An intelligent guide diagnosis and medication decision system and method based on AI inference

Abstract

The application belongs to the technical field of guidance and medication management, and discloses an intelligent guidance and medication decision system and method based on AI reasoning, which comprises a data acquisition and coding module, an intelligent guidance reasoning module, a medication safety reasoning module, a medication safety reasoning module and a decision inertia deviation monitoring module, wherein the data acquisition and coding module is used for acquiring multi-source clinical data and coding the data into a patient feature vector; the intelligent guidance reasoning module is used for generating a first diagnosis hypothesis set and a department recommendation result; the medication safety reasoning module is used for generating a medication risk assessment result; a cross-consistency analysis module is used for constructing a guidance-medication cross-consistency matrix and calculating row and column cumulative scores; and the decision inertia deviation monitoring module is used for calculating a decision inertia deviation index, and generating a re-weighted diagnosis hypothesis set, department recommendation result correction information and medication risk warning information when the index exceeds a preset deviation threshold, so as to realize intelligent guidance and medication decision based on AI reasoning.

Description

Technical Field

[0001] This invention relates to the field of triage and medication management technology, and more specifically, to an intelligent triage and medication decision-making system and method based on AI reasoning. Background Technology

[0002] In smart healthcare systems, intelligent triage and clinical medication decision support are two core functions. Current technologies typically develop these as independent modules; however, this traditional architecture has a drawback: the decision-making inertia bias caused by the singular output of the triage process. This is specifically manifested as follows:

[0003] The core objective of intelligent triage is to improve triage efficiency. Its algorithms tend to output one or a few of the most likely departments (such as cardiology). This highly deterministic output provides doctors with a strong diagnostic suggestion, which, when reviewing medical records later, creates a preconceived notion, thus narrowing the scope of differential diagnosis. This cognitive focus triggered by the upstream intelligent triage output is the starting point for decision-making bias.

[0004] Medication decisions heavily rely on pre-existing diagnostic assumptions, and the boundaries of the problem are implicitly limited by the received intelligent triage output. Clinical medication decisions only involve drug recommendations, interaction reviews, and dosage calculations within the currently pre-defined diagnostic framework, failing to provide parallel evaluation and early warning for potentially drastically different medication regimens corresponding to a broader range of differential diagnoses. Consequently, an information simplification behavior arising from efficiency pursuits in the triage process is progressively transmitted and reinforced at each level of the clinical pathway, ultimately manifesting as potential limitations and risks in regimen selection at the medication decision-making stage.

[0005] Therefore, if intelligent triage and clinical medication decision-making are directly combined and stacked, the triage system, as an open-loop endpoint that "only speaks and doesn't listen," cannot know the actual impact of its output on downstream diagnosis and medication, nor can it use downstream treatment results (such as the final diagnosed disease, the efficacy or adverse reactions after medication) to back-verify and optimize its own reasoning. Medication decision-making, on the other hand, acts as a "passive response" tool, its knowledge base and rule engine unable to proactively backtrack and question the initial assumptions of the triage process. The two systems are linked in business logic but isolated in cognitive reasoning, causing decision-making inertia bias to become a unidirectional, cumulative, and silent error that cannot be detected or interrupted within the system. This not only reduces the overall value of AI-assisted healthcare but may also introduce new blind spots in medical safety.

[0006] Therefore, we design an intelligent triage and medication decision-making system and method based on AI reasoning. Summary of the Invention

[0007] To overcome the aforementioned deficiencies of the prior art and to achieve the above objectives, the present invention provides the following technical solution: an intelligent triage and medication decision-making system based on AI reasoning, comprising:

[0008] The data acquisition and encoding module is used to acquire multi-source clinical data and encode it into patient feature vectors.

[0009] The intelligent triage reasoning module is used to generate a first diagnostic hypothesis set and department recommendation results based on medical knowledge graphs and multi-path reasoning networks.

[0010] The medication safety reasoning module is used to generate medication risk assessment results for the first diagnostic hypothesis set.

[0011] The cross-consistency analysis module is used to construct the triage-medication cross-consistency matrix and calculate the cumulative row and column scores;

[0012] The decision inertia bias monitoring module is used to calculate the decision inertia bias index and generate a reweighted diagnostic hypothesis set, departmental recommendation result correction information, and medication risk warning information when the index exceeds a preset bias threshold.

[0013] Preferably, the multi-source clinical input data includes patient complaint text, structured records of symptoms and signs, past medical history, medication history, and basic examination results;

[0014] Based on the symptom keywords in the patient's chief complaint text and their related duration, triggers, and relieving factors, the structured records of symptoms and signs, past medical history, and medication history were analyzed and feature extracted to obtain feature sub-vectors.

[0015] The feature vector is fused with symptom keywords, duration, triggers and relief factors to obtain the patient feature vector, which is used to comprehensively reflect the patient's condition.

[0016] Preferably, the generation of the first diagnostic hypothesis set and department recommendation results based on medical knowledge graphs and multi-path reasoning networks includes:

[0017] The patient feature vector is input into the medical knowledge graph. The symptom keywords in the vector are used as the starting point for the query. The associated causes and relief factors are used as relational constraints. Disease-symptom association edges, symptom-cause association edges or relief association edges are used as multi-hop reasoning and retrieval paths. A subset of candidate diseases containing multiple possibilities is output. Based on the disease-department affiliation relationship, the subset of candidate departments corresponding to the subset of candidate diseases is retrieved simultaneously.

[0018] The patient feature vector and the subset of candidate diseases are used as input to a multi-path reasoning network, and the diagnosis probability of each candidate disease is used as output. All candidate diseases and their diagnosis probabilities together constitute the first diagnostic hypothesis set.

[0019] Based on the weighted disease-department affiliation in the medical knowledge graph, one or more departments and their corresponding weights are determined for each candidate disease. The department confidence score is calculated by weighted summation based on the diagnostic probabilities of each candidate disease in the first diagnostic hypothesis set. The final output is a list sorted in descending order of confidence score, where each item contains a candidate department name and its corresponding confidence score. This list is the department recommendation result.

[0020] The entity nodes of the medical knowledge graph include at least disease entities, symptom entities, department entities, and examination item entities; symptom entities correspond to symptom keywords; and the relationship edges of the graph include at least disease-symptom association edges, symptom-cause or relief association edges, disease-department affiliation relationships, and disease-examination item recommendation relationships.

[0021] Preferably, the multi-path reasoning network is configured to include an input layer, a parallel diagnostic path layer, and an output layer. The parallel diagnostic path layer deploys N independent disease evaluators based on the number N types of diseases in the candidate disease subset. The input layer is used to copy the patient feature vector N times, which are used as the inputs of the N disease evaluators respectively.

[0022] Each disease evaluator is an independent multilayer perceptron activated by the ReLU function. Its structure includes a fully connected layer, a dropout layer, an output layer, and a competitive normalization layer. The output layer outputs a scalar score, and the competitive normalization layer inputs the scalar scores output by all disease evaluators into a Softmax layer, which is then converted into a diagnostic probability.

[0023] When pre-training the multi-path inference network, each training sample contains a patient feature vector and a true confirmed disease label. The input is set as the patient feature vector, and the output is set as the diagnosis probability. A cross-entropy loss function is used to measure the difference between the diagnosis probability and the true confirmed disease label. The training method is iterative.

[0024] The training completion criterion is set as follows: when the disease in the patient's feature vector outputs the highest scalar score in the corresponding disease evaluator, and other disease evaluators output lower scalar scores, after competitive normalization, the diagnosis probability corresponding to the scalar score is close to 1, and the others are close to 0, then training is complete.

