An ai-assisted diagnosis decision system for a rare bladder disease

Abstract

This invention relates to the field of medical artificial intelligence technology, and in particular to an AI-assisted diagnostic decision-making system for rare bladder diseases. The system includes modules for data acquisition, feature extraction and quantification, an AI rule-based reasoning engine, error warning, and treatment decision output. Its core lies in using natural language processing technology to transform unstructured clinical text into feature tags. The AI ​​engine then executes a multi-level identification process of "excluding infection – rapid screening of medical history – precise matching," outputting a list of candidate diseases sorted by probability. Simultaneously, the system can automatically warn of clinical errors such as antibiotic abuse and overtreatment, and generate a comprehensive report including primary care treatment plans and referral criteria. This invention combines a rule engine with intelligent reasoning, providing primary care physicians with a standardized and easy-to-use auxiliary diagnostic tool, which helps improve the efficiency and standardization of rare disease diagnosis and has high clinical translational value.

Description

Technical Field

[0001] This invention relates to the field of medical artificial intelligence technology, specifically to an AI-assisted diagnostic decision-making system for a rare bladder disease. Background Technology

[0002] Bladder diseases are common in the urinary system. Besides common bacterial cystitis, there are a series of relatively rare but complex non-infectious cystitis and other bladder diseases with varied presentations, such as interstitial cystitis, radiation cystitis, eosinophilic cystitis, glandular cystitis, vesicular endometriosis, and pelvic lipomatosis, among more than ten others. Although the overall incidence of these diseases is not high, their symptoms (such as urinary frequency, urgency, hematuria, and pelvic pain) often overlap with common urinary tract infections, making them prone to misdiagnosis, missed diagnosis, and inadequate treatment in primary healthcare practice. Traditional diagnosis heavily relies on the physician's personal experience, detailed medical history (such as history of radiotherapy, chemotherapy, and menstrual cycle correlation), and the interpretation of specific test results (such as characteristic findings under cystoscopy and blood / urine eosinophil counts), which poses a significant challenge to primary care physicians.

[0003] In recent years, the application of artificial intelligence (AI) technology in the medical field has become increasingly widespread, providing new tools for assisted diagnosis. In urology, AI has been explored for image recognition, pathological analysis, and prognostic prediction in prostate and bladder cancer. For example, research has developed an AI-based initial screening system for interstitial cystitis, which assists in diagnosis by integrating symptom text analysis and cystoscopic image recognition. Simultaneously, there are also AI-assisted pathology systems focusing on the identification of non-urothelial carcinoma subtypes of the bladder. These practices demonstrate the potential of AI in improving diagnostic efficiency and objectivity.

[0004] However, existing AI-assisted diagnostic systems and methods still have significant limitations and gaps when applied to the identification of rare bladder diseases in primary healthcare institutions: First, narrow disease coverage: existing systems mostly target single diseases (such as interstitial cystitis) or single categories of diseases (such as subtypes of bladder cancer), lacking an integrated diagnostic tool that can comprehensively cover more than ten rare bladder diseases with different etiologies and pathological mechanisms. Second, disconnect from primary healthcare scenarios: many AI systems rely on high-end examination equipment (such as high-resolution pathological slide scanning and multimodal imaging), while primary healthcare institutions typically only have the conditions for examinations such as urinalysis, ultrasound, and basic cystoscopy, making it impossible for existing systems to operate effectively based on this limited information. Third, lack of a closed-loop decision-making process: most systems only provide diagnosis or risk prediction, failing to deeply integrate with the complete process of primary healthcare, lacking real-time warning capabilities for common misconceptions such as "antibiotic abuse" and "overtreatment," and failing to output decision support reports that can directly guide the actions of primary care physicians, including treatment plans and clear referral indications. Fourth, insufficient interpretability and integration: The system relies on a single technical path (such as a pure rule engine or a pure deep learning model). When there are conflicting output results or when facing atypical cases, there is a lack of a collaborative adjudication mechanism that integrates rule certainty, large model semantic understanding and historical case empirical verification, which leads to a reduction in the credibility and clinical acceptance of the results.

