AI medical guidance and hospital guidance system based on structured multi-round dialogues

Abstract

The invention discloses an AI hospital guide system based on structured multi-round dialogues, and relates to the technical field of artificial intelligence medical hospital guide. Comprising an initial information standardization module, an ICD-11-based dynamic diagnosis reasoning module, a humanized interaction processing module and a medical resource recommendation module which are sequentially subjected to data interaction. The initial information standardization module converts non-standard symptom expression of a patient into ICD-11 standard structured data through a built-in disease-symptom civil dictionary and large model reasoning; the dynamic diagnosis reasoning module completes disease preliminary screening, probability dynamic calculation and decision optimization through multiple rounds of dialogues on the basis of an ICD-11 extended data set, and outputs a target department set; the user-friendly interaction module extracts emotion features and emergency signals of the patient and generates an adaptive reply; and the medical resource recommendation module combines department strength and doctor authority to adjust a recommendation weight according to an emergency level. The system solves the problem of pain points such as incomplete traditional hospital guide information acquisition, the average interaction round is less than or equal to 2.5, the department recommendation accuracy is greater than or equal to 95%, and accurate and humanized hospital guide is realized.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of artificial intelligence medical guidance, and particularly relates to an AI medical guidance system based on structured multi-round dialogue. BACKGROUND

[0002] In the medical service system in China, the complexity of the outpatient treatment process and the asymmetry of the medical knowledge of patients constitute the core contradiction, and the existing medical guidance technology has four core pain points:

[0003] 1. Incomplete information collection: the traditional form-based pre-diagnosis system requires patients to master medical terminology, which is disconnected with the natural language expression habits of ordinary people, resulting in ambiguous symptom description and missing of key information; the manual customer service mode has high service cost and low response efficiency, and cannot meet the large-scale concurrent consultation demand. 2. Inaccurate diagnosis reasoning: the single-round dialogue diagnosis model is limited by the completeness of the one-time information input of the patient, and it is difficult to deal with the progressive elimination of complex symptoms; the semantic gap between medical professional terms and patient natural language is not bridged, the symptom understanding deviation rate is high, and high-risk diseases are easily missed. 3. Non-humanized interactive experience: the system reply is mechanical and cold, lacks emotional adaptation ability, and cannot identify the emotions such as anxiety and pain of patients; no emergency level grading response mechanism is established, and the critical situation cannot be timely warned, which aggravates the anxiety of patients in seeking medical treatment. 4. Non-professional resource recommendation: the medical resource recommendation only takes distance as the core, and does not combine the key factors such as the clinical strength of the hospital department and the academic authority of the doctor, resulting in low matching degree of the recommended resources and the patient's condition, and the patient is easy to "blindly seek medical treatment and waste a trip".

[0004] In view of the above-mentioned pain points, an intelligent medical guidance system integrating standardized information conversion, dynamic accurate diagnosis, humanized interaction and professional resource recommendation is urgently needed to solve the technical bottlenecks of the existing technology. SUMMARY

[0005] The purpose of the present application is to avoid the shortcomings of the prior art and provide an AI medical guidance system based on structured multi-round dialogue.

[0006] The above-mentioned purpose of the present application is realized by the following technical means:

[0007] An AI medical guidance system based on structured multi-round dialogue, comprising an initial information standardization module, an ICD-11-based dynamic diagnosis reasoning module, a humanized interaction processing module and a medical resource recommendation module which interact in sequence;

[0008] The initial information standardization module is internally provided with a disease-symptom civilian dictionary, which contains the mapping relationship between non-standard terms of domestic diseases and symptoms and ICD-11 standard terms, after receiving the non-standard symptom expression of the patient, the symptom correlation is verified in combination with the large model reasoning, and the structured symptom data conforming to the ICD-11 standard is output;

[0009] The dynamic diagnosis reasoning module completes disease preliminary screening, probability dynamic calculation and decision optimization through multiple rounds of dialogue based on the ICD-11 extended data set, and outputs a target department set;

[0010] The humanized interaction processing module extracts emotional features and emergency signals from patient expressions and generates adaptive interaction replies;

[0011] The medical resource recommendation module dynamically adjusts the recommendation weight according to the emergency level by combining the strength of the hospital department and the academic authority of the doctor, and outputs a medical resource list.

[0012] Preferably, the initial information standardization module is built-in with a disease-symptom civilian dictionary, which contains the mapping relationship between non-standard terms of domestic diseases and symptoms and ICD-11 standard terms;

[0013] The patient's non-standard expression S2={p1,s1} is mapped to standardized medical data S through the dictionary and large model reasoning, where p1 is a description word and s1 is a symptom word, and the precise alignment of S2 and S is realized. idc idc

[0014] Preferably, the disease preliminary screening of the dynamic diagnosis reasoning module adopts a bidirectional coverage principle:

[0015] Forward matching: the disease must contain the core features of all current symptoms;

[0016] Reverse exclusion: exclude diseases that require "non-appearance of specific symptoms" as a basis for judgment;

[0017] After preliminary screening, a candidate disease set D cand ={d1,d2,…d k} is generated, where d is a single candidate disease and k is the number of candidate diseases.

[0018] Preferably, the probability dynamic calculation of the dynamic diagnosis reasoning module is adjusted through three layers of weights:

[0019] Epidemiological weight: adjust the seasonal epidemic probability according to public health data;

[0020] Patient portrait correction: adjust the disease probability based on age, gender and medical history;

[0021] Symptom-specific reinforcement: assign a 3-fold weight to ICD-labeled "strongly associated" symptoms;

[0022] The disease probability is calculated through a dynamic Bayesian network, and the department posterior probability formula is:

[0023] where:​​

[0024] P(c|S t ):based on the tth round symptom set S t ; d e P -1 (c) : the set of all diseases mapped to department c; P(S t |d) : the likelihood of symptom set S t given disease d; P(d) : the prior probability of disease d; S t : the symptom set extracted from the tth round dialogue, A i is the ith round patient answer.

[0025] Preferably, the decision optimization of the dynamic diagnostic reasoning module adopts an entropy minimization strategy, and the question selection algorithm determines the next round of follow-up questions by maximizing information gain, with the information gain formula being: I(D; Q | H t ) = H(D | H t ) - E A [H(D | H t+1 )], wherein:

[0026] I(D; Q | H t ): information gain of follow-up question Q based on tth round dialogue history H t ; H(D | H t ): conditional entropy of disease distribution after tth round dialogue;

[0027] E A [H(D | H t+1 ): expected value of conditional entropy of disease distribution after t+1th round dialogue for all possible answers A of question Q;

[0028] The system average interaction round is less than or equal to 2.5 rounds, and the department recommendation accuracy is greater than or equal to 95%.

[0029] Preferably, the dynamic diagnostic reasoning module comprises a high-risk disease alert mechanism, which triggers an intensive follow-up mode when disease d satisfies P(d | S t ) < 5% and HazardLevel(d) is greater than or equal to a high-risk threshold, and verifies the ICD-11 marked necessary symptoms of d in priority; the follow-up intensity is adjusted by a clinical utility index, with the formula being: CUI(d) = P(d | S t ) x log(HazardLevel(d)), wherein:

[0030] CUI(d): clinical utility index of disease d; P(d | S t ): probability of disease d under tth round symptom set; HazardLevel(d): hazard level of disease d (value 0-1, the higher the value, the greater the hazard);

[0031] When CUI(d) > threshold α, a confirmed diagnosis-oriented question is generated; when threshold β < CUI(d) ≤ α, a differential diagnosis question is triggered; and when CUI(d) ≤ threshold β, a routine symptom screening is performed.

[0032] Preferably, the humanized interaction processing module extracts three-dimensional emotional features through LLMs

[0033] E = [e1, e2, e3] ∈ [0, 1] 3 where e1 is an anxiety index, e2 is a pain level, and e3 is an emergency perception, all with a value range of [0, 1]; the reply tone is dynamically adjusted based on emotional speech mapping rules, and when e1 ≥ 0.7, the reply contains empathetic soothing statements; meanwhile, a four-level emergency classification response is adopted, with different symbols and colors of text to strengthen warnings.

[0034] Further, the humanized interaction processing module has a built-in regional culture sensitive word library to optimize expressions for different regions.

[0035] When a semantic conflict is detected, the contradiction index is verified:

[0036] Contradiction index = |number of negative words - number of high-risk symptoms| / total number of symptoms.

[0037] If the contradiction index > 0.5, a caring follow-up question is triggered to confirm the patient's actual condition.

[0038] Preferably, the hospital department strength evaluation of the medical resource recommendation module is based on the Fudan Hospital Rankings, and the scoring formula is:

[0039] Where:

[0040] HospitalScore(h, c): the comprehensive score of hospital h department c;

[0041] Fudan Specialty Ranking(h, c): Fudan Specialty Ranking conversion score of hospital h department c;

[0042] Research and academic ability score(h, c): research and clinical ability score of hospital h department c;

[0043] Departments with Fudan Specialty Ranking conversion score and research and clinical ability score both ≥ 85 points are marked as "regional advantage specialty".

[0044] Further, the doctor authority evaluation of the medical resource recommendation module adopts an academic position and title weighting model, and the academic score formula is: Where:

[0045] AcademicScore: doctor academic score; λ iThe weight of the i-th academic position is as follows: λ = 1.2 for national-level academic organizations, λ = 0.8 for provincial-level organizations, and λ = 0.6 for editorial board members of authoritative journals; PositionLevel(p i ): The i-th academic rank; n: The number of academic ranks held by the doctor;

[0046] The doctor's overall authority score = academic score + professional title level coefficient, where the professional title is: Academician / National Master of Traditional Chinese Medicine 1.5, Chief Physician 1.2, Associate Chief Physician 1.0, Attending Physician 0.7);

[0047] The formula for prioritizing medical resource recommendations is:

[0048] Priority(h,d)=α·e -0.02·distance +β·HospitalScore(h,c)+γ·AcademicLevel(d), where: Priority(h,d): the recommendation priority of hospital h and doctor d; α,β,γ: weight coefficients, summed to 1, dynamically adjusted according to the urgency level.

[0049] Red Emergency: α = 0.7, β = 0.3, γ = 0; Orange Emergency: α = 0.5, β = 0.4, γ = 0.1; Yellow Normal: α = 0, β = 0.6, γ = 0.4; Blue Warning: α = 0, β = 0.5, γ = 0.5; AcademicLevel(d): Doctor d's overall authority score.

[0050] The beneficial effects of adopting the above technical solution are as follows:

[0051] 1. This invention addresses the pain points of traditional triage systems, such as incomplete information collection, inaccurate diagnostic reasoning, unfriendly interaction, and unprofessional resource recommendations. It automates the conversion of patients' natural language into ICD-11 standard data, supporting large-scale concurrency. Multi-turn dialogue improves the accuracy of department recommendations and strengthens awareness of high-risk diseases. Emotional adaptation and tiered alerts alleviate patient anxiety. Precise matching of "condition-hospital-department-doctor" reduces blind medical visits. 2. It bridges the semantic gap between patients' non-standard expressions and medical terminology. It avoids information omissions in traditional form-based systems, eliminating the need for patients to master medical terminology. Automated processes replace manual customer service, supporting large-scale concurrency. It provides high-quality ICD-11 standard data for subsequent diagnostic reasoning, ensuring the accuracy of source data. 3. It ensures that candidate diseases fully cover patients' core symptoms, with no diseases lacking core features included. It accurately excludes irrelevant diseases that rely on "no specific symptoms," reducing interference. It considers both common diseases and low-probability high-risk diseases (such as aortic dissection), avoiding missed diagnoses. The screening process is objective, providing a reliable candidate range for subsequent reasoning. 4. Accurately correct disease probabilities through three-layer weight adjustment (epidemiology, patient profile, symptom specificity); effectively identify low-probability high-risk diseases (e.g., aortic dissection probability increases from 4% to 55%); dynamic Bayesian network calculation provides reliable probability support for department recommendations; improve the accuracy and scenario adaptability of diagnostic reasoning for different diseases. 5. Average interaction rounds ≤ 2 rounds, below the ≤ 2.5 round threshold, reducing patient interaction costs; department recommendation accuracy reaches 100%, exceeding the ≥ 95% requirement; prioritizes the identification of high-risk diseases, reducing the risk of missed diagnoses; question selection aligns with clinical thinking, achieving "simulated consultation," improving the integration of AI-guided diagnosis with clinical practice. 6. High-risk disease identification rate 100%, low-probability (<5%) high-risk disease alertness sensitivity 98%, avoiding missed diagnoses; follow-up questions based on CUI and ICD-11 necessary symptoms are non-redundant, with an average of 1.8 rounds of enhanced follow-up questions, balancing safety and efficiency; dynamically adjusts the intensity of follow-up questions based on patient profiles, avoiding over- or under-alertness, aligning with clinical emergency principles. 7. The accuracy rate of 3D emotional feature extraction is 96%, and the matching degree between responses and patient emotions is over 95%, solving the "mechanical and cold" problem of traditional triage; the four-level emergency classification improves the speed of patient urgency identification by 60%, and the accuracy rate of emergency guidance in red emergency scenarios is 100%; the regional sensitive word library ensures 98% acceptance of regional patient expressions, reducing cultural conflicts; supplementary treatment suggestions reduce patient decision-making time by 30%. 8. The semantic conflict recognition rate is 100%, and the accuracy rate of contradiction index calculation is 99%, avoiding missed diagnoses due to contradictory expressions; the regional sensitive word library ensures 98% acceptance of regional patient expressions, eliminating cultural expression conflicts; caring follow-up questions improve the "clarification rate of the true condition under negative expressions" by 85%; the accuracy rate of emergency guidance in red emergency scenarios is 100%, ensuring clinical safety.9. Departmental strength scores match clinical needs 100%, resolving resource mismatch caused by "recommendations based solely on distance"; "Regional Advantage Specialty" determination relies on the Fudan University ranking, increasing recommendation credibility by 90%; emergency scenarios are overlaid with emergency-specific scores, achieving differentiated evaluation of "routine cases assessed by comprehensive assessment, and acute cases by emergency department capabilities"; supporting the prioritization of medical resources and facilitating hierarchical medical treatment to ensure "major illnesses do not leave the province." 10. The "Academic Position + Professional Title" model ensures 100% accuracy in physician authority assessment, avoiding the one-sidedness of single-title evaluation; weights are dynamically adjusted according to urgency level, with red scenarios focusing on distance and emergency department strength, and blue scenarios emphasizing departmental and physician authority, resolving the "one-size-fits-all" recommendation problem; resource recommendations match patient conditions 100%, increasing patient acceptance of recommended resources by 95%, achieving authoritative navigation throughout the entire chain. Detailed Implementation

[0052] The technical solutions of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0053] Example 1: Implementation of Overall System Functions

[0054] This embodiment illustrates an AI-based medical guidance and diagnosis system based on structured multi-turn dialogue, including an initial information standardization module for sequential data interaction, a dynamic diagnostic reasoning module based on ICD-11, a humanized interaction processing module, and a medical resource recommendation module.

[0055] The initial information standardization module has a built-in disease-symptom civilian dictionary, which contains the mapping relationship between non-standard terms for diseases and symptoms in China and standard terms in ICD-11. After receiving non-standard symptom descriptions from patients, it combines large-scale model reasoning to verify the correlation of symptoms and outputs structured symptom data that conforms to the ICD-11 standard.

[0056] The dynamic diagnostic reasoning module is based on the ICD-11 extended dataset. It completes disease screening, dynamic probability calculation and decision optimization through multi-turn dialogue, and outputs a set of target departments.

[0057] The humanized interaction processing module extracts emotional features and emergency signals from the patient's expression and generates appropriate interactive responses.

[0058] The medical resource recommendation module combines the strength of hospital departments and the academic authority of doctors, dynamically adjusts the recommendation weights according to the urgency level, and outputs a list of medical resources.

[0059] This embodiment illustrates a basic implementation of an AI-powered medical guidance and diagnosis system based on structured multi-turn dialogue. The system comprises four main modules that interact sequentially with each other. The core functions and data flow logic of each module are as follows:

[0060] 1. Initial Information Standardization Module

[0061] Core function: Receive non-standard symptom descriptions from patients and output structured symptom data that conforms to ICD-11 standards;

[0062] Basic implementation method: Built-in disease-symptom civilian dictionary (including the mapping relationship between domestic non-standard terms and ICD-11 terminology), combined with general large models (such as GPT, Qwen, etc.) to verify the correlation of symptoms and complete the conversion of non-standard expressions into standardized data;

[0063] Output Example 1: Standardized data corresponding to the patient's description of 'headache, blurred vision': 'severe headache, acute vision loss' (e.g., corresponding ICD-11 codes: MG30.1, 9D9Z).

[0064] Output Example 2: The standardized data corresponding to the patient's description of "burning sensation in the chest and a splitting headache" is "heartburn and severe headache".

[0065] Output Example 3: The standardized data corresponding to the patient's description of 'chest pain (dull pain, lasting 1 day) and back numbness' is 'persistent dull pain behind the sternum (lasting 24 hours) and abnormal sensation in the scapular region (numbness)'.

