Forced compliance guarantee AI clinical decision suggestion generation method and system

Abstract

The invention relates to the cross technical field of artificial intelligence and medical informatics, and particularly discloses a compulsory compliance guarantee AI clinical decision suggestion generation method and system, and the method comprises the steps: S1, obtaining the symptom information of a patient, and converting the symptom information into coding information; s2, generating a preliminary diagnosis and treatment suggestion from the coded information through a large language model; s3, performing drug verification on the preliminary diagnosis and treatment suggestions; s4, performing guide index verification on the preliminary diagnosis and treatment suggestions passing the drug verification; s5, performing guide conformity verification on the preliminary diagnosis and treatment suggestions passing the guide index verification; and S6, displaying the preliminary diagnosis and treatment suggestions and verification results thereof. According to the method, a three-stage cascade check architecture is adopted, the diagnosis and treatment suggestions are ensured to strictly meet the requirements of clinical guidelines through irreversible check logic, and the compliance and safety of the AI medical decision suggestions are remarkably improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of artificial intelligence and medical informatics, and in particular to an AI-based clinical decision suggestion generation method and system with mandatory compliance guarantees. Background Technology

[0002] With the continuous development of artificial intelligence, Natural Language Processing (NLP) technology is widely used in medical text structuring. Currently, in the field of medical decision support, existing technologies have many shortcomings, specifically manifested in the following four aspects: 1. Lack of a mandatory enforcement mechanism for guideline provisions: Most existing medical decision support systems only have knowledge retrieval functions. They can provide doctors with relevant medical knowledge and clinical guideline information, but they cannot build an effective mechanism to ensure that these guideline provisions are effectively enforced in the actual medical decision-making process. This leads to doctors potentially failing to fully follow best practice guidelines when referring to these systems due to negligence or differences in personal judgment, thereby affecting the quality of medical care and patient safety.

[0003] 2. Loose Coupling Between LLM and Knowledge Base with Lack of Mandatory Rule-Blocking Mechanism: While some systems attempt to combine Large Language Models (LLMs) with knowledge bases, this combination is relatively loose. When generating medical recommendations, LLMs may not fully utilize accurate information from the knowledge base, leading to inaccurate or unreliable results. More critically, there is a lack of a mechanism to forcibly block rule-inconsistent recommendations. When LLM-generated recommendations conflict with clinical rules or guidelines, the system cannot promptly prevent these erroneous recommendations from entering the actual medical process, increasing medical risks.

[0004] 3. Follow-up management cannot be dynamically adjusted based on real-time adherence data: Follow-up is a crucial aspect of chronic disease management. However, existing follow-up management models cannot be dynamically adjusted based on real-time patient adherence data. Patient adherence during treatment, such as whether they take medication on time and follow dietary and exercise recommendations, has a significant impact on treatment outcomes. However, existing systems cannot collect and analyze this real-time data in a timely manner, thus failing to make targeted adjustments to treatment plans based on the patient's actual situation, affecting the effectiveness of chronic disease management.

[0005] 4. Insufficient interpretability, failing to meet clinical audit requirements: Decision recommendations generated by existing medical decision support systems often lack sufficient interpretability. When referring to these recommendations, physicians struggle to clearly understand the reasoning and basis behind them. This not only hinders physicians' evaluation and judgment of the recommendations but also fails to meet the stringent requirements of clinical auditing. Clinical auditing requires tracing and reviewing the entire medical decision-making process to ensure the rationality and standardization of medical practices. The lack of interpretable decision recommendations makes this process difficult and also lacks the function of recording auditable decision paths, failing to provide the necessary information support for auditing. Summary of the Invention

[0006] This invention aims to address the problems of non-mandatory guideline implementation, loose coupling between LLM and knowledge base, and insufficient risk control in existing medical decision support systems. To this end, this invention provides an AI-powered clinical decision recommendation generation method and system with mandatory compliance guarantees. It employs a three-level cascaded verification architecture and uses irreversible verification logic to ensure that treatment recommendations strictly comply with clinical guideline requirements, significantly improving the compliance and security of AI-driven medical decision recommendations. This is particularly suitable for the full-cycle management of chronic diseases such as diabetes and hypertension, effectively reducing medical risks and providing reliable compliance guarantees for clinical decision-making.

[0007] This invention provides an AI-based clinical decision suggestion generation method with mandatory compliance assurance, employing the following technical solution: including the following steps: S1: Obtain the patient's symptom information and convert it into coded information; S2: Encoded information is used to generate preliminary diagnostic and treatment suggestions through a large language model; S3: Validate the medication recommendations based on the initial treatment plan; S4: Verify the guideline indicators for the preliminary treatment recommendations that have passed drug validation; S5: Conduct guideline compliance verification on the preliminary treatment recommendations that have passed the guideline indicator verification. The specific process is as follows: S5.1: Calculate semantic similarity based on preliminary treatment recommendations and guideline-recommended protocols; S5.2: Determine the risk factor based on preliminary treatment recommendations and the patient's condition; S5.3: Determine the degree of deviation based on semantic similarity and risk coefficient; S5.4: Compare the deviation with the warning threshold and the lock threshold. If the deviation is greater than the warning threshold, a warning is triggered; if the deviation is greater than or equal to the lock threshold, a lock is triggered. S6: Displays preliminary treatment recommendations and their verification results.

[0008] Furthermore, in step S3, drug verification involves checking whether there are any drug incompatibilities in the preliminary treatment recommendations; If there are no drug incompatibilities, the drug verification is deemed successful, and the process proceeds to step S4. If drug incompatibility exists, the drug verification is deemed unsuccessful, triggering a lockout and proceeding to step S6.