[0005] Therefore, there is an urgent need for an AI-assisted system designed specifically for primary healthcare scenarios, capable of covering a variety of rare bladder diseases, integrating multi-source information and reasoning logic, and providing full-process decision support from differential diagnosis to misdiagnosis warning to treatment and referral, in order to make up for the shortcomings of existing technologies and effectively improve the quality of primary healthcare. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide an AI-assisted diagnostic decision-making system for rare bladder diseases.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an AI-assisted diagnostic decision-making system for rare bladder diseases, comprising: The data acquisition unit is configured to acquire the patient's basic information, core symptoms, auxiliary examination results and medical history text data, wherein the auxiliary examination results include at least urinalysis, urine culture, blood routine, renal function and cystoscopy description; The feature quantization unit, connected to the data acquisition unit, is configured to convert the medical history text data and unstructured cystoscopy description into AI-recognizable structured feature vectors through natural language processing and medical named entity recognition technology. The structured feature vectors include medical history labels, laboratory indicator Boolean values, and standardized labels for endoscopic features. The AI ​​inference unit, connected to the feature quantization unit, includes a rule engine, a large language model inference module, a typical case database, and a three-component collaborative adjudication module. The rule engine is configured to perform hard clinical logic triage based on a preset rule base for the identification of rare bladder diseases, the rule base containing a hierarchical decision tree for rare bladder diseases; The large language model reasoning module is configured to perform semantic understanding, probabilistic reasoning, and differential diagnosis ranking on unstructured medical record texts. The typical case database is configured to store structured feature vectors of confirmed cases and provide empirical comparisons based on similarity calculations; The three-component collaborative adjudication module is configured to collaboratively adjudicate and resolve conflicts in the outputs of the rule engine, the large language model inference module, and the typical case database according to a preset priority principle, and output candidate disease probability ranking, supporting evidence, excluded evidence, and high-priority warnings. The error warning unit, connected to the AI ​​inference unit, is configured to identify common errors in clinical decision-making and generate non-blocking warning prompts in real time when doctors save diagnostic conclusions, prescribe prescriptions, or formulate treatment plans, based on a preset error rule base. The diagnosis and treatment decision unit, connected to the AI ​​reasoning unit and the error warning unit, is configured to output executable treatment plans and mandatory referral indications based on the diagnosis results and preset primary care treatment rule templates and referral high-risk feature database. The output unit is configured to integrate the outputs of the AI ​​inference unit, the error warning unit, and the diagnosis and treatment decision unit to generate and display a decision report that includes candidate disease ranking, diagnostic basis, risk warning, primary care treatment plan, and referral suggestions.

[0008] In some embodiments, the rare bladder diseases include at least one of the following: radiation cystitis, chemical cystitis, hemorrhagic cystitis, toxoplasmosis cystitis, schistosomiasis cystitis, leathery cystitis, eosinophilic cystitis, glandular cystitis, interstitial cystitis, bladder leukoplakia, bladder endometriosis, bladder amyloidosis, and pelvic lipomatosis.

[0009] In some embodiments, the calling relationship and workflow of the three-component collaborative adjudication module follow a sequential process of "rule priority, intelligent supplementation, and empirical verification": In the first round, the rule engine performs hard rule matching and triage based on structured data, and outputs preliminary diagnostic clues or instructions to be identified; In the next round, the large language model reasoning module receives all patient information and clues output by the rule engine, performs deep identification reasoning, and outputs a list of diseases sorted by probability. Finally, the typical case database is used to search for similar cases based on the main candidate diagnoses and their key features output by the large language model reasoning module, in order to verify the rationality of the reasoning or indicate the possibility of rarity.

