradictory evidence. If strong contradictory evidence exists, the original classification result has high reliability; if no strong contradictory evidence exists, the confidence level of the original classification result is appropriately reduced to reflect the uncertainty of the classification result. This mechanism effectively reduces the overconfidence problem of the model in cases with ambiguous boundaries.

[0053] The confidence gradient value for the treatment phase is a binary tuple (C_t, ΔC / Δt), where C_t is the confidence of the system's classification result of the patient's treatment phase at the current time, and ΔC / Δt is the rate of change of confidence over time.

[0054] The calculation method for C_t is as follows: For the treatment phase classification task, the maximum probability value output by the classifier's softmax is taken as the confidence level at the current time. For example, if the probability distribution of each phase in the treatment phase classification output is [0.1, 0.7, 0.15, 0.05], then C_t takes the maximum value of 0.7.

[0055] The calculation method for ΔC / Δt is as follows: The system saves the patient's historical classification results and corresponding confidence score sequences. When a new patient visit results in a new classification, the current confidence score is added to the sequence. Then, linear regression analysis is performed on the confidence score sequence to calculate the slope of the confidence score over time; this slope is ΔC / Δt. If the patient has fewer than two historical classification records, ΔC / Δt is set to zero by default.

[0056] The confidence gradient value during the treatment phase not only measures the accuracy of classification, but more importantly, it measures the system's "self-awareness" of the reliability of its own judgment, providing a basis for subsequent decision-making and triage.

[0057] The decision-making and triage module is used to determine whether the preset automatic processing conditions are met based on the confidence gradient value of the treatment stage. If the conditions are met, the classification results are transmitted to the personalized medical order intelligent matching and generation engine. If the conditions are not met, the classification results are marked as pending review and pushed to the manual review channel.

[0058] The preset automatic processing conditions are: C_t is greater than or equal to a first threshold, and the absolute value of ΔC / Δt is less than or equal to a second threshold. In this embodiment, the first threshold is set to 0.9, and the second threshold is set to 0.01.

[0059] The preset automatic processing condition judgment logic includes the following cases: In the first scenario, when C_t is greater than or equal to the first threshold and the absolute value of ΔC / Δt is less than or equal to the second threshold, the system is considered to be in a high-confidence stable state, meeting the conditions for automatic processing. This indicates that the system's judgment on the patient's treatment stage is highly certain and stable, and the case enters a fully automated process. The classification results are directly transmitted to the personalized medical order intelligent matching and generation engine and the intelligent follow-up plan scheduling engine for further processing.

[0060] In the second scenario, when C_t is greater than or equal to the first threshold but ΔC / Δt is significantly negative, the system is considered to be in a state of uncertainty, automatic processing is paused, and manual review is triggered. This indicates that although the current confidence level is high, new diagnostic data is shaking the existing judgment, possibly due to a change in the treatment plan. In this embodiment, the criterion for a significantly negative value is ΔC / Δt less than -0.05. At this time, the system pauses automatic processing, marks the case as pending review, and pushes it to the manual review channel to ensure classification accuracy.

[0061] In the third scenario, when C_t is less than the first threshold, the case is classified as low-confidence and transferred to the manual review channel. This indicates that the system cannot reliably classify the case, which may be due to a rare treatment type, a highly non-standardized recording method, or a case with ambiguous boundaries. In this case, the case is directly marked as pending review and pushed to the manual review channel.

[0062] For cases requiring manual review, the system provides a dedicated review interface. This interface displays the system's classification results and their confidence levels, a list of key evidence influencing the classification decision (with the option to view the original records), historical cases similar to the current case and their classification results for reference, and a treatment pathway reference map. Reviewers can confirm the system's classification results or modify them to the correct classification. The manually confirmed classification results are then passed to subsequent modules for generating medical orders and arranging follow-up plans.

[0063] The personalized medical order intelligent matching and generation engine is used to retrieve matching medical order components from a pre-built medical order component library based on the classification results, and generate personalized postoperative medical order content through a dynamic assembly mechanism of medical order components.

[0064] This invention employs a component-based design approach to construct a medical order component library, deconstructing medical order content into the smallest semantic units, namely medical order components. Each medical order component is a structured object containing rich attributes, the main attributes of which include: a unique component identifier for indexing and referencing; multilingual support for the guidance content text; a category indicating whether it is dietary guidance, medication guidance, oral care guidance, activity restriction, abnormal symptom warning, or follow-up visit reminder, etc.; a list of applicable treatment types indicating which treatments the component is applicable to; a list of applicable treatment stages indicating which treatment stages the component is applicable to; an applicable postoperative time window indicating the effective time range of the component after surgery, such as 0 to 24 hours postoperatively, 1 to 7 days postoperatively, 1 to 3 months postoperatively, etc.; a priority indicating the importance of the component, divided into four levels: urgent, important, routine, and optional; a list of conflicting component identifiers indicating other components whose content contradicts this component; a list of pre-dependent component identifiers indicating other components that need to be presented before this component; and a personalized adjustment rule set containing rules for adjusting the component content based on individual patient factors.

