Large model collaborative reasoning and dynamic optimization method for oral clinical decision

Abstract

The invention provides a large model collaborative reasoning and dynamic optimization method for oral clinical decision, and belongs to the field of artificial intelligence. A structured thinking chain auditing reasoning mechanism is constructed, and multi-source evidence fusion and traceable output are realized through a planner, an actuator and a verifier; designing a dynamic expert routing and risk gating mechanism, packaging a large language model, a knowledge graph, image analysis and the like into pluggable experts, dynamically selecting and fusing according to context, and supporting security degradation when evidence is insufficient; a continuous optimization closed loop driven by multi-source feedback is established, doctor adoption and editing behaviors and patient follow-up results are converted into multi-dimensional rewards, a model is updated in combination with reinforcement learning and preference alignment, and meanwhile, elastic weight consolidation and knowledge distillation are introduced to prevent catastrophic forgetting. The method effectively solves the problems of uninterpretability, uncredibility, static solidification and single capability of the model, significantly improves the transparency, robustness and safety of decision making, and is suitable for orthodontics, implantation, maxillofacial surgery and other high-risk scenes.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of artificial intelligence and medical information processing, and specifically relates to a large model collaborative reasoning and dynamic optimization method for oral clinical decision-making, which is particularly suitable for intelligent auxiliary decision-making systems in high-risk scenarios such as orthodontics, implantation, and maxillofacial surgery. BACKGROUND

[0002] With the development of artificial intelligence and medical information processing technology, large language models have been widely explored for high-risk oral clinical decision-making scenarios such as orthodontics, implantation, and maxillofacial surgery. With its potential in understanding vast amounts of knowledge, complex logical reasoning, and multi-modal information fusion, large language models can integrate patient medical records, image data, clinical guidelines, and expert experience to generate structured diagnosis and treatment recommendations.

[0003] However, oral clinical decision-making often involves multi-modal data, cross-professional knowledge, and dynamic feedback. General large language models often exhibit untraceable reasoning processes, lack of credible sources of knowledge, static and unevolving output results in specialized tasks. Especially in high-risk complex cases, the "black box" characteristics and static knowledge system pose significant challenges to the reliability, safety, and explainability of clinical auxiliary systems.

[0004] Traditional oral medicine artificial intelligence methods mainly rely on single-task or single-modal models, such as image lesion recognition based on convolutional neural networks and diagnosis recommendation systems based on rule engines. Although they have certain effectiveness in specific scenarios, such methods generally lack explainability, collaboration between modules, and adaptive ability when faced with complex decisions involving multi-source information fusion, multi-step reasoning, and dynamic risk control. In high-level tasks such as orthodontic treatment planning and implant surgery, models often lack the ability to trace the reasoning process, quantify uncertainty, and continuously learn from new knowledge and clinical feedback, making it difficult to meet the requirements of modern precision medicine for safety and personalized services.

[0005] In recent years, retrieval-augmented generation (RAG) and thinking chain technologies have been introduced into specialized fields, which to some extent alleviate the model hallucination and improve the explainability by introducing external knowledge sources or structured reasoning steps. However, existing methods still have obvious limitations in complex oral medical decision-making scenarios: the dynamic fusion mechanism of multi-source evidence is not perfect, the collaborative scheduling of cross-modal expert capabilities lacks flexibility, and a closed loop from real clinical feedback to model parameter optimization has not been established. When faced with evidence conflicts, risk uncertainties, or knowledge updates, the model's ability to adaptively suppress error generation and implement safe degradation is still limited.

[0006] Therefore, an intelligent auxiliary method capable of realizing reliable reasoning, dynamic risk control and continuous capability evolution in complex oral clinical decision-making is urgently needed to meet the actual needs of precision medicine for safety, reliability and personalized services. SUMMARY

[0007] The present application provides a large model collaborative reasoning and dynamic optimization method for oral clinical decision-making. The method supports traceable, collaborative and sustainable evolution of intelligent auxiliary decision-making in high-risk oral diagnosis and treatment scenarios such as orthodontics, implantation, maxillofacial surgery, etc. by constructing a structured thinking chain reasoning framework, implementing a multi-source expert dynamic routing mechanism, and establishing a reinforcement learning closed loop based on clinical feedback.

[0008] To achieve the above-mentioned purpose, the present application provides a large model collaborative reasoning and dynamic optimization method for oral clinical decision-making, comprising the following steps:

[0009] S1, mapping the doctor's text, structured fields and image metadata to a unified semantic space through a multi-modal encoder to form a joint representation vector;

[0010] S2, based on the structured thinking chain, auditable reasoning is performed, the clinical problem is decomposed into reasoning steps by the planner, the multi-source evidence is called by the executor for step-by-step reasoning, and the evidence coverage and consistency of each step conclusion are verified by the verifier, and finally the clinical conclusion with complete evidence chain is output;

[0011] S3, based on the dynamic expert routing mechanism, the general large language model, the oral specialist knowledge graph, the image analysis model, the clinical term standardizer and the rule engine are packaged as pluggable expert modules, the expert combination is dynamically selected according to the case intention and risk level, and the safety degradation is implemented when the evidence coverage is insufficient or the risk is too high;

[0012] S4, a clinical feedback closed loop is constructed, the doctor's adoption, editing behavior and patient follow-up outcome of the suggestion are converted into multi-dimensional reward signals, the model parameters are continuously optimized through reinforcement learning and preference alignment, and a continuous learning mechanism containing an elastic weight consolidation and knowledge distillation mechanism is used to prevent catastrophic forgetting, and the safety boundary is not degraded.