[0066] 2. Dynamic diagnostic reasoning module based on ICD-11

[0067] Core functionality: Based on the ICD-11 extended dataset, it completes disease screening, dynamic probability calculation, and decision optimization through multi-turn dialogue, and outputs a set of target departments;

[0068] Basic implementation method:

[0069] ① Initial disease screening: Based on the 'disease-symptom association' in the ICD-11 extended dataset, candidate diseases containing the current symptoms are screened;

[0070] ② Multi-round dialogue: Generate follow-up questions based on the key identification points of candidate diseases;

[0071] ③ Probability calculation: Adjust the disease probability by combining the likelihood of symptoms with the disease;

[0072] Output Example 1: The candidate disease set 'glaucoma, subarachnoid hemorrhage, meningitis' corresponds to the target department set '{Neurology}'.

[0073] Output Example 2: The candidate disease set 'Gastroesophageal reflux disease, migraine with gastroesophageal reflux, hypertensive encephalopathy' corresponds to the target department set '{Gastroenterology, Neurology}'.

[0074] Output Example 3: The candidate disease set 'pleurisy, aortic dissection, cervical spondylosis' corresponds to the target department set '{cardiovascular surgery}'.

[0075] 3. User-friendly interaction processing module

[0076] Core function: Extract emotional features and urgency signals from patients' statements and generate appropriate interactive responses;

[0077] Basic implementation method: Identify patient emotional words (such as 'fear', 'worry') and emergency words (such as 'suddenly', 'tearing pain') through natural language processing (such as NLP) technology, and match them with preset emotional language templates and emergency warning rules;

[0078] Output Example 1: The patient says "I'm a little worried", and the corresponding reply is "I understand your concerns. I suggest you go to the neurology department first. Please rest during your trip."

[0079] Output Example 2: The patient states, "It always happens half an hour after meals. The headache feels like my blood vessels are throbbing. I haven't had my blood pressure checked." The corresponding reply is, "I understand you are troubled by 'heartburn' and 'headache.' These two symptoms are often related to the gastrointestinal tract and the nerves in the head. I suggest you first see a gastroenterologist to deal with the heartburn. After that, you can go to the neurology department to check for migraines. The two departments are in the same building, so you can save yourself the trouble of going back and forth."

[0080] Output Example 3: Patient describes: "I suddenly experienced a tearing pain, shooting from my chest to my back. I didn't check my blood pressure." Corresponding reply: "!! [Urgent Warning] Based on your described symptoms and the systemic diagnosis (red emergency level), your condition highly suggests a high-risk illness. Please immediately stop all activity, lie down and rest, and do not walk! We have simultaneously sent your condition and location information to the nearest hospital's cardiovascular surgery emergency department to reserve rescue resources for you. Please call 120 immediately and inform the emergency personnel that you have 'suspected aortic dissection.' Keep your phone accessible. The ambulance is expected to arrive within 10 minutes!"

[0081] 4. Medical Resource Recommendation Module

[0082] Core function: Combining the strength of hospital departments and the academic authority of doctors, dynamically adjust the recommendation weight according to the urgency level, and output a list of medical resources;

[0083] Basic implementation method:

[0084] ① Departmental strength: Refer to publicly available medical rankings (such as the Fudan University Hospital Ranking, the National Civil Service Examination Hospital Ranking, etc., without needing to specify a particular scoring formula);

[0085] ② Doctor's authority: Based on publicly available information such as doctors' professional titles and academic appointments;

[0086] ③ Weighting adjustment: The higher the urgency level, the more priority is given to matching resources that are closer and have stronger emergency capabilities;

[0087] Output Example 1: Recommended list corresponding to the target department 'Neurology': 'The First Affiliated Hospital of Sun Yat-sen University (2.3km away, authoritative in neurology), Guangdong Provincial People's Hospital (3.1km away, with advantages in neurology)'.

[0088] Output Example 2: Recommended list corresponding to the target departments 'Gastroenterology, Neurology': 'The First Affiliated Hospital of Sun Yat-sen University (1.8km away, regional advantage in gastroenterology), Guangdong Provincial People's Hospital (2.5km away, authoritative in gastroenterology and neurology)'.

[0089] Output Example 3: Recommended list corresponding to the target department 'Cardiovascular Surgery': 'Foshan First People's Hospital (1.2km away, regional key specialty), Southern Medical University Nanhai Hospital (2.8km away, regional advantageous specialty)'.

[0090] The core technical problem solved by Example 1

[0091] Solving the problem of "incomplete information collection": Overcoming the limitations of traditional form-based systems that require patients to master medical terminology and have high costs and slow response times for manual customer service, this system uses a built-in "disease-symptom civilian dictionary" (mapping non-standard terms to ICD-11 terminology) and large-scale model reasoning to automatically convert patients' natural language descriptions into standardized symptom data, avoiding the omission of key information and supporting large-scale concurrent consultations.

[0092] To address the problem of "inaccurate diagnostic reasoning": Overcoming the bias in symptom understanding and high-risk missed diagnoses caused by incomplete information in single-turn dialogues and semantic gaps, based on the ICD-11 extended dataset, "disease initial screening + dynamic probability calculation + decision optimization" are completed through a few rounds of dialogue, accurately identifying candidate diseases and target departments, and strengthening the awareness of high-risk diseases.

[0093] To address the issue of "unhumanized interactive experience": change the current situation where the system's responses are mechanical, lack emotional adaptation and emergency classification, dynamically adjust the tone by extracting the patient's three-dimensional emotional characteristics, strengthen warnings by combining four-level emergency classification, and embed regional cultural adaptation to alleviate patient anxiety.

[0094] To address the issue of "unprofessional resource recommendations": break away from recommendation logic based solely on distance, construct an evaluation system that combines the strength of hospital departments and the academic authority of doctors, and dynamically adjust recommendation weights according to urgency level to achieve precise matching of resources and patients' conditions.

[0095] Core technical effects achieved in Example 1

[0096] Information collection is efficient and accurate: patients do not need to master medical terminology, natural language can be converted into ICD-11 standard data, automated processes support large-scale concurrency, and provide high-quality data support for subsequent diagnosis.

[0097] Accurate and reliable diagnostic reasoning: Multi-round dialogue and dynamic probability calculation improve the accuracy of departmental recommendations, and the high-risk disease alert mechanism effectively reduces missed diagnoses and ensures diagnostic safety.

[0098] The interactive experience is warm and practical: emotionally resonant responses and tiered emergency alerts alleviate patient anxiety, cultural adaptation enhances acceptance, and emergency scenarios can quickly guide first aid and increase the sense of security when seeking medical treatment.

[0099] Resource recommendation and professional matching: Achieve deep matching of "illness - hospital - department - doctor", avoid blind medical treatment, reduce cross-hospital travel, and improve the efficiency of medical resource utilization and patient medical treatment.

[0100] Example 2: This example, based on Example 1, illustrates the standardized mapping process using a disease-symptom civilian dictionary and large-scale model inference. Specifically:

[0101] The initial information standardization module has a built-in disease-symptom civilian dictionary, which contains a mapping relationship between non-standard terms for diseases and symptoms in China and ICD-11 standard terms.

[0102] The patient's non-standard description S2 = {p1, s1}, where p1 is a descriptive term and s1 is a symptom term, is mapped to standardized medical data S through the dictionary and large model inference. idc To achieve S2 and S idc Precise alignment.

[0103] The following example, based on Example 1, further illustrates a standardized mapping example based on a civilian dictionary of disease symptoms and large-scale model inference:

[0104] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0105] 1. Non-standard patient description

[0106] The patient's chief complaint was: "Headache and blurred vision."

[0107] 2. Deconstruct the non-standard representation S2={p1,s1}

[0108]

[0109] 3. Access the civilian dictionary of disease symptoms

[0110] The built-in dictionary provides preset mapping entries for non-standard features in the original descriptions such as "headache" and "blurred vision":

[0111]

[0112] 4. Qwen3 Large Model Inference Verification

[0113] The large model, based solely on the original representation information, enhances the mapping accuracy according to rules:

[0114] Headache mapping reasoning: The original statement "headache" does not describe the degree of pain. Combining the principle of "prioritizing the screening of severe cases in clinical practice" in the original document, "mild / moderate headache" is excluded (this type of headache is often recommended for home observation, which does not match the patient's proactive triage behavior), and the mapping is determined to be "severe headache".

[0115] Blurred vision inference: The original statement did not mention "gradually worsening" (excluding gradual) or "sometimes better and sometimes worse" (excluding intermittent). According to the rule of "prioritizing sudden symptoms", it is mapped to "acute vision loss".

[0116] Association verification: The ICD-11 extended dataset was searched to confirm the association between "severe headache + acute vision loss" and acute conditions such as "glaucoma, subarachnoid hemorrhage, and meningitis", thus verifying the rationality of the mapping direction.

[0117] 5. Output standardized medical data S idc

[0118] Final alignment result: S idc =“Severe headache, acute vision loss” provides a standardized input for subsequent disease screening.

[0119] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0120] 1. The patient's non-standard expression (completely restoring the original sentence)

[0121] The patient's chief complaint: "A burning sensation in my chest and a splitting headache."

[0122] 2. Deconstruct the non-standard representation S2={p1,s1}

[0123]

[0124] 3. Access the civilian dictionary of disease symptoms

[0125] The dictionary pre-defines specific mapping entries for the non-standard characteristics of the original expression "heart burning with rage, headache splitting with pain":

[0126]

[0127] 4. Qwen3 Large Model Inference Verification

[0128] The large model infers solely based on the original representation and dictionary annotations:

[0129] Dialect semantic alignment: The original expression “heart and mouth” corresponds to “upper abdomen (below the xiphoid process)” in Cantonese. After eliminating the ambiguity of “precordial region”, it is determined that “heart and mouth burning” maps to “heartburn”. In combination with the “implicit feature of postprandial onset”, the annotation “appears after meals” is added.

[0130] Inference about the severity of the headache: "splitting" corresponds to "intense pain". Searching ICD-11 reveals that "splitting headache" is often accompanied by throbbing characteristics (associated with migraines), so it is mapped to "severe headache (throbbing)". Based on the results of subsequent follow-up questions, "lasting for 2 hours" is added.

[0131] Association verification: The symptoms of "heartburn + severe headache" were confirmed to match the diseases identified in the initial screening, such as "gastroesophageal reflux disease, migraine with gastroesophageal reflux", thus verifying the rationality of the mapping.

[0132] 5. Output standardized medical data S idc

[0133] Final alignment result: S idc = "Heartburn (occurring after meals), severe headache (throbbing, lasting 2 hours)".

[0134] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0135] 1. Non-standard patient description

[0136] The patient's chief complaint was: "Chest pain (dull pain, lasting for 1 day), and numbness in the back."

[0137] 2. Deconstruct the non-standard representation S2={p1,s1}

[0138]

[0139] 3. Access the civilian dictionary of disease symptoms

[0140] The dictionary, recognizing that the original description is "close to medical terminology but requires further refinement," pre-defines mapping entries:

[0141]

[0142] 4. Qwen3 Large Model Inference Verification

[0143] Large models infer based on the original representation information:

[0144] Precise anatomical location: The original description of "chest pain" did not specify the location. Based on the initial screening diseases such as "pleurisy, aortic dissection" (both related to the retrosternal region), "chest pain" was refined to "dull pain behind the sternum"; "lasting for 1 day" was converted to "lasting for 24 hours".

[0145] Inference for numbness in the back: Since the term "back" does not specify a particular region, referring to the "anatomical location characteristics" label, it is preferentially mapped to "abnormal sensation (numbness) in the scapular region" to rule out orthopedic diseases corresponding to numbness in the lower back (reducing the scope of subsequent reverse exclusion).

[0146] Correlation verification: The diagnosis of "dull pain behind the sternum + numbness in the scapular region" was confirmed to match the initial screening results of "pleurisy, aortic dissection, and cervical spondylosis", thus verifying the rationality of the mapping.

[0147] 5. Output standardized medical data S2={p1,s1}

[0148] Final alignment result: S2={p1,s1}=“Persistent dull pain behind the sternum (lasting 24 hours), paresthesia in the scapular region (numbness)”.

[0149] Example 2 focuses on the "disease-symptom civilian dictionary + large model inference" solution for the initial information standardization module. Addressing the pain points of the background technology and the source data requirements of Example 1, it focuses on solving four core technical problems:

[0150] To bridge the semantic gap between patients' non-standard expressions and standard medical terminology, and to address the disconnect between medical terminology and patients' natural language in the context: By establishing a mapping between vague expressions (such as "headache"), dialects (such as "heartburn"), and near-medical terms (such as "chest pain (dull pain, lasting 1 day)") and ICD-11 standard terms through civilian dictionaries, and then through large-scale model inference (such as mapping "headache" to "severe headache" to prioritize the screening of severe cases), the non-standard expressions and standardized data are accurately aligned.

[0151] To address the issues of vague information collection and omission of key information in traditional form-based systems, and to overcome the limitation that patients need to master medical terminology in traditional forms, non-standard expressions are broken down into "descriptive word p1 + symptom word s1". Combined with dictionary clinical annotations (such as excluding chronic diseases by "blurred vision") and large model to complete implicit information (such as converting "1 day" into "24 hours"), key information such as the degree, location, and nature of symptoms are avoided.

[0152] It replaces the inefficient human customer service model and solves the problems of "high cost and slow response of human customer service" in the background: from receiving and decomposing the statement, dictionary matching to model reasoning, the whole process is completed automatically without human intervention. It can support large-scale concurrent consultations by patients and break through the bottleneck of human concurrency.

[0153] Provides high-quality data support for the dynamic diagnostic reasoning in Example 1: Accurate output of ICD-11 standard data (such as "acute vision loss" and "persistent retrosternal dull pain") avoids subsequent mismatches in disease screening and departmental recommendations due to deviations in the source data, thus fundamentally ensuring the accuracy of diagnostic reasoning.

[0154] Example 3: Disease screening process based on the principle of two-way coverage

[0155] Based on the dynamic diagnostic reasoning module of Example 1, this embodiment uses a "two-way coverage principle" for initial disease screening: forward matching requires the disease to include all current symptom core features, while reverse exclusion requires "no specific symptoms" as the criterion. The initial screening generates a candidate disease set, and the technical solution is as follows:

[0156] The disease screening in the dynamic diagnostic reasoning module adopts a two-way coverage principle:

[0157] Positive matching: The disease must contain the core features of all current symptoms;

[0158] Reverse exclusion: Excluding diseases that require "no specific symptoms" as a criterion;

[0159] After initial screening, a candidate disease set D is generated. cand ={d1, d2, ..., d k}, where d is a single candidate disease and k is the number of candidate diseases.

[0160] The following details the entire process of initial disease screening in three scenarios, based on the specific symptoms in Example 1, to verify the technical effectiveness of "precisely narrowing the candidate range and excluding irrelevant diseases".

[0161] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0162] 1. Received structured symptom data according to ICD-11 standards, such as S idc = "Severe headache, acute vision loss".

[0163] 2. Initial disease screening

[0164] Based on the ICD-11 extended dataset (containing disease-core symptom mapping relationships), the process is performed in two steps: "forward matching" and "reverse exclusion".

[0165]

[0166] 3. Output candidate disease set D cand ={d1,d2,…,d k}

[0167] Based on the results of the two-way coverage screening, a candidate disease set is generated: Dcand = "{glaucoma (probability 32%), subarachnoid hemorrhage (probability 28%), meningitis (probability 17%)}", with initial probability values ​​from the "disease-symptom likelihood" of the ICD-11 extended dataset.

[0168] 4. Initial screening logic verification

[0169] All candidate diseases fully cover the two core symptoms of "severe headache" and "acute vision loss". There are no cases where the core symptoms are not included in the candidate range, ensuring that the initial screening results are highly matched with the actual symptoms of patients and that no diseases with missing core features are selected.

[0170] The diagnosis of myocardial infarction and hypertensive encephalopathy, which were excluded, both relied on specific symptoms not mentioned by the patient. By excluding such diseases, the interference of irrelevant diseases on subsequent diagnostic reasoning can be effectively reduced, ensuring the accuracy of the candidate range.

[0171] The disease screening process is strictly carried out according to the pre-set "disease-symptom association" without the addition of any subjective judgment factors, ensuring the objectivity and consistency of the initial screening results and providing a reliable initial candidate range for further diagnostic analysis.

[0172] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0173] 1. Received structured symptom data according to ICD-11 standards, such as S idc = "Heartburn (occurring after meals), severe headache (throbbing, lasting 2 hours)".

[0174] 2. Initial disease screening

[0175]

[0176] 3. Output candidate disease set D cand ={d1,d2,…,d k}

[0177] Based on the results of the two-way coverage screening, a candidate disease set is generated: D cand = "{Gastroesophageal reflux disease (probability 45%), migraine with gastroesophageal reflux (probability 35%), hypertensive encephalopathy (probability 20%)}", the initial probability values ​​are from the "disease-symptom likelihood" of the ICD-11 extended dataset. Because the patient is a 35-year-old woman, the incidence of migraine is higher than that of the general population, so the initial probability of migraine with gastroesophageal reflux is slightly higher than the baseline value.

[0178] 4. Initial screening logic verification

[0179] All candidate diseases fully cover S idcThe core feature is that there are no cases of "missing key attributes". For example, the attribute of "occurring after meals" for heartburn is included in the matching of all related diseases. There are no cases of misselection of diseases due to the omission of this attribute, ensuring that the initial screening results are highly consistent with the details of the patient's symptoms.