[0065] The system pre-builds a component library containing approximately two thousand medical order components, covering common treatment procedures in various subspecialties of dentistry. The component library was reviewed and compiled by experts from various dental specialties, and its content is based on the latest clinical guidelines and evidence-based medicine. The component library supports continuous updates and expansion.

[0066] The dynamic assembly mechanism for medical order components includes the following steps: The first step is to retrieve all matching medical order components from the medical order component library based on the treatment type and treatment stage in the classification results. The retrieval uses multi-condition joint matching, with query conditions including a list of applicable treatment types for the treatment type component and a list of applicable treatment stages for the treatment stage component. The retrieval result is a set of all medical order components that meet the criteria.

[0067] The second step involves triggering personalized adjustment rules within the medical order component based on the patient's individualized risk stratification in the classification results, modifying or supplementing the component's content. Each medical order component can contain multiple personalized adjustment rules, each including triggering conditions and adjustment actions. Triggering conditions are based on the patient's risk factors, such as "diabetes equals true" or "age less than 12." Adjustment actions include adding content, replacing content, and modifying parameters. For example, for diabetic patients, the relevant component will add blood glucose monitoring reminders and infection risk warnings; for pregnant patients, the relevant component will adjust medication recommendations to avoid using contraindicated drugs; for pediatric patients, the relevant component will translate medical terminology into language that is easy for parents to understand.

[0068] The third step involves performing conflict detection and dependency sorting between components to eliminate contradictory content and establish a presentation order. Conflict detection is achieved by querying the list of conflicting components for each component, identifying component pairs with contradictory content. For conflicting components, the following rules are applied for resolution: components with higher priority are retained first, components with higher specificity are retained first, and components with more recent postoperative time windows are retained first. Complex conflicts that cannot be resolved automatically are marked for manual review. Dependency sorting establishes a reasonable presentation order based on the pre-dependencies between components, using a topological sorting algorithm. For example, the component "If you experience severe pain, please seek medical attention immediately" should be placed after the component "It is normal to experience mild discomfort after surgery" to avoid causing unnecessary panic among patients.

[0069] The fourth step involves filtering a subset of time-fitting components based on the current postoperative time window. The system calculates the patient's current postoperative time window based on the time elapsed since the surgery, and then filters out applicable components that match the current postoperative time window. For example, the medical instructions sent on the first day after surgery should include the component "apply ice within 24 hours after surgery," while the medical instructions sent one week after surgery do not need to include this component.

[0070] The fifth step involves prioritizing components based on the principle of optimizing the utility density of medical order information. The utility density of medical order information is defined as the ratio of the amount of information in the medical order information pushed to the patient that is truly relevant to the current stage of treatment and has practical guiding significance for their behavior, to the total amount of information pushed. Based on this principle, the system prioritizes the filtered components, ensuring that the most important information is placed first. Simultaneously, components are truncated as necessary based on the information capacity limitations of the push channel. During truncation, components of urgent and important levels are retained first to ensure that critical information is not overlooked.

[0071] The sixth step is to render the assembled component set into patient-readable text. The rendering process considers the patient's language preferences and the formatting requirements of the delivery channel. For SMS channels, it is rendered as plain text with controlled word count; for WeChat official accounts, it is rendered as rich text, supporting styles such as titles, lists, and bolding; for hospital client channels, it supports mixed text and image layouts and short video embedding. The rendering also generates personalized greetings and opening remarks based on patient information.

[0072] The system also includes a parallel treatment conflict resolution module, which is used to handle situations where multiple parallel treatments are detected for the same patient at the same time.

[0073] The specific functions of the parallel treatment conflict resolution module include the following four aspects: First, the module scans for conflicts between all medical order components to be assembled and selects those that conflict according to preset rules. It first identifies the sets of medical order components generated for each parallel treatment, then scans for conflicts between components from different treatments. For conflicting components, selection is based on treatment importance and timeliness. For example, if a post-implantation medical order component requires "no chewing on the surgical site for three months" conflicts with a root canal treatment medical order component requiring "normal chewing is allowed, but avoid hard foods," the post-implantation requirement is retained based on treatment importance.

[0074] Second, common guidance components that are repeated across multiple treatments are merged. For common guidance content that is included in multiple treatments, such as "take medication on time" and "pay attention to oral hygiene," the module merges them into one item to avoid sending duplicate information to patients.

[0075] Third, the module generates zoning guidelines for parallel treatments involving different oral regions. When a patient's parallel treatment involves different oral regions, the module generates zoning guidelines that clearly distinguish the different requirements for each region. For example, if a patient has undergone implant surgery on the right side and root canal treatment on the left side, the zoning guidelines will clearly state that "chewing is strictly prohibited in the right implant area, while chewing is allowed in the left root canal treatment area, but hard foods should be avoided."