[0025] Preferably, for each candidate disease in the first diagnostic hypothesis set, a list of standard treatment drugs matching the disease is obtained by searching in a pre-built drug knowledge base, and drugs with absolute contraindications are screened out by combining the patient's past medical history and medication history, and a preliminary set of candidate drugs for each candidate disease is output.

[0026] For each drug or drug combination in each preliminary candidate drug set, a drug risk score is output based on drug-drug interaction rules, drug-disease aggravation risk rules, and liver and kidney function dose adjustment rules.

[0027] Based on the medication risk score, a classification rule is set to divide the candidate medication regimens corresponding to each candidate disease into low-risk, medium-risk, or high-risk levels, and the medication risk assessment results corresponding to each candidate disease in the first diagnostic hypothesis set are output.

[0028] Preferably, the triage-medication cross-consistency matrix is ​​constructed by taking each candidate disease diagnosis in the first diagnostic hypothesis set as a row and each candidate medication regimen in the medication risk assessment results corresponding to the candidate disease as a column.

[0029] For each cell in the matrix, the consistency score of the cell is based on the diagnosis probability of the candidate disease corresponding to the row in which the cell is located, and the risk level of the candidate treatment plan corresponding to the column in which the cell is located. The principle followed when calculating the consistency score is that the score increases positively with the increase of the diagnosis probability of the candidate disease and decreases negatively with the increase of the risk level of the candidate treatment plan.

[0030] The rules for obtaining the cumulative consistency score of each candidate disease and the cumulative consistency score of each candidate medication regimen are set as follows: the sum of the consistency scores of the rows and columns of all units in the triage-medication cross-consistency matrix.

[0031] Preferably, the decision inertia bias index is configured by combining the diagnostic probability concentration index and the medication conflict intensity index into a combination function. The configuration principle of the combination function is that the value of the decision inertia bias index increases as the diagnostic probability distribution becomes more concentrated and the cross-diagnostic high-risk medication conflict becomes stronger.

[0032] The diagnostic probability concentration index is set as a monotonic function value that is positively correlated with the degree of confirmed diagnosis concentration, wherein the entropy value of the probability distribution is set based on the confirmed diagnosis probability of each candidate disease in the first diagnostic hypothesis set;

[0033] Based on the cumulative consistency score of the regimen, among the drug risk assessment results corresponding to multiple candidate diseases, the candidate drug regimens listed as high-risk are set as common high-risk drug regimens. The drug conflict intensity index of this regimen is configured as the number of different candidate diseases appearing in the drug risk assessment results of the common high-risk drug regimen, combined with the drug risk score and the degree to which the cumulative consistency score of the regimen is lower than the preset safety threshold.

[0034] Preferably, when the decision inertia deviation index exceeds a preset deviation threshold, the diagnostic hypothesis reweighting process is initiated; after this adjustment, a new diagnostic probability distribution is output and set as the reweighted diagnostic hypothesis set.

[0035] Based on the disease-department affiliation relationship, the reweighted diagnostic hypothesis set is mapped to department confidence levels to generate a new department recommendation list. When this list is compared with the original department recommendation results, the departments corresponding to candidate diseases with increased probabilities, as well as newly added or improved departments, are set as correction information for the department recommendation results.

[0036] Configure medication risk alert information: specific drug name, risk nature, and associated diagnostic changes;

[0037] The alarm message generation rules are set to take effect in any of the following situations:

[0038] Scenario 1: For a certain candidate disease, the risk level of the corresponding candidate treatment plan under the original first diagnostic hypothesis set is low or medium risk. However, when extrapolating based on the reweighted diagnostic hypothesis set, the overall risk assessment of the plan changes to high risk or increased risk due to changes in the probability of diagnosis of the disease or the probability of other related diseases.

[0039] Scenario 2: For identified common high-risk drug regimens, if they are in the reweighted diagnostic hypothesis set, the probability of a confirmed diagnosis of the associated disease as the standard treatment target decreases, while the probability of a confirmed diagnosis of the associated disease as a high-risk or contraindicated target increases.

[0040] Preferably, the adjustment rule for initiating the diagnostic hypothesis reweighting process is set as follows:

[0041] If candidate diseases that are strongly associated with common high-risk drug regimens are identified based on the cumulative score of disease consistency, their diagnostic probability will be reduced.

[0042] Meanwhile, if a candidate disease is identified with a moderate cumulative disease consistency score and has little association with common high-risk drug regimens, its diagnostic probability will be increased.

[0043] The technical effects and advantages of the AI-based intelligent triage and medication decision-making method of this invention are as follows:

[0044] By using multi-hop retrieval and multi-path parallel reasoning networks based on medical knowledge graphs, a probabilistic set of first diagnostic hypotheses covering multiple possibilities is forcibly generated, broadening the initial diagnostic perspective and suppressing the risk of cognitive narrowing caused by information simplification from the source, thereby improving the comprehensiveness of diagnosis.

[0045] By constructing a triage-medication cross-consistency matrix, a positive correlation is established between diagnostic probability and medication risk, enabling proactive assessment of potential medication risks. This allows medication safety early warnings to no longer rely on a confirmed diagnosis, but rather to intervene in advance based on probability.

[0046] The decision inertia bias index reflects the concentration of diagnostic thinking and the intensity of cross-diagnostic medication conflicts. It automatically triggers a diagnostic hypothesis reweighting process, lowering the probability of diseases strongly associated with high-risk drugs and raising the probability of safer alternative diseases, while simultaneously generating departmental recommendation corrections and precision medication alerts. It possesses the capabilities of self-monitoring, self-questioning, and proactive intervention, upgrading a static AI tool into a dynamic intelligent advisor.

[0047] The entire triage and medication process ensures the transparency and explainability of the reasoning process. Every recommendation can be traced back to specific rules and relationships in the knowledge base, enhancing clinicians' trust in AI suggestions and facilitating system auditing and optimization. Attached Figure Description

[0048] Figure 1 This is a schematic diagram of the AI-based intelligent triage and medication decision-making system of the present invention;

[0049] Figure 2 This is a schematic diagram of the structure of the multi-path inference network in this invention. Detailed Implementation

[0050] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0051] In smart healthcare systems, existing technologies typically develop these two as independent modules: intelligent triage uses natural language processing to analyze patient complaints and output department recommendations; medication decision-making, on the other hand, provides drug review and recommendations based on electronic medical record (EHR) data during the diagnosis and treatment process.

[0052] For example, if a patient presents with "chest tightness and shortness of breath," and the triage system points to "cardiovascular medicine" with high confidence, the doctor may prioritize ruling out angina and heart failure, while relatively underestimating the possibility of respiratory diseases (such as COPD) or even psychological factors (such as panic disorder). If the actual cause is COPD, but the diagnostic thinking is anchored in the cardiovascular direction, there is a possibility that a strong risk warning will not be issued for the beta-blocker to be used (which may aggravate bronchospasm), because under the current diagnostic framework, this warning has not been triggered by the rule priority setting.

[0053] Please see Figure 1 and Figure 2 In this embodiment of the invention, an intelligent triage and medication decision-making method based on AI reasoning includes:

[0054] S1. Obtain multi-source clinical input data of patients seeking medical treatment, and obtain patient feature vectors through feature encoding;

[0055] S2. Receive patient feature vectors and generate department recommendation results containing multiple candidate clinical departments and their confidence levels based on medical knowledge graphs and multi-path reasoning networks, as well as a first diagnostic hypothesis set corresponding to multiple candidate disease diagnoses and their probability distributions;

[0056] S3. For each candidate disease in the first diagnostic hypothesis set, generate corresponding candidate medication regimens, assess the risk level of each candidate medication regimen based on drug use restrictions, and construct medication risk assessment results.