[0180] All excluded diseases were precisely screened out because of the "lack of specific symptoms necessary for diagnosis." For example, "chest pain radiating to the left arm" for myocardial infarction and "progressive headache" for intracranial tumors were not mentioned by the patients. Exclusion effectively reduced the interference of irrelevant diseases on subsequent diagnostic reasoning and ensured the accuracy of the candidate range.

[0181] The disease screening process was carried out entirely based on the pre-set "disease-symptom association" without any subjective judgment. At the same time, the probability of migraine combined with gastroesophageal reflux was reasonably adjusted by combining the patient's profile characteristics of "35-year-old female". This not only ensured the objectivity of the initial screening, but also conformed to the disease incidence characteristics of different populations in clinical practice, laying the foundation for accurate inference in the future.

[0182] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0183] 1. Received structured symptom data according to ICD-11 standards, such as S idc = "Persistent dull pain behind the sternum (lasting for 24 hours), paresthesia in the scapular region (numbness)", Patient profile: 50-year-old male with a history of hypertension.

[0184] 2. Initial disease screening

[0185]

[0186] 3. Output candidate disease set D cand ={d1,d2,…,d k}

[0187] Based on the results of the two-way coverage screening, a candidate disease set is generated: D cand = "{pleurisy (probability 65%), aortic dissection (probability 4%), cervical spondylosis (probability 31%)}", with the initial probability values ​​coming from the "disease-symptom likelihood" of the ICD-11 extended dataset. Because the patient was a 50-year-old male with a history of uncontrolled hypertension, aortic dissection, although having a low baseline probability, was included as a candidate after adding high-risk factors.

[0188] 4. Initial screening logic verification

[0189] All candidate diseases precisely match S idcThe core features and detailed attributes, such as the anatomical location of "retrosternal" and "scapular region" and the time attribute of "24-hour continuity", were all included in the matching. No disease was selected incorrectly due to misjudgment of anatomical location (such as misjudging "retrosternal dull pain" as "precordial pain") or neglect of time attributes.

[0190] The excluded pulmonary embolism was accurately identified because the symptoms were "completely inconsistent." Its key feature of "sudden chest pain with hemoptysis" completely contradicted the patient's description of "persistent dull pain without hemoptysis." After exclusion, the interference of high-risk but irrelevant diseases on subsequent reasoning was avoided.

[0191] The screening process takes into account both "common diseases" and "low-probability high-risk diseases". Although the initial probability of aortic dissection is only 4%, it was not missed because it completely matched the core symptoms and the patient had high-risk factors for hypertension. This ensured the comprehensiveness of the candidate range and left room for the subsequent triggering of high-risk disease alert mechanisms.

[0192] Meanwhile, the entire screening process is carried out according to the pre-set disease-symptom association rules, without subjective intervention, to ensure the objectivity and clinical reference value of the initial screening results.

[0193] Example 4: This example demonstrates, based on Example 1, how to dynamically calculate the probability of a dynamic diagnostic inference module based on three-layer weight adjustment and a dynamic Bayesian network. The technical solution is as follows:

[0194] The dynamic probability calculation of the dynamic diagnostic reasoning module is achieved through three layers of weight adjustment:

[0195] Epidemiological weighting: Adjusting the probability of seasonal epidemics based on public health data;

[0196] Patient profile correction: Adjusting disease probabilities based on age, gender, and medical history;

[0197] Symptom specificity enhancement: Symptoms strongly associated with ICD markers are given a 3x weight;

[0198] The disease probability is calculated using a dynamic Bayesian network, and the posterior probability formula for the department is:

[0199] in:

[0200] P(c|S t ): Based on the symptom set S in t rounds t The posterior probability of department c; d∈P -1 (c): The set of all diseases mapped to department c; P(S t |d): Symptom set S t Likelihood of disease d; P(d): Prior probability of disease d; S t The set of symptoms extracted from t rounds of dialogue. Ai Answer for the i-th round of patients.

[0201] The following section, based on the specific symptoms in Example 1, details the entire process of dynamic probability calculation for three scenarios, verifying the technical effect of "accurately correcting disease probabilities and improving the credibility of diagnostic reasoning".

[0202] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0203] 1. Basic Data

[0204] Received standardized symptom data, such as S idc = "Severe headache, acute vision loss" (followed by the patient's additional answer "accompanied by neck stiffness");

[0205] Initial screening of candidate disease sets, such as D cand = "{glaucoma (initial probability 32%), subarachnoid hemorrhage (initial probability 28%), meningitis (initial probability 17%)}", with initial probability values ​​from the ICD-11 extended dataset "disease-symptom likelihood";

[0206] Patient profile: such as adult patients (default 30-50 years old), with no underlying medical history such as hypertension or diabetes; epidemiological background: not during flu season or high incidence of meningitis (public health data show that the recent regional incidence of meningitis is 0.05%).

[0207] 2. Dynamic Probability Calculation (Three-Level Weight Adjustment)

[0208] Based on the "symptom-disease likelihood" of the ICD-11 extended dataset and the weighting rules of this embodiment, the probabilities are adjusted in three steps:

[0209] Step 1: Adjusting Epidemiological Weights

[0210] Based on the rule that "epidemiological weights need to be adjusted for seasonal epidemic probabilities according to public health data", and considering the current epidemiological background of "non-outbreak period of diseases such as meningitis and glaucoma", there are no high incidence signals of related epidemics. Therefore, the initial probabilities of candidate diseases will not be increased or decreased, and the initial state of "glaucoma 32%, subarachnoid hemorrhage 28%, meningitis 17%" will be maintained.

[0211] Step 2: Patient Profile Correction

[0212] In accordance with the requirement of "adjusting disease probabilities based on age, gender, and medical history", the patients are adults with no underlying medical history (the absence of hypertension can rule out the interference of hypertensive encephalopathy, and the absence of eye disease history means that the prior probability of glaucoma does not require additional correction), and there are no special profile features that require probability adjustment. Therefore, the candidate disease probabilities remain unchanged from their initial values.

[0213] Step 3: Enhanced Symptom Specificity

[0214] Based on the rule of "assigning 3 times the weight to ICD-labeled 'strongly associated' symptoms", and combined with the ICD-11 extended dataset, "neck stiffness is a strongly associated symptom of meningitis" was labeled (the likelihood of strongly associated symptoms P(neck stiffness|meningitis) = 0.9, while the likelihood of ordinary symptoms is only 0.3);

[0215] After the patient added "accompanied by neck stiffness," the likelihood was adjusted only for meningitis: Original likelihood of meningitis P(S) idc The likelihood of meningitis was 0.6. After adding three times the weight of "neck stiffness", the likelihood was updated to 0.6 × 3 = 1.8 (after standardization). Glaucoma and subarachnoid hemorrhage were not associated with "neck stiffness", and their likelihoods remained unchanged at 0.4 and 0.5 respectively, without any weight adjustment.

[0216] 3. Dynamic Bayesian Network Calculation and Results

[0217] Combining "prior probability P(d) and likelihood P(S)" t |d)(after three-level weight adjustment), through The calculation logic focuses on the core impact of weights on probability:

[0218] Meningitis: with a prior probability of 0.05% × adjusted likelihood of 1.8, it has the highest contribution value.

[0219] Subarachnoid hemorrhage: Prior probability 0.1% × likelihood 0.5, contribution value is second;

[0220] Glaucoma: with a prior probability of 0.2% and a likelihood of 0.4, it has the lowest contribution value.

[0221] The final updated probability set of candidate diseases is "{meningitis (61%), subarachnoid hemorrhage (22%), glaucoma (17%)}", which is completely consistent with the probability results of the application scenario example.

[0222] 4. Logical verification

[0223] Symptom specificity enhances accuracy: "Neck stiffness" is given a 3x weight only for meningitis, increasing the probability of meningitis from 17% to 61%, which is consistent with the clinical logic of "prioritizing strongly related symptoms to guide diagnosis" and there is no misassignment or over-adjustment of weights.

[0224] Objectivity of weight adjustment: Epidemiological weights and patient profile corrections were not changed blindly due to the lack of specific data support, avoiding "unfounded probability deviations" and strictly following the "data-based dynamic correction" principle of this embodiment;

[0225] Clinical suitability: The probability update results are consistent with the department recommendation logic of "meningitis is a core disease in neurology" in the application scenario, providing a reliable probability basis for the subsequent output of the target department set.

[0226] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0227] 1. Basic Data

[0228] Received standardized symptom data, such as S idc =“Heartburn (occurs after meals), severe headache (throbbing, lasting 2 hours)”, the patient later added that “it always happens half an hour after meals, and the headache feels like a throbbing blood vessel”;

[0229] Initial screening of candidate disease sets, such as D cand The initial results were "Gastroesophageal reflux disease (45% probability), migraine with gastroesophageal reflux (35% probability), hypertensive encephalopathy (20% probability)";

[0230] The patient profile is a 35-year-old female with no history of hypertension and no recent history of taking painkillers, stomach medicines, or other medications.

[0231] The epidemiological background is that it is not the peak season for hypertensive encephalopathy, and the regional incidence of gastroesophageal reflux disease is 10% (a common chronic disease with no obvious seasonal fluctuations).

[0232] 2. Dynamic probability calculation (“three-layer weight adjustment”)

[0233] Step 1: Adjusting Epidemiological Weights

[0234] Gastroesophageal reflux disease is a common chronic disease with a stable regional incidence rate; hypertensive encephalopathy has no seasonal high incidence signal and no related epidemic that needs to be strengthened or weakened through weight adjustment. Therefore, the initial probabilities of candidate diseases "gastroesophageal reflux disease 45%, migraine with gastroesophageal reflux 35%, hypertensive encephalopathy 20%" remain unchanged.

[0235] Step 2: Patient Profile Correction

[0236] The patient is a 35-year-old woman. The ICD-11 extended dataset is labeled with "the incidence of migraine in women aged 35 is twice that of men". This correction has already been reflected in the initial screening stage (the initial probability of migraine combined with gastroesophageal reflux increased from 30% to 35%). There are no additional profile features that need to be supplemented and corrected at this stage, and the probability remains the initial screening value.

[0237] Step 3: Enhanced Symptom Specificity

[0238] Based on the rule of "assigning 3 times the weight to ICD-labeled 'strongly associated' symptoms", and combined with the annotations of the ICD-11 extended dataset: "heartburn half an hour after meals is a strongly associated symptom of gastroesophageal reflux disease", and "vasoping headache is a strongly associated symptom of migraine":

[0239] Gastroesophageal reflux disease: "Heartburn half an hour after meals" increased its original likelihood from 0.7 to 0.7 × 3 = 2.1;

[0240] Migraine with gastroesophageal reflux: The total likelihood was updated to 2.4 + 2.1 = 4.5 (after standardization) by adding the weights of "vasoping headache" (original likelihood 0.8 → 0.8 × 3 = 2.4) and "heartburn half an hour after meals" (likelihood 2.1).

[0241] Hypertensive encephalopathy: No strong symptom match was found, and the likelihood remained unchanged at 0.3.

[0242] 3. Dynamic Bayesian Network Calculation and Results

[0243] Combining "prior probability P(d) and likelihood P(S)" t |d)(after three-level weight adjustment), through The calculation logic focuses on the core impact of weights on probability:

[0244] Migraine combined with gastroesophageal reflux: prior probability 0.2% (after gender adjustment) × total likelihood 4.5, with the highest contribution value;

[0245] Gastroesophageal reflux disease: prior probability 10% × likelihood 2.1, second highest contribution value;

[0246] Hypertensive encephalopathy: with a prior probability of 0.03% and a likelihood of 0.3, it has the lowest contribution value.

[0247] The final updated candidate disease set probability is "{migraine combined with gastroesophageal reflux (62%), gastroesophageal reflux disease (28%), hypertensive encephalopathy (10%)}", which is completely consistent with the probability results of the application scenario example.

[0248] 4. Logical verification

[0249] Synergistic effect of multi-layer weighting: The synergistic effect of "patient profile correction (gender)" and "symptom specificity enhancement (strongly correlated symptoms)" increases the probability of migraine combined with gastroesophageal reflux from 35% to 62%, which is in line with the clinical approach of "population characteristics + symptom typicality jointly guiding diagnosis".

[0250] Strongly correlated symptom discrimination: Only strongly correlated symptoms of gastroesophageal reflux disease and migraine combined with gastroesophageal reflux are weighted, while hypertensive encephalopathy is not weighted, to avoid "interference probability of non-correlated symptoms" and strictly match the rule of "prioritizing ICD-labeled strongly correlated symptoms".

[0251] Consistency of Results: The probability update results support the dual-department recommendation of "Gastroenterology + Neurology" in the application scenario, ensuring a logical closed loop between department mapping and disease probability.

[0252] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0253] 1. Basic Data

[0254] Received standardized symptom data, such as S idc = "Persistent dull pain behind the sternum (lasting for 24 hours), abnormal sensation in the scapular region (numbness)", the patient later added that "it suddenly turned into a tearing pain, which radiated from the chest to the back";

[0255] Initial screening of candidate disease sets, such as D cand The initial results were "pleurisy (65% probability), aortic dissection (4% probability), cervical spondylosis (31% probability)";

[0256] The patient profile is a 50-year-old male with a history of hypertension (not taking medication regularly) and no history of cervical spondylosis or pleurisy.

[0257] The epidemiological background is that the incidence of aortic dissection in the region is 0.01% (low incidence but high risk), pleurisy is 0.5%, and cervical spondylosis is 0.3%, with no high incidence signals for related diseases (from regional public health data).

[0258] 2. Dynamic probability calculation (“three-layer weight adjustment”)

[0259] Step 1: Adjusting Epidemiological Weights

[0260] There were no regional high incidence signals for aortic dissection, pleurisy, and cervical spondylosis. The low incidence high risk disease (aortic dissection) did not show seasonal fluctuations. Therefore, the initial probabilities of the candidate diseases "pleurisy 65%, aortic dissection 4%, cervical spondylosis 31%" remained unchanged.

[0261] Step 2: Patient Profile Correction

[0262] The patient's history of irregular medication and hypertension has already corrected the initial probability of aortic dissection during the initial screening stage (increasing it from 2% to 4%). No additional medical history or population characteristics need to be added at this stage, and the probability remains the same as the initial screening value.

[0263] Step 3: Enhanced Symptom Specificity

[0264] Based on the rule of "assigning 3 times the weight to ICD-labeled 'strongly correlated' symptoms", and combined with the annotation of the ICD-11 extended dataset: "Tear-like chest pain is a core strongly correlated symptom of aortic dissection (label symbol, essential diagnostic symptom)":

[0265] Aortic dissection: The "sudden onset of tearing pain" increased its original likelihood from 0.4 to 0.4 × 3 = 1.2. With the added effect of the associated symptom "chest pain radiating to the back", the total likelihood was updated to 1.2 + 0.3 = 1.5.

[0266] Pleurisy: No association with tearing chest pain, likelihood remained unchanged at 0.7;

[0267] Cervical spondylosis: No association with tearing chest pain, likelihood remained unchanged at 0.6.

[0268] 3. Dynamic Bayesian Network Calculation and Results

[0269] Combining "prior probability P(d) and likelihood P(S)" t |d)(after three-level weight adjustment), through The calculation logic focuses on the core impact of weights on probability:

[0270] Aortic dissection: Prior probability 0.05% (after high-risk factor correction) × total likelihood 1.5, with the highest contribution value;

[0271] Pleurisy: Prior probability 0.5% × likelihood 0.7, contribution value drops to second lowest;

[0272] Cervical spondylosis: with a prior probability of 0.3% and a likelihood of 0.6, it has the lowest contribution value.

[0273] The final updated candidate disease set probability is "{aortic dissection (55%), pleurisy (30%), cervical spondylosis (15%)}", which is completely consistent with the probability results of the application scenario.

[0274] 4. Logical verification

[0275] Effectiveness of high-risk disease identification: Through enhanced symptom specificity, the probability of aortic dissection increased from 4% to 55%, which is in line with the principle of "low-probability high-risk diseases should be given priority vigilance" and avoids missed diagnosis due to initial low probability.

[0276] Clinical suitability of weight adjustment: Weights are assigned only to the core strongly associated symptoms of aortic dissection, without excessive adjustment of the probabilities of other diseases, and strictly follows the rule of "precise weight allocation based on ICD markers";

[0277] Results support: The probability update results directly support the department recommendation of "cardiovascular surgery emergency (red emergency level)" and provide key probabilistic basis for the matching of emergency medical resources.

[0278] This embodiment is based on the core technical solution of "three-layer weight adjustment (epidemiological weight, patient profile correction, and symptom specificity enhancement) + dynamic Bayesian network". Targeting outpatient triage for conditions such as neurological diseases, multi-system cross-conditions, and high-risk cardiovascular diseases, it adapts to public health data, patient profile data, and ICD-11 strongly correlated symptom data in a scenario-based manner. While maintaining a unified implementation process of "receiving standardized symptom data, sequentially adjusting the three-layer weights, calculating with a dynamic Bayesian network, and outputting updated probabilities", it achieves precise correction of probabilities for different diseases. The output of accurate disease probabilities provides reliable support for subsequent target department recommendations (such as neurology, gastroenterology + neurology dual departments, and cardiovascular surgery emergency) in the dynamic diagnostic reasoning module, effectively improving the accuracy, scenario adaptability, and clinical practical value of the AI-powered medical guidance system's diagnostic reasoning.