[0076] Fourth, the follow-up timelines of multiple treatments are merged into a unified follow-up calendar. The module retrieves the follow-up nodes generated for each treatment separately, and then merges nodes with similar times onto the same day as much as possible, reducing the number of times patients need to come to the hospital and improving the convenience of follow-up visits. After merging, a unified follow-up calendar is generated, clearly showing the examination items required for each follow-up visit.

[0077] The intelligent follow-up plan orchestration engine generates personalized follow-up plans with multiple follow-up visit nodes based on classification results and the built-in standard dental treatment pathway map.

[0078] The standard pathway map for oral treatment is a structured description of the standard treatment procedures for each subspecialty, validated by evidence-based medicine. Each pathway defines a complete timeline from the start of treatment to long-term maintenance and key follow-up appointments.

[0079] Taking the standard pathway for implant treatment as an example, it includes the following stages and milestones: the preoperative assessment and treatment plan design stage includes the initial assessment milestone and the treatment plan confirmation milestone, with an interval of approximately three to seven days between the two milestones; the first-stage implant surgery stage includes the milestone on the day of surgery and the milestone of suture removal seven days after surgery; the osseointegration waiting period is usually three to six months, including the healing assessment milestone one month after surgery and the osseointegration assessment milestone three months after surgery; the second-stage implant surgery stage includes the abutment connection milestone and the milestone of suture removal seven days after surgery; the repair period includes the milestones of impression taking, trial fitting, and final adhesion, with an interval of approximately two weeks between each milestone; the long-term maintenance period includes the first maintenance check-up milestone three months after surgery and the regular maintenance milestones every six months thereafter.

[0080] Taking the standard path of root canal treatment as an example, it includes the following stages and milestones: the pulp chamber opening and drainage stage includes the milestone on the day of pulp chamber opening; the root canal preparation stage includes the milestone of the follow-up visit seven to fourteen days after the operation; the root canal filling stage includes the milestone on the day of root canal filling; the crown restoration assessment stage includes the milestone of the crown restoration assessment two to four weeks after the operation; and the long-term maintenance period includes the milestone of the annual regular check-up.

[0081] The follow-up plan generation process includes the following steps: The first step is standard node localization. Based on the treatment type and current stage in the classification results, the patient's current location is located in the standard path map, and then all key follow-up visit nodes after the current location and their standard time intervals are obtained.

[0082] The second step is complexity correction. Based on the treatment complexity in the classification results, the time for standard nodes is adjusted. For example, for implant treatment with maxillary sinus lift, the osseointegration waiting period is extended from the standard three months to four to five months; for implant treatment with guided bone regeneration surgery, the osseointegration waiting period is extended to five to six months; and the osseointegration waiting period for immediate implantation can be appropriately shortened.

[0083] The third step is risk factor correction. Based on the individualized risk stratification of patients in the classification results, the time points are further adjusted. For diabetic patients, the waiting period for osseointegration is extended by one to three months depending on blood glucose control, and the frequency of periodontal follow-up is increased; for smoking patients, the frequency of periodontal follow-up is increased from the standard every three months to every 1.5 to two months; for patients using bisphosphonates long-term, the healing assessment point after tooth extraction is postponed from the standard one month to two to three months, and a bone necrosis risk monitoring point is added.

[0084] The fourth step is node merging optimization. For merging requests from the parallel treatment conflict resolution module, follow-up nodes for multiple treatments are merged into a unified follow-up calendar, and nodes with similar times are scheduled for visits on the same day as much as possible.

[0085] The fifth step is to configure the reminder strategy. Configure a corresponding reminder strategy for each follow-up stage, including the timing, frequency, and escalation mechanism. For general follow-up appointments, send the initial reminder seven days in advance; for surgical appointments, send the initial reminder three days in advance; and for imaging examinations, send the reminder two weeks in advance to allow patients to schedule appointments. For important appointments, set up multiple reminders, and for patients who have missed their follow-up appointments, gradually escalate the reminder frequency and channels.

[0086] The multi-channel intelligent push engine is used to send personalized postoperative medical orders and follow-up reminders to patients through selected push channels. This engine supports multiple push channels, including SMS, WeChat official accounts, hospital clients, AI-powered voice calls, and email.

[0087] A smart strategy is employed for channel selection. First, patient preferences are considered; if a patient explicitly specifies a preferred channel, that channel is used first. Second, patient age and digital literacy are taken into account: younger patients are prioritized for push notifications via WeChat official accounts or the hospital's mobile app; elderly patients are prioritized for SMS or AI-powered voice calls; and medical orders for children are sent to their guardians. Third, the complexity and importance of the information are considered: lengthy and complex medical orders are prioritized for channels with large information capacity, such as WeChat official accounts or the hospital's mobile app; urgent and important reminders are pushed through multiple channels in parallel to ensure delivery.