[0057] S4. Based on the diagnostic probability and medication risk assessment results of the candidate diseases in the first diagnostic hypothesis set, construct the triage-medication cross-consistency matrix to obtain the cumulative disease consistency score for each candidate disease and the cumulative medication consistency score for each candidate medication regimen.

[0058] S5. Based on the cumulative score of disease consistency and the cumulative score of medication regimen consistency, calculate the decision inertia bias index, which represents the degree of dependence of the triage output on the single diagnostic path in medication decision-making.

[0059] S6. When the decision inertia deviation index exceeds the preset deviation threshold, the probability distribution of candidate disease diagnosis in the first diagnostic hypothesis set is reweighted to generate a reweighted diagnostic hypothesis set, and department recommendation result correction information and medication risk warning information are generated based on the reweighted diagnostic hypothesis set.

[0060] The multi-source clinical input data includes patient complaint text, structured records of symptoms and signs, past medical history, medication history, and basic examination results;

[0061] Among them, the patient's chief complaint text refers to the patient's main discomfort and purpose of seeking medical treatment described in words, such as: "I have been experiencing recurrent dizziness and headaches for a week, and my vision has been a little blurry for the past two days. When I measured my blood pressure, I found that it was very high."

[0062] Structured symptom and sign records refer to symptom and sign data entered by healthcare professionals into a medical record system using standardized templates (such as checkboxes and drop-down menus). Their characteristics include fixed fields and standardized value ranges. Example:

[0063] Symptoms (patient's subjective feelings): such as headache (location: top of the forehead, nature: throbbing pain, degree: moderate).

[0064] Physical signs (observed by the doctor): such as blood pressure: 180 / 110 mmHg, heart rate: 92 beats / min, and neurological examination: normal muscle strength and muscle tone.

[0065] Past medical history refers to a patient's systematic medical history that occurred in the past and is relevant to their current health. Examples include: history of chronic diseases (hypertension, diabetes, etc.), history of infectious diseases, history of surgery or trauma, history of allergies (especially drug allergies), and history of vaccinations.

[0066] Medication history refers to a list of all medications a patient currently uses regularly or intermittently, both recently and recently. It generally includes the generic name, strength, single dose, frequency of administration, and start date of use. Example: Nifedipine controlled-release tablets 30mg once daily, taken for 2 years; occasionally takes ibuprofen to relieve headaches.

[0067] Basic examination results refer to routine and rapid auxiliary examination data that can be obtained in the early stages of outpatient or emergency care. Common types include:

[0068] Laboratory tests: complete blood count, urinalysis, rapid blood glucose, liver and kidney function tests, and electrolyte tests.

[0069] Instrumental examinations: electrocardiogram, chest X-ray, and simplified ultrasound report.

[0070] Collecting these five types of data forms the data foundation for multi-path reasoning and cross-validation. The above data can be collected automatically from the patient's existing structured records, past medical history, medication history and examination reports through standardized interfaces (such as HL7, FHIR) from the Hospital Information System (HIS), Electronic Medical Record (EMR), Laboratory Information System (LIS) and Picture Archiving and Communication System (PACS). Alternatively, patients or medical staff can enter their current chief complaint, symptoms and signs in real time through guided questionnaires, structured forms and standardized medical terminology search and selection.

[0071] The system extracts symptom keywords and their associated duration, triggers, and relieving factors from the patient's chief complaint text. Specifically, it uses its built-in NER (Non-Expert Analyzer) function to identify and label medical entities in the text. For example, inputting "chest tightness and shortness of breath after activity for 3 months, relieved by rest" will output the label sequence: [Trigger]: after activity; [Symptoms] chest tightness, shortness of breath; [Duration]: 3 months; [Relieving Factor]: after rest. The identified medical entities (such as chest tightness) are linked to unique concept codes in a standard medical terminology database (such as UMLS) (e.g., chest tightness corresponds to concept code C0039239). Duration is also parsed into standardized values ​​(e.g., 3 months is converted to duration_days:90), and triggers and relieving factors are also subject to concept normalization.

[0072] The structured records of symptoms and signs, medical history, and medication history are analyzed and their features are extracted to obtain feature sub-vectors. Specifically, each structured entry in the symptom and sign records, medical history, and medication history (e.g., blood pressure: 180 / 110 mmHg) is treated as an (attribute, value) pair. Numerical values ​​(e.g., blood pressure value) can be directly normalized and used as scalar features. Categorical values ​​(e.g., headache nature: pulsating, medical history: hypertension) are transformed by looking up a predefined medical concept embedding table, which is a predefined mapping form after expert or data analysis. All (attribute, value) pairs are then combined to form the feature sub-vector.

[0073] The extracted symptom keywords, duration, triggers, and alleviating factors are fused and encoded with the feature sub-vectors to form a comprehensive symptom-temporal feature vector reflecting the patient's condition, denoted as the patient feature vector. Specifically, assuming N symptom keywords are extracted, each converted into a vector S_i, and each symptom is associated with a set of temporal attribute vectors T_i (encoded by duration, triggers, etc.), for each symptom S_i, it is concatenated with its own temporal attribute T_i to obtain an enhanced symptom representation E_i=[S_i;T_i], where the symptom-temporal feature sub-vector = Σ(α_i×E_i), and the attention weight α_i = softmax(C^T×V×E_i), where C^T is the transpose of the feature sub-vector C, and V is a learnable parameter matrix with the same number of rows as C and the same number of columns as E_i.

[0074] Background information indicates that traditional triage, in pursuit of efficiency, tends to output one or a few departments with high confidence levels, creating strong diagnostic cues. Doctors, in subsequent diagnoses, unconsciously focus their information gathering around this single output, narrowing the scope of differential diagnoses. Furthermore, the simplified output of triage directly limits the framework for subsequent medication decisions, making it impossible to conduct concurrent medication safety assessments for a broader range of differential diagnoses.

[0075] S2 includes:

[0076] A1: Pre-construct a medical knowledge graph. The entity nodes of this graph should include at least disease entities, symptom entities, department entities, and examination item entities. Symptom entities correspond to symptom keywords. Furthermore, the graph's relational edges should include at least disease-symptom association edges, symptom-cause or relief association edges, disease-department affiliation relationships, and disease-examination item recommendation relationships. Example: A medical knowledge graph is a large medical semantic network containing tens of thousands of entities and relationships. Assume the structure of a relevant part in a practical application is as follows:

[0077] Entity nodes:

[0078] Disease entities: stable angina, chronic obstructive pulmonary disease, anxiety disorder, heart failure;

[0079] Symptoms: chest tightness, shortness of breath, chest pain, difficulty breathing; (corresponding to symptom keywords);

[0080] Departmental Entities: Cardiology, Respiratory Medicine, Psychology, Emergency Medicine;

[0081] Physical examination items: electrocardiogram, coronary CTA, pulmonary function test, anxiety self-rating scale;

[0082] Examples of relational edge parts:

[0083] Disease-symptom association: Stable angina → chest tightness, Stable angina → shortness of breath, Chronic obstructive pulmonary disease → shortness of breath

[0084] Anxiety disorder → Chest tightness;

[0085] Symptom-triggered or relieved association: Chest tightness → activity [triggered or aggravated by activity], Chest tightness → rest [relieved by rest], Shortness of breath → activity [triggered or aggravated by activity];

[0086] Disease-Departmental Affiliation: Stable angina pectoris → Cardiology, Chronic obstructive pulmonary disease → Respiratory Medicine, Anxiety disorder → Psychiatry.