[0013] Further, the information in S1 is mapped to a unified semantic space and a joint representation vector is formed, which specifically includes the following contents:

[0014] S1.1, receiving multi-source clinical input data, including doctor's diagnosis and treatment inquiry text submitted in natural language form covering clinical problem intention, decision target and focus; structured case fields , including patient age, gender, chief complaint, medical history, oral examination results and contraindication information, used to provide objective clinical state constraints; image metadata , including CBCT or intraoral scanner generated scan type, resolution, voxel size, spatial coordinate system parameters and key anatomical region identification, providing geometric context support for subsequent image evidence precise positioning and reference;

[0015] S1.2, modal-specific encoding is performed on each type of input to generate dense vector representations of each modality. First, for doctor text inquiries, a pre-trained language model is used to extract context semantic vectors, capturing semantic structure and medical concept associations. Second, for structured case fields, embedding layers or feature encoders are used to convert them into dense vectors, achieving unified representation of categorical and numerical variables. Finally, for image metadata, a lightweight meta-information encoder is used to generate geometric context vectors after normalization, preserving spatial scale and position relationship information.

[0016] S1.3, the encoding results of each modality are input into a multi-modal fusion module, which aligns and fuses them through concatenation, weighted summation or cross-modal attention mechanisms to generate joint representation vectors in a unified semantic space , which comprehensively reflects the doctor's decision intention, the patient's objective clinical state and the image spatial context, serving as the entry representation of the planner in the subsequent S2 module, used to start the auditable reasoning process of the structured thinking chain.

[0017] Further, the S2 based on structured thinking chain for clinical decision-making auditable reasoning specifically includes the following contents:

[0018] S2.1, the planner parses the joint representation vector into a series of ordered reasoning steps , each step corresponds to an independently verifiable clinical sub-problem, ensuring that the entire reasoning process has clear logical decomposition and traceability; this parsing process follows the verifiability constraint, i.e. all steps must satisfy , where is the evidence subset supporting the current reasoning step , is a judgment function for the logical consistency between the conclusion and the evidence and the coverage of the evidence.

[0019] S2.2, the executor calls the multi-source evidence acquisition module for each reasoning step , retrieves or generates relevant evidence from the knowledge graph , clinical guideline library and image analysis model , and the query intention of the current step Fusing evidence subsets from multiple experts into a unified context representation whose computation process satisfies:

[0020] ;

[0021] wherein, denotes the participating step of the expert set, the evidence subset provided for the kth expert, denotes the multi-source information fusion function.

[0022] S2.3, in the generation stage, a dynamic mask is applied to the output vocabulary, only allowing general large language models to decode in the union set of the retrieved evidence vocabulary set , the preset clinical term whitelist and the terminator set , and the generation process is optimized by a language modeling loss function, and the output probability distribution satisfies:

[0023] ;

[0024] wherein, is the word generated in the tth step, is the unified context representation fused in S2.2, is the allowed vocabulary set for generation, is the logits value of the word , denotes the large language model training parameter used to calculate the logits value, is the allowed vocabulary in the vocabulary , is the logits value of the vocabulary.

[0025] S2.4, the generated free text sequence is subjected to clinical term standardization processing, and non-standard expressions are mapped to a standard vocabulary set in the standard terminology system of stomatology , and the mapping process is defined by the term standardization function :

[0026] ;

[0027] wherein, denotes the standardized form of the ith word and , word-level replacement is realized based on the term synonym dictionary and the context semantic alignment.

[0028] S2.5, the verifier verifies each key conclusion sequence Implementing double-checking:

[0029] The evidence coverage check is performed by attention weight matrix Measuring alignment strength between conclusions and evidences The sequence requires that any conclusion has at least one corresponding evidence with attention weight no less than a preset threshold , i.e.

[0030] ;

[0031] While the semantic consistency check is to introduce a natural language inference module to calculate the contradiction score between conclusions and evidences , and force it not to exceed the allowed maximum tolerance threshold , i.e.

[0032] ;

[0033] wherein is the attention weight of the conclusion on the evidence , and is the contradiction score calculated by the NLI module.

[0034] S2.6, if any key conclusion does not meet the evidence coverage constraint or the semantic consistency constraint, trigger the fallback mechanism to prohibit outputting the conclusion, and perform the following operations according to the failure type: if the coverage fails due to insufficient evidence, return the executor, re-call the multi-source evidence acquisition module to expand the evidence set from the knowledge graph, clinical guidelines or image analysis model to enhance support; if the consistency fails due to logical contradiction, downgrade the conclusion to an evidence summary form , wherein only contains direct reference fragments from the evidence sequence , and does not introduce any inferred expressions, the fallback mechanism ensures the clinical safety and explainability of the system output in uncertain or high-risk scenarios.

[0035] S2.7, finally output a structured clinical conclusion set and its corresponding complete evidence chain , wherein each conclusion is associated with an evidence trace record , which explicitly marks at least one of the following: the clinical guideline clause number relied on, in the format of ; the identification of the similar case cited, in the format of​ ; the position of the image key sign relied on, in the format of ;

[0036] wherein, , represents a continuous three-dimensional physical space in millimeters, obtained by CBCT or intraoral scanner calibration; k and m are positive integer indexes; the evidence chain supports the whole-process backtracking, manual review and responsibility positioning, and realizes the auditable and explainable of the reasoning results.