[0009] Furthermore, in step S4, the process of verifying the guideline indicators is as follows: Based on the initial diagnosis and treatment recommendations, obtain the corresponding laboratory indicators and find the corresponding guideline safety thresholds; Compare laboratory indicators with guideline safety thresholds; if laboratory indicators meet guideline safety thresholds, proceed to step S5; if indicator data do not meet guideline indicator standards, increase the risk coefficient and proceed to step S5.

[0010] Furthermore, in step S4, if laboratory indicators cannot be obtained, it is determined whether the laboratory indicators are critical indicators; critical indicators are those that are essential for diagnosis and treatment recommendations and directly determine the safety and compliance of the decision. If the laboratory indicator is a critical indicator, the indicator verification of the guideline is deemed to have failed, triggering a lock and proceeding to step S6. If the laboratory indicator is not a critical indicator, the indicator verification of the guideline is deemed to have passed, and the process proceeds to step S5.

[0011] Furthermore, in step S5.1, if the guidelines do not mention it, the semantic similarity is 0.5; if the guidelines explicitly do not recommend it, a lock is triggered, and the process proceeds to step S6.

[0012] Furthermore, in step S5.2, the baseline value of the risk coefficient is determined based on the operation type corresponding to the preliminary treatment recommendation; The patient population is determined based on the patient's condition, and then the first adjustment value for the risk factor is determined. The urgency of the disease is determined based on the patient's condition, and then a second adjustment value for the risk factor is determined. The risk coefficient is obtained by adding the base value of the risk coefficient, the first adjustment value of the risk coefficient, and the second adjustment value of the risk coefficient.

[0013] Furthermore, the types of procedures include routine oral medication adjustments, injectable medication use, high-risk medication use, and recommendations for invasive procedures; The patient population includes elderly patients, pediatric patients, patients with multiple organ dysfunction, and pregnant / lactating women; The urgency of the disease includes management of the chronic stable phase, management of the acute exacerbation phase, and recommendations for critical care.

[0014] Furthermore, in step S5.4, the formula for calculating the warning threshold is: T = base threshold × (1 + α + β) Where T is the warning threshold, α is the complication coefficient, and β is the age coefficient. The baseline threshold, complication coefficient, and age coefficient were determined based on initial treatment recommendations and the patient's condition.

[0015] Furthermore, in step S5.4, the locking threshold is determined based on the risk coefficient.

[0016] This invention also provides an AI-powered clinical decision suggestion generation system with mandatory compliance guarantees, employing the following technical solution: It includes an input processing module, a large language model module, a rule engine verification module, and an execution control module. The input processing module is used to acquire the patient's symptom information and convert it into encoded information; The large language model module is used to encode information and generate preliminary diagnostic suggestions through the large language model; The rule engine verification module is used to perform drug verification on preliminary treatment recommendations, guideline indicator verification on preliminary treatment recommendations that pass drug verification, and guideline compliance verification on preliminary treatment recommendations that pass guideline indicator verification. The execution control module is used to display preliminary diagnosis and treatment suggestions and their verification results.

[0017] The above-described one or more technical solutions in the embodiments of the present invention have at least one of the following technical effects: This invention designs an irreversible three-level verification chain, where the rule engine's output weight is mandated to be higher than the initial treatment recommendations generated by the LLM. This means that throughout the decision-making process, the rule engine has the final review authority over the initial treatment recommendations generated by the LLM. Once the LLM generates recommendations, they must undergo rigorous verification by the rule engine. If the rule engine finds that a recommendation does not comply with clinical guidelines or relevant rules, even if the LLM deems the recommendation feasible, the rule engine's result will prevail, and the recommendation will be corrected or blocked, thereby ensuring that the final output treatment recommendations comply with medical standards and safety requirements.

[0018] This invention ensures that the output conforms to the latest clinical guidelines through mandatory verification by a rule engine. The rule engine has a rich set of built-in clinical guideline rules and logical judgments, enabling a comprehensive and detailed review of recommendations generated by LLM (Limited Management Model). Whether it's drug use, treatment plan selection, or the arrangement of examinations, everything is verified against the latest clinical guidelines, effectively avoiding medical risks caused by following outdated or incorrect guidelines and improving the consistency and reliability of medical quality.

[0019] This invention configures different tolerance standards based on disease type / risk level. Different chronic diseases, and even the same disease at different risk levels, require varying levels of accuracy and safety in medical decision-making. For example, for high-risk surgical procedures or treatment of special populations (such as pregnant women, children, and the elderly), the accuracy requirements for decision-making are higher, and the tolerance standards should be set more strictly; while for decisions regarding routine medications, the tolerance standards can be relatively lenient. By dynamically adjusting the thresholds, this invention can more flexibly adapt to various medical scenarios, improving the efficiency and applicability of decision-making while ensuring medical safety.

[0020] The system-level locking mechanism of this invention prevents high-risk recommendations from being mistakenly executed. When the rule engine detects that a recommendation generated by the LLM (Low-Level Medical Record) carries a high risk and does not comply with clinical guidelines or safety standards, this invention automatically triggers a read-only locking mechanism. In read-only mode, related medical operations (such as electronic medical record editing and prescription issuance) are prohibited, preventing doctors from mistakenly executing high-risk recommendations due to negligence or other reasons. Only after the doctor confirms understanding of the risk through strict identity verification methods such as biometric authentication will the lock be opened, resuming normal operations. This effectively reduces medical risks and ensures patient safety.

[0021] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in this invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0023] Figure 1 This is a flowchart of the method provided by the present invention.

[0024] Figure 2 This is a structural block diagram of the system provided by the present invention.

[0025] Figure label: 1. Input processing module; 2. Large language model module; 3. Rule engine validation module; 4. Execution control module. Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The following embodiments are used to illustrate this invention but should not be used to limit the scope of this invention.