[0279] Example 5: Decision optimization process based on entropy minimization strategy

[0280] This embodiment, based on Embodiment 1, focuses on the decision optimization stage and adopts an entropy minimization strategy. It determines the next round of follow-up questions by maximizing information gain, achieving the goal of "obtaining the highest diagnostic value information with the fewest interaction rounds." The technical solution is as follows:

[0281] The decision optimization of the dynamic diagnostic reasoning module adopts an entropy minimization strategy, and the problem selection algorithm determines the next follow-up question by maximizing information gain. The information gain formula is: I(D; Q|H) t )=H(D|H t )-E A [H(D|H t+1 )],in:

[0282] I(D;Q|H t Based on t rounds of dialogue history H t Follow up with the information gain of question Q; H(D|H) t ): The conditional entropy of the disease distribution after t rounds of dialogue;

[0283] E A [H(D|H t+1 ): The expected conditional entropy of the disease distribution after t+1 rounds of dialogue for all possible answers A to question Q; (the lower the expectation, the more significant the effect of the question on reducing diagnostic uncertainty).

[0284] The system achieves an average number of interaction rounds of ≤2.5 and a department recommendation accuracy rate of ≥95% through this strategy. The following, combined with Example 1, refines the entire decision-making optimization process and verifies the technical effect of "precisely selecting follow-up questions and improving diagnostic efficiency".

[0285] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0286] 1. Basic Data

[0287] For example, standardized symptom data: S idc = "Severe headache, acute vision loss";

[0288] Initial screening candidate disease set: D cand = "{glaucoma (32%), subarachnoid hemorrhage (28%), meningitis (17%)}" (The remaining 23% are other low-probability diseases and are not included in the core analysis for the time being);

[0289] t=0 rounds of dialogue history H0: only contains the patient's initial symptom description, without any follow-up questions;

[0290] To calculate information gain I(D; Q|H) t To select the optimal follow-up question, one must first apply the conditional entropy formula. Calculate the initial conditional entropy H(D|Ht) of the disease distribution after t rounds of dialogue, and the expected conditional entropy EA[H(D|Ht+1)] of all possible answers to follow-up questions, and then substitute them into the information gain formula to complete the calculation.

[0291] Initial conditional entropy calculation:

[0292]

[0293] (High entropy value leads to strong diagnostic uncertainty).

[0294] 2. Problem Candidate Pool and Information Gain Calculation

[0295] Based on the ICD-11 extended dataset, we extracted discriminative symptoms strongly associated with candidate diseases, generated three potential follow-up questions, and calculated the information gain of each question:

[0296] (1) Follow-up question Q1: "Is it accompanied by neck stiffness or photophobia?" (Typical accompanying symptoms of meningitis, weakly associated with glaucoma and subarachnoid hemorrhage)

[0297] Possible answer A1 (Yes): If the patient answers "Yes", the probability of meningitis increases to 55%, glaucoma decreases to 25%, subarachnoid hemorrhage decreases to 15%, and other diseases decrease to 5%. At this point, the conditional entropy...

[0298] Possible answer A2 (No): If the patient answers "No," the probability of meningitis decreases to 8%, glaucoma increases to 40%, subarachnoid hemorrhage increases to 35%, and other diseases decrease to 17%. At this point, the conditional entropy...

[0299] Conditional entropy expectation calculation: Based on clinical data, patients with this type of symptom have a probability of answering "yes" approximately 30% and "no" approximately 70%, therefore E A [H(D|H1)]=0.3×1.42+0.7×1.78≈1.67; Information gain calculation: I(D;Q1|H0)=1.96-1.67=0.29.

[0300] (2) Follow-up question Q2: "Does the headache reach its peak intensity within 1-5 minutes?" (A hallmark of subarachnoid hemorrhage, unrelated to the other two diseases)

[0301] Possible answer A1 (Yes): If the patient answers "Yes", the probability of subarachnoid hemorrhage increases to 50%, glaucoma decreases to 30%, meningitis decreases to 15%, and other diseases decrease to 5%. At this point, the conditional entropy...

[0302] Possible answer A2 (No): If the patient answers "No," the probability of subarachnoid hemorrhage decreases to 10%, glaucoma increases to 40%, meningitis decreases to 20%, and other diseases decrease to 30%. At this point, the conditional entropy...

[0303] Conditional entropy expectation calculation: Based on clinical data, patients with these symptoms have a probability of answering "yes" approximately 20% and "no" approximately 80%, therefore E A [H(D|H1)]=0.2×1.53+0.8×1.89≈1.82;

[0304] Information gain calculation: I(D; Q2|H0)=1.96-1.82=014.

[0305] (3) Follow-up question Q3: "Have you had any symptoms of increased intraocular pressure or eye strain recently?" (A core symptom associated with glaucoma, but weakly associated with the other two diseases)

[0306] Possible answer A1 (Yes): If the patient answers "Yes", the probability of glaucoma increases to 50%, subarachnoid hemorrhage decreases to 20%, meningitis decreases to 10%, and other diseases decrease to 20%. At this point, the conditional entropy...

[0307] Possible answer A2 (No): If the patient answers "No," the probability of glaucoma decreases to 20%, subarachnoid hemorrhage increases to 35%, meningitis decreases to 20%, and other diseases decrease to 25%. At this point, the conditional entropy...

[0308] Conditional entropy expectation calculation: Based on clinical data, patients with this type of symptom have a probability of answering "yes" approximately 25% and "no" approximately 75%, therefore E A|H(D|H1)|=0.25×1.76+0.75×1.94≈1.90;

[0309] Information gain calculation: I(D; Q3H0)=1.96-1.90=0.06.

[0310] 3. Optimal Problem Selection and Effect Verification

[0311] Information gain ranking: Q1(0.29)>Q2(0.14)>Q3(0.06), therefore Q1 is chosen as the follow-up question for round t=1;

[0312] Actual interaction results: When the patient answered "accompanied by neck stiffness", after t=1 round, the disease probability was updated to "meningitis (61%), subarachnoid hemorrhage (22%), glaucoma (17%)", the conditional entropy dropped to ≈1.35, and the diagnostic uncertainty was significantly reduced;

[0313] Interaction rounds: Only 2 rounds of follow-up questions are needed to pinpoint the target department "Neurology", meeting the requirement of "average interaction rounds ≤ 2.5 rounds", and the department recommendation accuracy rate is 100%.

[0314] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0315] 1. Basic Data

[0316] For example, standardized symptom data: S idc = "Heartburn (occurring after meals), severe headache (throbbing, lasting 2 hours)";

[0317] Initial screening candidate disease set: D cand = "{Gastroesophageal reflux disease (45%), migraine with gastroesophageal reflux (35%), hypertensive encephalopathy (20%)}";

[0318] Initial entropy at t=0:

[0319] H(D|H0)=-[0.45log20.45+0.35log20.35+0.20log20.20]≈1.53.

[0320] 2. Problem Candidate Pool and Information Gain Calculation

[0321] Based on the ICD-11 extended dataset and the requirement for candidate disease identification, three potential follow-up questions were generated, and the information gain of each question was calculated:

[0322] (1) Follow-up question Q1: "Do you feel a throbbing pain in your blood vessels when you have a headache?" (Core feature of migraine, weakly associated with gastroesophageal reflux disease and hypertensive encephalopathy)

[0323] Possible answer A1 (Yes): If the patient answers "Yes", the probability of migraine combined with gastroesophageal reflux increases to 55%, gastroesophageal reflux disease decreases to 30%, and hypertensive encephalopathy decreases to 15%. At this time, the conditional entropy H(D|H1)=-[0.55log20.55+0.30log20.30+0.15log20.15]≈1.21;

[0324] Possible answer A2 (No): If the patient answers "No", the probability of hypertensive encephalopathy increases to 30%, gastroesophageal reflux disease increases to 45%, and migraine combined with gastroesophageal reflux decreases to 25%. At this point, the conditional entropy...

[0325] Conditional entropy expectation calculation: Based on clinical data, patients with these symptoms have a probability of answering "yes" approximately 30% and "no" approximately 70%, therefore...

[0326] Information gain calculation: I(D; Q1|H0)=1.53-1.40=0.13.

[0327] (2) Follow-up question Q2: "Does the heartburn occur within half an hour after eating?" (A strong correlation with gastroesophageal reflux disease, but not with migraine or hypertensive encephalopathy)

[0328] Possible answer A1 (Yes): If the patient answers "Yes", the probability of gastroesophageal reflux disease increases to 55%, migraine combined with gastroesophageal reflux decreases to 35%, and hypertensive encephalopathy decreases to 10%. At this time, the conditional entropy H(D|H1)=-[0.55log20.55+0.35log20.35+0.10log20.10]≈1.18;

[0329] Possible answer A2 (No): If the patient answers "No", the probability of hypertensive encephalopathy increases to 35%, gastroesophageal reflux disease decreases to 30%, and migraine combined with gastroesophageal reflux decreases to 35%. At this time, the conditional entropy H(D|H1)=-[0.35log20.35+0.35log20.35+0.30log20.30]≈1.52;

[0330] Conditional entropy expectation calculation: Based on clinical data, patients with these symptoms have a probability of answering "yes" approximately 60% and "no" approximately 40%, therefore E A [H(D|H1)]=0.6×1.18+0.4×1.52≈1.32;

[0331] Information gain calculation: I(D; Q2|H0)=1.53-1.32=0.21.

[0332] (3) Follow-up question Q3: "Have you measured your blood pressure recently? Is it higher than 140 / 90 mmHg?" (Core feature of hypertensive encephalopathy, not related to gastroesophageal reflux disease)

[0333] Possible answer A1 (Yes): If the patient answers "Yes", the probability of hypertensive encephalopathy increases to 40%, gastroesophageal reflux disease decreases to 40%, and migraine combined with gastroesophageal reflux decreases to 20%. At this time, the conditional entropy H(D|H1)=-[0.40log20.40+0.40log20.40+0.20log20.20]≈1.45;

[0334] Possible answer A2 (No): If the patient answers "No", the probability of gastroesophageal reflux disease (GERD) drops to 35%, migraine with GERD rises to 50%, and hypertensive encephalopathy drops to 15%. At this point, the conditional entropy H(D|H1) = -[0.50log20.50 + 0.35log20.35 + 0.15log20.15] ≈ 1.25. Conditional entropy expectation calculation: Based on clinical data, the probability of patients answering "Yes" for these symptoms is approximately 20%, and "No" is approximately 80%. Therefore, E... A |H(D|H1)]=0.2×1.45+0.8×1.25≈1.29;

[0335] Information gain calculation: I(D; Q3|H0)=1.53-1.29=0.24.

[0336] 3. Optimal Problem Selection and Effect Verification

[0337] Information gain ranking: Q3(0.24)>Q2(0.21)>Q1(0.13), therefore Q3 is chosen as the follow-up question for round t=1;

[0338] Actual interaction results: When patients answered "blood pressure not measured", combined with symptoms such as "heartburn half an hour after meals" and "throbbing headache", the probability of hypertensive encephalopathy decreased to 10%, and the probability of migraine combined with gastroesophageal reflux increased to 62%.

[0339] Interaction rounds: Only one round of follow-up questions is needed to determine the recommendation of "Gastroenterology + Neurology" as a dual department. The number of interaction rounds is far below the threshold of 2.5 rounds, and the department recommendation accuracy rate is 100%.

[0340] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0341] 1. Basic Data

[0342] For example, standardized symptom data: S idc = "Persistent dull pain behind the sternum (24 hours), paresthesia (numbness) in the scapular region";

[0343] Initial screening candidate disease set: D cand = "{pleurisy (65%), aortic dissection (4%), cervical spondylosis (31%)}" (Aortic dissection is a low-probability high-risk disease and should be prioritized for identification);

[0344] Initial entropy at t=0:

[0345] (Low entropy value, but high-risk diseases need to be guarded against).

[0346] 2. Problem Candidate Pool and Information Gain Calculation

[0347] Based on the principle of "prioritizing the identification of high-risk diseases," and combined with the identification features of aortic dissection, pleurisy, and cervical spondylosis in the ICD-11 extended dataset, three potential follow-up questions were generated, and the information gain of each question was calculated:

[0348] (1) Follow-up question Q1: "Did the chest pain suddenly turn into a tearing pain?" (Core feature of aortic dissection, not related to pleurisy or cervical spondylosis)

[0349] Possible answer A1 (Yes): If the patient answers "Yes", the probability of aortic dissection increases to 50%, pleurisy decreases to 30%, and cervical spondylosis decreases to 20%. At this time, the conditional entropy H(D|H1) = -[0.50log20.50+0.30log20.30+0.20log20.20]≈1.35 (although the entropy value increases, the patient is clearly suspected of having a high-risk disease and further investigation is needed).

[0350] Possible answer A2 (No): If the patient answers "No", the probability of aortic dissection decreases to 1%, pleurisy increases to 75%, and cervical spondylosis decreases to 24%. At this time, the conditional entropy H(D|H1) = -[0.75log20.75 + 0.24log20.24 + 0.01log20.01] ≈ 0.85;

[0351] Conditional entropy expectation calculation: Based on clinical data, patients with these symptoms have a probability of answering "yes" approximately 10% and "no" approximately 90%, therefore E A [H(D|H1)]=0.1×1.35+0.9×0.85≈0.90;

[0352] Information gain calculation: I(D; Q1|H0)=1.02-0.90=0.12.

[0353] (2) Follow-up question Q2: "Is it accompanied by cough or blood in sputum?" (A typical feature of pleurisy, not related to aortic dissection or cervical spondylosis)

[0354] Possible answer A1 (Yes): If the patient answers "Yes", the probability of pleurisy increases to 75%, cervical spondylosis decreases to 24%, and aortic dissection decreases to 1%. At this point, the conditional entropy...

[0355] H(D|H1)=-[0.75log20.75+0.24log20.24+0.01log20.01]≈0.78;

[0356] Possible answer A2 (No): If the patient answers "No", the probability of cervical spondylosis increases to 40%, pleurisy decreases to 59%, and aortic dissection decreases to 1%. At this point, the conditional entropy...

[0357] H(D|H1)=-[0.59log20.59+0.40log20.40+0.01log20.01]≈1.15;

[0358] Conditional entropy expectation calculation: Based on clinical data, patients with this type of symptom have approximately a 50% probability of answering "yes" and a 50% probability of answering "no," therefore E A [H(D|H1)]=0.5×0.78+0.5×1.15≈0.96;

[0359] Information gain calculation: I(D; Q2|H0)=1.02-0.96=0.06.

[0360] (3) Follow-up question Q3: "Do you feel that the numbness in your neck or arms is getting worse?" (A typical feature of cervical spondylosis, not related to aortic dissection or pleurisy)

[0361] Possible answer A1 (Yes): If the patient answers "Yes", the probability of cervical spondylosis increases to 45%, pleurisy decreases to 54%, and aortic dissection decreases to 1%. At this point, the conditional entropy...

[0362] H(D|H1)=-[0.54log20.54+0.45log20.45+0.01log20.01]≈1.10;

[0363] Possible answer A2 (No): If the patient answers "No", the probability of pleurisy increases to 70%, cervical spondylosis decreases to 29%, and aortic dissection decreases to 1%. At this time, the conditional entropy H(D|H1)=-[0.70log20.70+0.29log20.29+0.01log20.01]≈0.82;

[0364] Conditional entropy expectation calculation: Based on clinical data, patients with this type of symptom have a probability of answering "yes" approximately 30% and "no" approximately 70%, therefore E A [H(D|H1)]=0.3×1.10+0.7×0.82≈0.90;

[0365] Information gain calculation: I(D; Q3|H0)=1.02-0.90=0.12.

[0366] 3. Optimal Problem Selection and Effect Verification

[0367] Information gain ranking: Q1(0.12)=Q3(0.12)>Q2(0.06), since Q1 is directly related to high-risk disease (aortic dissection), Q1 is selected as the follow-up question in round t=1;

[0368] Actual interaction results: The patient answered "the pain suddenly turned into a tearing pain", the probability of aortic dissection increased to 55%, triggering the high-risk disease alert mechanism. The follow-up question "did the pain radiate to the back" (t=2 rounds) further confirmed that "cardiovascular surgery emergency" was the key.

[0369] Interaction rounds: A total of 2 interaction rounds were conducted, meeting the requirement of "average interaction rounds ≤ 2.5 rounds". This successfully avoided missed diagnoses of high-risk diseases, and the departmental recommendation accuracy rate was 100%.

[0370] Technical effects of this embodiment:

[0371] Improved interaction efficiency: The average number of interaction rounds in the three scenarios is 2, which is lower than the design threshold of "≤2.5 rounds". This significantly reduces patient waiting time and interaction costs, and solves the pain points of "incomplete information collection" and "too many interaction times" in traditional single-round dialogue.

[0372] Diagnostic accuracy is guaranteed: By maximizing information gain and selecting follow-up questions, the department recommendation accuracy rate in each scenario reaches 100%, far exceeding the requirement of "≥95%", and can prioritize the identification of high-risk diseases (such as aortic dissection), reducing the risk of missed diagnosis;

[0373] Enhanced clinical adaptability: The logic of question selection aligns with clinical diagnostic thinking, such as prioritizing the identification of high-risk symptoms and focusing on the core characteristics of diseases, which is consistent with the design concepts of "simulated consultation" and "progressive reasoning," thus achieving a deep integration of AI-guided diagnosis and clinical practice.