[0037] Further, the S3 multi-source capability collaboration and risk control based on dynamic expert routing includes the following contents:

[0038] S3.1, the general large language model , the oral specialty knowledge graph , the image analysis model , the clinical term standardizer and the rule engine These five expert modules are encapsulated as a set of pluggable expert module collection, namely:

[0039] ;

[0040] wherein K=5; each expert module provides a standardized interface, and its output includes conditional probability distribution , confidence score and evidence sufficiency weight based on the evidence set .

[0041] S3.2, the input case context is processed by the entry shunt, and the task intent category and the basic risk score are output by the intent recognition module and the risk prediction module respectively, wherein:

[0042] ;

[0043] ;

[0044] wherein, the task intent category includes orthodontic planning, implant evaluation, maxillofacial surgery decision and other oral specialty sub-tasks; the basic risk score considers the patient's medical history complexity, contraindication conflict and suggestion reversibility, and is used for subsequent expert combination screening and safety degradation triggering.

[0045] S3.3, based on task intent category Primary screening candidate expert subset , and comprehensive base risk score , estimated evidence coverage and computing overhead , each candidate expert is ranked by a weighted scoring function , and the top-K experts with the highest scores are selected to form the active set , whose scoring function is defined as:

[0046]

[0047] ;

[0048] where represents the set of task intents supported by the expert ; is an indicator function that is 1 if the condition is true, otherwise 0; estimated by the lightweight evidence retriever, reflecting the sufficiency of the evidence that the expert can provide; is the maximum value of the overhead of all candidate experts ; is a configurable weight; if the estimated evidence coverage of any candidate expert and the base risk score , a safety degradation is triggered, and only the rule engine and knowledge graph are enabled ; ensure the safety and controllability of the output results in high-risk scenarios. S3.4, fusion of outputs of each expert in the active expert set , first, if the current evidence coverage

[0049] , only allow non-generated experts (knowledge graph , rule engine ) to supplement evidence, prohibit generated experts from outputting suggestions, to prevent unreliable inferences when evidence is insufficient; second, under the condition of sufficient evidence, based on the confidence of each expert in S3.1 and the evidence sufficiency weight , the conditional distribution is weighted and fused to obtain the integrated output distribution: ;

[0050] ;

[0051] Then, a consistency vote is implemented on the key conclusions: if the proportion of experts in the top-K experts supporting the same option is less than the threshold ​​​or a composite risk score If so, trigger safety degradation, only output evidence summary and risk alert, no executable clinical advice is generated.

[0052] wherein, is the output confidence of the kth expert, estimated by its internal model or calibration module, is the sufficiency weight of the evidence relied on by the kth expert, estimated by the evidence retriever, is the output probability distribution of the kth expert.

[0053] Further, the S4 based on multi-source clinical feedback for reinforcement learning and safety constraint continuous optimization, specifically includes the following content:

[0054] S4.1, in the process of clinical use, record the following multi-dimensional feedback signals and carry out structured storage: doctor's interaction behavior to model suggestion, including adoption mark , editing track and editing time-consuming ; System running index, including inference cost and time delay ; Patient medium and long-term follow-up outcome, including complication occurrence mark , recurrence state and pain score ;

[0055] All data are desensitized, and the request-evidence-advice-feedback-outcome five-tuple is archived to the training warehouse for subsequent reward shaping and strategy optimization.

[0056] S4.2, based on the structured feedback five-tuple , define a multi-dimensional reward shaping function , which is in the form of weighted sum:

[0057]

[0058] ;

[0059] Among them, represents the adoption reward, which gives positive incentive when the doctor adopts the model suggestion; represents the editing penalty, which calculates the modified cost according to the editing track and time-consuming , the greater the value, the less accurate the suggestion; represents the outcome reward, which is determined by the patient follow-up result, reflecting the actual effect of treatment; represents the cost reward, which negatively punishes the inference resource consumption; ​This indicates a security reward; if the system triggers a security downgrade or outputs a risk warning, a positive reward is given to reinforce conservative behavior. These are preset non-negative weighting coefficients used to adjust the relative importance of each dimension in the total reward, and can be dynamically configured according to the needs of clinical scenarios.

[0060] S4.3, Reward with Discount To monitor the signal, the proximal policy optimization (PPO) algorithm is used to update the collaborative inference policy. Its objective function is:

[0061] ;

[0062] in, Indicating in strategy The expected value of the generated state-action trajectory at time step t; The policy ratio at step t reflects the current policy. Compared to the old strategy The probability ratio; For the advantage estimation at step t, based on the discount report Perform calculations; The pruning threshold limits the magnitude of policy updates;

[0063] At the same time, the doctor preference alignment loss is introduced. Based on paired samples Constructing binary cross-entropy:

[0064] ;

[0065] in Recommendations that doctors are more likely to adopt Suggestions that have been edited or rejected; For the sigmoid function, This is a temperature coefficient used to adjust the intensity of preference. For strategy In the input Output suggestions The probability of;

[0066] Furthermore, the online violation rate is explicitly constrained to not exceed a threshold using the Lagrange duality method. ,Right now The joint optimization objective is:

[0067] ;

[0068] in To balance the hyperparameters; This is a safety violation indicator variable, taking a value of 1 when the model outputs a violation of clinical safety rules, and 0 otherwise.