[0027] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., refer to specific features, structures, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.

[0028] The following is combined with Figures 1 to 2 The present invention will be further described in detail below, describing an AI-based clinical decision suggestion generation method and system with mandatory compliance assurance: In this embodiment, as Figure 1 As shown, an AI-based clinical decision suggestion generation method with mandatory compliance assurance is provided, including the following steps: S1: Obtain the patient's symptom information and convert it into coded information.

[0029] Commonly, the symptom information includes one or more of voice, text, and images. Symptom information is standardized into ICD-10 encoding using technologies such as ASR (Automatic Speech Recognition), NLP (Natural Language Processing), and OCR (Optical Character Recognition). For example, a patient can describe their symptoms verbally; ASR technology converts the speech to text, and NLP technology analyzes and processes the text to extract key information, converting it into ICD-10 encoding for unified recognition and processing in the medical information system. Medical images uploaded by patients can have their text information extracted using OCR technology, and then converted into ICD-10 encoding using NLP technology.

[0030] S2: Encoded information is used to generate preliminary diagnostic and treatment suggestions through a large language model.

[0031] The large language model in this embodiment employs a knowledge-enhanced large language model module, which is fine-tuned according to medical guidelines and coupled with a RAG retrieval tool. Based on medical knowledge and clinical guidelines, it can generate preliminary treatment suggestions according to the encoded information of the patient's symptoms. The training, fine-tuning, and coupling of the knowledge-enhanced large language model module with the RAG retrieval tool are all implemented using existing technologies, and will not be elaborated upon here.

[0032] The initial diagnosis and treatment recommendations are then entered into the rules engine, where a three-level verification is enforced in steps S3-S5.

[0033] In order to ensure that the large language model can correctly generate preliminary diagnosis and treatment suggestions and that the rule engine can perform three-level verification smoothly, this embodiment designs a separate and dynamically updated structured guide library and rule library, which are independent of each other but work together.

[0034] The rules originate from the transformation of guideline clauses: the core rules of the rule base are machine-executable translations of clauses in the structured guideline library. For example, the clause "Metformin is contraindicated when eGFR < 30 ml / min" in the guideline library is transformed into "IF drug code = metformin (RxNorm: 6809) AND patient eGFR < 30 ml / min THEN trigger contraindication rule" in the rule base, and each rule is bound to the corresponding guideline clause number.

[0035] Reverse correlation guide during verification: After the rule engine verifies the data, it retrieves the original text from the guide library by clause number and generates an explanatory graph to ensure that doctors understand the basis for verification.

[0036] Dynamic update linkage: When the guide library is updated, such as when a new version of the guide is released, the rules in the rule library that are bound to the old clauses are automatically compared and a "rule invalidation warning" is triggered. The administrator can regenerate the rules based on the new clauses to ensure that the two are synchronized.

[0037] The clinical guideline repository stores the latest and most authoritative medical clinical guidelines (including national / international clinical guidelines, expert consensus, and operational procedures), which are structured for easy retrieval and retrieval by the system. The rule repository contains various medical rules, such as drug contraindication rules and indicator threshold rules. Both repositories support dynamic updates, ensuring that knowledge and rules are updated promptly in response to advancements in medical research, feedback from clinical practice, and changes in policies and regulations, guaranteeing that the system's decision-making basis is always up-to-date and accurate.

[0038] S3: Validate the initial treatment recommendations with appropriate medications.

[0039] Drug validation is a drug-level validation process that examines drug interactions. Its purpose is to rule out potential drug incompatibilities that may exist in the initial treatment recommendations. Drug validation utilizes RxNorm and the FDA drug incompatibilities database, using RxNorm coding to verify drug interactions and ensure the safety of drug use.

[0040] Drug verification is used to check whether there are any drug incompatibilities in the initial treatment recommendations. If there are no drug incompatibilities, the drug verification is deemed successful, and the process proceeds to step S4. If drug incompatibility exists, the drug verification is deemed unsuccessful, triggering a lockout and proceeding to step S6.

[0041] Verification failure triggers a read-only lock on the initial treatment recommendation until unlocked by physician authentication. When the rules engine detects that the initial treatment recommendation generated by the LLM fails drug verification, it immediately triggers the read-only lock mechanism. At this time, the electronic medical record editing function will be disabled to prevent erroneous recommendations from being recorded and executed. Simultaneously, the initial treatment recommendation and the violated guideline clause number will be displayed to alert the physician to the problem. Physicians must unlock the recommendation using biometric authentication (such as fingerprint or facial recognition) combined with two-factor authentication based on their employee ID number to modify or regenerate it, ensuring that only authorized physicians who understand the risks can perform the operation. Locks triggered in subsequent steps are also unlocked using the same method.

[0042] S4: Verify the guideline indicators for the preliminary treatment recommendations that have passed drug validation.

[0043] The guideline indicator verification process compares laboratory indicators with guideline safety thresholds to determine whether preliminary treatment recommendations align with the patient's physiological condition.

[0044] The process of verifying the guidelines' indicators is as follows: Based on the initial diagnosis and treatment recommendations, the corresponding laboratory indicators are obtained from the laboratory LIS system, and the corresponding guideline safety thresholds are found from the guideline threshold table in the rule base. The laboratory indicators are compared with the guideline safety thresholds. If the laboratory indicators meet the guideline safety thresholds, proceed to step S5; if the indicator data does not meet the guideline indicator standards, the risk coefficient is increased, and proceed to step S5. In this embodiment, when the indicator data does not meet the guideline indicator standards, the risk coefficient used in step S5 is multiplied by 2.

[0045] In some cases, there may be no corresponding laboratory indicators. If key laboratory indicators are missing, directly passing the guideline indicator verification may lead to the risk of misjudgment. Therefore, this embodiment implements blocking prompts and tiered processing.