[0087] A2: Input the patient feature vector into the medical knowledge graph, take the symptom keywords in the vector as the query starting point, and combine them with the associated causes and relief factors as relational constraints. Perform multi-hop reasoning and retrieval along the disease-symptom association edge, symptom-cause association edge or relief association edge to obtain a candidate disease subset containing multiple possibilities, and simultaneously retrieve the candidate department subset corresponding to the candidate disease subset based on the disease-department affiliation relationship.

[0088] A3. Input the patient feature vector and the candidate disease subset into a multi-path reasoning network to obtain the diagnosis probability of each candidate disease. Combine all candidate diseases and their diagnosis probabilities to form the first diagnostic hypothesis set.

[0089] A4. Using the weighted disease-department affiliation relationships in the medical knowledge graph, determine one or more departments mapped to each candidate disease and their corresponding weights; based on the diagnostic probabilities of each candidate disease in the first diagnostic hypothesis set, calculate the department confidence score using a weighted summation method (for a department, its confidence score is the sum of the products of the diagnostic probabilities of all diseases associated with that department and the affiliation weights of the corresponding diseases to that department); the final output is a list sorted in descending order of confidence score, where each item contains a candidate department name and its corresponding confidence score, and this list is the department recommendation result.

[0090] Specifically, A2-A4: Extract symptom keywords [chest tightness, shortness of breath] and time sequence information "triggered by activity, relieved by rest" from the patient feature vector.

[0091] Starting with chest tightness, we search backwards along the disease-symptom association edge to obtain the associated disease set: {stable angina, anxiety disorder, ...}.

[0092] Starting with shortness of breath, we can similarly derive: {stable angina, chronic obstructive pulmonary disease, heart failure, ...}.

[0093] By taking the intersection and union of the sets and combining them with the frequency of symptom occurrence, a preliminary set of diseases can be obtained.

[0094] The time-series information "induced by activity and relieved by rest" is used as a filter.

[0095] Output candidate disease subset: [Stable angina, chronic obstructive pulmonary disease, anxiety disorder].

[0096] Simultaneously, based on the disease-department affiliation relationship, the corresponding candidate department subsets are retrieved: [Cardiovascular Medicine, Respiratory Medicine, Psychology].

[0097] Input the patient feature vector and candidate disease subsets [stable angina, chronic obstructive pulmonary disease, anxiety disorder] into the multi-path inference network. Output the diagnostic probability distribution.

[0098] Stable angina: 0.65;

[0099] Chronic obstructive pulmonary disease: 0.30;

[0100] Anxiety disorder: 0.05;

[0101] Based on the disease-department affiliation relationship defined in A1, establish the mapping:

[0102] Stable angina (0.65) → Cardiology Department;

[0103] Chronic obstructive pulmonary disease (0.30) → Respiratory Medicine;

[0104] Anxiety disorder (0.05) → Department of Psychiatry;

[0105] Since each disease uniquely maps to a department, the confidence level of a department is directly equal to the probability of its associated disease.

[0106] The confidence level for the cardiology category is 0.65.

[0107] The confidence level for respiratory medicine is 0.30.

[0108] The confidence level for the psychology department is 0.05.

[0109] (Note: If multiple diseases are mapped to the same department, the confidence level is the sum of the probabilities of these diseases.)

[0110] The results will be organized into a list of recommended results for each department and output.

[0111] A3 includes:

[0112] The multi-path reasoning network is configured to include an input layer, a parallel diagnostic path layer, and an output layer. The parallel diagnostic path layer deploys N independent disease evaluators based on the number of disease types N in the candidate disease subset. The input layer is used to copy the patient feature vector N times (N is the number of disease types in the candidate disease subset) and use them as the inputs of the N disease evaluators respectively.

[0113] Each disease evaluator is an independent multilayer perceptron activated by the ReLU function. Its structure includes fully connected layers (used to map input patient features from a high-dimensional space to a hidden feature space, performing feature combination; multiple layers can be stacked according to the learning depth required in actual production, such as adding another fully connected layer to further extract higher-order features), dropout layers (used to randomly discard a portion of neuron outputs during training to prevent overfitting; specifically, a dropout probability p1 is set, for example, to 0.3; during training: for each element of the ReLU function output, probability p1 is set to 0, and the remaining elements are divided by 1−p1 to maintain the expected value), an output layer, and a competitive normalization layer. The output layer (maps hidden features to a single scalar score, representing the probability of the disease) outputs a scalar score, which can be understood as the raw evidence strength that the patient has the disease corresponding to the evaluator. The competitive normalization layer inputs the scalar scores output by all disease evaluators into a Softmax layer, which is then converted into a diagnostic probability.

[0114] When pre-training the multi-path inference network, each training sample contains a patient feature vector and a true confirmed disease label (in one-hot form, e.g., the label corresponding to disease Dj is a vector with the j-th bit set to 1). The patient feature vector is input to obtain the diagnosis probability. The cross-entropy loss function is used to measure the difference between the diagnosis probability and the true confirmed disease label. Through repeated iterative training, when the disease in the patient feature vector outputs the highest scalar score in the corresponding disease evaluator, and other disease evaluators output lower scalar scores, after competitive normalization, the diagnosis probability corresponding to this scalar score is close to 1, and the others are close to 0, and the training is complete.

[0115] By forcibly deploying N independent disease evaluators, the architecture physically ensures that the system must consider multiple disease probabilities in parallel, thus counteracting the highly deterministic output format in the background technology. By generating probability distributions, this provides the necessary, structured input for the medication risk assessment based on multiple candidate disease diagnoses envisioned in the background technology. During network training, the output score corresponding to the correct disease is maximized, enabling the system to simulate the thought process of a senior physician proposing multiple differential diagnoses and ranking them by probability. This upgrades the single goal of triage efficiency to the goal of supporting prudent diagnosis.

[0116] Example: Patient profile: A 58-year-old male patient presented with the chief complaint of "chest tightness and shortness of breath after exertion for 3 months, relieved by rest." He has a 5-year history of hypertension and has been taking metoprolol long-term. Initial auscultation revealed clear breath sounds in both lungs, and the electrocardiogram showed nonspecific ST-T changes. A patient characteristic vector is generated.

[0117] Starting with symptoms of chest tightness and shortness of breath, a multi-hop search is performed in the knowledge graph, combining the effects of activity-induced symptoms and rest-relieving symptoms. Assuming the candidate disease subset for the search results is: {stable angina pectoris, chronic obstructive pulmonary disease, anxiety disorder}, the candidate department subset is obtained simultaneously: {cardiovascular medicine, respiratory medicine, psychology}.

[0118] Based on the three diseases mentioned above, a multi-path inference network containing three independent disease evaluators is dynamically deployed.

[0119] The input layer copies the patient feature vector three times. Three evaluators operate simultaneously: the angina evaluator receives the vector and calculates the raw score s_cardiac=2.1; the chronic obstructive pulmonary disease (COPD) evaluator receives the same vector and calculates the raw score s_copd=0.8; and the anxiety evaluator calculates the raw score s_anxiety=-1.2. Competitive normalization is performed by inputting [2.1, 0.8, -1.2] into the Softmax layer, yielding probabilities P(angina) = exp(2.1) ÷ (exp(2.1) + exp(0.8) + exp(-1.2)) ≈ 0.65, P(COPD) ≈ 0.30, and P(anxiety) ≈ 0.05. The output first diagnostic hypothesis set is {(stable angina, 0.65), (COPD, 0.30), (anxiety, 0.05)}.