[0069] S4.4. In the incremental training phase, a continuous learning regularization term is introduced to prevent catastrophic forgetting of historical tasks. First, based on the old strategy... Calculate the parameter importance matrix (Approximate Fisher information matrix), which reflects the sensitivity of each model parameter on the old task, i.e., the magnitude of the impact of parameter changes on the output; subsequently, the Elastic Weight Consolidation (EWC) loss is constructed:

[0070] ;

[0071] in, These are the current model parameters; These are the optimal parameters for the old strategy; The diagonal elements represent the importance of the i-th parameter. The larger the value, the more critical the parameter is to the old task and should be protected during the update.

[0072] Secondly, utilize the old strategy For the current input Generate soft label distribution This serves as the output of the teacher model; simultaneously, based on the current model parameters... Output Prediction Distribution The student model output is then used as the model output; subsequently, the new model output is aligned with it using knowledge distillation loss constraints.

[0073] ;

[0074] in, For the old strategy For the current input Generate soft label distribution; For the current model parameters In the same input The predicted distribution is as follows; This represents the KL divergence, used to measure the difference between two probability distributions;

[0075] Finally, the regularization term is incorporated into the joint optimization objective to form a continuously learning and enhanced objective function:

[0076] ;

[0077] in The joint optimization objective is defined in S4.3. This is a regularization strength coefficient used to adjust the degree of retention of historical knowledge; through this mechanism, while absorbing new clinical knowledge, backward compatibility with existing disease types, safety rules and terminology standards is strictly maintained.

[0078] Beneficial Effects: This invention constructs a structured thought chain reasoning framework, explicitly decomposing clinical decision-making into a traceable three-stage process of planning, execution, and verification. It enforces evidence coverage verification and semantic consistency verification at each step, significantly improving the auditability and logical reliability of the output. Through a dynamic expert routing mechanism, it encapsulates a general-purpose large language model, a dental specialty knowledge graph, and an image analysis model into pluggable capability units. The optimal expert combination is scheduled in real-time based on task intent and risk level, and automatic safety degradation is triggered when evidence is insufficient or risk exceeds limits, effectively ensuring decision-making safety in high-risk scenarios. Furthermore, a multi-source feedback loop is constructed based on physician adoption behavior, editing trajectory, and long-term patient follow-up outcomes, transforming it into multi-dimensional reward signals. Combined with PPO reinforcement learning, preference alignment, and elastic weight consolidation, this continuous learning mechanism absorbs new knowledge while preventing catastrophic forgetting, ensuring the model's capabilities can continuously evolve and its safety boundaries do not degrade. In summary, this invention effectively solves key problems of existing large-scale oral medicine models in complex scenarios such as orthodontics, implantology, and maxillofacial surgery, including untraceable reasoning, static and fixed knowledge, weak risk control, and lack of continuous evolution capabilities. It significantly improves the interpretability, safety, adaptability, and clinical applicability of intelligent assisted decision-making systems. In particular, it demonstrates stronger robustness and reliability under challenging conditions such as conflicting evidence, contraindications, or insufficient experience at the grassroots level. It can be widely applied to high-value clinical scenarios such as intelligent oral diagnosis and treatment, high-risk surgical planning, empowerment of primary healthcare, and remote consultation. Attached Figure Description

[0079] Figure 1 This is a schematic diagram of the overall process of the present invention;

[0080] Figure 2 This is a flowchart for the joint representation generation of multimodal clinical information;

[0081] Figure 3 It is an auditable reasoning flowchart of a structured thinking chain;

[0082] Figure 4 It is a flowchart of dynamic expert routing and risk gating decision-making;

[0083] Figure 5 It is a flowchart for reinforcement learning and continuous optimization; Detailed Implementation

[0084] The invention will now be further described with reference to the accompanying drawings.

[0085] Example

[0086] Furthermore, such as Figure 1 As shown, a large-scale collaborative reasoning and dynamic optimization method for oral clinical decision-making includes the following steps:

[0087] S1. The doctor's text, structured fields, and image metadata are mapped to a unified semantic space through a multimodal encoder to form a joint representation vector;

[0088] S2. Based on the structured thinking chain, auditable reasoning is performed. The planner decomposes the clinical problem into reasoning steps, the executor calls multi-source evidence to perform step-by-step reasoning, and the verifier verifies the evidence coverage and consistency of each step's conclusion. Finally, a clinical conclusion with a complete chain of evidence is output.

[0089] S3. Based on the dynamic expert routing mechanism, the general large language model, oral specialty knowledge graph, image analysis model, clinical terminology standardizer and rule engine are encapsulated into pluggable expert modules. The expert combination is dynamically selected according to the case intent and risk level, and a safety downgrade is implemented when the evidence coverage is insufficient or the risk is too high.

[0090] S4. Construct a clinical feedback loop, transforming doctors' adoption of suggestions, editing behavior, and patient follow-up outcomes into multi-dimensional reward signals. Continuously optimize model parameters through reinforcement learning and preference alignment, and employ a continuous learning mechanism that includes elastic weight consolidation and knowledge distillation to prevent catastrophic forgetting and ensure that the safety margin does not degrade.

[0091] Furthermore, such as Figure 2 As shown, in step S1, information is mapped to a unified semantic space and a joint representation vector is formed. The specific steps are as follows:

[0092] S1.1 Receive multi-source clinical input data, including medical inquiry texts submitted by doctors in natural language format. Structured case field This includes patient age, gender, chief complaint, past medical history, oral examination results, and contraindications; image metadata. This includes the scan type, resolution, voxel size, spatial coordinate system parameters, and key anatomical region markers generated by CBCT or intraoral scanners.