[0010] In some embodiments, the conflict resolution mechanism of the three-component collaborative adjudication module includes: When the rule engine and the large language model inference module output conflict, the rule engine result is written as the primary diagnosis in the main diagnosis field, and the high-risk objection raised by the large language model is encapsulated as a high-priority warning field. When the large language model inference module conflicts with the output of the typical case database, the disease ranking of the large language model is maintained, and the differences found in the case database are used as key identification prompts. When the outputs of the three components are inconsistent, the internal confidence of each component is calculated and weighted and fused. At the same time, following the "safety first" circuit breaker mechanism, the suspicion pointing to a disease with serious consequences or characteristics of emergency intervention is used as the highest level warning item.

[0011] In some embodiments, the weighted fusion employs a confidence-weighted algorithm, with the rule engine having a weight of 0.6, the large language model inference module having a weight of 0.3, and the typical case database having a weight of 0.1; the total confidence score is calculated using the following formula: Total confidence level = (0.6 × rule score) + (0.3 × LLM score) + (0.1 × case matching degree); When all three are inconsistent and the large language model inference module or typical case database points to a disease with serious consequences, the system ignores the regular weights and directly places the suspected disease at the top of the report as the highest level warning item.

[0012] In some embodiments, the hierarchical decision tree of the rule engine adopts a multi-level decision structure: The first layer is the infection triage layer, which prioritizes excluding acute bacterial cystitis based on urine culture results, urine leukocyte / nitrite index, and antibiotic efficacy data. The second layer is a rapid screening layer for medical history and triggers. Under the premise of non-infection, it scans for specific high-weight medical history and triggers, including history of pelvic radiotherapy, history of chemotherapy drug use, history of contact with epidemic areas, history of allergies, and menstrual-related history, in order to quickly anchor specific rare disease categories. The third layer is the feature and endoscopy precision matching layer, which is subdivided and verified based on specific laboratory indicators and cystoscopic morphological features. The cystoscopic morphological features include Huntner's ulcer, glomerulation, sand-like plaques, white plaques, follicular hyperplasia, villous hyperplasia, bluish-purple nodules, and external pressure changes.

[0013] In some embodiments, the probability ranking of the AI ​​inference unit adopts a mechanism combining "dynamic weighted scoring" and "gold standard one-vote triggering": Each candidate disease starts with a score of 0. If a supporting feature is matched, the score is added according to a preset weight. If an exclusion feature is matched, the score is reset to zero or significantly deducted. When pathological or endoscopic gold standard features are captured, including positive Congo red staining, discovery of parasite eggs, or histological evidence for diagnosis, they are given the highest priority and directly become the first candidate diagnosis. The total score for each disease is converted into a probability percentage after Softmax normalization, and the final candidate diseases are generated by ranking them according to their probability.

[0014] In some embodiments, the natural language processing in the feature quantization unit includes: By using a specialized fine-tuned large language model combined with medical named entity recognition technology, word segmentation and entity locking were performed on the cystoscopy description text, and site features and pathological modifiers were extracted. The extracted unstructured phrases are matched with a pre-defined standard dictionary of rare cystoscopy features using vector similarity. After identifying specific words, they are automatically converted into discrete, standardized Boolean labels, which are then used as calculation factors to input into the multi-level decision tree of the AI ​​inference unit.

[0015] In some embodiments, the system further includes a one-click "infection exclusion" judgment module, which uses a dynamic scoring system rather than a fixed set of multiple rules to satisfy: A negative urine culture is worth +3 points; ineffective standard course of antibiotic treatment is worth +2 points; normal white blood cell count in routine urine tests is worth +1 point; presence of non-infectious precipitating factors is worth +1 point; presence of microscopic hematuria but no pyuria is worth +0.5 points. A high-risk warning is triggered when the total score is ≥5; a medium-risk warning is triggered when the total score is ≤3 and <5; and the standard treatment procedure is maintained when the total score is <3. When some conditions are missing but strong evidence exists, the prompt content is dynamically adjusted and a screening path for non-infectious cystitis is triggered.