[0046] If laboratory indicators cannot be obtained, determine whether they are critical indicators. If they are critical indicators, i.e., critical indicators are missing, the guideline indicator verification is directly deemed unsuccessful, triggering a lock, locking the preliminary treatment recommendations, and proceeding to step S6, prompting "[Indicator Name] is missing, please supplement the test"; after the laboratory indicators are supplemented, the guideline indicator verification is performed again.

[0047] If the laboratory indicator is not a critical indicator, i.e., a non-critical indicator is missing, the indicator verification of the guideline is deemed successful, and the process proceeds to step S5. However, the display in step S6 indicates "It is recommended to supplement [indicator name] within 24 hours".

[0048] Definitions of key and non-key metrics: 1) Key indicators are those that are essential for specific treatment recommendations (such as the use or operation of specific drugs) and directly determine the safety and compliance of the decision. The absence of such indicators will lead to extremely high risks in the treatment recommendations (such as drug contraindications, dosage errors, etc.), so it is mandatory to supplement them; otherwise, the verification will fail.

[0049] Example: eGFR (estimated glomerular filtration rate) before using metformin: Metformin is excreted by the kidneys, and a low eGFR (e.g., <30 ml / min) can lead to drug accumulation and lactic acidosis. Therefore, eGFR is a "key indicator" for the use of this drug.

[0050] INR (International Normalized Ratio) before using anticoagulants (such as warfarin): INR directly reflects coagulation function and is the core basis for adjusting anticoagulant dosage and avoiding the risk of bleeding or thrombosis, therefore it is a "key indicator".

[0051] 2) Non-critical indicators are those that have a minimal impact on the safety and compliance of the current treatment recommendations, and whose absence will not immediately lead to serious risks. Even if they are missing, the treatment recommendations are allowed to proceed, but supplementation will be requested, usually within a specified timeframe.

[0052] Example: Blood lipid indicators were missing when adjusting antihypertensive drugs during follow-up: The direct correlation between blood lipids and antihypertensive drugs is weak. The core basis for adjusting antihypertensive drugs is blood pressure value. Therefore, blood lipid indicators are considered "non-critical indicators".

[0053] Thyroid function indicators are often missing in routine diabetes follow-up: Thyroid function has little impact on routine diabetes medications (such as metformin and insulin), and is therefore considered a "non-critical indicator".

[0054] The validity of laboratory indicators decays over time. For example, blood glucose and electrolytes fluctuate rapidly, while glycated hemoglobin (HbA1c) is relatively stable. Dynamic validity periods need to be set according to the type of indicator. Expired indicators are considered "invalid indicators" and treated as "no indicators". See Table 1 for details.

[0055] Table 1

[0056] The knowledge base used by the large language model to generate preliminary treatment suggestions includes a structured guideline library, and the core rules of the rule base are machine-executable translations of the clauses in the structured guideline library. Therefore, there will be no situation where the corresponding guideline safety threshold cannot be found in the rule base.

[0057] S5: Perform guideline compliance verification on the preliminary treatment recommendations that have passed the guideline indicator verification.

[0058] The purpose of guideline compliance verification is to ensure consistency between preliminary treatment recommendations and clinical guidelines, which requires access to structured guideline libraries and treatment terminology libraries.

[0059] The specific process is as follows: S5.1: Calculate semantic similarity based on preliminary treatment recommendations and guideline recommendations; in this embodiment, the semantic similarity between preliminary treatment recommendations and guideline recommendations is calculated using the BERT model, with a value ranging from 0 to 1.

[0060] This step focuses on the consistency between treatment details (such as frequency, dosage, monitoring cycle, etc.) and guidelines. The initial value of semantic similarity is determined by the matching degree between the suggestions generated by knowledge-enhanced LLM and the guideline text (range 0~1, the higher the value, the more consistent). When there are significant differences between details such as medication frequency and dosage and guidelines (such as frequency difference > 50%), the corresponding score is deducted (such as 0.3), and finally the 'adjusted semantic similarity' is obtained.

[0061] When a corresponding guideline recommendation is not found in the guideline compliance check, such as for emerging treatment technologies, treatment suggestions for rare complications, or areas not covered by the guideline, semantic similarity cannot be calculated. This embodiment, based on the principle of "clinical safety first + manual confirmation as a fallback," initiates the following mechanism: For cases not mentioned in the guidelines (such as novel drug combinations): the semantic similarity is set to 0.5 by default.

[0062] If the guidelines do not explicitly recommend a dose (e.g., exceeding the recommended dosage): trigger lockout and proceed to step S6.

[0063] S5.2: Determine the risk factor based on preliminary treatment recommendations and the patient's condition.

[0064] The role of the risk coefficient is to amplify the basic differences. The higher the risk of the scenario, the larger the coefficient, and the greater the deviation under the same basic differences, enabling the system to sensitively capture even "small deviations" in high-risk scenarios. In this embodiment, the risk coefficient setting is combined with the risk level of the medical scenario, such as the invasiveness of the operation and the toxicity of the drug, and refers to the strength of safety warnings in clinical guidelines.

[0065] The risk factor in this embodiment is determined based on the type of operation, the patient population, and the urgency of the disease.

[0066] Based on the type of procedure corresponding to the initial treatment recommendation, a baseline risk coefficient was determined. The procedure types include recommendations for routine oral medication adjustments, injectable medication use, high-risk medication use, and invasive procedures; the corresponding baseline risk coefficients are shown in Table 2.

[0067] Table 2

[0068] The patient population was determined based on the patients' conditions, and then the first adjusted risk factor was determined. The patient population included elderly patients, pediatric patients, patients with multiple organ dysfunction, and pregnant / lactating women, and the corresponding first adjusted risk factors are shown in Table 3.