[0093] S1.2. Modal-specific encoding is performed on various types of inputs to generate dense vector representations for each modality. First, for doctor text queries, the pre-trained language model BERT is used to extract their context semantic vectors to capture the semantic structure and its association with medical concepts. Second, for structured case fields, a feature encoder is used to convert them into dense vectors to achieve a unified representation of categorical and numerical variables. Finally, for image metadata, after normalization, a lightweight metadata encoder generates geometric context vectors.

[0094] S1.3 Input the encoding results of each modality into the multimodal fusion module, and align and fuse them through concatenation, weighted summation or cross-modal attention mechanisms to generate a joint representation vector in a unified semantic space. This serves as the entry point representation for the planner in the subsequent S2 module.

[0095] Furthermore, such as Figure 3 As shown, the auditable reasoning for clinical decision-making based on structured thought chains in S2 specifically includes the following:

[0096] S2.1, The planner will combine the joint representation vectors. It can be analyzed as a series of ordered reasoning steps. Each step It corresponds to an independently verifiable clinical sub-problem and satisfies ,in, To support the steps A subset of evidence This represents the verifiability determination function.

[0097] S2.2, The executor performs each reasoning step. Call the multi-source evidence acquisition module from the knowledge graph Clinical Guideline Database and image analysis models Retrieve or generate relevant evidence and include the query intent of the current step. Integrate subsets of evidence from multiple experts into a unified contextual representation. Its calculation process satisfies:

[0098] ;

[0099] in, Indicate participation steps A group of experts in reasoning. The subset of evidence provided by the k-th expert. This represents a multi-source information fusion function.

[0100] S2.3. During the generation phase, a dynamic mask is applied to the output vocabulary, allowing only general-purpose large language models to use it. In the retrieved vocabulary of evidence Pre-defined whitelist of clinical terms Set of terminators The decoding is performed centrally, and the generation process is optimized using a language modeling loss function. The output probability distribution satisfies:

[0101] ;

[0102] in, For the word generated in step t, This is the unified context representation integrated in S2.2. To allow the generation of a set of words, For words The logits value, This represents the training parameters of a large language model, used to calculate the logits value. To allow vocabulary The vocabulary in the text This is the logits value of the word.

[0103] S2.4, the generated free text sequence The clinical terminology is standardized by mapping non-standard expressions to the set of standardized terms in the standard terminology system of oral medicine. Its mapping process is determined by the terminology standardization function. definition:

[0104] ;

[0105] in, Describes the normalized form of the i-th word and , Word-level replacement is achieved by aligning terminology with a thesaurus and contextual semantics.

[0106] S2.5, The checker verifies each key conclusion. sequence Implement dual verification.

[0107] Evidence coverage verification is performed using the attention weight matrix. Measurement conclusions Each lexical element and evidence sequence The system calculates the maximum attention weight for each conclusion lexical unit based on the alignment strength between the conclusions, and requires that any conclusion... There is at least one corresponding piece of evidence. The attention weight is not less than a preset threshold. ,Right now:

[0108] ;

[0109] Semantic consistency verification introduces a natural language reasoning module to calculate a contradiction score between the conclusion and the evidence. And force it to not exceed the maximum allowable tolerance threshold. ,Right now:

[0110] ;

[0111] in The preset coverage threshold; This represents the maximum permissible threshold for conflict tolerance. For conclusion On the evidence Attention weights; The contradiction score is calculated by the NLI module.

[0112] S2.6 If the verifier determines the key conclusion If the evidence coverage constraint or semantic consistency constraint is not met, a fallback mechanism is triggered, prohibiting the output of the conclusion, and performing the following operations based on the failure type. If coverage fails due to insufficient evidence, the process returns to the executor to perform the current reasoning step. The multi-source evidence acquisition module is reactivated to expand the evidence set from knowledge graphs, clinical guidelines, or image analysis models to enhance support; if consistency fails due to logical contradictions, the conclusion is downgraded to an evidence summary. ,in Only includes evidence sequences It directly quotes fragments and only outputs a structured list of evidence, without generating inferential conclusions.

[0113] S2.7 The final output includes a set of structured clinical conclusions. and its corresponding complete chain of evidence Each conclusion Each is associated with an evidence tracing record. The evidence tracing record clearly indicates at least one of the following: the clinical guideline clause number on which it is based, in the format of The format of the similar case identifiers cited is as follows: The location of the key image features relied upon is in the following format: ;

[0114] in, , Represents a continuous three-dimensional physical space in millimeters, calibrated by CBCT or an intraoral scanner; k and m are positive integer indices; the chain of evidence. It supports full-process backtracking, manual review, and accountability, ensuring the auditability and interpretability of reasoning results.

[0115] Furthermore, such as Figure 4 As shown, the multi-source capability coordination and risk gating based on dynamic expert routing in S3 specifically includes the following:

[0116] S3.1, The general-purpose large language model Dental Specialty Knowledge Map Image analysis model Clinical terminology standardizer and rules engine These five expert modules are encapsulated into a set of pluggable expert modules, namely:

[0117] ;

[0118] Where K=5, each expert module Provides a standardized interface whose output includes conditional probability distributions. , represents the module's prediction result for the current inference task; confidence score This reflects the reliability of the module's output; based on the evidence set. The obtained weight of evidence sufficiency .