[0016] In some embodiments, the error warning unit focuses on identifying the error types and triggering conditions, including: Antibiotic abuse: Triggered when a urine culture is negative, there are no signs of systemic infection, and there is a prescription for antibiotics; Overtreatment: Triggered when the AI ​​diagnosis ranking shows a high probability of benign or precancerous lesions and the treatment recommendation is major surgery, while most similar cases are treated conservatively; Risk of missed diagnosis: This is triggered when the system identifies specific keywords in the cystoscopy description but the doctor's initial diagnosis does not include the corresponding disease; Delayed referral: Triggered when the AI ​​candidate diagnosis includes a disease that requires in-depth specialist treatment or multidisciplinary collaboration, and there is no recent referral or consultation plan in the medical record; Insufficient examination: Triggered when the AI ​​has a high probability of diagnosing a disease that requires a biopsy for confirmation, and there is no corresponding biopsy scheduled in the examination plan.

[0017] In some embodiments, the diagnostic decision unit includes: The primary care intervention prompt generation module is configured to call a fixed rule template to output standardized lifestyle guidance, symptomatic relief plan and follow-up visit nodes when the primary disease falls within the scope of initial intervention at the primary level. The mandatory referral red line module is configured to immediately trigger the referral template and highlight the referral suggestion when the following conditions are identified: persistent gross hematuria, hydronephrosis of the upper urinary tract, suspected malignancy on ultrasound, significantly reduced bladder capacity, need for biopsy for diagnosis but local hospitals cannot perform the corresponding examination, or unclear diagnosis but symptoms continue to exceed the preset threshold.

[0018] In some embodiments, the case data in the typical case database includes structured fields: basic information field, clinical characteristics field, laboratory test field, imaging and endoscopic characteristics field, cystoscopy description field, and final diagnosis and identification field; The typical case database is configured to perform cosine similarity calculation between the structured feature vector of the current patient and the cases in the database, and return the Top-K most similar past cases and their diagnoses, providing empirical probabilistic support for AI reasoning.

[0019] Compared with the prior art, the beneficial effects of the present invention are: 1. Comprehensive diagnostic coverage, specifically addressing pain points at the grassroots level: This invention is the first to systematically integrate thirteen clinically rare bladder disease differentiation rules, effectively filling the gaps in existing AI tools in this field. By constructing a dedicated differentiation rule engine, it can accurately distinguish between diseases that are most easily confused or missed by grassroots doctors (such as interstitial cystitis and infection, glandular cystitis and tumors), directly applying the effectiveness of AI to the most pressing clinical needs at the grassroots level.

[0020] 2. Highly adaptable to primary care settings, enabling lightweight and efficient decision-making: The system design uses readily available examination information (such as symptoms, urinalysis, and basic cystoscopy descriptions) as the minimum input requirement, providing services through lightweight front-end applications (such as mini-programs), greatly reducing the barrier to entry. Its built-in "one-click infection exclusion" and "automatic warning of common misconceptions" functions can immediately correct habitual errors such as antibiotic abuse and overtreatment, significantly improving the standardization and safety of primary care diagnosis and treatment.

[0021] 3. Forming a closed-loop decision-making system and providing full-process action guidance: This invention goes beyond simple diagnostic prompts, constructing a complete decision support chain of "differential diagnosis - misdiagnosis warning - treatment plan - referral criteria". The system not only outputs disease probability rankings, but also generates standardized treatment plans that can be implemented at the primary care level, and triggers mandatory referral reminders based on clear "red line" rules (such as refractory hematuria, upper urinary tract hydronephrosis), realizing full-process empowerment from assisting thinking to assisting action.