[0069] Table 3

[0070] This embodiment also considers the impact of disease urgency and designs a second adjustment value for the risk coefficient. This second adjustment value is added to the base risk coefficient and the first adjustment value to obtain the risk coefficient. The urgency of the disease is determined based on the patient's condition, and then the second adjustment value for the risk coefficient is determined. Disease urgency includes management of chronic stable phase, management of acute exacerbation phase, and recommendations for critical care. The corresponding second adjustment values ​​for the risk coefficient are shown in Table 4.

[0071] Table 4

[0072] After calculating the risk coefficient, it is necessary to consider whether to increase the risk coefficient based on the situation in step S4 to determine the final risk coefficient. In this embodiment, if the indicator data does not meet the guideline indicator standards, the risk coefficient is multiplied by 2 to obtain the final risk coefficient. Furthermore, when the risk coefficient is increased due to guideline indicator verification, the final value must not exceed 2.0.

[0073] Example: If a patient's serum potassium level is 6.0 mmol / L (>5.5 mmol / L threshold), which does not meet the guideline criteria, and LLM recommends the use of spironolactone (a high-risk drug); Basic semantic similarity: The match between the LLM recommendations and the guidelines is 0.7 (i.e., "1 - semantic similarity = 0.3"). Original risk factor: Spironolactone is a high-risk drug with a conventional risk factor of 1.2; Level 2 verification triggers 'risk coefficient × 2': the adjusted risk coefficient is 2.0; Deviation calculation: Deviation = 0.3 × 2.0 = 0.6; Result: The warning threshold for this scenario is 0.3, so 0.6 > 0.3, triggering a warning.

[0074] Dynamic adjustment mechanism of risk coefficient: (1) Linkage with rule base: When a certain type of scenario has a high frequency of violations, such as the increase in adverse reaction rate when elderly patients use high-risk drugs, the administrator can increase the corresponding coefficient through the rule base, such as from 1.2 to 1.4, to enhance the strictness of verification. (2) Collaboration with deviation threshold: The higher the risk coefficient, the easier it is for the deviation calculation result to exceed the warning threshold and lock threshold, thereby triggering the warning and lock mechanism earlier. Adjust the value of risk coefficient appropriately according to actual usage.

[0075] S5.3: Determine the deviation based on semantic similarity and risk coefficient.

[0076] In this embodiment, the formula for calculating the deviation is: Deviation = (1 - semantic similarity) × risk coefficient.

[0077] "1-Semantic similarity" represents the basic difference between preliminary treatment suggestions and guideline recommendations, with a value ranging from 0 to 1. The larger the value, the more significant the difference.

[0078] S5.4: Compare the deviation with the warning threshold and the lock threshold. If the deviation is greater than the warning threshold, a warning is triggered; if the deviation is greater than or equal to the lock threshold, a lock is triggered. For the same scenario, the lock threshold is greater than the warning threshold; therefore, if a lock is triggered, a warning will inevitably be triggered.

[0079] The formula for calculating the warning threshold is: T = base threshold × (1 + α + β) Where T is the warning threshold, α is the complication coefficient, and β is the age coefficient. The baseline threshold, complication coefficient, and age coefficient were determined based on initial treatment recommendations and the patient's condition.

[0080] Based on preliminary treatment recommendations and the patient's condition, scenarios are determined and categorized into routine medication, high-risk procedures, and special populations. Different scenarios correspond to different baseline thresholds. In this embodiment, the baseline thresholds are: 0.15 for routine medication, 0.05 for high-risk procedures, and 0.10 for special populations.

[0081] Definitions of routine medication use, high-risk procedures, and special populations: (1) Conventional medication: refers to oral medications that are clinically mature, have high safety, a clear dosage adjustment range, and controllable adverse reactions, and meet the following requirements: Drug type: Non-controlled, non-highly toxic drugs (such as common antihypertensive drugs, regular-dose hypoglycemic drugs, common antibiotics, etc.); Use cases: Maintenance treatment of chronic diseases (such as amlodipine 5mg qd for patients with hypertension, and glimepiride 2mg qd for patients with type 2 diabetes); Risk characteristics: Dosage error or improper short-term use will not lead to fatal adverse reactions (such as mild hypotension or blood sugar fluctuations).

[0082] (2) High-risk procedures: These refer to diagnostic and treatment behaviors that require strict adherence to guidelines due to the high toxicity of the drugs, the high invasiveness of the procedures, or the serious consequences of dosage errors, and must meet the following requirements: Procedure types include: high-risk drug use (such as warfarin, chemotherapy drugs, immunosuppressants), invasive procedures (such as insulin pump implantation, central venous catheterization), and off-label drug use; Risk characteristics: Incorrect execution may directly lead to serious adverse events (such as bleeding, organ damage, anaphylactic shock); Guidelines are based on the requirement that the drugs must meet the requirements of "double verification" and "mandatory monitoring indicators" in the "Administrative Measures for High Alert Drugs" and "Clinical Diagnosis and Treatment Operation Specifications".

[0083] (3) Special populations: These refer to groups with low drug tolerance and significantly increased risk of adverse reactions due to special physiological / pathological conditions, including: Age-related: ≥75 years old (liver and kidney function decline), ≤14 years old (organs not fully developed); Physiological status: Pregnant / lactating women (risk of drug-induced teratogenicity); Pathological conditions: multiple organ dysfunction (such as liver failure, end-stage renal disease), allergic constitution, malignant tumor patients, etc.; Key features: Special dosage adjustments and contraindications must be separately noted in the guidelines (e.g., the "Expert Consensus on the Safety Management of Medication Use in the Elderly in China" clearly states that the dosage of medication for the elderly should be reduced by 30%).