[0088] The push notification timing employs a personalized optimization strategy. The system records the distribution of historical message opening times for each patient and analyzes the time window in which they are most likely to read the messages. Non-urgent messages are scheduled to be pushed within this time window as much as possible to improve the message open and read rates. Urgent messages are pushed immediately without time window restrictions. At the same time, non-urgent information is avoided being pushed late at night and in the early morning to avoid disturbing patients.

[0089] For detailed medical orders, a progressive information presentation strategy is adopted. The initial push only displays the three to five highest priority core points as a summary layer. Patients can expand to view the full content by clicking the link or the "View Details" button. This design ensures that key information is noticed by patients immediately, while the complete information is easily accessible when needed.

[0090] The system also includes a continuous learning module, which collects the review results from the manual review channel as labeled data to incrementally train the classification model in the multi-dimensional treatment classification engine.

[0091] The implementation of the continuous learning module includes the following: First, a manual review feedback loop is established. The decision-making triage module pushes low-confidence cases or cases with wavering judgments to the manual review channel, where reviewers confirm or correct the system's classification results. The corrected classification results, along with the original diagnostic semantic residual feature vectors, are stored as new labeled data in the labeling database.

[0092] Second, perform incremental model training regularly. The system periodically extracts newly accumulated labeled data from the labeled database and performs incremental training on the classification model. Incremental training adopts an online learning strategy, learning patterns from new data while retaining historical data memory, thus avoiding catastrophic forgetting. After training is completed, the new model is evaluated, and once performance improvement is confirmed, the new model is deployed online.

[0093] Third, the system tracks clinical outcomes to optimize the medical order component library. It performs correlation analysis between clinical outcomes such as the doctor's assessment of healing progress, the occurrence of complications, and patient satisfaction during follow-up visits, and the medical order components previously received by the patient. Medical order components correlated with poor clinical outcomes are marked as pending review, and experts decide whether revisions are necessary. Medical order components correlated with good clinical outcomes and high patient satisfaction can have their priority weight appropriately increased.

[0094] The system also implements a closed-loop management process, including the following: First, the system automatically sends follow-up appointment reminders to patients before each scheduled follow-up appointment. The system monitors the time of each follow-up appointment in the schedule and automatically sends reminders before the appointed appointment based on the configured reminder strategy. The reminders include the appointment time, purpose, required documents, and precautions.

[0095] Second, the system automatically reads new medical records and returns them to the processing flow after a patient's follow-up visit. The system monitors the hospital information system for new patient records. When a follow-up visit is detected, it automatically retrieves the new medical data and returns it to the semantic residue extraction engine for re-semantic parsing. Then, the treatment stage classification is updated, and based on the new classification results, medical order generation and follow-up plan updates are triggered again. If treatment enters a new stage, the system generates the corresponding postoperative medical orders and subsequent follow-up plans. This creates a complete closed-loop management system, ensuring that patients receive accurate postoperative guidance and timely follow-up reminders throughout the entire treatment cycle.

[0096] Third, an escalated reminder mechanism is activated for patients who fail to return for follow-up appointments. If a patient still fails to return for a follow-up appointment after the scheduled time, the system first increases the frequency of reminders, then upgrades the reminder channel (e.g., from WeChat official account messages to SMS, and then to AI voice outbound calls). For patients who still do not respond after multiple reminders, the system marks them and pushes them to a list for manual follow-up, where a dedicated person will conduct a telephone follow-up.

[0097] To more clearly illustrate the technical solution of the present invention, specific implementation examples are provided below.

[0098] Implementation Example 1: Single Treatment Program Mr. Zhang, 45 years old, sought endodontic treatment for caries in his lower right first molar and underwent root canal treatment. After the dentist completed the root canal filling, the hospital information system generated a medical record: the diagnosis was pulpitis, the charge was for root canal treatment of three posterior root canals, the prescription was ibuprofen sustained-release capsules 300 mg as needed, and the medical record stated: "Treatment of the lower right sixth root canal completed, filled with heated gutta-percha, postoperative X-ray showed proper root canal filling, elective crown restoration recommended."

[0099] The system automatically retrieves and processes the medical record. The treatment semantic residue extraction engine standardizes "lower right sixth tooth" to FDI code "46" using the tooth position labeling standardization module. It identifies "root canal treatment completed" and "root canal filling completed" as the root canal filling completion stage using the oral terminology synonym disambiguation module. The treatment operation sequence extraction module, combined with historical medical records, extracts the complete root canal treatment operation sequence. The implicit information reasoning module infers that crown restoration is the next step based on "recommended elective crown restoration." The final generated treatment semantic residue feature vector includes: treatment type: root canal treatment; treatment stage: root canal filling completed, awaiting crown restoration; anatomical location: 46; surgical complexity: conventional; patient risk: low risk.