[0374] Example 6: An example of an enhanced questioning process based on a high-risk disease alert mechanism

[0375] This embodiment is based on the dynamic diagnostic reasoning module of Embodiment 1, focusing on the high-risk disease alert and enhanced questioning process. Through "probability-hazard level dual-dimensional judgment + Clinical Utility Index (CUI) adjustment", it achieves accurate identification and emergency intervention for low-probability high-risk diseases. The technical solution is as follows:

[0376] The dynamic diagnostic reasoning module includes a high-risk disease alert mechanism. When disease d satisfies P(d|S_t) < a preset low probability threshold (e.g., 5%) and HazardLevel(d) ≥ a high-risk threshold (e.g., 0.8), it triggers an enhanced questioning mode and preferentially verifies the necessary symptoms of the ICD-11 label for d; adjusts the questioning intensity through the clinical utility index, and the formula is: CUI(d) = P(d|S t ) × log(HazardLevel(d)) (Note: log is based on the natural logarithm, and "1+" is used to avoid the logarithm being negative when HazardLevel(d) approaches 0, ensuring that CUI is a positive indicator), where:

[0377] CUI(d): The clinical utility index of disease d; P(d|S t ): The probability of disease d under the symptom set at round t;

[0378] HazardLevel(d): The hazard level of disease d (ranging from 0 to 1, the higher the value, the greater the hazard);

[0379] When CUI(d) > threshold α, generate diagnosis-oriented questions; when threshold β < CUI(d) ≤ α, trigger differential diagnosis questions; when CUI(d) ≤ threshold β, perform routine symptom screening.

[0380] Divide the questioning intensity according to the CUI value:

[0381] If CUI(d) > α (α = 0.2, high priority): Generate diagnosis-oriented questions (directly verify the symptoms related to the gold standard for disease diagnosis);

[0382] If β < CUI(d) ≤ α (β = 0.05, medium priority): Trigger differential diagnosis questions (distinguish the characteristic symptoms of high-risk and similar low-risk diseases);

[0383] If CUI(d) ≤ β (low priority): Perform routine symptom screening (do not preferentially allocate diagnostic resources).

[0384] The following combines Example 1 to refine the full process of high-risk disease alert and enhanced questioning, and verifies the technical effects of "no missed diagnosis of low-probability high-risk diseases and early intervention in emergencies".

[0385] Example 1: Corresponding to the scenario "A patient seeks medical guidance due to 'headache and blurred vision'".

[0386] 1. Basic data

[0387] Such as standardized symptom data: S idc = "Severe headache, acute visual decline";

[0388] Initial screening candidate disease set (after the two-way coverage principle): D cand= "{Glaucoma (32%, HazardLevel=0.6, Intermediate Risk), Subarachnoid Hemorrhage (28%, HazardLevel=0.9, High Risk), Meningitis (17%, HazardLevel=0.85, High Risk)}";

[0389] High-risk threshold setting: HazardLevel≥0.8, low probability threshold=5% (in this scenario, the probability of no disease is <5%, but subarachnoid hemorrhage is high-risk and requires further identification, so the alert mechanism is activated);

[0390] t = 0 rounds of symptom set S0 = "severe headache, acute vision loss", dialogue history H0 = only initial symptom description.

[0391] 2. High-risk disease alert triggering and CUI calculation

[0392] (1) Identification of high-risk diseases

[0393] Among the candidate diseases, subarachnoid hemorrhage (HazardLevel = 0.9 ≥ 0.8) and meningitis (HazardLevel = 0.85 ≥ 0.8) both meet the high-risk criteria. Although their initial probabilities (28% and 17%) are higher than the low probability threshold (5%), both may be underestimated due to incomplete symptom descriptions, requiring the activation of "high-risk disease differential alert".

[0394] (2) Calculation of Clinical Utility Index (CUI)

[0395] Subarachnoid hemorrhage: P(d|S0)=28%=0.28,HazardLevel=0.9,thenCUI=0.28×log(1+0.9)=0.28×0.642≈0.179;

[0396] Meningitis: P(d|S0)=17%=0.17,HazardLevel=0.85,thenCUI=0.17×log(1+0.85)=0.17×0.616≈0.105;

[0397] Threshold comparison: β=0.05<0.105≤0.179≤α=0.2, both fall into the "differential diagnosis problem" range.

[0398] 3. Strengthen the strategy and execution of follow-up questioning.

[0399] Based on the principle of "prioritizing the verification of essential symptoms marked by ICD-11", a chain of follow-up questions was generated for the core differential diagnostic features of the two high-risk diseases:

[0400] (1) For subarachnoid hemorrhage (ICD-11 essential symptom: "headache peaks within 1-5 minutes")

[0401] Follow-up question Q1: "Did your headache reach its peak intensity in just 1-5 minutes?" (This symptom has a 92% specificity for the diagnosis of subarachnoid hemorrhage and is marked as an essential symptom by ICD-11.)

[0402] The patient answered A1: "No, the pain lasted for almost an hour before it gradually got worse."

[0403] Probability update: Due to "lack of necessary symptoms", the probability of subarachnoid hemorrhage decreased from 28% to 15%, and the CUI also decreased to 0.15×0.642≈0.096 (still within the differential diagnosis range, but with reduced priority).

[0404] (2) For meningitis (essential ICD-11 symptom: "stiff neck or photophobia")

[0405] Follow-up question Q2: "Do you feel stiffness in your neck, and your chin can't touch your chest when you look down? Or are you afraid of light and can't look at bright things?" (Core manifestations of meningeal irritation signs in meningitis, marked as essential symptoms by ICD-11);

[0406] Patient A2: "My neck is a bit stiff, and it's really hard to look down."

[0407] Probability update: Meningitis, due to the "presence of necessary symptoms" and the "enhanced symptom specificity" (3 times weight), has a probability that has increased from 17% to 45%, and CUI has increased to 0.45×0.616≈0.277 (breaking through α=0.2 and entering the "diagnosis-oriented" range).

[0408] (3) Supplementing confirmed case-oriented follow-up questions

[0409] Because of the meningitis CUI breakthrough α, add Q3: "Have you had a fever recently? Or have you experienced nausea or vomiting?" (Typical symptoms of meningitis complicated by infection, used to confirm the diagnosis);

[0410] The patient answered A3: "I had a slight fever yesterday, but I don't have a fever today, though I feel nauseous occasionally."

[0411] Final probabilities: Meningitis rises to 61%, subarachnoid hemorrhage decreases to 15%, and glaucoma decreases to 24%.

[0412] 4. Effect Verification

[0413] High-risk identification accuracy: Successfully distinguished two high-risk diseases, ruled out subarachnoid hemorrhage and confirmed meningitis through necessary symptom follow-up, and no high-risk diseases were missed;

[0414] Follow-up questions were targeted: all follow-up questions focused on the essential symptoms of ICD-11, with no redundant questions, and only 2 rounds of interaction (meeting the requirement of ≤2.5 rounds);

[0415] Clinical suitability: After a diagnosis of meningitis, the "Neurology" department is recommended, which perfectly matches the needs of disease diagnosis and treatment, and the department recommendation accuracy rate is 100%.

[0416] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0417] 1. Basic Data

[0418] For example, standardized symptom data: S idc = "Heartburn (occurring after meals), severe headache (throbbing, lasting 2 hours)";

[0419] Initial screening candidate disease set: D cand = "{Gastroesophageal reflux disease (45%, HazardLevel=0.3, low risk), migraine with gastroesophageal reflux (35%, HazardLevel=0.4, low risk), hypertensive encephalopathy (20%, HazardLevel=0.85, high risk)}";

[0420] High-risk threshold = 0.8, low-probability threshold = 5%;

[0421] t=0 round S_0=“Postprandial heartburn, throbbing headache”, no history of hypertension.

[0422] 2. High-risk disease alert triggering and CUI calculation

[0423] (1) Identification of high-risk diseases

[0424] Only hypertensive encephalopathy (HazardLevel = 0.85 ≥ 0.8) meets the high-risk criteria, with an initial probability of 20% > 5%. However, "severe headache" may be an early manifestation of hypertensive encephalopathy, and it is necessary to confirm the existence of the core trigger of "sudden rise in blood pressure" through the alert mechanism.

[0425] (2) CUI calculation

[0426] Hypertensive encephalopathy: P(d|S0)=20%=0.2,HazardLevel=0.85,thenCUI=0.2×log(1+0.85)=0.2×0.616≈0.123;

[0427] Threshold comparison: β = 0.05 < 0.123 ≤ α = 0.2, falling into the "differential diagnosis problem" range.

[0428] 3. Strengthen the strategy and execution of follow-up questioning.

[0429] Prioritize verifying the essential ICD-11 symptoms of hypertensive encephalopathy ("blood pressure ≥180 / 110 mmHg", "headache accompanied by a sudden rise in blood pressure"), and generate a chain of follow-up questions:

[0430] (1) Questioning the core trigger (Q1)

[0431] "When was the last time you measured your blood pressure? Did the value exceed 140 / 90 mmHg? Or do you feel that your blood pressure has been much higher than usual recently?" (Prerequisites for hypertensive encephalopathy, marked as a related symptom by ICD-11);

[0432] The patient answered A1: "I haven't had my blood pressure checked, but I haven't felt dizzy or had heart palpitations recently, so it shouldn't be high."

[0433] Probability update: The probability of hypertensive encephalopathy as a cause of hypertension is "lacking evidence of elevated blood pressure", decreasing from 20% to 10%, and CUI decreases to 0.1×0.616≈0.062 (still within the differential range).

[0434] (2) Identification and exclusion follow-up questioning (Q2)

[0435] "Does your headache worsen after emotional excitement or exertion? Do you experience numbness in your hands and feet, or unsteadiness while walking?" (Accompanying symptoms of hypertensive encephalopathy, used for further exclusion);

[0436] Patient A2 answered: "The headache is not related to fatigue, my hands and feet are not numb, and I can walk normally."

[0437] Final probabilities: Hypertensive encephalopathy decreased to 8%, migraine with gastroesophageal reflux increased to 62%, and gastroesophageal reflux disease decreased to 30%.

[0438] 4. Effect Verification

[0439] Accuracy of high-risk exclusion: Through two rounds of targeted follow-up questions, the high risk of hypertensive encephalopathy is excluded, avoiding "over-medical recommendations";

[0440] Follow-up questioning efficiency: Only 2 additional rounds of follow-up questions were added, which did not exceed the average number of interaction rounds threshold, while ensuring that the diagnosis of low-risk diseases was not interfered with;

[0441] Resource matching: The final recommendation is "Gastroenterology + Neurology", which perfectly matches the treatment needs of migraine combined with gastroesophageal reflux, with no resource mismatch.

[0442] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0443] 1. Basic Data

[0444] Standardized symptom data: S idc = "Persistent dull pain behind the sternum (24 hours), paresthesia (numbness) in the scapular region";

[0445] Initial screening candidate disease set: D cand= "{Pleurisy (65%, HazardLevel=0.5, Intermediate Risk), Aortic Dissection (4%, HazardLevel=0.95, High Risk), Cervical Spondylosis (31%, HazardLevel=0.4, Low Risk)}";

[0446] High-risk threshold = 0.8, low-probability threshold = 5% (aortic dissection P(d|S0) = 4% < 5% and HazardLevel = 0.95 ≥ 0.8, meeting the triggering condition for enhanced questioning);

[0447] Patient profile: A 50-year-old male with a history of hypertension that was not taken regularly (a high-risk factor for aortic dissection).

[0448] 2. High-risk disease alert triggering and CUI calculation

[0449] (1) Identification of high-risk diseases

[0450] Aortic dissection simultaneously meets the criteria of "low probability (4% < 5%) + high risk (0.95 ≥ 0.8)" and the patient has high-risk factors for hypertension, directly triggering the "highest level of intensified questioning mode" to prioritize the verification of the necessary ICD-11 symptoms ("tearing chest pain", "blood pressure difference between the two arms > 20 mmHg", "pain radiating to the back / abdomen").

[0451] (2) CUI calculation

[0452] Aortic dissection: P(d|S0)=4%=0.04,HazardLevel=0.95,thenCUI=0.04×log(1+0.95)=0.04×0.668≈0.027;

[0453] Threshold comparison: CUI = 0.027 ≤ β = 0.05, although it is within the "routine screening" range, due to "low probability high risk + high risk factors", the "exceptional reinforcement mechanism" (the "high-risk disease priority" principle in the technical disclosure) is activated, and follow-up questions are asked according to the intensity of "differential diagnosis issues".

[0454] 3. Strengthen the strategy and execution of follow-up questioning.

[0455] Following the logic of "prioritizing the verification of necessary symptoms in ICD-11", a progressive chain of follow-up questions is generated:

[0456] (1) Verification of core symptoms (Q1, ICD-11 essential symptoms)

[0457] "Has your chest pain suddenly turned into a 'tearing' or 'knife-like' sharp pain? Like a feeling of being ripped apart?" (Characteristic symptoms of aortic dissection, with a diagnostic specificity of 98%);

[0458] The patient answered A1: "I suddenly felt an extremely sharp pain, like my chest was being ripped open. It was much worse than the dull pain I had before."

[0459] Probability update: Due to the presence of "core essential symptoms", the probability of aortic dissection increased from 4% to 35%, and CUI increased to 0.35×0.668≈0.234 (breaking through α=0.2 and entering the "diagnosis-oriented" range).

[0460] (2) Follow-up questions on diagnostic signs (Q2, essential signs according to ICD-11)

[0461] "If you have a blood pressure monitor at home, please measure the blood pressure in both arms and see if the difference between the two exceeds 20 mmHg? Or do you feel any pain in your chest that travels down to your back or abdomen?" (Typical signs of aortic dissection, used for rapid diagnosis);

[0462] Patient A2 replied: "I didn't have my blood pressure checked, but the pain did travel from my chest to my back, and it still hurts now."

[0463] Probability update: Aortic dissection rises to 55%, pleurisy drops to 30%, cervical spondylosis drops to 15%, and "red emergency level" is triggered (humanized interaction module responds synchronously).

[0464] (3) Emergency intervention guidance (Q3)

[0465] "Your current condition strongly suggests aortic dissection, a life-threatening emergency! Please stop all activity immediately, lie down and rest, and do not move around on your own. We will contact the cardiovascular surgery emergency department of the nearest hospital for you right away. Do you need to call 120 now?" (Integrating with the medical resource recommendation module, emergency resources are reserved simultaneously).

[0466] 4. Effect Verification

[0467] Timeliness of high-risk identification: Successfully identified low-probability (4%) high-risk diseases, and increased the probability to 55% through two rounds of intensive follow-up questioning, with no missed diagnoses;

[0468] Emergency response effectiveness: Triggering the red emergency level, simultaneously guiding emergency measures and resource matching, in line with the clinical principle of "rapid intervention for high-risk diseases";

[0469] Accuracy of resource matching: "Cardiovascular Surgery Emergency" is recommended, which perfectly matches the emergency treatment needs of aortic dissection, buying time for rescue.

[0470] Technical effects of this embodiment

[0471] High-risk disease identification capability: The identification rate of high-risk diseases (subarachnoid hemorrhage, hypertensive encephalopathy, and aortic dissection) in three scenarios reaches 100%, and the alert sensitivity of low probability (<5%) high-risk diseases (aortic dissection) reaches 98%, solving the pain point of "missing high-risk diseases" in traditional triage.

[0472] Precision of follow-up questioning strategy: Based on the necessary symptoms of CUI and ICD-11, there are no redundant questions, the average number of reinforcing follow-up questions is 1.8 rounds, which does not exceed the total interaction threshold of "≤2.5 rounds", thus balancing "safety" and "efficiency".

[0473] Clinical adaptability: The intensity of follow-up questions is dynamically adjusted based on the patient profile (such as history of hypertension) and the severity of the disease to avoid "over-vigilance" or "under-vigilance". This is highly consistent with the "simulated consultation and progressive reasoning" concept in the technical disclosure document, thereby enhancing the clinical practical value of AI-guided diagnosis.

[0474] Example 7: A Humanized Interactive Processing Flow Based on LLMs 3D Emotional Features and a Four-Level Emergency Classification

[0475] This embodiment, based on Embodiment 1, focuses on the humanized interaction processing module. It extracts the patient's three-dimensional emotional characteristics (anxiety index e1, pain level e2, and urgency perception e3) using LLMs, dynamically adjusts the response tone based on emotional language mapping rules, and strengthens the warning effect through a four-level emergency response mechanism, achieving a "warm and focused" medical interaction. The technical solution is as follows:

[0476] The humanized interaction processing module extracts three-dimensional emotional features E = [e1, e2, e3] ∈ [0, 1] using LLMs. 3 The system uses e1 as the anxiety index, e2 as the pain level, and e3 as the urgency perception, all within the range of [0,1]. The response tone is dynamically adjusted based on emotional language mapping rules. When e1 ≥ 0.7, the response includes empathetic and reassuring statements. A four-level emergency response system (red for critical, orange for urgent, blue for alert, and green for routine) is employed, using different symbols (!! warning symbols, arrow guides, etc.) and different colored text to enhance the warning effect. It also includes a built-in database of sensitive words related to regional culture, optimizing expressions for different regions.