[0119] S3.2, Input case context Ingress traffic is handled through the intent recognition module. Risk prediction module Output the task intent category respectively and basic risk score :

[0120] ;

[0121] ;

[0122] Among them, the task intent category This includes sub-tasks in oral specialty medicine such as orthodontic planning, implant assessment, and maxillofacial surgery decision-making; the aforementioned basic risk score Taking into account the complexity of the patient's medical history, contraindications, and the reversibility of recommendations, it is used for subsequent expert group selection and safety downgrade triggering.

[0123] S3.3, Based on task intent category Preliminary selection of candidate experts And take into account the basic risk score Estimated evidence coverage and calculation costs Each candidate expert is evaluated using a weighted scoring function. Sort the experts and select the top-K experts with the highest scores to form the activation set. Its scoring function is defined as:

[0124]

[0125] ;

[0126] in Experts Supported set of task intents; This is an indicator function; it returns 1 if the condition is true, and 0 otherwise. The lightweight evidence retrieval tool estimates the sufficiency of the evidence the expert can provide. For all candidate experts Expenses The maximum value in; Configurable weights; if the estimated evidence coverage of any candidate expert... And basic risk score This will trigger a security downgrade, enabling only the rules engine. With knowledge graph Ensure the safety and controllability of output results in high-risk scenarios.

[0127] S3.4, Activating the expert set Chinese experts The outputs are then fused. First, if the current evidence coverage... Only non-generative expert knowledge graphs are allowed. Rule Engine Supplement evidence, prohibit generative experts The system provides recommendations to prevent unreliable inferences when evidence is insufficient; secondly, when evidence is sufficient, recommendations are made based on the confidence levels of each expert as outlined in S3.1. Weight of sufficiency of evidence conditional distribution Weighted fusion is performed to obtain the integrated output distribution:

[0128] ;

[0129] Subsequently, a consensus vote is conducted on key conclusions: if the proportion of top-K experts supporting the same option is below a threshold... or comprehensive risk score If this is not the case, a safety downgrade will be triggered, and only an evidence summary and risk warning will be output, without generating any actionable clinical recommendations.

[0130] Furthermore, such as Figure 5 As shown, the continuous optimization based on multi-source clinical feedback and safety constraints in S4 specifically includes the following:

[0131] S4.1 During clinical use, record and structure the following multi-dimensional feedback signals: doctor's interaction with model suggestions, including adoption markers. Editing trajectory and editing time System performance metrics, including inference cost With delay Long-term follow-up outcomes of patients, including markers of complication occurrence. Relapse status and pain score ;

[0132] All data has been anonymized and then arranged according to the five-tuple of request-evidence-suggestion-feedback-outcome. Archived in the training pod for subsequent reward shaping and strategy optimization;

[0133] in, The input case context is from S1; This is the chain of evidence, originating from S2; The model outputs suggestions from S2; For doctor interaction and feedback, from the above , , ; The patient follow-up outcome is derived from the above. , , .

[0134] S4.2, Based on Structured Feedback Five-Tube Define a multi-dimensional reward shaping function Its form is a weighted sum:

[0135]

[0136] ;

[0137] in This indicates an adoption reward; doctors receive a positive incentive when they adopt the model's suggestions, and zero otherwise. This indicates an editing penalty, based on the editing history. and time consumption Calculate the correction cost; a higher value indicates that the recommendation is less accurate. The outcome reward is determined by the patient's follow-up results; This represents a cost-based reward, while negatively penalizing the consumption of reasoning resources to encourage efficient reasoning. This indicates a security reward; if the system triggers a security downgrade or outputs a risk warning, a positive reward is given to reinforce conservative behavior. These are preset non-negative weighting coefficients used to adjust the relative importance of each dimension in the total reward, and can be dynamically configured according to the needs of clinical scenarios.

[0138] S4.3, Reward with Discount To monitor the signal, the proximal policy optimization (PPO) algorithm is used to update the collaborative inference policy. Its objective function is:

[0139] ;

[0140] in, Indicating in strategy The expected value of the generated state-action trajectory at time step t; The policy ratio at step t reflects the current policy. Compared to the old strategy The probability ratio; For the advantage estimation at step t, based on the discount report Perform calculations; The pruning threshold limits the magnitude of policy updates;

[0141] At the same time, the doctor preference alignment loss is introduced. Based on paired samples Constructing binary cross-entropy:

[0142] ;

[0143] in Recommendations that doctors are more likely to adopt. Suggestions that have been edited or rejected; For the sigmoid function, Temperature coefficient; For strategy In the input Output suggestions The probability of;

[0144] Furthermore, the online violation rate is explicitly constrained to not exceed a threshold using the Lagrange duality method. ,Right now The joint optimization objective is:

[0145] ;

[0146] in To balance the hyperparameters; This is a safety violation indicator variable, taking a value of 1 when the model outputs a violation of clinical safety rules, and 0 otherwise.

[0147] S4.4 In the incremental training phase, a continuous learning regularization term is introduced to prevent catastrophic forgetting of historical tasks: First, based on the old strategy... Calculate the parameter importance matrix (Approximate Fisher information matrix), construct the Elastic Weighted Consolidation (EWC) loss:

[0148] ;

[0149] in, These are the current model parameters; These are the optimal parameters for the old strategy; The diagonal elements represent the importance of the i-th parameter. The larger the value, the more critical the parameter is to the old task and should be protected during the update.