[0100] The multi-dimensional treatment classification engine performs five-level classification based on feature vectors: main treatment type classification (root canal treatment, confidence level 0.98), treatment stage classification (root canal filling completed, confidence level 0.97), treatment complexity classification (standard, confidence level 0.99), patient risk stratification classification (low risk, confidence level 0.98), and treatment progress deviation detection classification (no deviation, confidence level 0.95). The treatment stage confidence gradient value is (0.97, 0), representing a high-confidence stable state.

[0101] The decision-making and triage module determines that C_t=0.97 is greater than the first threshold of 0.9 and the absolute value of ΔC / Δt is equal to the second threshold of 0.01, which meets the automatic processing conditions, and the case enters the fully automatic process.

[0102] The personalized medical order intelligent matching and generation engine retrieved applicable medical order components, including: a component to avoid chewing on the affected side after root canal treatment, a component stating that mild postoperative swelling and pain is normal, a component reminding patients to seek medical attention if pain persists, and a component suggesting crown restoration. Since the patient is considered low-risk, no personalized rules were triggered. After conflict detection, dependency sorting, and priority sorting, the following medical orders were generated: Avoid chewing with the affected tooth for two hours post-surgery; warm or cool liquids are acceptable; mild postoperative swelling and pain is normal; ibuprofen sustained-release capsules can be taken; if pain persists for more than three days or worsens, please contact your doctor immediately; please schedule crown restoration within two to four weeks to prevent tooth fracture.

[0103] The intelligent follow-up plan scheduling engine generates a follow-up plan based on the standard path map of root canal treatment: a text message is sent three days after the operation to inquire about pain, a crown restoration appointment reminder is pushed two weeks after the operation, and if no appointment is made within two weeks, the reminder will be upgraded to four weeks after the operation.

[0104] The multi-channel intelligent push engine selects the WeChat official account channel based on the patient's preferences and pushes medical advice and first follow-up visit reminders in the evening when the patient usually reads messages.

[0105] Implementation Example 2: Multidisciplinary Combined Sequential Therapy Ms. Li, a 55-year-old patient with type 2 diabetes and a glycated hemoglobin level of 7.5%, underwent a combined periodontal and implant restoration treatment for chronic periodontitis and multiple missing teeth in the upper right posterior region. This was her sixth visit, during which she underwent delayed implant placement at the edentulous sites, along with simultaneous maxillary sinus lift and guided bone regeneration surgery.

[0106] The system automatically retrieves the current visit record and backtracks the patient's historical data. The treatment semantic residue extraction engine identifies the complete treatment sequence through the treatment operation sequence extraction module: the first to third visits were for basic periodontal treatment, the fourth visit was for periodontal follow-up to confirm stability, the fifth visit was for implant planning, and the sixth visit (this visit) was for implant placement with maxillary sinus lift and guided bone regeneration. The implicit information inference module infers a medium-to-high risk of diabetes based on the HbA1c score of 7.5%, and infers a complex implant surgery based on the "maxillary sinus lift" and "Bio-Oss bone powder" in the surgical record. The final generated treatment semantic residue feature vector includes: treatment type: implant treatment; treatment stage: osseointegration waiting period after first-stage implant surgery; anatomical sites: 14, 15, and 16; surgical complexity: high (with maxillary sinus lift and guided bone regeneration); and patient systemic factors: medium-to-high risk of diabetes.

[0107] The multi-dimensional treatment classification engine outputs the following classification results: Primary treatment type: implantation treatment (confidence 0.99); Treatment stage: completion of first-stage implantation surgery (confidence 0.96); Treatment complexity: highly complex (confidence 0.98); Patient risk stratification: intermediate-high risk (confidence 0.94); Treatment progress deviation detection: no bias (confidence 0.91). The confidence gradient for treatment stage is (0.96, +0.02), indicating high confidence with a slight increase in confidence.

[0108] Once the decision-making and triage module determines that the conditions for automatic processing are met, the case enters the fully automated process.

[0109] The personalized medical order intelligent matching and generation engine retrieved multiple applicable medical order components. Because the patient was at medium to high risk of diabetes and underwent complex implant surgery, several personalized adjustment rules were triggered: After the diabetes rule was triggered, blood glucose monitoring reminders and infection risk warnings were added to the relevant components; after the maxillary sinus lift rule was triggered, sinus cavity protection precautions were added, such as avoiding blowing the nose, avoiding drinking water with a straw, and avoiding air travel; after the multiple implant rule was triggered, a time requirement for prohibiting chewing in the surgical area was added. After conflict detection, dependency sorting, and priority sorting, a structured personalized medical order was generated, including emergency warning sections, special reminders for maxillary sinus protection, special reminders for diabetes management, postoperative care guidance, medication guidance, and follow-up appointment scheduling.