[0022] 4. Reliable Technological Integration Enhances Results Credibility and Acceptability: The system innovatively employs a collaborative working mechanism that prioritizes a rule engine, utilizes large-scale model reasoning as its core, and employs case database verification as supporting evidence. This integrated architecture ensures the certainty and interpretability of results based on hard clinical rules (such as gold standard test results), while leveraging the semantic understanding capabilities of the large-scale model to handle complex and atypical case descriptions, and then performs empirical calibration through similar historical cases. In the event of output conflicts, this mechanism intelligently adjudicates based on the "safety first" principle, thereby outputting a comprehensive report that combines high accuracy, strong interpretability, and clinical applicability, making it easily understood and adopted by primary care physicians.

[0023] Details of one or more embodiments of this application are set forth in the following drawings and description to make other features, objects and advantages of this application more readily apparent. The embodiments of this application will provide a detailed description and understanding of the application. Attached Figure Description

[0024] Figure 1 This is a structural block diagram of the AI-assisted diagnostic decision-making system for rare bladder diseases of the present invention; Figure 2 This is a flowchart illustrating the system call sequence and workflow of the present invention. Figure 3 This is a flowchart illustrating the multi-level identification and diagnosis workflow executed by the AI ​​rule reasoning engine in this invention. Detailed Implementation

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

[0026] like Figures 1 to 3As shown in the figure, this embodiment of the invention provides an AI-assisted diagnostic decision-making system for rare bladder diseases. It adopts a lightweight architecture design for primary healthcare institutions, with the front end deployed as a WeChat mini-program or web page, and the back end supported by a cloud server cluster for core computing. The system as a whole includes: a data acquisition unit, a feature quantification unit, an AI inference unit, a misdiagnosis warning unit, a diagnosis and treatment decision-making unit, and a result output unit.

[0027] The data acquisition unit is configured to acquire the patient's basic information, core symptoms, auxiliary examination results, and medical history text data. Basic information includes gender, age, and disease duration; core symptoms are collected through a multi-selection method and cover urinary frequency, urgency, dysuria, gross hematuria, microscopic hematuria, bladder area pain, lower abdominal distension, fever, and allergic reactions; auxiliary examination results include at least urinalysis, urine culture, complete blood count, renal function tests, and cystoscopy descriptions; medical history text data includes past medical history such as radiotherapy, medication history (especially the use of chemotherapy drugs such as cyclophosphamide), surgical history, allergy history, gynecological / pelvic surgery history, and history of residence or contact with epidemic areas.

[0028] The feature quantization unit is connected to the data acquisition unit and is configured to convert the medical history text data and unstructured cystoscopy descriptions into AI-recognizable structured feature vectors using Natural Language Processing (NLP) and Medical Named Entity Recognition (NER) technologies. The structured feature vectors include medical history tags (e.g., "Pelvic radiotherapy history_true", "Cyclophosphamide use history_true", "Contact history in epidemic areas_true", "Menstrual-related_true"), Boolean values ​​of laboratory indicators (e.g., "Negative urine culture_true", "Elevated blood eosinophil count_true", "Positive urine parasite eggs_true"), and standardized endoscopic feature tags (e.g., "Hunner ulcer_true", "Psammomatous plaque_true", "White plaque_true", "Follicular hyperplasia_true", "Blue-purple nodule_true", "External pressure change_true").

[0029] The AI ​​inference unit, connected to the feature quantization unit, includes a rule engine, a large language model inference module, a typical case database, and a three-component collaborative adjudication module. The rule engine is configured to perform hard clinical logic triage based on a preset rule base for identifying rare bladder diseases. This rule base contains hierarchical decision trees for 13 rare bladder diseases. The large language model inference module is configured to perform semantic understanding, probabilistic inference, and differential diagnosis ranking on unstructured medical record text. The typical case database is configured to store structured feature vectors of confirmed cases and provide empirical comparisons based on similarity calculations. The three-component collaborative adjudication module is configured to collaboratively adjudicate and resolve conflicts according to a preset priority principle, outputting a probability ranking of candidate diseases, supporting evidence, excluded evidence, and high-priority warnings.

[0030] The error warning unit is connected to the AI ​​inference unit and is configured to identify common errors in clinical decision-making and generate non-blocking warning prompts in real time when doctors save diagnostic conclusions, prescribe prescriptions or formulate treatment plans, based on a preset error rule base.