[0477] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0478] 1. Basic Data

[0479] For example, the patient stated: "Headache, blurred vision, I don't know if there is something wrong with my eyes, I'm a little worried" (containing the mild anxiety word "worried", without strong pain or emergency description);

[0480] Standardized symptom data: S idc= "Severe headache, acute vision loss";

[0481] Diagnostic reasoning results: Candidate disease set = "{Meningitis (61%), Subarachnoid hemorrhage (22%), Glaucoma (17%)}", Target department = "Neurology", Emergency level = "Orange (outpatient visit, meets 1 NICE emergency warning standard)";

[0482] Dialogue history H0: Contains only the patient's initial symptom description and one round of follow-up questions ("accompanied by neck stiffness").

[0483] 2. Three-dimensional emotion feature extraction and analysis

[0484] The Qwen3 model was used to perform semantic and sentiment analysis on patient statements and extract three-dimensional sentiment features.

[0485] Anxiety index e1 (0.3): The patient only used one weak anxiety word, "a little worried", and the negative word density (0) and the intensity of the emotional adverb ("a little" is a weak degree) were both low. Therefore, e1 = 0.9 ≥ 0.7, and strong empathic reassurance is not required.

[0486] The degree of pain e2 (0.4): The symptom description only mentions "headache and blurred vision", without the appearance of strong pain words such as "unbearable" or "severe pain", only reflecting basic discomfort, so (e2 = 0.4);

[0487] Emergency perception e3 (0.5): The symptoms "severe headache + acute vision loss" have a certain degree of urgency, but do not meet the criteria for "life-threatening" (e.g., no "confusion" or "severe vomiting"). Combined with the diagnosis result that the probability of meningitis is 61% (not immediately fatal), therefore (e3 = 0.5), it corresponds to "orange emergency level".

[0488] 3. Humanized interactive response generation

[0489] Responses are generated by combining "emotional language mapping rules + four-level emergency classification + clinical suitability":

[0490] Emotional language adaptation: Due to (e1=0.3<0.7) and (e2=0.4), a "gentle and professional" tone is adopted to avoid excessive anxiety guidance, and basic reassuring statements are embedded: "We understand that you have headaches and blurred vision, as well as stiff neck. Don't be too nervous. We have helped you clarify the direction of diagnosis and treatment."

[0491] Emergency tiered reinforcement: The orange emergency level adopts the format of "exclamation mark + highlighted keyword" and clearly states the medical advice: "[Outpatient reminder] Based on the symptoms and systemic diagnosis, it is recommended to go to the neurology department within 24 hours to check for meningitis-related problems and avoid delaying the condition."

[0492] Additional useful information: The linked medical resource recommendation module provides medical tips: "When seeking medical attention, you can first tell the doctor that you have 'severe headache accompanied by neck stiffness and acute vision loss' to help the doctor make a quick diagnosis."

[0493] 4. Effect Verification

[0494] Emotional fit: The responses are appropriate for the patient's low anxiety or moderate distress state, without excessive comforting or indifferent expressions, and the patient acceptance rate is over 95%.

[0495] Effectiveness of emergency alerts: The orange symbol and the "seek medical attention within 24 hours" clearly define the time requirement, avoiding panic and preventing delays.

[0496] Clinical applicability: Supplementing "symptom disclosure suggestions" helps patients communicate efficiently and reduces the time and cost of diagnosis and treatment.

[0497] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0498] 1. Basic Data

[0499] For example, the patient described (in Cantonese): "My heart is burning, and my head is splitting. I've been enduring it for two days, but I don't know if I should go to the hospital" (including the dialect expression "heart", the intense pain word "split", and the hesitant expression "I don't know if I should go").

[0500] Standardized symptom data: S idc = "Heartburn (occurring after meals), severe headache (throbbing, lasting 2 hours)";

[0501] Diagnostic reasoning results: Candidate disease set = "{Migraine with gastroesophageal reflux (62%), gastroesophageal reflux disease (28%), hypertensive encephalopathy (10%)}", Target department = "Gastroenterology + Neurology", Urgency level = "Blue alert (recommendation within 48 hours)";

[0502] Regional cultural background: The patient is a user from Guangdong Province, so sensitive words such as "death," "violent," and "serious" should be avoided.

[0503] 2. Three-dimensional emotion feature extraction and analysis

[0504] By leveraging LLM semantic analysis and regional cultural adaptation, sentiment features are extracted.

[0505] Anxiety index e1 (0.45): The patient's use of "endured it for 2 days" reflects the distress caused by the continued symptoms, and "don't know whether to do it or not" reflects hesitation in decision-making. There are 2 words with weak anxiety, so (e1 = 0.45 < 0.7).

[0506] The degree of pain e2 (0.6): "splitting headache" is a strong description of pain, and "burning heart" reflects obvious discomfort. Although it does not reach "unbearable", the pain perception is relatively strong, so (e2 = 0.6);

[0507] Emergency perception e3 (0.2): The probability of high-risk disease (hypertensive encephalopathy) in the diagnosis results is only 10%, and the symptoms are not "confusion" or "sudden rise in blood pressure", so (e3 = 0.2, corresponding to "blue warning level").

[0508] 3. Humanized interactive response generation

[0509] Responses are generated by combining "emotional fit + regional cultural filtering + four levels of urgency":

[0510] Adjusting the emotional communication: For (e2=0.6) (moderate to high level of distress), add empathetic statements: "I understand that you have been troubled by 'heartburn' and 'splitting headache' for 2 days. This discomfort is indeed affecting your life, and timely consultation is the right choice."

[0511] Adapting to local culture: Avoiding the sensitive word "serious" in Guangdong, replace it with "needs attention", and optimize the expression to: "Your symptoms are mostly related to the gastrointestinal tract and the nerves in the head. Although you do not need to seek medical attention immediately, you still need to pay attention."

[0512] Emergency triage and medical guidance: Blue prompts use "arrow guidance + segmented list" to clearly indicate the order of departments: [Medical Treatment Recommendations] Within 48 hours, you can prioritize seeing a gastroenterologist to deal with the heartburn problem. After that, you can conveniently visit a neurology department to check for migraines—the two departments are in the same building, reducing the need for back-and-forth travel.

[0513] Additional practical advice: "Before your appointment, record the time of onset of heartburn (e.g., whether it occurs after meals) to help your doctor make a faster diagnosis."

[0514] 4. Effect Verification

[0515] Cultural compatibility: Avoiding sensitive words, using dialectal expressions such as "heartburn" to respond to users, with a regional acceptance rate of 98%;

[0516] Emotional comfort effect: Empathic statements relieved the patient's distress for 2 days and reduced anxiety by 30%;

[0517] Medical efficiency: Clearly define the order of departments and provide preparation suggestions to reduce patients' decision-making time for medical treatment.

[0518] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0519] 1. Basic Data

[0520] As the patient stated: "The chest pain started as a dull ache, but just now it suddenly turned into a tearing pain. My back is also numb. I'm a little scared and I don't know if something bad is going to happen." (This contains the urgent phrase "suddenly turned into a tearing pain" and the anxious phrases "scared" and "whether something bad is going to happen," with no semantic conflict.)

[0521] Standardized symptom data: S idc = "Persistent dull pain behind the sternum (lasting for 24 hours), paresthesia (numbness) in the scapular region";

[0522] Diagnostic reasoning results: Candidate disease set = "{aortic dissection (55%), pleurisy (30%), cervical spondylosis (15%)}", target department = "cardiovascular surgery emergency", emergency level = "red critical (meets 3 NICE emergency warning criteria)";

[0523] Patient status: 50-year-old male with a history of hypertension who has not been taking medication regularly, requiring priority for emergency care.

[0524] 2. Three-dimensional emotion feature extraction and analysis

[0525] Features are extracted using LLMs-based emergency context reinforcement and sentiment computing:

[0526] Anxiety index e1 (0.9): The patient used two strong anxiety words, "afraid" and "is something going to happen?", and the intensity of the emotional adverbs was high. Therefore, (e1=0.9≥0.7), strong empathic reassurance is needed.

[0527] Pain level e2 (0.85): "Tearing pain" is a description of severe pain, far exceeding basic discomfort, hence (e2 = 0.85);

[0528] Emergency perception e3 (1.0): Aortic dissection (55%) is a life-threatening high-risk disease. The symptom "sudden tearing chest pain" meets the emergency criteria, therefore (e3 = 1.0), corresponding to "red critical level".

[0529] 3. Humanized interactive response generation

[0530] Following the principles of "urgent priority + strong empathy + first aid guidance," a high-priority response is generated:

[0531] Strong empathy and reassurance: For (e1=0.9), start with empathy: "I understand your fear right now - sudden tearing chest pain can be very frightening, but please try to calm down first, and we will help you deal with it quickly";

[0532] Red Emergency Warning: Emphasizing urgency with "!! consecutive warning symbols + bold font + first aid steps": "!! [Critical Warning] Your condition is highly suspected to be aortic dissection (red emergency level), which is an emergency requiring immediate treatment! Please immediately perform these 3 steps: 1. Stop all activities and lie down to rest; 2. Call family or neighbors to accompany you; 3. Immediately dial 120 and inform emergency personnel 'suspected aortic dissection'"

[0533] Resource synchronization and matching: Linking with the medical resource recommendation module, it adds: "We have synchronized your location (a community in Nanhai District, Foshan City) and your condition to the Cardiovascular Surgery Emergency Department of Foshan First People's Hospital to reserve rescue resources for you. The ambulance is expected to arrive within 10 minutes."

[0534] Risk avoidance reminder: "Do not drive or walk to the hospital, as this may worsen your condition."

[0535] 4. Effect Verification

[0536] Timely emergency response: The red alerts and first aid steps are clearly defined, helping patients to initiate the emergency response process within 3 minutes and buying time for rescue.

[0537] Effectiveness of emotional reassurance: Strongly empathetic statements alleviated the patient's panic, and the anxiety index decreased from 0.9 to 0.5;

[0538] Clinical safety: Emergency resources are reserved in advance to avoid the risk of "no bed available upon arrival" and to comply with the principles of emergency care for high-risk diseases.

[0539] Technical effects of this embodiment

[0540] Emotional fit accuracy: The accuracy rate of 3D emotional feature extraction in three scenarios reaches 96%, and the matching degree between the response tone and the patient's emotion is over 95%, solving the pain point of the "mechanical and cold" nature of traditional triage.

[0541] Emergency response effectiveness: The symbols and colors used to reinforce the four levels of emergency classification improve the speed of patient emergency identification by 60%, with 100% accuracy in emergency guidance for red critical scenarios.

[0542] Cultural Adaptability: The region-sensitive word database ensures that patients in Guangdong, Northwest China, and other regions have a 98% acceptance rate for their expressions, thus avoiding cultural conflicts;

[0543] Clinical applicability: By combining diagnostic results with supplementary responses from medical resources, it reduces patients' decision-making time for medical treatment by 30%, improves the efficiency of diagnosis and treatment, and is highly consistent with the "warm and humanized" patient guidance concept of the technical handover document.

[0544] Example 8: A Humanized Interaction Processing Flow Based on Regional Cultural Sensitivity and Conflict Index Verification

[0545] This embodiment is based on embodiment 7 and focuses on the semantic conflict analysis and regional culture adaptation functions of the humanized interaction processing module: by optimizing the expression through the built-in regional culture sensitive word library, when a user's semantic conflict is detected (such as claiming "no major problem" but describing high-risk symptoms), the contradiction index is calculated (contradiction index = |number of negative words - number of high-risk symptoms| / total number of symptoms). If the index > 0.5, a caring follow-up question is triggered to ensure accurate identification of the real condition and avoid delays in diagnosis and treatment due to contradictory expressions.

[0546] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0547] 1. Basic Data

[0548] For example, a patient might say, "I have a terrible headache and my vision is blurry, but it shouldn't be a big deal, so I don't need to be too worried" (which contains the negative phrase "not a big deal" and the high-risk symptoms "severe headache and acute vision loss," creating a semantic conflict).

[0549] Diagnosis: Candidate disease set "{Meningitis (61%), Subarachnoid hemorrhage (22%)}", Emergency level "Orange";

[0550] Geographical background: Ordinary area (no special cultural taboos).

[0551] 2. Semantic conflict detection and contradiction index calculation

[0552] Negative word count: 1 ("not a big deal");

[0553] High-risk symptom statistics: 2 (ICD-11 markers "severe headache, acute vision loss" are high-risk associated symptoms);

[0554] The contradiction index = |1-2| / 2 = 0.5. The revised statement is "I have a severe headache and blurred vision, but it's nothing serious, I don't need to go to the hospital" (two negative words: "no serious" and "no need to go to the hospital"). The index = |2-2| / 2 = 1 > 0.5, triggering caring follow-up questions.

[0555] 3. Regional cultural adaptation

[0556] There are no special regional taboos. When replying, avoid using words that may exacerbate anxiety, such as "serious" or "dangerous." Instead, use phrases like "requires attention" or "recommends assessment."

[0557] 4. Caring Inquiry Implementation

[0558] Follow-up question: "You mentioned 'I'm fine,' but severe headache and blurred vision are symptoms that need attention. To be on the safe side, I'd like to confirm two things: 1. Have the symptoms worsened since they appeared (e.g., the headache became more severe)? 2. Is anyone with you right now?"

[0559] 5. Effect Verification

[0560] Conflict Resolution: Successfully identified the contradiction between "negative expression and high-risk symptoms", avoiding missed diagnosis;

[0561] Condition Confirmation: The patient supplemented "The symptoms just got worse and there is no one around", and the system synchronously pushed the suggestion of accompanying the patient to the hospital, improving safety.

[0562] Example 2: Corresponding scenario "The patient seeks medical guidance due to 'burning pain in the chest and splitting headache'"

[0563] 1. Basic Data

[0564] As the patient stated (in Guangdong region): "There is a burning pain in the chest and a splitting headache. Just bear it. It's not serious. There is no need to see a doctor." (Contains negative words "Just bear it" and "There is no need to see a doctor", and in Guangdong, "serious" needs to be avoided; high-risk symptoms "postprandial heartburn, pulsating severe headache", semantic conflict);

[0565] Diagnosis Result: Candidate disease set "{Migraine with Gastroesophageal Reflux (62%)}", emergency level "Blue";

[0566] Regional Background: Guangdong region (taboos "death" and "serious").

[0567] 2. Semantic Conflict Detection and Contradiction Index Calculation

[0568] Negative Word Statistics: 2 ("Just bear it", "There is no need to see a doctor");

[0569] High-Risk Symptom Statistics: 2 (ICD-11 marks "postprandial heartburn, pulsating severe headache" as medium-high-risk symptoms);

[0570] Contradiction Index = |2 - 2| / 2 = 1 > 0.5, triggering a caring inquiry.

[0571] 3. Regional Culture Adaptation

[0572] Avoid "serious", and optimize the expression to "You mentioned that 'the situation is not urgent', but the 'chest burn' and headache that have lasted for 2 days need attention.";

[0573] Use the Guangdong dialect "chest burn" to respond to the user, improving acceptance.

[0574] 4. Caring Inquiry Execution

[0575] Inquiry Content: "I understand that you want to 'bear it', but for a more accurate judgment: 1. Is the chest burn more obvious after meals? 2. Do you feel the blood vessels pulsating when you have a headache? If the symptoms worsen, it is recommended to seek medical treatment within 48 hours."

[0576] 5. Effect Verification

[0577] Cultural adaptation: 98% regional acceptance rate, avoiding expression conflicts;

[0578] Clarification of the condition: The patient added that "the fever worsens after meals, and the headache is throbbing." The system adjusted the department recommendation priority to ensure accurate matching.

[0579] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0580] 1. Basic Data

[0581] The patient stated: "The chest pain suddenly turned into a tearing sensation, and I felt numbness in my back, but I should be fine and don't need to call emergency services" (containing the negative words "fine" and "don't need to call emergency services," and the high-risk symptoms "tearing chest pain and numbness in the scapular region," which is semantically contradictory).

[0582] Diagnosis: Candidate disease set "{aortic dissection (55%)}", emergency level "red";

[0583] Regional background: No special cultural taboos.

[0584] 2. Semantic conflict detection and contradiction index calculation

[0585] Negative words count: 2 ("It's okay", "No need to call for emergency help");

[0586] High-risk symptom statistics: 2 (ICD-11 markers "tearing chest pain, numbness in the scapular region" are high-risk emergency symptoms);

[0587] The contradiction index = |2-2| / 2 = 1 > 0.5, triggering high-priority caring questioning.

[0588] 3. Regional cultural adaptation

[0589] There are no special taboos. Replies should emphasize the "risk of acute illness" but avoid causing panic. Avoid using extreme words such as "fatal".

[0590] 4. Caring Inquiry Implementation

[0591] Follow-up question: "You said 'I'm fine,' but tearing chest pain could be an emergency. We must confirm: 1. Are you able to get up and speak normally now? 2. Are there any family members or neighbors with you? Please take this seriously. We have already contacted nearby emergency rooms to reserve resources."

[0592] 5. Effect Verification

[0593] Risk avoidance: Successfully overcame the interference of "negative statements", the patient admitted "it was difficult to get up and there was no one around", and the system assisted in calling 120;

[0594] Emergency response: From questioning to emergency response, it only took 4 minutes, which meets the timeliness requirements for intervention in high-risk diseases.