[0084] Handling mechanisms for cross-cutting scenarios or scenarios not classified as Class III: When a scenario meets the definitions of two or three categories, the approach is to "take the high-risk scenario as the base threshold and calculate the coefficients by adding them together": Example: A 76-year-old patient (special population, age coefficient β=0.3) using warfarin (high-risk procedure, baseline threshold 0.05) and also has heart failure (complication coefficient α=0.2). Early warning threshold calculation: T = Basic threshold for high-risk operations × (1 + α + β) = 0.05 × (1 + 0.2 + 0.3) = 0.05 × 1.5 = 0.075; Logic: The basic threshold for high-risk operations is more stringent (0.05 < 0.15 for routine medication). After adding the coefficient for special populations, the threshold is further reduced (0.075), making the system more rigorous in verifying cross-scenarios.

[0085] Handling mechanisms for non-Class III scenarios (extremely rare): This refers to scenarios that are neither routine medication use nor high-risk procedures, nor involve special populations, such as routine physical examination recommendations for healthy adults, vaccination guidance, etc. The handling rules are as follows: The preset "general basic threshold 0.2" is lenient for routine medication, and α=0, β=0 (no complications / age coefficient); The warning threshold T = 0.2 × (1 + 0 + 0) = 0.2, and the warning is triggered only when the deviation is greater than 0.2 (without locking the system), taking into account the efficiency of low-risk scenarios.

[0086] Complication coefficient α: 0.2 is added for each additional serious complication.

[0087] Serious complications specifically include: Cardiovascular system: Heart failure (NYHA class II or above), acute myocardial infarction, severe arrhythmia; Kidney system: CKD stage 4-5 (eGFR < 30), end-stage renal disease, diabetic nephropathy (massive proteinuria); Metabolic system: diabetic ketoacidosis, hyperosmolar hyperglycemic syndrome, recurrent episodes of severe hypoglycemia; Other: severe infection, advanced malignant tumor, liver failure (Child-Pugh C).

[0088] Age coefficient β: Add 0.3 for patients ≥75 years old.

[0089] The specific values ​​for the baseline threshold, complication coefficient, and age coefficient can be adjusted according to the actual situation.

[0090] Definition of Lockout Threshold: The lockout threshold is the critical value used to determine that "the recommended solution has serious security risks and must be suspended and forcibly reviewed." Its value is positively correlated with the risk coefficient. The higher the risk of the scenario, the lower the lockout threshold, meaning it is easier to trigger interception. The warning threshold triggers an alert, while the lockout threshold is the mandatory interception threshold, triggering a lockout. These two thresholds form a 'dual threshold control' system, and the lockout threshold is always higher than the warning threshold in the same scenario. The lockout threshold formula is expressed as: Lock-in threshold = Lock-in base threshold × Risk coefficient weight In this embodiment, the basic locking threshold is set to 0.3. The locking thresholds for different scenario risk levels are shown in Table 5.

[0091] Table 5

[0092] Dynamic adjustment mechanism of locking threshold: (1) Linkage with rule base: When a serious adverse event occurs in a certain scenario, such as the patient's death due to a deviation of 0.4 when using high-risk drugs, the administrator can lower the locking threshold of the corresponding scenario through the rule base, such as from 0.45 to 0.4, to enhance the interception sensitivity. (2) Synchronization with clinical guidelines: If the guidelines are updated and the risk level of a certain operation is redefined, such as the original "routine medication" being upgraded to "high-risk drug", the locking threshold will be automatically updated with the adjustment of the risk coefficient, such as from 0.3 to 0.45.

[0093] S6: Displays preliminary treatment recommendations and their verification results.

[0094] The verification results include the results of the three verifications, as well as information such as lockout, warning, and prompt.

[0095] Based on the generated treatment recommendations, this embodiment can also construct an interpretive atlas according to the medical knowledge and clinical guidelines used in the three verifications. The interpretive atlas links relevant clinical guideline clauses. Doctors can quickly view the corresponding guideline clauses by clicking on specific content in the preliminary treatment recommendations, understanding the basis and standards for the recommendations, facilitating rapid and accurate review of the recommendations, and improving the scientific rigor and reliability of medical decisions.

[0096] All preliminary treatment recommendations are accompanied by traceable guideline clause numbers. This design greatly facilitates clinical auditing. When referring to preliminary treatment recommendations, physicians can not only see the specific content of the recommendations but also clearly identify the specific clause number of the clinical guideline on which the preliminary treatment recommendations are based. During clinical audits, auditors can quickly trace back to the corresponding guideline clauses using these numbers to review whether the decision-making process complies with guideline requirements, improving the efficiency and accuracy of audits and enhancing the interpretability and transparency of medical decisions.

[0097] Interpretive atlases help reduce doctors' search time. When reviewing initial treatment recommendations, doctors no longer need to manually consult numerous clinical guidelines; they can quickly understand the basis and source of the recommendations simply by using interpretive atlases. This significantly saves search time, improves work efficiency, and allows doctors to focus more on diagnosing and optimizing patient treatment plans.

[0098] This embodiment uses a type 2 diabetes patient as an example to demonstrate the workflow of this method: Step 1: Convert symptom information into coded information.

[0099] 1) Voice input: "My fasting blood sugar was 10 mmol / L recently, and I took glimepiride." The input is converted to text via ASR, and the NLP module uses the BioBERT model to extract the entities: blood sugar level 10 mmol / L, and the drug "glimepiride". 2) Image input: Blood routine test report, OCR recognition and conversion into structured data: eGFR=58ml / min, HbA1c=8.2%; 3) Electronic medical record data: Previous diagnoses "Type 2 diabetes (ICD-10:E11.9)" and "Hypertension (ICD-10:I10)".