[0031] The diagnosis and treatment decision unit is connected to the AI ​​reasoning unit and the error warning unit, and is configured to output executable treatment plans and mandatory referral indications based on the diagnosis results and preset primary care treatment rule templates and referral high-risk feature database.

[0032] The result output unit is configured to integrate the outputs of the AI ​​inference unit, the misconception warning unit, and the diagnosis and treatment decision unit to generate and display a decision report that includes candidate disease ranking, diagnostic basis, risk warning, primary care treatment plan, and referral suggestions.

[0033] The error warning unit is connected to the AI ​​inference unit and is configured to identify common errors in clinical decision-making and generate non-blocking warning prompts in real time when doctors save diagnostic conclusions, prescribe prescriptions or formulate treatment plans, based on a preset error rule base.

[0034] The diagnosis and treatment decision unit is connected to the AI ​​inference unit and the error warning unit, and is configured to output executable treatment plans and mandatory referral indications based on the diagnosis results and preset primary care treatment rule templates and referral high-risk feature database.

[0035] The result output unit is configured to integrate the outputs of the AI ​​inference unit, the error warning unit, and the diagnosis and treatment decision unit to generate and display a decision report that includes candidate disease ranking, diagnostic basis, risk warning, primary care treatment plan, and referral suggestions.

[0036] The 13 rare bladder diseases include: radiation cystitis, chemical cystitis, hemorrhagic cystitis, toxoplasmosis cystitis, schistosomiasis cystitis, leathery cystitis, eosinophilic cystitis, glandular cystitis, interstitial cystitis, bladder leukoplakia, bladder endometriosis, bladder amyloidosis, and pelvic lipomatosis.

[0037] Specific implementation of feature quantization unit The core task of the feature quantization unit is to transform unstructured clinical input text into machine-computable structured feature vectors. This is achieved by combining a specialized, finely tuned general-purpose language model with medical named entity recognition technology.

[0038] Specifically, when the system receives a natural language text describing a cystoscopy, such as "endoscopically, the triangular area showed mucosal congestion with several scattered, sand-like yellow patches, with inflammatory exudate adhering to the surface," the natural language processing module first uses an attention mechanism to segment and lock the text, extracting location features ("triangular area") and pathological modifiers ("mucosal congestion," "sand-like," "yellow patches," "inflammatory exudate"). Subsequently, the system performs vector similarity matching between the extracted unstructured phrases and a pre-defined "Standard Dictionary of Rare Cystoscopic Features." This dictionary pre-includes 13 specific endoscopic descriptive standard entries for rare bladder diseases and their vectorized representations. When the vector similarity between "sand-like patches" and the dictionary entry "schistosomiasis cystitis - sand-like patches" exceeds a pre-defined threshold (e.g., cosine similarity ≥ 0.85), the system automatically converts this feature into a discrete, standardized Boolean label `cystoscopy_sandy_patches:True`, which is then used as a calculation factor input into the subsequent multi-level decision tree.

[0039] Similarly, for the quantification of medical history text, the system scans the present illness and chief complaint of the medical record through natural language processing. When urinary symptom keywords (such as "hematuria" and "frequent urination") and time / gynecological related terms (such as "worsening during menstruation", "onset with menstrual cycle", and "dysmenorrhea") are extracted from the text, the system determines the "menstrual correlation" Boolean value to True and maps it as a high-weighted calculated feature supporting "bladder endometriosis". For the mapping of the "antibiotics ineffective" feature, the system reads previous prescriptions or medical history records. If it identifies that the patient has recently used cephalosporin or quinolone antibiotics and has completed the standard course of treatment (such as 7-14 days), and the current follow-up chief complaint includes "symptoms not relieved" or the urinalysis shows no significant decrease in white blood cells, then the antibiotics_ineffective:True tag is activated, automatically adding points to the "exclude infection" condition and triggering the screening path for non-infectious cystitis.