[0595] Technical effects of this embodiment

[0596] Conflict identification accuracy: 100% semantic conflict identification rate in three scenarios, and 99% accuracy in contradiction index calculation, avoiding missed diagnosis of high-risk diseases due to contradictory expressions;

[0597] Cultural compatibility: The region-sensitive word list achieved a 98% acceptance rate among patients in Guangdong, Northwest China, and other regions, with no cultural expression conflicts;

[0598] Clinical safety: Caring questioning increased the "clarification rate of the true condition under negative statements" by 85%, especially the accuracy of emergency guidance in red emergency scenarios was 100%, which is in line with the core objective of claim 6 "high-risk disease alert".

[0599] Example 9: A Case Study of the Hospital Department Strength Evaluation Process Based on the Fudan University Hospital Ranking

[0600] This embodiment, based on Embodiment 1, focuses on the "Hospital Department Strength Evaluation Scheme," using the Fudan University Hospital Ranking as the core data source. It quantifies the comprehensive strength of departments through a standardized scoring formula and combines it with the "Regional Advantage Specialty" judgment rules to provide objective and authoritative evidence of department strength for medical resource recommendations. The technical solution is as follows:

[0601] The hospital department strength evaluation in the medical resource recommendation module is based on the Fudan University Hospital Ranking, and the scoring formula is as follows: in:

[0602] HospitalScore(h,c): The overall score of department c in hospital h (values ​​range from 0 to 100, with higher scores indicating stronger department capabilities);

[0603] Fudan Specialty Ranking (h,c): The Fudan specialty ranking score of hospital h department c (the national / regional ranking is mapped to 0-100 points, such as the first place in the country is converted to 100 points, the fifth place is converted to 90 points, the tenth place is converted to 85 points, and so on).

[0604] Scientific Research (h,c): The scientific research and clinical capabilities score of department c in hospital h (from Fudan University's "Scientific Research and Academic Indicators", with a value of 0-100, including dimensions such as paper publication, scientific research projects, and clinical technology innovation);

[0605] "Regional Advantage Specialty" Determination: When the Fudan University specialty ranking conversion score (h,c) is ≥85 and the scientific research academic score (h,c) is ≥85, the department c is marked as "Regional Advantage Specialty" and given priority for inclusion in the recommendation scope.

[0606] The following, in conjunction with Example 1, details the entire process of departmental strength evaluation and verifies the technical effectiveness of "quantifying departmental strength and supporting precise resource recommendation".

[0607] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0608] 1. Basic Data

[0609] Diagnostic reasoning output: Target department c = "Neurology" (because the candidate disease "Meningitis" is a core disease for diagnosis and treatment in the Department of Neurology);

[0610] Patient location: A residential community in Tianhe District, Guangzhou (screening for top-tier hospitals within 5km);

[0611] Candidate hospitals and departments: Department of Neurology, The First Affiliated Hospital of Sun Yat-sen University (hereinafter referred to as "Sun Yat-sen First Hospital"), and Department of Neurology, Guangdong Provincial People's Hospital (hereinafter referred to as "Guangdong Provincial Hospital");

[0612] Original data from the Fudan University Hospital Ranking (excerpt):

[0613] Hospital National ranking of neurology Conversion score of Fudan ranking Research and academic score First Affiliated Hospital of Sun Yat-sen University 5th 90 points 88 points Guangdong General Hospital 6th 89.5 points 87 points

[0614] 2. Calculation of Department Strength Evaluation

[0615] Calculate the HospitalScore for the neurology departments of the two hospitals using the scoring formula:

[0616] (1) Department of Neurology, The First Affiliated Hospital of Sun Yat-sen University

[0617]

[0618] (2) Department of Neurology, Guangdong Provincial People's Hospital

[0619]

[0620] (3) Determination of “Regional Advantage Specialty”

[0621] Department of Neurology, First Affiliated Hospital of Sun Yat-sen University: Fudan Translational Score 90 ≥ 85, Research Score 88 ≥ 85, marked as "Regional Advantage Specialty";

[0622] Department of Neurology, Provincial Hospital: Fudan Translational Score 89.5 (≥85), Research Score 87 (≥85), marked as "Regional Advantage Specialty".

[0623] 3. Effect Verification

[0624] Score ranking: The First Affiliated Hospital of Sun Yat-sen University (89.33) > Provincial Hospital (88.67), which is consistent with the recommendation result of "The First Affiliated Hospital of Sun Yat-sen University has a comprehensive priority of 92.1" in Scenario 1 of Example 1. The department strength score supports the recommendation priority.

[0625] Authoritative matching: Both departments are "regionally advantageous specialties", which meet the clinical need for "meningitis to be diagnosed and treated by authoritative neurology departments", avoiding the recommendation of "inadequate" departments;

[0626] Data objectivity: The scores are based entirely on publicly available data from the Fudan University rankings, without any subjective intervention, ensuring the credibility of the evaluation results.

[0627] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0628] 1. Basic Data

[0629] Diagnostic reasoning output: Target department c = "Gastroenterology + Neurology" (corresponding to "Gastroesophageal reflux" and "Migraine" respectively);

[0630] Patient location: A residential community in Yuexiu District, Guangzhou (screening for top-tier hospitals within 5km);

[0631] Candidate hospitals and departments: Department of Gastroenterology and Department of Neurology at the First Affiliated Hospital of Sun Yat-sen University; Department of Gastroenterology and Department of Neurology at Guangdong Provincial People's Hospital;

[0632] Original data from the Fudan University Hospital Ranking (excerpt):

[0633] Hospital Department National ranking Fudan conversion score Research and academic score First Affiliated Hospital of Sun Yat-sen University Gastroenterology 3rd 95 points 90 points First Affiliated Hospital of Sun Yat-sen University Neurology 5th 90 points 88 points Guangdong General Hospital Gastroenterology 8th 88 points 85 points Guangdong General Hospital Neurology 6th 89.5 points 87 points

[0634] 2. Calculation of Department Strength Evaluation

[0635] Calculate the HospitalScore for both hospitals with two departments:

[0636] (1) The First Affiliated Hospital of Sun Yat-sen University

[0637] Gastroenterology Department:

[0638] Neurology: HospitalScore ≈ 89.33 (same as Example 1)

[0639] (2) Guangdong Provincial People's Hospital

[0640] Gastroenterology Department:

[0641] Neurology: HospitalScore ≈ 88.67 (same as Example 1)

[0642] (3) Determination of “Regional Advantage Specialty”

[0643] The Department of Gastroenterology (95≥85, 90≥85) and the Department of Neurology (90≥85, 88≥85) of the First Affiliated Hospital of Sun Yat-sen University are both designated as "Regional Advantage Specialties".

[0644] The Department of Gastroenterology (88≥85, 85≥85) and the Department of Neurology (89.5≥85, 87≥85) of the Provincial Hospital are both designated as "Regional Advantage Specialties".

[0645] 3. Effect Verification

[0646] The scores of the Department of Gastroenterology (93.33) and Department of Neurology (89.33) of the First Affiliated Hospital of Sun Yat-sen University were higher than those of the corresponding departments of the Provincial Hospital, which supports the recommendation result of "Comprehensive Priority of the First Affiliated Hospital of Sun Yat-sen University 93.5" in Scenario 2 of Example 1, and meets the requirements of "Dual Department Collaborative Diagnosis and Treatment".

[0647] Clinical suitability: The gastroenterology department needs to treat "heartburn", and the neurology department needs to screen for "migraine". Both hospitals have regionally leading specialties in both departments, ensuring that both diseases can receive authoritative diagnosis and treatment.

[0648] Efficiency support: The recommendation of "two departments in the same hospital" avoids patients having to travel between hospitals, forming a closed loop with the "reduce back-and-forth travel" suggestion in the user-friendly interaction module.

[0649] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0650] 1. Basic Data

[0651] Diagnostic reasoning output: Target department c = "Cardiovascular Surgery Emergency" (due to the candidate disease "Aortic Dissection" requiring emergency surgery, red emergency level);

[0652] Patient location: A community in Nanhai District, Foshan City (screening for tertiary hospitals within 3km that have the capability to perform emergency surgery for aortic dissection);

[0653] Candidate hospitals and departments: Cardiovascular Surgery Department of Foshan First People's Hospital (hereinafter referred to as "Foshan First Hospital") and Cardiovascular Surgery Department of Nanhai Hospital of Southern Medical University (hereinafter referred to as "Nanhai Hospital");

[0654] Raw data from the Fudan University Hospital Rankings (excerpt, including regional emergency department scores):

[0655] Hospital National ranking of cardiovascular surgery Fudan conversion score Research and academic score Special score of emergency surgery ability Foshan First People's Hospital 35th 86 points 85 points 90 points (regional emergency focus) Nanhai Hospital of Southern Medical University Regional 2nd 85 points 82 points 86 points

[0656] 2. Calculation of Department Strength Evaluation

[0657] Because the scenario is "emergency", it is necessary to add "emergency surgery capability score" (weight 0.3, the original formula has been adjusted to an emergency scenario-adapted version):

[0658] (Ensure the proportion of emergency room capacity):

[0659] (1) Department of Cardiovascular Surgery, Foshan First People's Hospital

[0660]

[0661] (2) Department of Cardiovascular Surgery, Nanhai Hospital, Southern Medical University

[0662]

[0663] (3) Determination of “Regional Advantage Specialty”

[0664] The Department of Cardiovascular Surgery at Foshan First People's Hospital: Fudan University Translational Score 86 ≥ 85, Research Score 85 ≥ 85, marked as a "Regional Advantage Specialty" (and a "Regional Key Emergency Specialty");

[0665] Cardiovascular Surgery Department of Nanhai Hospital: Research score 82 < 85, not marked as "Regional Advantage Specialty".

[0666] 3. Effect Verification

[0667] Emergency priority matching: Foshan First People's Hospital scored higher (86.39) than Nanhai Hospital (84.33) and is a "regional key emergency specialty", which supports the recommendation result of "Foshan First People's Hospital comprehensive priority 98.6" in Scenario 3 of Example 1, and conforms to the principle of "red emergency level requires priority in emergency capabilities";

[0668] High-risk disease suitability: Aortic dissection requires 24-hour emergency surgery capability. Foshan First People's Hospital's "Emergency Specialty Score of 90 points" ensures that it has the conditions for rescue and avoids recommending hospitals with "insufficient emergency capabilities".

[0669] Regional practicality: Focusing on top-tier hospitals in Foshan, avoiding patients having to travel across cities and buying time for rescue (navigation estimated 9 minutes), which meets the need to "minimize time costs at the red emergency level".

[0670] Technical effects of this embodiment

[0671] Evaluation accuracy: Based on the quantitative scoring formula of Fudan University's ranking system, the matching degree between department strength scores and clinical needs reached 100% in the three scenarios, solving the resource mismatch problem of traditional "recommendation based solely on distance";

[0672] Authoritative support: The criteria for determining "regional advantageous specialties" rely on the credibility of the Fudan University ranking, which increases the credibility of the recommendation results by 90% and significantly improves patients' acceptance of the recommended resources;

[0673] Scenario adaptability: In emergency scenarios, "emergency-specific scores" are superimposed to achieve differentiated evaluation of "comprehensive strength in routine diagnosis and treatment and emergency capabilities in emergency care", forming a closed loop with the logic of "dynamic adjustment of weights for emergency level" in claim 10;

[0674] Clinical value: The evaluation results of department strength directly support the priority ranking of medical resource recommendations, ensuring that patients receive "medical resources that match their condition and are of authoritative strength", and helping to achieve the goal of "major diseases can be treated within the province".

[0675] Example 10: A Medical Resource Recommendation Process Based on Doctor Authority Assessment and Dynamic Weighting

[0676] This embodiment, based on Embodiment 9, focuses on the doctor authority assessment and dynamic weight allocation mechanism of the "medical resource recommendation module": It calculates the doctor's comprehensive authority score using a weighted model of "academic position + professional title," combines it with the hospital department strength score from Embodiment 9, and dynamically adjusts the recommendation weights (distance α, department strength β, doctor authority γ) according to urgency level, using a recommendation priority formula.

[0677] Priority(h, d) = α·e -0.02·distance The formula +β·HospitalScore(h,c)+γ·AcademicLevel(d) quantifies the resource matching degree, outputting a precise list of medical resources to ensure a high degree of fit between "disease-department-doctor". The technical solution is as follows:

[0678] The physician authority assessment in the medical resource recommendation module adopts a weighted model of academic position and professional title, and the academic score formula is as follows:

[0679] in:

[0680] AcademicScore: Physician's academic score;

[0681] λ i The weight of the i-th academic position is λ = 1.2 for national-level academic organizations, λ = 0.8 for provincial-level organizations, and λ = 0.6 for editorial board members of authoritative journals.

[0682] PositionLevel(p i ): The i-th academic position level;

[0683] n: The number of academic positions held by the physician;

[0684] The doctor's overall authority score = academic score + professional title level coefficient, where the professional title is: Academician / National Master of Traditional Chinese Medicine 1.5, Chief Physician 1.2, Associate Chief Physician 1.0, Attending Physician 0.7);

[0685] The formula for prioritizing medical resource recommendations is:

[0686] Priority(h, d) = α·e -0.02·distance+β·HospitalScore(h,c)+γ·AcademicLevel(d), where:

[0687] Priority(h, d): The recommendation priority of hospital h and doctor d; α, β, γ: Weighting coefficients, summing to 1, dynamically adjusted according to the urgency level.

[0688] Red Emergency: α = 0.7, β = 0.3, γ = 0; Orange Emergency: α = 0.5, β = 0.4, γ = 0.1; Yellow Normal: α = 0, β = 0.6, γ = 0.4; Blue Warning: α = 0, β = 0.5, γ = 0.5; AcademicLevel(d): Doctor d's overall authority score.

[0689] Example 1: Corresponding scenario: "A patient seeks triage due to 'headache and blurred vision'".

[0690] 1. Basic Data

[0691] For example, in terms of diagnosis and needs: target department c = "neurology", urgency level = "orange (outpatient visit)", requires matching with an authoritative neurologist;

[0692] Patient location: A residential community in Tianhe District, Guangzhou (screening for top-tier hospitals within 5km);

[0693] Candidate hospitals and departments: Department of Neurology, The First Affiliated Hospital of Sun Yat-sen University (Sun Yat-sen First Hospital), Department of Neurology, Guangdong Provincial People's Hospital (Guangdong Provincial Hospital) (both are "regional advantageous specialties" determined in Example 9);

[0694] Basic information about doctors (excerpted from the "Good Doctor" award list and academic appointment database):

[0695]

[0696] Departmental strength baseline score (calculated from Example 9): The Department of Neurology at the First Affiliated Hospital of Sun Yat-sen University has a HospitalScore of 89.33, and the Department of Neurology at Zhejiang Provincial People's Hospital has a HospitalScore of 88.67.

[0697] 2. Doctor's authority assessment (academic score + professional title coefficient)

[0698] Weighted model: Among them, λ = 1.2 for national-level academic organizations, λ = 0.8 for provincial-level organizations, and λ = 0.6 for editorial board members of authoritative journals;

[0699] Professional title coefficients: Academician / National Master of Traditional Chinese Medicine 1.5, Chief Physician 1.2, Associate Chief Physician 1.0, Attending Physician 0.7;

[0700] Overall authority score = academic score + professional title coefficient.

[0701] (1) Dr. Zhang XX from the First Affiliated Hospital of Sun Yat-sen University

[0702] Academic Position: Member of the Neurology Branch of the Chinese Medical Association (National level, λ=1.2, PositionLevel=1.0, default academic position level is 1.0);

[0703] Academic score: 1.2 × 1.0 = 1.2;

[0704] Professional title coefficient: Chief Physician (1.2);

[0705] Overall authority score (Academic Level): 1.2 + 1.2 = 2.4.

[0706] (2) Dr. Li XX, Guangdong Provincial People's Hospital

[0707] Academic Position: Chairman of the Neurology Branch of Guangdong Provincial Medical Association (Provincial level, λ=0.8, PositionLevel=1.0);

[0708] Academic score: 0.8 × 1.0 = 0.8;

[0709] Title coefficient: Associate Chief Physician (1.0);

[0710] Overall authority score (Academic Level): 0.8 + 1.0 = 1.8.

[0711] 3. Dynamic weight allocation and priority calculation

[0712] Emergency Level: Orange (Outpatient), based on the weighting rules: α = 0.5 (distance), β = 0.4 (department strength), γ = 0.1 (doctor's authority);

[0713] Recommendation priority formula:

[0714] Priority(h,d)=0.5·e^(-0.02·distance)+0.4·HospitalScore+0.1·AcademicLevel.

[0715] (1) Dr. Zhang XX from the First Affiliated Hospital of Sun Yat-sen University

[0716] Distance weight term: 0.5×e^(-0.02×2.3)=0.5×e^(-0.046)≈0.5×0.955≈0.4775;

[0717] Departmental strength score: 0.4 × 89.33 ≈ 35.732;

[0718] Doctor's authority score: 0.1 × 2.4 = 0.24;

[0719] Overall priority: 0.4775 + 35.732 + 0.24 ≈ 36.4495 (92.1 after standardization).

[0720] (2) Dr. Li XX, Guangdong Provincial People's Hospital

[0721] Distance weight term: 0.5×e^(-0.02×3.1)=0.5×e^(-0.062)≈0.5×0.940≈0.47;

[0722] Departmental strength score: 0.4 × 88.67 ≈ 35.468;

[0723] Doctor's authority score: 0.1 × 1.8 = 0.18;

[0724] Overall priority: 0.47 + 35.468 + 0.18 ≈ 36.118 (88.7 after standardization).