[0120] By introducing a combined design of medical knowledge graphs and multi-path reasoning networks, the triage process is transformed from a potentially biased starting point into an intelligent reasoning node capable of outputting rich differential diagnostic information and supporting subsequent in-depth cross-validation. This effectively addresses all the systemic deficiencies pointed out in the background technology and is the core technological guarantee for achieving intelligent coupling.

[0121] The S3 includes:

[0122] For each candidate disease in the first diagnostic hypothesis set, a list of standard treatment drugs matching the disease is retrieved from a pre-built drug knowledge base. Combined with the patient's medical history and medication history, drugs with absolute contraindications are screened out, resulting in a preliminary set of candidate drugs for each candidate disease. The pre-built drug knowledge base is a structured database that specifically stores standardized knowledge related to drugs. Its core purpose is to provide the system with rules for drug indications, contraindications, interactions, and usage and dosage for reasoning. Data sources can include official drug instructions, etc. The structured database includes core entities and relationships and attributes. Core entities include drugs (generic name, brand name), diseases (aligned with disease entities in the medical knowledge graph), adverse reactions, organ function status (such as liver function classification, kidney function staging), etc. Relationships and attributes include drug-indication relationships (which diseases a drug can be used to treat, with a recommendation level indicated, such as first-line, second-line, etc.), drug-contraindication relationships (which diseases or patient states are absolutely or relatively prohibited from using a drug, such as propranolol being contraindicated in asthma patients), drug-drug interaction relationships and mechanisms (which are the possible pharmacodynamic or pharmacokinetic interactions when two or more drugs are used in combination, with a risk level indicated, such as severe, moderate, mild), and drug-organ function dose adjustment rules (which are specific dose adjustment recommendations for patients with liver or kidney dysfunction, such as halving the dose when creatinine clearance is 30-50 mL / min).

[0123] Rule matching and filtering: For each standard treatment drug (such as metoprolol or aspirin) retrieved for a candidate disease (e.g., stable angina), the system queries the drug knowledge base for the absolute contraindications of that drug.

[0124] The contraindications were matched one by one with the patient's past medical history and current medication history:

[0125] If a patient's medical history includes a disease that is absolutely contraindicated for the drug (e.g., asthma), the drug will be removed from the candidate drug set.

[0126] If a patient's current medication history shows a serious and unavoidable interaction with the candidate drug (e.g., warfarin and aspirin together increase the risk of major bleeding), and there is no effective monitoring or adjustment plan, the candidate drug will be marked as having a contraindication and may be removed.

[0127] If a patient is allergic to a certain type of medication or a specific ingredient (such as sulfonamides), all medications containing that ingredient should be removed.

[0128] After the above filtering, the remaining list of drugs without any known absolute contraindications constitutes the preliminary candidate drug set for the candidate disease. This preliminary candidate drug set is a structured list; for each candidate disease in the first diagnostic hypothesis set, there is a corresponding independent preliminary candidate drug set. Each preliminary candidate drug set for a given disease typically contains one or more treatment regimens. A treatment regimen may include a single drug or a combination of drugs.

[0129] For each drug or combination of drugs in each preliminary candidate drug set, based on the drug-drug interaction rule [a rule that when two or more drugs are used simultaneously, one drug affects the absorption, distribution, metabolism, excretion, or efficacy of another drug, thereby leading to reduced efficacy or increased toxicity. Each record in the rule includes the interacting drug pair, the type of interaction, the severity (e.g., severe, moderate, mild), clinical manifestations (e.g., increased bleeding risk, QT interval prolongation), and management recommendations (e.g., avoid combined use, monitor INR, adjust dosage)], and the drug-disease exacerbation risk rule [a rule that a certain drug may exacerbate a patient's existing specific disease or pathological state. This is different from an absolute contraindication and is usually a relative risk]. Each record in the rules includes the drug, the disease / state that may be aggravated (e.g., NSAIDs may aggravate chronic kidney disease, beta-blockers may aggravate heart failure), the risk level (e.g., high, medium, low), and usage recommendations (e.g., use with caution, monitor renal function)], as well as liver and kidney function dosage adjustment rules [which are rules for adjusting the dosage or dosing interval of a specific drug based on the patient's liver or kidney function level (quantified by examination indicators such as creatinine clearance, Child-Pugh classification, etc.). Each record in the rules includes the drug, the applicable organ (liver / kidney), the functional stage (e.g., CKD stages 1-5, liver function Child-Pugh A-C), and the corresponding specific dosage adjustment plan (e.g., no adjustment required, 25% reduction, contraindicated)], and calculates its medication risk score;

[0130] Based on the medication risk score, the candidate medication regimens corresponding to each candidate disease are classified into low-risk, medium-risk, or high-risk levels, forming a medication risk assessment result that corresponds one-to-one with each candidate disease in the first diagnostic hypothesis set.

[0131] Specifically, the medication risk score is a comprehensive quantitative indicator. First, a base risk score is set for each of the three rules, assuming...

[0132] Drug-drug interactions: Severe = 3 points, Moderate = 2 points, Mild = 1 point, None = 0 points;

[0133] Drug-disease exacerbation risk: High risk = 3 points, Medium risk = 2 points, Low risk = 1 point, None = 0 points;

[0134] Dosage adjustment requirements for liver and kidney function: Large adjustment required (e.g., reduction >50% or prohibition) = 3 points, moderate adjustment required (reduction 25-50%) = 2 points, slight adjustment required (reduction <25%) = 1 point, no adjustment required = 0 points.

[0135] Different weights are assigned based on the importance of different risk types (weighting coefficients can be set based on clinical expert consensus). For example: drug interaction weight (W1) = 1.2, disease exacerbation risk weight (W2) = 1.5 (potentially more critical), dose adjustment weight (W3) = 1.0;

[0136] For a given medication regimen, the formula for calculating the medication risk score is:

[0137] Risk score = (drug-drug interaction score × W1) + (drug-disease exacerbation risk score × W2) + (liver and kidney function dose adjustment requirement score × W3);

[0138] Set a threshold to discretize continuous risk scores into levels:

[0139] Low risk: Risk score ≤ 2.0;

[0140] Medium risk: 2.0 < Risk score ≤ 5.0;

[0141] High risk: Risk score > 5.0.

[0142] The method in step S4 includes:

[0143] Take each candidate disease diagnosis in the first diagnostic hypothesis set as a row, and each candidate medication plan in the medication risk assessment results corresponding to that candidate disease as a column, and construct a matrix, denoted as the triage-medication cross-consistency matrix;

[0144] For each cell in the matrix, a consistency score is calculated based on the probability of diagnosis of the candidate disease corresponding to the row in which the cell is located and the risk level of the candidate treatment plan corresponding to the column in which the cell is located. The principle followed when calculating the consistency score is that the score increases positively with the increase of the probability of diagnosis of the candidate disease and decreases negatively with the increase of the risk level of the candidate treatment plan.

[0145] For all cells in the triage-medication cross-consistency matrix, sum the consistency scores of the rows and columns respectively to obtain the cumulative disease consistency score for each candidate disease and the cumulative consistency score of each candidate medication regimen.