[0100] The above data is mapped to standardized codes: the drug "glimepiride" corresponds to the RxNorm code "104081", and the blood glucose value is associated with the "fasting blood glucose" index code in the Diabetes Guidelines.

[0101] Step 2: Generate preliminary suggestions using knowledge-enhanced LLM.

[0102] Knowledge-enhanced large language model (fine-tuned based on LLaMA-2-7B, training data includes the full text of the 2023 edition of the "Chinese Guidelines for Type 2 Diabetes" and 100,000 diabetes diagnosis and treatment cases) execution: 1) Using the RAG search engine: For the search guide fragment “eGFR=58+glimepiride+HbA1c=8.2%”, the result is “SGLT-2 inhibitors can be combined when HbA1c>7.5% and eGFR≥45”; 2) Recommendation generated: "Continue glimepiride 2mg qd, combined with dapagliflozin 10mg qd, fasting blood glucose target <7.2mmol / L, monitor blood glucose every 2 weeks."

[0103] Step 3: Three-level verification by the rule engine.

[0104] 1) Level 1 (Drug Validation): The rule engine calls the RxNorm drug database and inputs "glimepiride (104081) + dapagliflozin (153185)". It finds that there are no contraindications for either drug, but dapagliflozin requires monitoring of renal function, thus triggering Level 2 validation.

[0105] 2) Level 2 (Guideline Indicator Verification): The patient's eGFR=58ml / min was compared with the threshold of "eGFR≥45ml / min can be used" in the "Expert Consensus on Clinical Application of SGLT-2 Inhibitors", and it met the requirements; at the same time, the deviation of HbA1c=8.2% from the target value (<7.2%) was checked and recorded as "intervention indicator".

[0106] 3) Level 3 (Guideline Compliance Verification): Semantic Similarity: The semantic similarity between the LLM recommendation and the guideline recommendation calculated using the BERT model is 0.85 (the guideline recommends "combining SGLT-2 inhibitors when HbA1c is not up to standard"). The rule engine calls the "Guideline for Glimepiride Dosing Frequency" (recommended "qd") in the "Chinese Guidelines for Type 2 Diabetes" and compares it with the "Glimepiride 2mg qd" recommended by LLM. There is no difference in frequency, so semantic similarity is not deducted.

[0107] If the recommendation is "glimepiride 2mg bid" (100% different from the frequency of the guideline), then the semantic similarity is reduced by 0.3. The adjusted semantic similarity = initial value (assumed to be 0.8) - 0.3 = 0.5, which is then substituted into the deviation formula for further calculation.

[0108] Risk factor: Since the patient does not have severe renal insufficiency and the medication is used in a routine setting, the risk factor is set at 0.6.

[0109] Deviation = (1-0.85)×0.6 = 0.15×0.6 = 0.09.

[0110] The warning threshold is configured as "0.15 for routine medication". 0.15 > 0.09, so the guideline compliance check passes.

[0111] Step 4: Doctor confirmation and system feedback.

[0112] Preliminary treatment suggestions and explanatory graphs were pushed to the doctor's terminal, showing: "Deviation degree 0.09 < conventional medication threshold 0.15, the suggestion meets the guideline requirements and can be implemented directly." The doctor, considering the patient's individual situation of eGFR=58ml / min (mild renal function decline), independently optimized the dapagliflozin dose to 5mg qd (originally recommended 10mg qd) and confirmed it. The system recorded this proactive modification (not a mandatory adjustment) and fed the "initial dose preference for patients with mild renal insufficiency" into the rule base for future optimization of suggestions in similar scenarios.

[0113] This embodiment also provides an AI-based clinical decision suggestion generation system with mandatory compliance guarantees, such as... Figure 2 As shown, the technical solution adopted is as follows: including: input processing module 1, large language model module 2, rule engine verification module 3, and execution control module 4.

[0114] The input processing module is used to acquire the patient's symptom information and convert it into encoded information; The large language model module is used to encode information and generate preliminary diagnostic suggestions through the large language model; The rule engine verification module is used to perform drug verification on preliminary treatment recommendations, guideline indicator verification on preliminary treatment recommendations that pass drug verification, and guideline compliance verification on preliminary treatment recommendations that pass guideline indicator verification. The specific process of guideline compliance verification is as follows: calculate semantic similarity based on preliminary treatment recommendations and guideline recommended schemes; determine risk coefficient based on preliminary treatment recommendations and patient conditions; determine deviation based on semantic similarity and risk coefficient; compare the deviation with warning threshold and locking threshold; if the deviation is greater than the warning threshold, a warning is triggered; if the deviation is greater than or equal to the locking threshold, a locking is triggered. The execution control module is used to display preliminary diagnosis and treatment suggestions and their verification results.

[0115] For locked preliminary treatment recommendations, doctors can unlock them in the execution control module using biometric authentication (such as fingerprint recognition, facial recognition, etc.) combined with two-factor authentication methods such as employee ID, and then modify or regenerate the recommendations.

[0116] Example of locking and unlocking in a high-risk scenario: When the system processes a diabetic patient with an eGFR of 28, the large language model module mistakenly generates "metformin 500mg bid". The first-level check of the rule engine verification module finds that "eGFR < 30ml / min + metformin" matches the contraindications in the "Guidelines for the Use of Nephrology", with a deviation of 0.95, which is greater than the locking threshold, triggering a lock. The execution control module disables the "prescription" button on the electronic medical record. After receiving a warning, the doctor views the explanatory graph generated by the system, and a pop-up window displays the violated guidelines and specific clauses. The doctor needs to unlock the record through "fingerprint + employee ID" two-factor authentication. After unlocking, the doctor modifies the treatment plan to a more suitable insulin plan for the patient with impaired renal function based on the patient's actual situation and guideline requirements, ensuring the safety and effectiveness of the treatment. After the doctor unlocks the record, the system automatically pushes alternative solutions, such as insulin treatment recommendations, which the doctor can refer to when modifying the treatment plan. In addition, the system also records "lock time, unlocking doctor, and modification content" to form a traceable log for auditing purposes.