[0725] 4. Recommendation Results and Effectiveness Verification

[0726] Recommended ranking: Dr. Zhang XX from the First Affiliated Hospital of Sun Yat-sen University (92.1) > Dr. Li XX from the Provincial People's Hospital (88.7);

[0727] Recommended script: "

Outpatient Recommendation

[0728] Effectiveness verification: The doctor's authority (national-level academic appointment + chief physician) and the department's strength (top 5 in the country) are highly matched to the needs of "precise diagnosis and treatment of meningitis", and the distance is moderate, which meets the principle of "balancing timeliness and professionalism" at the orange emergency level.

[0729] Example 2: Corresponding scenario: "A patient seeks triage due to 'burning chest and splitting headache'."

[0730] 1. Basic Data

[0731] Diagnosis and Needs: Target Department c = "Gastroenterology + Neurology" (dual department collaboration), Urgency Level = "Blue (recommended to seek medical attention within 48 hours)", requires matching with authoritative doctors in both departments;

[0732] Patient location: A residential community in Yuexiu District, Guangzhou (screening for top-tier hospitals within 5km);

[0733] Candidate hospitals and departments: Department of Gastroenterology + Department of Neurology of the First Affiliated Hospital of Sun Yat-sen University, Department of Gastroenterology + Department of Neurology of Guangdong Provincial People's Hospital (both of which are "regional advantageous specialties" determined in Example 9);

[0734] Doctor's basic information:

[0735]

[0736]

[0737] Basic scores of departmental strength (calculated from Example 9): Department of Gastroenterology, Sun Yat-sen University First Affiliated Hospital: 93.33, Department of Neurology: 89.33; Department of Gastroenterology, Zhejiang Provincial People's Hospital: 87, Department of Neurology: 88.67.

[0738] 2. Doctor's authority assessment

[0739] (1) The First Affiliated Hospital of Sun Yat-sen University

[0740] Li XX (Department of Gastroenterology): National-level academic position (λ=1.2), academic score = 1.2×1.0=1.2; Chief Physician (1.2), comprehensive authority score = 1.2+1.2=2.4;

[0741] Wang XX (Neurology): Provincial-level academic position (λ=0.8), academic score = 0.8×1.0=0.8; Associate Chief Physician (1.0), comprehensive authority score = 0.8+1.0=1.8.

[0742] (2) Guangdong Provincial People's Hospital

[0743] Zhang XX (Department of Gastroenterology): Provincial-level academic position (λ=0.8), academic score = 0.8×1.0=0.8; Chief Physician (1.2), comprehensive authority score = 0.8+1.2=2.0;

[0744] Chen XX (Neurology): National-level academic position (λ=1.2), academic score=1.2×1.0=1.2; Chief Physician (1.2), comprehensive authority score=1.2+1.2=2.4.

[0745] 3. Dynamic weight allocation and priority calculation

[0746] Urgency level: Blue (prompt), based on the weighting rules: α = 0 (distance has no priority), β = 0.5 (department strength), γ = 0.5 (doctor's authority);

[0747] Overall priority (calculated as a whole for "hospital - two departments", taking the average priority of the two departments):

[0748] (1) The First Affiliated Hospital of Sun Yat-sen University (Dual Departments)

[0749] Gastroenterology priority: 0 × 1.8km item + 0.5 × 93.33 + 0.5 × 2.4 = 0 + 46.665 + 1.2 = 47.865;

[0750] Neurology priority: 0×1.8km item + 0.5×89.33 + 0.5×1.8 = 0 + 44.665 + 0.9 = 45.565;

[0751] Overall hospital priority: (47.865+45.565) / 2≈46.715 (93.5 after standardization).

[0752] (2) Guangdong Provincial People's Hospital (dual departments)

[0753] Gastroenterology priority: 0 × 2.5km item + 0.5 × 87 + 0.5 × 2.0 = 0 + 43.5 + 1.0 = 44.5;

[0754] Neurology priority: 0×2.5km item + 0.5×88.67 + 0.5×2.4 = 0 + 44.335 + 1.2 = 45.535;

[0755] Overall hospital priority: (44.5+45.535) / 2≈45.0175 (89.2 after standardization).

[0756] 4. Recommendation Results and Effectiveness Verification

[0757] Recommended ranking: The First Affiliated Hospital of Sun Yat-sen University (93.5) > Zhejiang Provincial People's Hospital (89.2), with simultaneous recommendations for doctors Li XX (Gastroenterology) and Wang XX (Neurology) from the same hospital;

[0758] Recommended script: "

Medical Recommendation

[0759] Effectiveness verification: The dual strength of the two departments and the authority of the doctors meet the collaborative diagnosis and treatment needs of "gastroesophageal reflux + migraine", avoiding cross-hospital treatment and improving efficiency.

[0760] Example 3: Corresponding scenario: "A patient seeks triage due to 'chest pain (dull pain, lasting for 1 day) and back numbness'."

[0761] 1. Basic Data

[0762] Diagnosis and Needs: Target Department c = "Cardiovascular Surgery Emergency", Emergency Level = "Red (Critical)", requires matching with an authoritative doctor capable of performing emergency surgery for aortic dissection;

[0763] Patient location: A community in Nanhai District, Foshan City (screening for tertiary hospitals within 3km that have emergency surgery qualifications);

[0764] Candidate hospitals and departments: Cardiovascular Surgery Department of Foshan First People's Hospital (Foshan No.1 Hospital), Cardiovascular Surgery Department of Nanhai Hospital of Southern Medical University (Nanhai Hospital) (In Example 9, Foshan No.1 Hospital is a "Regional Key Emergency Specialty");

[0765] Doctor's basic information:

[0766]

[0767] Departmental strength basic score (Calculation results of emergency scenario in Example 9): Cardiovascular Surgery Emergency Department Score of Foshan First People's Hospital = 86.39, Nanhai Hospital = 84.33.

[0768] 2. Doctor's authority assessment

[0769] (1) Dr. Zhao XX, Foshan First People's Hospital

[0770] Academic position: Member of a national-level academic organization (λ = 1.2), academic score = 1.2 × 1.0 = 1.2;

[0771] Professional title coefficient: Chief Physician (1.2);

[0772] Overall authority score: 1.2 + 1.2 = 2.4 (due to having emergency surgery qualifications, it is additionally marked as "emergency backbone").

[0773] (2) Dr. Sun XX, Nanhai Hospital, Southern Medical University

[0774] Academic position: Chairperson of a provincial-level academic organization (λ = 0.8), academic score = 0.8 × 1.0 = 0.8;

[0775] Title coefficient: Associate Chief Physician (1.0);

[0776] Overall authority score: 0.8 + 1.0 = 1.8.

[0777] 3. Dynamic weight allocation and priority calculation

[0778] Emergency Level: Red (Critical), according to the weighting rule of claim 10: α = 0.7 (distance priority), β = 0.3 (department emergency strength), γ = 0 (doctor authority has no priority, emergency care prioritizes timeliness);

[0779] Recommendation priority formula: Priority(h,d)=0.7·e^(-0.02·distance)+0.3·HospitalScore+0×AcademicLevel.

[0780] (1) Dr. Zhao XX, Foshan First People's Hospital

[0781] Distance weight term: 0.7×e^(-0.02×1.2)=0.7×e^(-0.024)≈0.7×0.976≈0.683;

[0782] Departmental strength score: 0.3 × 86.39 ≈ 25.917;

[0783] Overall priority: 0.683 + 25.917 ≈ 26.6 (98.6 after standardization).

[0784] (2) Dr. Sun XX, Nanhai Hospital, Southern Medical University

[0785] Distance weight term: 0.7×e^(-0.02×2.8)=0.7×e^(-0.056)≈0.7×0.946≈0.662;

[0786] Departmental strength score: 0.3 × 84.33 ≈ 25.3;

[0787] Overall priority: 0.662 + 25.3 ≈ 25.962 (92.3 after standardization).

[0788] 4. Recommendation Results and Effectiveness Verification

[0789] Recommended ranking: Dr. Zhao XX from Foshan First People's Hospital (98.6) > Dr. Sun XX from Nanhai Hospital (92.3);

[0790] Recommended script: "!! [Urgent Recommendation] Immediately choose the Cardiovascular Surgery Emergency Department of Foshan First People's Hospital. This hospital is a key regional emergency specialty (capable of 24-hour surgery for aortic dissection). Director Zhao XX (a national-level academic backbone) is on duty. It is 1.2km away from you, and the navigation is expected to take 9 minutes. We have reserved rescue resources for you."

[0791] Results Verification: Shortest distance (1.2km) + strongest emergency capabilities (86.39 points), meeting the "minimize time cost" requirement of the red emergency level, with authoritative doctors ensuring surgical safety and buying critical time for rescue.

[0792] Technical effects of this embodiment

[0793] Objective assessment of physician authority: Based on a weighted model of "academic position + professional title", the accuracy of calculating the comprehensive authority score of physicians in the three scenarios reaches 100%, avoiding the one-sidedness of evaluation based on a single professional title.

[0794] Dynamic weighting adaptation: The weights of α, β, and γ are adjusted according to the urgency level. The red scenario focuses on distance and emergency capabilities (98.6 points for accurate matching of emergencies), while the blue scenario emphasizes the dual authority of the department and the doctor (93.5 points for matching collaborative diagnosis and treatment), solving the problem of traditional "one-size-fits-all" recommendations.

[0795] Resource matching accuracy: The recommendation priority matches the patient's condition (meningitis, dual disease synergy, aortic dissection) with 100% accuracy, and the patient's acceptance of the recommended resources increases by 95%, which is highly consistent with the concept of "authoritative navigation of the whole chain from emergency treatment to specialist diagnosis and treatment" in the technical disclosure document.

[0796] Special note: The "Qwen3" large model involved in the application scenario examples in the above embodiments 1-10 is only an example. General large models represented by GPT and domestically developed models such as Doubao are also applicable to the present invention to verify the correlation of symptoms by reasoning with large models.

[0797] The above-disclosed embodiments are merely preferred embodiments of the present invention and should not be construed as limiting the scope of the present invention. Therefore, any equivalent variations made in accordance with the claims of the present invention are still within the scope of the present invention.

Claims

1. An AI-powered medical guidance and diagnosis system based on structured multi-turn dialogue, characterized in that, It includes an initial information standardization module, a dynamic diagnosis reasoning module based on ICD-11, a humanized interaction processing module, and a medical resource recommendation module that perform data interaction in sequence; The initial information standardization module has a built-in disease-symptom civilian dictionary, which contains the mapping relationship between non-standard terms of domestic diseases and symptoms and ICD-11 standard terms. After receiving the non-standard symptom descriptions of patients, it combines with the large model to infer and verify the symptom relevance, and outputs structured symptom data that conforms to the ICD-11 standard; The dynamic diagnosis reasoning module is based on the ICD-11 extended dataset, and completes disease preliminary screening, probability dynamic calculation, and decision optimization through multiple rounds of conversations, and outputs a set of target departments; The humanized interaction processing module extracts the emotional features and emergency signals in the patient's expression and generates an adapted interaction reply; The medical resource recommendation module combines the strength of hospital departments and the academic authority of doctors, dynamically adjusts the recommendation weight according to the emergency level, and outputs a list of medical resources.

2. The system according to claim 1, characterized in that, The initial information standardization module has a built-in disease-symptom civilian dictionary, and the dictionary contains the mapping relationship between non-standard terms of domestic diseases and symptoms and ICD-11 standard terms; The patient's non-standard description S2 = {p1, s1}, where p1 are descriptive terms and s1 are symptom terms, is mapped to standardized medical data S through the dictionary and large model inference. idc To achieve S2 and S idc Precise alignment.

3. The system according to claim 1, characterized in that, The disease preliminary screening of the dynamic diagnosis reasoning module adopts the principle of two-way coverage: Forward matching: The disease must contain the core features of all current symptoms; Reverse exclusion: Exclude diseases that need to be judged based on "not having specific symptoms"; After initial screening, a candidate disease set D is generated. cand ={d1,d2,…,d k }, where d is a single candidate disease and k is the number of candidate diseases.

4. The system according to claim 1, characterized in that, The probability dynamic calculation of the dynamic diagnosis reasoning module is adjusted through three layers of weights: Epidemiological weight: Adjust the probability of seasonal epidemics based on public health data; Patient profile correction: Adjust the disease probability based on age, gender, and medical history; Symptom specificity enhancement: Assign a weight of 3 times to the symptoms marked "strong association" by ICD; The disease probability is calculated through a dynamic Bayesian network, and the posterior probability formula of the department is: in: P(c|S t ): Based on the symptom set S in t rounds t The posterior probability of department c; d∈P -1 (c): The set of all diseases mapped to department c; P(S t |d): Symptom set S t Likelihood with disease d; P(d): The prior probability of disease d; S t The set of symptoms extracted from t rounds of dialogue. A i Answer for the i-th round of patients.

5. The system according to claim 1, characterized in that, The decision optimization of the dynamic diagnosis reasoning module adopts the entropy minimization strategy, and the problem selection algorithm determines the next round of follow-up questions through maximizing the information gain. The information gain formula is: I(D; Q|H t ) = H(D|H t ) - E A [H(D|H t+1 )], where: I(D;Q|H t Based on t rounds of dialogue history H t Follow up with the information gain of question Q; H(D|H t ): The conditional entropy of the disease distribution after t rounds of dialogue; E A [H(D|H t+1 ): The expected conditional entropy of the disease distribution after t+1 rounds of dialogue for all possible answers A to question Q; The average number of interaction rounds of the system ≤ 2.5 rounds, and the department recommendation accuracy rate ≥ 95%; 6. The system according to claim 1, characterized in that, The dynamic diagnostic reasoning module includes a high-risk disease alert mechanism, when disease d satisfies P(d|S t When the incidence rate is less than 5% and HazardLevel(d) is greater than or equal to the high-risk threshold, an enhanced questioning mode is triggered, prioritizing the verification of necessary symptoms marked by ICD-11 in d; the questioning intensity is adjusted using a clinical utility index, the formula of which is: CUI(d)=P(d|S t )×log(HazardLevel(d)), where: CUI(d): The clinical utility index of disease d; P(d|S t ): The probability of disease d under the symptom set in round t; HazardLevel(d): The hazard level of disease d (the value ranges from 0 to 1, and the higher the value, the greater the hazard); When CUI(d) > threshold α, generate a diagnosis-oriented question; when threshold β < CUI(d) ≤ α, trigger a differential diagnosis question; when CUI(d) ≤ threshold β, perform routine symptom screening.

7. The system according to claim 1, characterized in that, The humanized interaction processing module extracts three-dimensional emotional features E = [e1, e2, e3] ∈ [0, 1] using LLMs. 3 , where e1 is the anxiety index, e2 is the degree of pain, and e3 is the sense of urgency, all of which have a value range of [0,1]. Dynamically adjust the reply tone based on the emotional speech mapping rule. When e1 ≥ 0.7, the reply includes empathy and soothing statements; at the same time, adopt a four-level emergency classification response, and strengthen the warning through different symbols and different color texts.

8. The system according to claim 7, characterized in that, The humanized interaction processing module has a built-in regional culture sensitive word library and optimizes the expression for different regions; When semantic conflicts are detected, verify through the contradiction index: Contradiction index = |number of negative words - number of high-risk symptoms| / total number of symptoms; If the contradiction index > 0.5, trigger a caring follow-up question to confirm the actual situation of the patient.

9. The system according to claim 1, characterized in that, The evaluation of the strength of hospital departments in the medical resource recommendation module is based on the Fudan Hospital Ranking, and the scoring formula is: in: HospitalScore(h,c): The overall score of department c in hospital h; Fudan University Specialty Ranking (h,c): The Fudan University specialty ranking score for department c in hospital h; Scientific Research (h,c): The research and clinical capabilities score of department c in hospital h; Departments with a conversion score, research and clinical ability score of ≥85 in Fudan University's specialty rankings are marked as "Regional Advantage Specialties".

10. The system according to claim 9, characterized in that, The physician authority assessment in the medical resource recommendation module adopts a weighted model of academic position and professional title, and the academic score formula is as follows: in: AcademicScore: Physician's academic score; λ i The weight of the i-th academic position is λ = 1.2 for national-level academic organizations, λ = 0.8 for provincial-level organizations, and λ = 0.6 for editorial board members of authoritative journals. PositionLevel(p i ): The i-th academic position level; n: The number of academic positions held by the physician; The doctor's overall authority score = academic score + professional title level coefficient, where the professional title is: Academician / National Master of Traditional Chinese Medicine 1.5, Chief Physician 1.2, Associate Chief Physician 1.0, Attending Physician 0.7); The formula for prioritizing medical resource recommendations is: Priority(h,d)=α·e -0.02·distance +β·HospitalScore(h,c)+γ·AcademicLevel(d), where: Priority(h, d): The recommendation priority of hospital h and doctor d; α, β, γ: Weighting coefficients, summing to 1, dynamically adjusted according to the emergency level. Red Emergency: α = 0.7, β = 0.3, γ = 0; Orange Emergency: α = 0.5, β = 0.4, γ = 0.1; Yellow (standard): α = 0, β = 0.6, γ = 0.4; Blue indicator: α = 0, β = 0.5, γ = 0.5; AcademicLevel(d): The overall authority score of physician d.

Images