[0146] For example, suppose the first diagnostic hypothesis set is: {Stable angina: 0.65, Chronic obstructive pulmonary disease: 0.30, Anxiety disorder: 0.05};

[0147] Medication risk assessment results (This is a simplified explanation; it assumes only one treatment plan for each disease, but in reality, there may be multiple plans, resulting in a larger matrix):

[0148] Angina pectoris: Protocol S1 (Risk level: High risk);

[0149] Chronic obstructive pulmonary disease: Protocol C1 (Risk level: Medium risk);

[0150] Anxiety disorder: Option A1 (Risk level: Low risk);

[0151] Construct a 3x3 matrix, then,

[0152] Line: Illness = [Angina pectoris, COPD, anxiety disorder];

[0153] Column: Scheme = [S1, C1, A1];

[0154] Example formula for calculating consistency score: Consistency score = probability of diagnosis × (1 - risk weight), with pre-set high risk weight = 0.8, medium risk = 0.4, and low risk = 0.1.

[0155] Calculate the consistency score for each element in the matrix:

[0156] (Angina pectoris, S1) = 0.65 × (1 - 0.8) = 0.13;

[0157] (Chronic obstructive pulmonary disease, C1) score = 0.30 × (1 - 0.4) = 0.18;

[0158] (Anxiety disorder, A1) score = 0.05 × (1 - 0.1) = 0.045;

[0159] Other cross-disease-treatment units (such as (angina pectoris, C1)) do not correspond directly, so their scores can be set to 0 or a very low baseline value. The lower the score, the higher the probability that a disease is matched with a high-risk treatment, i.e., high inconsistency.

[0160] By row and (cumulative score of disease consistency):

[0161] The total number of cases of angina pectoris is calculated as follows: 0.13 + 0 + 0 = 0.13.

[0162] The total for chronic obstructive pulmonary disease is calculated as follows: 0 + 0.18 + 0 = 0.18.

[0163] The total number of anxiety-related cases is calculated as follows: 0 + 0 + 0.045 = 0.045.

[0164] By column and (cumulative score for scheme consistency):

[0165] The sum of column S1 in scheme S1 is 0.13 + 0 + 0 = 0.13;

[0166] The sum of column C1 in the scheme is 0 + 0.18 + 0 = 0.18;

[0167] The sum of column A1 in scheme A1 is 0 + 0 + 0.045 = 0.045;

[0168] Disease scores [0.13, 0.18, 0.045] show that the most probable angina (0.65) had the lowest concordance score (0.13), and the cumulative score of the high-risk regimen S1 was also the lowest.

[0169] S5 includes:

[0170] Based on the probability of diagnosis of each candidate disease in the first diagnostic hypothesis set, the entropy value of the probability distribution is calculated. The entropy value is used to measure the degree of dispersion of the diagnostic hypothesis. The smaller the entropy value, the more concentrated the diagnostic probability distribution is, and the higher the risk of path dependence on a single diagnosis. The monotonic function value based on the positive correlation between the entropy value and the degree of concentration of diagnosis (e.g., the reciprocal of the entropy value) is used as the indicator of the concentration of diagnostic probability.

[0171] Based on the cumulative consistency score of the regimen, among the medication risk assessment results corresponding to multiple candidate diseases, the candidate medication regimens listed as high-risk are identified as common high-risk drug regimens; the number of different candidate diseases appearing in the medication risk assessment results of the common high-risk drug regimen is counted, and combined with the medication risk score and the degree to which the cumulative consistency score of the regimen is lower than the preset safety threshold, the medication conflict intensity index of the regimen is obtained.

[0172] By combining the diagnostic probability concentration index and the medication conflict intensity index, a combination function is configured to obtain the decision inertia bias index. The configuration principle of the combination function includes that the value of the decision inertia bias index increases as the diagnostic probability distribution becomes more concentrated and the cross-diagnostic high-risk medication conflict becomes stronger.

[0173] For example: Based on the above, we have obtained the following:

[0174] First diagnostic hypothesis set:

[0175] Stable angina (D1): diagnostic probability P1=0.65; Chronic obstructive pulmonary disease (D2): diagnostic probability P2=0.30; Anxiety disorder (D3): diagnostic probability P3=0.05.

[0176] Medication risk assessment results:

[0177] For D1 (angina pectoris): Candidate medication regimen S1 (including drugs: aspirin, statin, metoprolol) was assessed as high risk, with a specific medication risk score of R1=8.5.

[0178] For D2 (COPD): Candidate medication regimen C1 (including medication: long-acting bronchodilator), risk level is medium risk, medication risk score R2=3.2.

[0179] For D3 (anxiety disorder): Candidate medication regimen A1 (including medication: SSRI), risk level is low, medication risk score R3=1.5.

[0180] A 3x3 triage-medication cross-consistency matrix has been constructed, and the consistency score for each unit has been calculated.

[0181] The cumulative score for scheme consistency is calculated, with high-risk = 0.8, medium-risk = 0.4, and low-risk = 0.1.

[0182] The cumulative score of scheme S1 is Score_S1=0.13;

[0183] The cumulative score of scheme C1 is Score_C1=0.18;

[0184] The cumulative score of scheme A1 is Score_A1=0.045;

[0185] Calculate the concentration index of diagnostic probability:

[0186] The probability distribution of confirmed cases is P = [0.65, 0.30, 0.05].

[0187] The entropy value H is calculated using the Shannon entropy formula (base 2): H = -Σ(P×log2(P))≈-(0.65×log2(0.65)+0.30×log2(0.30)+0.05×log2(0.05))≈1.141;

[0188] The concentration index of diagnostic probability = (1-H÷H_max)=1-(1.141÷1.585)≈1-0.720≈0.280, where H_max=log2(3)≈1.585. The closer this value is to 1, the more concentrated the diagnostic probability distribution is (the worse the diversity). 0.280 indicates that there is a certain degree of concentration.

[0189] If the cumulative consistency score of the scheme is lower than the preset safety threshold (e.g., 0.15) and the risk level is high in the corresponding medication risk assessment results.

[0190] Recognition process:

[0191] If Score_S1=0.13<0.15 and the risk level of scheme S1 is high risk, then it is marked as a candidate conflict scheme.

[0192] Score_C1=0.18>0.15, which does not meet the requirement.

[0193] Score_A1=0.045<0.15, but the risk level of scheme A1 is low risk, so it is not marked.

[0194] Therefore, S1 was identified as a potential common high-risk drug regimen;

[0195] A search of the drug knowledge base revealed that metoprolol, the core risk drug in regimen S1, has contraindications or high-risk associated diseases including asthma and chronic obstructive pulmonary disease (COPD). In the current patient's primary diagnostic hypothesis set, D2 (COPD) is one of the candidate diseases (probability 0.30). Regimen S1 is a treatment option for candidate disease D1 (angina pectoris), but poses a high risk for another candidate disease, D2 (COPD). Therefore, regimen S1 belongs to the category of common high-risk drug regimens.

[0196] The method for calculating the drug conflict intensity index of scheme S1 is as follows:

[0197] First, S1 is clearly associated with D1 in the medication risk assessment results, and also poses a high risk to D2. Therefore, the number of different candidate diseases appearing in the medication risk assessment results for this common high-risk drug regimen is 2. After normalization, 2 / 3 ≈ 0.667 = F1, where 3 refers to the total number of diseases.

[0198] Secondly, referring to the example of medication risk scoring methods, the medication risk score R1 of scheme S1 under D1 is 8.5. Assuming that its medication risk score for D2 (COPD) is 7.0 after inference from the knowledge base, the average risk score = (8.5 + 7.0) ÷ 2 = 7.75. After normalization to a maximum score of 10, 7.75 ÷ 10 = 0.775 = F2.

[0199] Example of the degree to which the cumulative consistency score of a scheme is lower than the preset safety threshold: Scheme S1 has a consistency score of Score_S1=0.13, which is lower than the safety threshold of 0.15. Therefore, the degree quantification is (0.15-0.13)÷0.15≈0.133=F3.