[0040] Specific implementation of AI inference unit I. Hierarchical Decision Tree of the Rule Engine The hierarchical decision tree of the rule engine adopts a multi-level judgment structure, and its design follows the conventional thinking path of clinical differential diagnosis, gradually narrowing the scope of differentiation from macro to micro.

[0041] The first layer is the infection triage layer. This layer prioritizes excluding acute bacterial cystitis based on urine culture results, urinalysis leukocyte / nitrite levels, and antibiotic efficacy data. The system's strict rules include: if urine culture detects common pathogens (such as Escherichia coli, Proteus, etc.) with a colony count ≥10^5 CFU / mL, or if urinalysis shows significant leukocyturia (≥5 / HP) combined with positive nitrite, then "acute bacterial cystitis" is directly output as the primary diagnosis, and subsequent rare diseases are ranked lower. Unless treatment is ineffective or there are strong contradictions in the medical history or examinations, the case will not proceed to the in-depth differential diagnosis process.

[0042] The second layer is a rapid screening layer for medical history and predisposing factors. Under non-infectious conditions, the system scans for specific high-weight medical history and predisposing factors to quickly identify specific rare disease categories. The rapid screening rules set in this layer include: matching "history of pelvic radiotherapy" → prioritizing radiation cystitis; matching "history of use of chemotherapy drugs such as cyclophosphamide" → prioritizing chemical / hemorrhagic cystitis; matching "history of residence in an epidemic area" or "history of contact with schistosomiasis" → prioritizing the schistosomiasis cystitis screening process; matching "history of allergies" or "history of asthma" → prioritizing eosinophilic cystitis; matching "dysmenorrhea" or "menstrual hematuria" → prioritizing bladder endometriosis.

[0043] The third layer is the feature and endoscopy precise matching layer. This layer is based on specific laboratory indicators and cystoscopic morphological features for detailed verification. The cystoscopic morphological features include Huntner's ulcers, glomerulation (petechial hemorrhage), sand-like plaques, white plaques, follicular hyperplasia, villous hyperplasia, bluish-purple nodules, and external pressure changes. The matching rules in this layer have extremely high disease specificity: Identifying "Hunner's ulcer" or "glomerulation" greatly increases the probability of interstitial cystitis; identifying "sand-like plaques" or "calcifications" greatly increases the probability of schistosomiasis-related cystitis; identifying "white plaques" greatly increases the probability of leukoplakia of the bladder mucosa; identifying "follicular" or "villous hyperplasia" greatly increases the probability of glandular cystitis; identifying "blue-purple nodules" greatly increases the probability of vesicular endometriosis; identifying "external pressure changes" or "elongation and deformation" greatly increases the probability of pelvic lipomatosis, and a CT scan is recommended for confirmation; a description of "diffuse congestion and edema, fragile mucosa" supports chemical or radiation-induced cystitis.

[0044] II. Large Language Model Reasoning Module Given the limitations of response speed and computing power in grassroots systems, this system adopts a lightweight implementation approach of "General Large Language Model (LLM) + Specialty Rule Base Constraints (PromptEngineering)," without in-depth local parameter fine-tuning. The LLM reasoning module mainly performs two core functions: first, medical record text structuring, intelligently converting unstructured chief complaints or cystoscopy text input by doctors into standardized labels that can be recognized by AI features; second, deep discriminative reasoning, receiving all patient information (including clues output by the rule engine and unstructured text descriptions), performing semantic understanding and probabilistic reasoning, outputting a list of diseases sorted by probability, and listing supporting and excluding evidence for each candidate diagnosis.

[0045] In the deep discriminative reasoning phase, LLM is injected with specialist rule base constraint prompts, requiring it to simulate the discriminative thinking of a senior urologist and evaluate 13 rare bladder diseases one by one. LLM needs to analyze hidden clues in unstructured text, such as inferring the possibility of interstitial cystitis from "the patient reported that the pain was not significantly relieved after urination and that he got up 5-6 times at night," and identifying "some white spots seen under cystoscopy" as possibly "white patches" rather than ordinary inflammatory exudate.