[0110] The intelligent follow-up planning engine modifies the standard implant follow-up pathway based on the patient's complexity and risk factors: an AI-powered phone consultation is scheduled one day post-surgery, followed by a text message inquiry about swelling and blood sugar control three days post-surgery, a suture removal follow-up examination seven days post-surgery, and a healing assessment one month post-surgery. Because the patient was at medium-to-high risk of diabetes and underwent complex implant surgery, the osseointegration assessment was postponed from the standard three months post-surgery to four months. At six months post-surgery, the decision on whether to perform a second-stage implant surgery is made based on the osseointegration assessment results.

[0111] Implementation Example 3: Borderline Cases with Insufficient Classification Confidence The patient, Mr. Wang, is 30 years old. His medical record only briefly describes "treatment of the lower right seventh dentistry patient". The charges are for composite resin filling and pulp devitalization, and the prescription is eugenol.

[0112] The diagnostic semantic residue extraction engine detected contradictory signals during analysis: the oral terminology synonym disambiguation module identified "filling" as usually indicating that caries filling treatment was completed, but "pulp devitalization" suggested that this might be the first step in root canal treatment. The implicit information reasoning module inferred that the current "filling" might only be a temporary sealing filling rather than a final restorative filling.

[0113] The multi-dimensional treatment classification engine outputs two possible classification results: Hypothesis 1, simple filling treatment completed, with a confidence level of 0.35; and Hypothesis 2, first step of root canal treatment, pulp devitalization plus temporary filling, with a confidence level of 0.58. The counterfactual reasoning verification layer in the evidence fusion hierarchical attention network examines both hypotheses and finds no strongly contradictory evidence; therefore, the confidence levels for either hypothesis are not increased. The confidence gradient value for the final treatment stage is (0.58, 0), indicating a moderately confident stable state.

[0114] The decision-making and triage module determines that C_t=0.58 is less than the first threshold of 0.9, which is considered a low-confidence state. The case is marked as pending review and pushed to the manual review channel.

[0115] The review interface displays the system's classification judgment, the confidence level of each hypothesis, key evidence (fees include pulp devitalization, prescription of eugenol, etc.), and suggests that the nurse confirm the patient's current treatment stage. After reviewing the information, the nurse confirms it is in the root canal devitalization stage, and after manual confirmation, the case proceeds to the next processing step.

[0116] The personalized medical order intelligent matching and generation engine generates medical orders based on the confirmed classification results: Do not chew hard objects with the affected tooth using the temporary sealing material; if the sealing material falls off, please seek medical attention promptly for resealing; if severe pain occurs during the devitalization medication period, please contact your doctor immediately; please return for a follow-up visit within seven to fourteen days to continue root canal treatment.

[0117] The continuous learning module collects the case and the results of manual confirmation as new labeled data and stores them in the labeled database for subsequent incremental training of the model, thereby improving the system's classification accuracy for similar cases.

[0118] Implementation Example 4: Parallel Treatment The patient, Ms. Zhao, is 40 years old and is undergoing implant treatment in the upper right posterior region (the first stage of implant surgery has been completed) and root canal treatment in the lower left first molar (root canal filling has been completed).

[0119] The system detected that the same patient was receiving two parallel treatments at the same time, triggering the parallel treatment conflict resolution module.

[0120] The module first scans for conflicts between the medical order components generated for the two treatments. The post-implantation medical order component requires "no chewing on the surgical area for three months," while the root canal treatment medical order component requires "normal chewing is allowed, but avoid hard foods." These two orders conflict. The module determines the conflict based on the importance and location of the treatments: the implant surgery is on the right side, and the root canal treatment is on the left side, placing them in different oral regions.

[0121] The module generates zoning guidance content according to the processing rules: Chewing is strictly prohibited in the right implant area (upper right 4, 5, 6 positions) for three months after surgery, please eat on the left side; Chewing is normal in the left root canal area (lower left 6 position), but avoid hard foods.

[0122] The module merges common guidance components that are repeated in two treatments, such as "take medication on time" and "pay attention to oral hygiene," and retains only one of them.

[0123] The module merges the follow-up timelines of the two treatments into a unified follow-up calendar: suture removal is performed seven days after the procedure, and the recovery of root canal treatment can be checked at the same time; crown restoration is evaluated two weeks after the procedure; and bone integration assessment is performed three months after the procedure.

[0124] The final generated medical orders and follow-up plan integrate the requirements of both treatments, with clear divisions and merged nodes, making it easy for patients to understand and follow.

[0125] The technical solution of this invention has been verified and applied in multiple dental medical institutions, achieving significant technical results.

[0126] In terms of classification accuracy, the system achieves an accuracy rate of over 97% in classifying primary treatment types and over 91% in classifying treatment stages, significantly outperforming traditional methods based on billing code matching. Particularly in complex scenarios such as multidisciplinary sequential treatment, the system can accurately identify the patient's current treatment stage, avoiding classification errors caused by coding confusion in traditional methods.