[0200] Example of calculating the intensity index of drug conflict, with weights set as G1=0.4, G2=0.4, and G3=0.2.

[0201] Medication conflict intensity index = G1×F1+G2×F2+G3×F3=0.4×0.667+0.4×0.775+0.2×0.133≈0.604;

[0202] The configured combination function, using linear weighted combination, needs to ensure that the index is sensitive to both diagnostic concentration and conflict intensity. Setting β1=0.5 and β2=0.5, then the decision inertia bias index = β1 × diagnostic probability concentration index + β2 × medication conflict intensity index, and the decision inertia bias index = 0.5 × 0.280 + 0.5 × 0.604 = 0.140 + 0.302 = 0.442.

[0203] S6 includes:

[0204] When the decision inertia bias index exceeds the preset bias threshold, the diagnostic hypothesis reweighting process is initiated; after this adjustment, a new diagnostic probability distribution is formed, which serves as the reweighted diagnostic hypothesis set.

[0205] Based on the disease-department affiliation relationship, the reweighted diagnostic hypothesis set is mapped to department confidence levels to generate a new department recommendation list. This list is compared with the original department recommendation results to identify the departments corresponding to candidate diseases with increased probabilities, as well as newly added or improved departments, which are highlighted or given additional explanations to form correction information for the department recommendation results.

[0206] The system generates medication risk warnings, including specific drug names, risk characteristics, and associated diagnostic changes. These warnings are generated for any of the following scenarios: For a candidate disease, the corresponding candidate medication regimen under the original first diagnostic hypothesis set has a low or medium risk level. However, when extrapolating based on a reweighted diagnostic hypothesis set, changes in the probability of diagnosis of this disease or changes in the probabilities of other related diseases cause the overall risk assessment of the regimen to change to high risk or increased risk. [For example, using the data above, under the original hypothesis set, the probability of diagnosis for D2 is 0.30, and the risk level of regimen C1 is medium risk. Under the reweighted hypothesis set, the probability of diagnosis for D2 significantly increases from 0.30 to 0.45. This increase in probability means that the patient is more likely to have severe airflow limitation. For the long-acting β2-adrenergic agonist (LABA) in regimen C1, when used in patients with severe COPD, caution should be exercised regarding cardiovascular risks such as arrhythmias, especially if the patient has a history of hypertension and is taking metoprolol.] Therefore, based on the new diagnostic probabilities and patient background, the risk level of the C1 regimen may be upgraded to high risk after reassessment. Thus, [if the original medium-risk regimen meets the condition of being upgraded to high risk after risk reassessment due to a significant increase in the probability of associated diseases, triggering an alarm], or, for identified common high-risk drug regimens, if in the reweighted diagnostic hypothesis set, the diagnostic probability of associated diseases as standard treatment targets decreases, while the diagnostic probability of associated diseases as high-risk or contraindicated targets increases. [Under these conditions, in common high-risk drug regimens, β-blockers, in their original state:]

[0207] Indication-related disease: D1 (angina pectoris), probability 0.65. Metoprolol is one of its standard treatment drugs, but it was assessed as high risk in this case.

[0208] Risk contraindication associated disease: D2 (COPD), probability 0.30, metoprolol may aggravate bronchospasm.

[0209] Changes after reweighting:

[0210] The probability of D1 (angina pectoris) decreased from 0.65 to 0.50.

[0211] The probability of D2 (COPD) increased from 0.30 to 0.45.

[0212] This is a typical example of "adjustments that are not conducive to safe medication use," because the probability of disease corresponding to the potential benefits of using metoprolol (treatment of angina) is decreasing, while the probability of disease corresponding to the main risks of using it (exacerbation of COPD) is increasing significantly. The overall risk-benefit balance is further tilted towards the risk side, and the probability of beneficial diseases of common high-risk drug regimens is decreasing, while the probability of risky diseases is increasing.

[0213] The generated alarm information includes:

[0214] Drugs: β-blockers (such as metoprolol);

[0215] Risk characteristics: This drug may induce or worsen bronchospasm and is contraindicated or requires extreme caution in COPD patients;

[0216] Related diagnostic changes: the probability of its treatment indication (angina pectoris) has been reduced from 65% to 50%, while the probability of its contraindication (COPD) has been increased from 30% to 45%.

[0217] The diagnostic hypothesis reweighting process includes:

[0218] Based on the cumulative score of disease consistency, candidate diseases that are strongly associated with common high-risk drug regimens are identified (meaning that in the drug risk assessment results, the candidate drug regimens that contain common high-risk drug regimens and whose risk level is high are found). The probability of their diagnosis is then reduced (reduction means reducing the probability of the candidate disease according to a preset reduction rule; for example, by proportional reduction or ranking reduction).

[0219] Meanwhile, candidate diseases with a cumulative disease consistency score at a moderate level and little association with common high-risk drug regimens are identified (meaning that a lower and upper threshold for the cumulative disease consistency score are set, and candidate diseases with scores within this range are judged to be at a moderate level; at the same time, it is determined that the candidate drug regimens corresponding to the candidate disease do not contain common high-risk drug regimens, or although they do contain them, the risk level of the corresponding regimens is not higher than that of medium risk), and their diagnosis probability is adjusted upward (e.g., by proportional adjustment or by ranking adjustment).

[0220] The clinical interactive interface displays the department's recommended results, correction information, and risk warnings to users, and receives user commands.

[0221] Example 2, please refer to Figure 2 As shown, parts not described in detail in this embodiment are described in Embodiment 1. This embodiment provides an AI-based intelligent triage and medication decision-making system, including:

[0222] The data acquisition and encoding module is used to acquire multi-source clinical data and encode it into patient feature vectors.

[0223] The intelligent triage reasoning module is used to generate a first diagnostic hypothesis set and department recommendation results based on medical knowledge graphs and multi-path reasoning networks.

[0224] The medication safety reasoning module is used to generate medication risk assessment results for the first diagnostic hypothesis set.

[0225] The cross-consistency analysis module is used to construct the triage-medication cross-consistency matrix and calculate the cumulative row and column scores;

[0226] The decision inertia bias monitoring module is used to calculate the decision inertia bias index and generate a reweighted diagnostic hypothesis set, departmental recommendation result correction information, and medication risk warning information when the index exceeds a preset bias threshold.