[0117] This invention is particularly applicable to decision support for chronic diseases requiring long-term management (such as diabetes and hypertension).

[0118] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; 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.

Claims

1. A method for generating AI-based clinical decision recommendations with mandatory compliance assurance, characterized in that, Includes the following steps: S1: Obtain the patient's symptom information and convert it into coded information; S2: Encoded information is used to generate preliminary diagnostic and treatment suggestions through a large language model; S3: Validate the medication recommendations based on the initial treatment plan; S4: Verify the guideline indicators for the preliminary treatment recommendations that have passed drug validation; S5: Conduct guideline compliance verification on the preliminary treatment recommendations that have passed the guideline indicator verification. The specific process is as follows: S5.1: Calculate semantic similarity based on preliminary treatment recommendations and guideline-recommended protocols; S5.2: Determine the risk factor based on preliminary treatment recommendations and the patient's condition; S5.3: Determine the degree of deviation based on semantic similarity and risk coefficient; S5.4: Compare the deviation with the warning threshold and the lock threshold. If the deviation is greater than the warning threshold, a warning is triggered. If the deviation is greater than or equal to the locking threshold, locking is triggered; S6: Displays preliminary treatment recommendations and their verification results.

2. The AI-based clinical decision suggestion generation method with mandatory compliance assurance as described in claim 1, characterized in that, In step S3, drug verification involves checking whether there are any drug incompatibilities in the preliminary treatment recommendations; If there are no drug incompatibilities, the drug verification is deemed successful, and the process proceeds to step S4. If drug incompatibility exists, the drug verification is deemed unsuccessful, triggering a lockout and proceeding to step S6.

3. The AI-based clinical decision suggestion generation method with mandatory compliance assurance as described in claim 1, characterized in that, In step S4, the process of verifying the guideline indicators is as follows: Based on the initial diagnosis and treatment recommendations, obtain the corresponding laboratory indicators and find the corresponding guideline safety thresholds; Compare laboratory indicators with guideline safety thresholds; if laboratory indicators meet guideline safety thresholds, proceed to step S5; if indicator data do not meet guideline indicator standards, increase the risk coefficient and proceed to step S5.

4. The AI-based clinical decision suggestion generation method with mandatory compliance assurance as described in claim 3, characterized in that, In step S4, if laboratory indicators cannot be obtained, it is determined whether the laboratory indicators are critical indicators. Critical indicators are those that are essential for diagnosis and treatment recommendations and directly determine the safety and compliance of the decision. If the laboratory indicator is a critical indicator, the indicator verification of the guideline is deemed to have failed, triggering a lock and proceeding to step S6. If the laboratory indicator is not a critical indicator, the indicator verification of the guideline is deemed to have passed, and the process proceeds to step S5.

5. The AI-based clinical decision suggestion generation method with mandatory compliance assurance as described in claim 1, characterized in that, In step S5.1, if the guidelines do not mention it, the semantic similarity is 0.5; if the guidelines explicitly do not recommend it, a lock is triggered, and the process proceeds to step S6.

6. The AI-based clinical decision suggestion generation method with mandatory compliance assurance as described in claim 1, characterized in that, In step S5.2, the baseline value of the risk coefficient is determined based on the operation type corresponding to the preliminary treatment recommendation; The patient population is determined based on the patient's condition, and then the first adjustment value for the risk factor is determined. The urgency of the disease is determined based on the patient's condition, and then a second adjustment value for the risk factor is determined. The risk coefficient is obtained by adding the base value of the risk coefficient, the first adjustment value of the risk coefficient, and the second adjustment value of the risk coefficient.

7. The AI ​​clinical decision suggestion generation method with mandatory compliance assurance as described in claim 6, characterized in that, The types of procedures include routine oral medication adjustments, injectable medication use, high-risk medication use, and recommendations for invasive procedures; The patient population includes elderly patients, pediatric patients, patients with multiple organ dysfunction, and pregnant / lactating women; The urgency of the disease includes management of the chronic stable phase, management of the acute exacerbation phase, and recommendations for critical care.

8. The AI-based clinical decision suggestion generation method with mandatory compliance assurance as described in claim 1, characterized in that, In step S5.4, the formula for calculating the warning threshold is: T = base threshold × (1 + α + β) Where T is the warning threshold, α is the complication coefficient, and β is the age coefficient. The baseline threshold, complication coefficient, and age coefficient were determined based on initial treatment recommendations and the patient's condition.

9. The AI-based clinical decision suggestion generation method with mandatory compliance assurance as described in claim 1, characterized in that, In step S5.4, the locking threshold is determined based on the risk coefficient.

10. An AI-powered clinical decision recommendation generation system with mandatory compliance guarantees, characterized in that, An AI-based clinical decision suggestion generation method for implementing mandatory compliance assurance as described in any one of claims 1 to 9 includes: an input processing module, a large language model module, a rule engine verification module, and an execution control module. The input processing module is used to acquire the patient's symptom information and convert it into encoded information; The large language model module is used to encode information and generate preliminary diagnostic suggestions through the large language model; The rule engine verification module is used to perform drug verification on preliminary treatment recommendations, guideline indicator verification on preliminary treatment recommendations that pass drug verification, and guideline compliance verification on preliminary treatment recommendations that pass guideline indicator verification. The execution control module is used to display preliminary diagnosis and treatment suggestions and their verification results.

Images