[0150] Secondly, utilize the old strategy For the current input Generate soft label distribution This serves as the output of the teacher model; simultaneously, based on the current model parameters... Output Prediction Distribution The student model output is then used as the model output; subsequently, the new model output is aligned with it using knowledge distillation loss constraints.

[0151] ;

[0152] in, For the old strategy For the current input Generate soft label distribution; For the current model parameters In the same input The predicted distribution is as follows; This represents the KL divergence, used to measure the difference between two probability distributions;

[0153] Finally, the regularization term is incorporated into the joint optimization objective to form a continuously learning and enhanced objective function:

[0154] ;

[0155] in The joint optimization objective is defined in S4.3. This is the regular intensity coefficient.

[0156] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention. The scope of protection of the present invention should be determined by the scope of protection of the appended claims.

Claims

1. A large model collaborative reasoning and dynamic optimization method for oral clinical decision-making, characterized in that, Comprising the following steps: S1, mapping the doctor text, structured fields and image metadata to a unified semantic space through a multi-modal encoder to form a joint representation vector; S2, based on the structured thinking chain, performing auditable reasoning, decomposing the clinical problem into reasoning steps through the planner, calling multi-source evidence through the executor for step-by-step reasoning, and verifying the evidence coverage and consistency of each step conclusion through the verifier, and finally outputting the clinical conclusion with complete evidence chain; S3, based on a dynamic expert routing mechanism, encapsulating general large language models, oral specialist knowledge graphs, image analysis models, clinical term standardizers and rule engines into pluggable expert modules, dynamically selecting expert combinations according to case intent and risk level, and implementing safety degradation when evidence coverage is insufficient or risk is too high; S4, constructing a clinical feedback loop, converting the doctor's adoption, editing behavior and patient follow-up outcome of the suggestion into a multi-dimensional reward signal, continuously optimizing model parameters through reinforcement learning and preference alignment, and using a continuous learning mechanism containing an elastic weight consolidation and knowledge distillation mechanism to prevent catastrophic forgetting, and ensuring that the safety boundary does not degrade.

2. The method of claim 1, wherein, The S1 comprises the following steps: S1.1, receiving multi-source clinical input data, including doctor-submitted diagnosis query text in natural language form structured case fields and image metadata ; S1.2, modality-specific encoding is performed on each type of input, first, for doctor text inquiries, a pre-trained language model is used to extract semantic vectors; in addition, for structured case fields, an embedding layer or feature encoder is used to convert them into dense vector representations; finally, for image metadata, after normalization, a lightweight meta-information encoder generates geometric context vectors; S1.3, input the encoding results of each modality into a multi-modal fusion module, align to a unified semantic space through splicing, weighted summation or cross-modal attention mechanism, and generate a joint representation vector .

3. The method of claim 1, wherein, The S2 comprises the following steps: S2.1, by the planner, resolve the joint representation vector into a series of ordered inference steps such that each step corresponds to an independently verifiable clinical sub-problem and satisfies where is a subset of evidence supporting step , denotes a verifiability decision function; S2.

2. For each inference step, the executor invokes the multi-source evidence acquisition module to retrieve or generate relevant evidence from the knowledge graph , the clinical guideline library , and the image analysis model , and fuses the query intent of the current step with the multi-source evidence into a unified contextual representation , the calculation process of which satisfies: ； wherein denotes the participating step a set of experts of reasoning, a subset of evidence provided for the kth expert, denotes a multi-source information fusion function; S2.3, in the generation stage, impose a dynamic mask on the output vocabulary, only allow general large language models In the retrieved evidence vocabulary set , the preset clinical term white list and the terminator set , the decoding is carried out in the union set, and the generation process is optimized by a language modeling loss function, and the output probability distribution satisfies: ； wherein, is the word generated for the t-th step, is the unified contextual representation fused in S2.2, is the generated vocabulary set, is the word is the logits value of the word represents the large language model training parameters used to calculate the logits value, is the vocabulary in the vocabulary , and is the logits value of the vocabulary. S2.4, on the generated free text sequence performing clinical term normalization, mapping non-standard expressions to a set of canonical lexemes in a standard terminology system for dentistry whose mapping process is defined by a term normalization function Definition: ； wherein, denotes the normalized form of the ith word and , Word-level replacement is achieved based on term synonym dictionaries and contextual semantic alignment; S2.5, the verifier checks each critical conclusion sequence Implementing double-checking: Evidence coverage check is by attention weight matrix Measuring conclusions With evidence Sequence Alignment strength between, and forced to meet the minimum coverage constraints: ； And semantic consistency check is to introduce natural language reasoning module to calculate the contradiction score between the conclusion and the evidence and impose consistency constraints: ； wherein, is a preset coverage threshold, is a maximum allowed contradiction tolerance threshold; S2.6, if any of the key conclusions is not satisfied, a fallback mechanism is triggered, the conclusion is prohibited from being output, and the following operations are performed according to the failure type: if the coverage failure is caused by insufficient evidence, the executor is returned to the current reasoning step , the multi-source evidence acquisition module is re-invoked to expand the evidence set; if the consistency failure is caused by logical contradiction, the conclusion is downgraded to an evidence summary form , wherein only contains direct reference fragments from the evidence sequence , and does not introduce any inferential expression; S2.