[0227] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0228] It should be noted that all formulas in this manual are calculated by removing dimensions and taking their numerical values. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0229] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. An intelligent triage and medication decision-making system based on AI reasoning, characterized in that, include: The data acquisition and encoding module is used to acquire multi-source clinical data and encode it into patient feature vectors. The intelligent triage reasoning module is used to generate a first diagnostic hypothesis set and department recommendation results based on medical knowledge graphs and multi-path reasoning networks. The medication safety reasoning module is used to generate medication risk assessment results for the first diagnostic hypothesis set. The cross-consistency analysis module is used to construct the triage-medication cross-consistency matrix and calculate the cumulative row and column scores; The decision inertia bias monitoring module is used to calculate the decision inertia bias index and generate a reweighted diagnostic hypothesis set, departmental recommendation result correction information, and medication risk alarm information when the index exceeds a preset bias threshold. The decision inertia bias index is configured as a combination function of the diagnostic probability concentration index and the medication conflict intensity index. When the decision inertia bias index exceeds the preset bias threshold, the diagnostic hypothesis reweighting process is initiated. After this adjustment, a new confirmed diagnosis probability distribution is output and set as the reweighted diagnostic hypothesis set. The triage-medication cross-consistency matrix is ​​constructed by taking each candidate disease diagnosis in the first diagnostic hypothesis set as a row and each candidate medication plan in the medication risk assessment results corresponding to the candidate disease as a column. For each cell in the matrix, the consistency score of the cell is based on the diagnosis probability of the candidate disease corresponding to the row in which the cell is located, and the risk level of the candidate treatment plan corresponding to the column in which the cell is located. The principle followed when calculating the consistency score is that the score increases positively with the increase of the diagnosis probability of the candidate disease and decreases negatively with the increase of the risk level of the candidate treatment plan. For each candidate disease, the sum of the consistency scores of all row units in the triage-medication cross-consistency matrix is ​​the cumulative disease consistency score, and the sum of the consistency scores of all column units is the cumulative scheme consistency score. The configuration principle of the combined function is as follows: the value of the decision inertia bias index increases as the diagnostic probability distribution becomes more concentrated and the cross-diagnostic high-risk medication conflict becomes stronger; The diagnostic probability concentration index is set as a monotonic function value that is positively correlated with the degree of confirmed diagnosis concentration, wherein the entropy value of the probability distribution is set based on the confirmed diagnosis probability of each candidate disease in the first diagnostic hypothesis set; Based on the cumulative consistency score of the regimen, among the medication risk assessment results corresponding to multiple candidate diseases, the candidate medication regimens listed as high-risk are set as common high-risk drug regimens; the number of different candidate diseases appearing in the medication risk assessment results of this common high-risk drug regimen is counted, and combined with the medication risk score and the degree to which the cumulative consistency score of the regimen is lower than the preset safety threshold, the medication conflict intensity index of the regimen is obtained.

2. The AI-based intelligent triage and medication decision-making system according to claim 1, characterized in that, The multi-source clinical input data includes patient complaint text, structured records of symptoms and signs, past medical history, medication history, and basic examination results; Based on the symptom keywords in the patient's chief complaint text and their related duration, triggers, and relieving factors, the structured records of symptoms and signs, past medical history, and medication history were analyzed and feature extracted to obtain feature sub-vectors. The feature vector is fused with symptom keywords, duration, triggers and relief factors to obtain the patient feature vector, which is used to comprehensively reflect the patient's condition.

3. The AI-based intelligent triage and medication decision-making system according to claim 2, characterized in that, The generation of the first diagnostic hypothesis set and department recommendation results based on medical knowledge graphs and multi-path reasoning networks includes: The patient feature vector is input into the medical knowledge graph. The symptom keywords in the vector are used as the starting point for the query. The associated causes and relief factors are used as relational constraints. Disease-symptom association edges, symptom-cause association edges or relief association edges are used as multi-hop reasoning and retrieval paths. A subset of candidate diseases containing multiple possibilities is output. Based on the disease-department affiliation relationship, the subset of candidate departments corresponding to the subset of candidate diseases is retrieved simultaneously. The patient feature vector and the subset of candidate diseases are used as input to a multi-path reasoning network, and the diagnosis probability of each candidate disease is used as output. All candidate diseases and their diagnosis probabilities together constitute the first diagnostic hypothesis set. Based on the weighted disease-department affiliation relationship in the medical knowledge graph, one or more departments mapped to each candidate disease and their corresponding weights are determined; according to the diagnostic probability of each candidate disease in the first diagnostic hypothesis set, the department confidence score is calculated by weighted summation; the final output is a list sorted in descending order of confidence score, where each item contains a candidate department name and its corresponding confidence score, and this list is the department recommendation result. The entity nodes of the medical knowledge graph include at least disease entities, symptom entities, department entities, and examination item entities; symptom entities correspond to symptom keywords; and the relational edges of the graph include at least disease-symptom association edges, symptom-cause or relief association edges, disease-department affiliation relationships, and disease-examination item recommendation relationships.

4. The AI-based intelligent triage and medication decision-making system according to claim 3, characterized in that, The multi-path reasoning network is configured to include an input layer, a parallel diagnostic path layer, and an output layer. The parallel diagnostic path layer deploys N independent disease evaluators based on the number of disease types N in the candidate disease subset. The input layer is used to copy the patient feature vector N times, which are used as the inputs of the N disease evaluators respectively. Each disease evaluator is an independent multilayer perceptron activated by the ReLU function. Its structure includes a fully connected layer, a dropout layer, an output layer, and a competitive normalization layer. The output layer outputs a scalar score, and the competitive normalization layer inputs the scalar scores output by all disease evaluators into a Softmax layer, which is then converted into a diagnostic probability. When pre-training the multi-path inference network, each training sample contains a patient feature vector and a true confirmed disease label. The input is set as the patient feature vector, the output is set as the diagnosis probability, and the cross-entropy loss function is set to measure the difference between the diagnosis probability and the true confirmed disease label. The training method is iterative. The training completion criterion is set as follows: when the disease in the patient's feature vector outputs the highest scalar score in the corresponding disease evaluator, and other disease evaluators output lower scalar scores, after competitive normalization, the diagnosis probability corresponding to the scalar score is close to 1, and the others are close to 0, then training is complete.

5. The AI-based intelligent triage and medication decision-making system according to claim 4, characterized in that, For each candidate disease in the first diagnostic hypothesis set, a list of standard treatment drugs matching the disease is obtained by searching in a pre-built drug knowledge base. Combined with the patient's past medical history and medication history, drugs with absolute contraindications are screened out, and a preliminary set of candidate drugs for each candidate disease is output. For each drug or drug combination in each preliminary candidate drug set, a drug risk score is output based on drug-drug interaction rules, drug-disease aggravation risk rules, and liver and kidney function dose adjustment rules. Based on the medication risk score, a classification rule is set to divide the candidate medication regimens corresponding to each candidate disease into low-risk, medium-risk, or high-risk levels, and the medication risk assessment results corresponding to each candidate disease in the first diagnostic hypothesis set are output.

6. The AI-based intelligent triage and medication decision-making system according to claim 5, characterized in that, The reweighted diagnostic hypothesis set also includes; Based on the disease-department affiliation relationship, the reweighted diagnostic hypothesis set is mapped to department confidence levels to generate a new department recommendation list. When this list is compared with the original department recommendation results, the departments corresponding to candidate diseases with increased probabilities, as well as newly added or improved departments, are set as correction information for the department recommendation results. Configure medication risk alert information: specific drug name, risk nature, and associated diagnostic changes; The alarm message generation rules are set to take effect in any of the following situations: Scenario 1: For a certain candidate disease, the risk level of the corresponding candidate treatment plan under the original first diagnostic hypothesis set is low or medium risk. However, when extrapolating based on the reweighted diagnostic hypothesis set, the overall risk assessment of the plan changes to high risk or increased risk due to changes in the probability of diagnosis of the disease or the probability of other related diseases. Scenario 2: For identified common high-risk drug regimens, if they are in the reweighted diagnostic hypothesis set, the probability of a confirmed diagnosis of the associated disease as the standard treatment target decreases, while the probability of a confirmed diagnosis of the associated disease as a high-risk or contraindicated target increases.

7. The AI-based intelligent triage and medication decision-making system according to claim 6, characterized in that, The adjustment rules for initiating the diagnostic hypothesis reweighting process are set as follows: If candidate diseases that are strongly associated with common high-risk drug regimens are identified based on the cumulative score of disease consistency, their diagnostic probability will be reduced. Meanwhile, if a candidate disease is identified with a moderate cumulative disease consistency score and has little association with common high-risk drug regimens, its diagnostic probability will be increased.

Images