[0127] Regarding the personalization of medical orders, after blind review and evaluation by specialist physicians, the system-generated medical orders scored significantly higher than the general template solutions in all three dimensions: content accuracy, stage appropriateness, and personalization. Through a dynamic assembly mechanism of medical order components and personalized adjustment rules, the system can generate highly personalized medical order content based on the patient's treatment type, stage, complexity, and risk factors.

[0128] In terms of efficiency improvement, the time for generating and sending postoperative medical orders for each patient has been reduced from several minutes in manual mode to within seconds in automatic mode, reducing nurses' postoperative management workload by approximately 70%. The system frees medical staff from tedious and repetitive tasks, allowing them to devote more energy to core medical work that requires professional judgment.

[0129] Regarding patient adherence, after using this system, the on-time follow-up rate increased from 62% to 85% from baseline, and the proportion of patients who missed a follow-up appointment for more than 30 days decreased from 23% to 7%. Personalized medical advice and timely follow-up reminders significantly improved patient adherence, helping to ensure treatment effectiveness and reduce the risk of complications.

[0130] In terms of human-machine collaboration, the confidence gradient index during treatment phases enables the system to assess the reliability of its own judgments. The system achieves fully automated processing for cases in a high-confidence, stable state, and actively requests human intervention for cases in a low-confidence state or where judgment is wavering, thus finding the optimal balance between efficiency and safety.

[0131] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic classification and management system for postoperative medical orders and follow-up of oral patients, characterized in that, include: The diagnostic semantic residue extraction engine is used to obtain multi-source diagnostic data of patients from the electronic medical record system of dentistry. It performs semantic parsing on the unstructured text in the multi-source diagnostic data through a natural language processing pipeline dedicated to the dental field, extracts diagnostic semantic residue features, and encodes them into structured diagnostic semantic residue feature vectors. The natural language processing pipeline dedicated to the dental field includes a tooth position annotation standardization module, a dental terminology synonym disambiguation module, a treatment operation sequence extraction module, and a latent information reasoning module. A multi-dimensional treatment classification engine is used to perform multi-level classification processing based on the residual feature vector of the diagnosis and treatment semantics, using an evidence fusion hierarchical attention network. The output includes classification results including main treatment type, treatment stage, treatment complexity and patient individualized risk stratification, and calculates the treatment stage confidence gradient value corresponding to each level of classification results. The decision-making and triage module is used to determine whether the preset automatic processing conditions are met based on the confidence gradient value of the treatment stage. If the conditions are met, the classification result is transmitted to the personalized medical order intelligent matching and generation engine. If the conditions are not met, the classification result is marked as pending review and pushed to the manual review channel. A personalized medical order intelligent matching and generation engine is used to retrieve matching medical order components from a pre-built medical order component library based on the classification results, and generate personalized postoperative medical order content through a dynamic assembly mechanism of medical order components. The intelligent follow-up plan orchestration engine is used to generate a personalized follow-up plan containing multiple follow-up visit nodes based on the classification results and the built-in standard oral treatment path map. A multi-channel intelligent push engine is used to send the personalized postoperative medical orders and follow-up reminders to the patient through the selected push channels.

2. The automatic classification and management system for postoperative medical orders and follow-up of oral patients according to claim 1, characterized in that, The evidence fusion hierarchical attention network includes: The evidence source weighted attention layer is used to evaluate the contribution of information from different data sources to the current classification task through learnable attention weights; The temporal decay attention layer is used to apply temporal decay weights to the historical medical records of the same patient, where the decay rate is dynamically adjusted according to the treatment type. The counterfactual reasoning verification layer is used to, after outputting the classification result, assume the classification result is another category and check whether there is evidence in the current feature vector that strongly contradicts the assumption. The confidence of the original classification result is adjusted according to the check result.

3. The automatic classification and management system for postoperative medical orders and follow-up of oral patients according to claim 1, characterized in that, The confidence gradient value for the treatment stage is a binary tuple (C_t, ΔC / Δt), where C_t is the confidence of the system in classifying the patient's treatment stage at the current moment, and ΔC / Δt is the rate of change of confidence over time. The preset automatic processing conditions are: C_t is greater than or equal to the first threshold, and the absolute value of ΔC / Δt is less than or equal to the second threshold.

4. The automatic classification and management system for postoperative medical orders and follow-up of oral patients according to claim 1, characterized in that, The dynamic assembly mechanism for the medical order components includes: Based on the treatment type and treatment stage in the classification results, retrieve all matching medical order components from the medical order component library; Based on the individualized risk stratification of patients in the classification results, the personalized adjustment rules within the medical order component are triggered to modify or supplement the component content. Perform conflict detection and dependency sorting between components, eliminate conflicting content, and establish a presentation order; Filter the component subsets that match the timeliness based on the current postoperative time window; The components are prioritized based on the principle of optimizing the utility density of medical order information. Render the assembled component set into patient-readable text.