7. Final output contains structured clinical conclusion set and its corresponding complete evidence chain where each conclusion is associated with one evidence provenance record which explicitly annotates at least one of the following: the clinical guideline clause number it is based on, in the format of ; the similar case identification it cites, in the format of ; the image key finding location it depends on, in the format of ; wherein, , represents a continuous three-dimensional physical space in millimeters, calibrated by CBCT or intraoral scanner; k and m are positive integer indices; the evidence chain thereof supports the whole-process backtracking, manual review and responsibility positioning, and realizes the auditable and explainable of the reasoning results.

4. The method of claim 1, wherein, The S3 comprises the following steps: S3.1, a general large language model , an oral specialty knowledge graph , an image analysis model , a clinical term standardizer , and a rules engine These five expert modules are encapsulated as a set of pluggable expert modules, namely: ； where K = 5; each expert module provides a standardized interface whose output includes conditional probability distributions , confidence scores , and evidence sufficiency weights based on the set of evidence ; S3.2, on input case context performing entry triage processing by an intent recognition module with a risk prediction module outputting a task intent category and a base risk score wherein: ; ; The task intent category Including orthodontic planning, implant evaluation, maxillofacial surgery decision, etc. Subtasks of oral specialist; The basic risk score Considering the complexity of patient history, contraindication conflict and suggestion reversibility, for subsequent expert combination screening and safety degradation trigger; S3.3, based on task intent category preliminary screening of candidate experts , and integrate the base risk score , estimate the evidence coverage , and calculate the overhead , rank each candidate expert by a weighted scoring function , and select the top-K experts with the highest scores as the active set , whose scoring function is defined as: ; wherein represents an expert supported task intent set, is an indicator function, estimated by a lightweight evidence retriever, is the maximum value among the overheads of all candidate experts , is a configurable weight; if the estimated evidence coverage of any candidate expert and , a security downgrade is triggered and only the rule engine is enabled and the knowledge graph ; S3.4, Fusion of outputs from the activated experts The outputs of each expert are fused: first, if the current evidence coverage is not sufficient, only allow non-generative experts (knowledge graph , rule engine ) to supplement the evidence, and prohibit generative experts from outputting suggestions; second, under the condition of sufficient evidence, weight the conditional distribution based on the confidence of each expert in S3.1 and the evidence sufficiency weight to obtain the integrated output distribution: ; Subsequently, a consistency vote is performed on the key conclusions: if the proportion of experts in the top-K experts supporting the same option is lower than a threshold , or the integrated risk score , then a security downgrade is triggered, and only the evidence summary and risk alert are output, without generating executable clinical recommendations.

5. The method of claim 1, wherein, The S4 comprises the following steps: S4.

1. During clinical use, record and structure store the following multi-dimensional feedback signals: physician interaction behavior on model recommendations, including adoption flag , edit trajectory , and edit time consumption ; system running indicators, including inference cost and latency ; patient mid-long term follow-up outcomes, including complication occurrence flag , relapse status , and pain score ; All data is de-identified and structured into a request-evidence- recommendation-feedback-outcome quintuple Archived to the training bin for subsequent reward shaping and policy optimization; S4.2, Structured Feedback Five-Tuple , defining a multi-dimensional reward shaping function in the form of a weighted sum: ； wherein represents an adoption reward; represents an edit penalty; represents an outcome reward; represents a cost reward; represents a safety reward; is a preset non-negative weight coefficient; S4.3, Discounted return For the supervision signal, the proximal policy optimization (PPO) algorithm is used to update the collaborative inference policy with the objective function ； wherein, represents the policy generated state-action trajectory on time step t; is the policy ratio for the t-th step, reflecting the current policy versus the old policy ; is the advantage estimate for the t-th step, computed based on the discounted returns ; is the clipping threshold, limiting the magnitude of policy updates; Meanwhile, introduce the doctor preference alignment loss , based on the pair of samples Construct binary cross-entropy: ； wherein is a recommendation more likely to be adopted by a physician, is a recommendation that is edited or rejected, is a sigmoid function, is a temperature coefficient; is a strategy inputs outputs a probability of a recommendation . Further, by Lagrangian duality method, the constraint that the violation rate on the constraint line does not exceed the threshold value is explicitly expressed i.e. The joint optimization objective is: ； wherein is a balanced hyperparameter; is a safety violation indicator variable that takes the value 1 when the model output violates a clinical safety rule, and 0 otherwise; S4.4, In the incremental training phase, introduce a continual learning regularization term to prevent catastrophic forgetting of the history tasks: First, based on the old policy Computing parameter importance matrix (approximate Fisher information matrix), construct an Elastic Weight Consolidation (EWC) loss: ; wherein, is the current model parameter; is the old policy optimal parameter; is the diagonal element, indicating the importance of the ith parameter, the larger the value, the more critical the parameter to the old task, which should be protected during updating; Second, use the old policy on the current input to generate a soft label distribution as the teacher model output; meanwhile, output a prediction distribution as the student model output based on the current model parameters ; then constrain the new model output to align with it through the knowledge distillation loss: ； in, For the old strategy For the current input Generate soft label distribution; For the current model parameters In the same input The predicted distribution is as follows; This represents the KL divergence, used to measure the difference between two probability distributions; Finally, the regularization term is included in the joint optimization objective to form a continuous learning enhanced objective function: ； wherein is the joint optimization objective defined in S4.3, is the regular intensity coefficient; through this mechanism, while absorbing new clinical knowledge, strict backward compatibility is maintained for existing diseases, safety rules and terminology specifications.

Images