5. The automatic classification and management system for postoperative medical orders and follow-up of oral patients according to claim 1, characterized in that, Also includes: The parallel treatment conflict resolution module is used to scan the conflict relationship between all the medical order components to be assembled and select them according to preset rules when multiple parallel treatments are detected for the same patient at the same time. It merges the duplicate general guidance components in multiple treatments, generates zonal guidance content for parallel treatments involving different oral regions, and merges the follow-up timelines of multiple treatments into a unified follow-up calendar. The continuous learning module is used to collect the review results from the manual review channel as labeled data and to incrementally train the classification model in the multi-dimensional treatment classification engine.

6. An automatic classification and management method for postoperative medical orders and follow-up of oral patients, characterized in that, Includes the following steps: S1: Obtain multi-source diagnosis and treatment data of the patient from the electronic medical record system of the Department of Stomatology, and retrieve the patient's historical medical records; S2: Semantic parsing of unstructured text in the multi-source diagnostic and treatment data is performed through a dedicated natural language processing pipeline for the oral field, extracting diagnostic and treatment semantic residual features and encoding them into structured diagnostic and treatment semantic residual feature vectors; The semantic parsing includes tooth position labeling standardization, oral terminology synonym disambiguation, treatment operation sequence extraction, and implicit information reasoning. S3: Based on the residual feature vector of the diagnosis and treatment semantics, a multi-level classification process is performed using an evidence fusion hierarchical attention network. The output includes classification results including main treatment type, treatment stage, treatment complexity and patient individualized risk stratification. The confidence gradient value of the treatment stage corresponding to each level classification result is calculated. S4: Determine whether the preset automatic processing conditions are met based on the confidence gradient value of the treatment stage. If the conditions are met, proceed to step S5. If the conditions are not met, mark the classification result as pending review and push it to the manual review channel. Proceed to step S5 after manual confirmation. S5: Based on the classification results, retrieve matching medical order components from the pre-built medical order component library, and generate personalized postoperative medical order content through the dynamic assembly mechanism of medical order components; S6: Based on the classification results and the built-in standard oral treatment pathway map, generate a personalized follow-up plan that includes multiple follow-up visit nodes; S7: Send the personalized postoperative medical instructions and first follow-up visit reminder to the patient through the selected push channel.

7. The method for automatic classification and management of postoperative medical orders and follow-up for oral patients according to claim 6, characterized in that, The multi-level classification process described in step S3 includes: The first level of primary treatment type classification is used to determine the types of primary treatment items; The second level of treatment stage classification is used to determine the current stage in the overall treatment process; The third level of treatment complexity classification is used to determine the level of treatment complexity. The fourth level of patient-specific risk stratification is used to identify individualized risk factors for each patient. The fifth layer of treatment progress deviation detection is used to detect the deviation between the actual treatment progress and the standard treatment path.

8. The method for automatic classification and management of postoperative medical orders and follow-up for oral patients according to claim 6, characterized in that, In step S4, the confidence gradient value for the treatment phase is a binary tuple (C_t, ΔC / Δt), and the judgment logic for the preset automatic processing conditions includes: When C_t is greater than or equal to the first threshold and the absolute value of ΔC / Δt is less than or equal to the second threshold, it is determined to be a high-confidence stable state, which meets the automatic processing conditions. When C_t is greater than or equal to the first threshold but ΔC / Δt is significantly negative, it is judged as a wavering state, automatic processing is suspended and manual review is triggered; When C_t is less than the first threshold, it is determined to be a low confidence state and is transferred to the manual review channel.

9. The method for automatic classification and management of postoperative medical orders and follow-up for oral patients according to claim 6, characterized in that, The dynamic assembly mechanism of the medical order components in step S5 includes: Retrieve matching medical order components based on the treatment type and treatment stage in the classification results; Personalized adjustment rules within the component are triggered based on the patient's individualized risk stratification in the classification results; Perform conflict detection and dependency ordering between components; Components with suitable timeliness are selected based on the current postoperative time window; Prioritize the information based on the principle of optimizing the utility density of medical orders; Render the assembled component set into patient-readable text.

10. The method for automatic classification and management of postoperative medical orders and follow-up for oral patients according to claim 6, characterized in that, It also includes step S8: Automatically send follow-up reminders to patients before each scheduled visit; After a patient's follow-up visit, the system automatically reads the new medical records and returns to step S2, updates the treatment stage classification, and re-triggers steps S5 and S6 to generate medical orders and update the follow-up plan, thus forming a closed-loop management system. Collect the review results from the manual review channel as labeled data, and incrementally train the classification model in step S3.

Images