[0046] In the deep discriminative reasoning phase, LLM is injected with specialist rule base constraint prompts, requiring it to simulate the discriminative thinking of a senior urologist and evaluate 13 rare bladder diseases one by one. LLM needs to analyze hidden clues in unstructured text, such as inferring the possibility of interstitial cystitis from "the patient reported that the pain was not significantly relieved after urination and that he got up 5-6 times at night," and identifying "some white spots seen under cystoscopy" as possibly "white patches" rather than ordinary inflammatory exudate.

[0047] III. Typical Case Database The case data in the typical case database includes the following structured fields: basic information fields (case ID, disease label / gold standard diagnosis, diagnosis method, case typicality score 1-5), clinical feature fields (multiple-selection labels for core symptoms, disease course classification, special precipitating factors text keywords), laboratory test fields (urine routine red blood cells / white blood cells / nitrite, presence or absence of eosinophils in urine, presence or absence and type of parasite eggs in urine, blood routine eosinophil count, renal function creatinine value), imaging and endoscopic feature fields (multiple-selection labels for cystoscopy keywords, original description of cystoscopy), and final diagnosis and identification fields.

[0048] The typical case database is configured to calculate the cosine similarity between the current patient's structured feature vector and the cases in the database. The specific algorithm is as follows: after normalizing the current patient's feature vector and the feature vectors of each case in the database, the cosine value of the angle between them is calculated. The system returns the Top-K (e.g., the top 5-10) most similar previous cases and their diagnoses, providing empirical probabilistic support for AI inference. For example, if LLM initially suggests interstitial cystitis, but the case database search shows that the 5 most similar cases are all glandular cystitis, the system will add the difference found in the case database as a "key identification tip" to the report, suggesting a biopsy to clarify the distinction.

[0049] The typical case database is configured to calculate the cosine similarity between the current patient's structured feature vector and the cases in the database. The specific algorithm is as follows: after normalizing the current patient's feature vector and the feature vectors of each case in the database, the cosine value of the angle between them is calculated. The system returns the Top-K (e.g., the top 5-10) most similar previous cases and their diagnoses, providing empirical probabilistic support for AI inference. For example, if LLM initially suggests interstitial cystitis, but the case database search shows that the 5 most similar cases are all glandular cystitis, the system will add the difference found in the case database as a "key identification tip" to the report, suggesting a biopsy to clarify the distinction.

[0050] IV. Three-Component Collaborative Adjudication Module The calling relationship and workflow of the three-component collaborative adjudication module follow a sequential process of "rule priority, intelligent supplementation, and empirical verification": In the first round, the rule engine performs hard rule matching and triage based on structured data, outputting preliminary diagnostic clues or instructions for further identification. If a clear rule is triggered (such as a positive urine culture), a diagnosis is output directly, and the process may end prematurely. Otherwise, an instruction for further identification is output, along with confirmed clues (such as "infection ruled out" or "history of chemotherapy").

[0051] In the next round, the large language model reasoning module receives all patient information and clues output by the rule engine, performs deep discrimination reasoning, and outputs a list of diseases sorted by probability.

[0052] Finally, the typical case database is used to search for similar cases based on the main candidate diagnoses and their key features output by the large language model reasoning module, in order to verify the rationality of the reasoning or indicate the possibility of rarity.

[0053] V. Conflict Resolution and Weighted Integration Mechanism When the outputs of the three components diverge, the system makes a comprehensive decision based on the priority principle of "rules as the basis, LLM as the main focus, and cases as supplementary evidence," and clearly marks the points of divergence in the report.

[0054] When the rule engine's output conflicts with the large language model's inference module's output, the rule engine's result is written as the primary diagnosis in the `primary_diagnosis` field, while the high-risk objection raised by the large language model is encapsulated in the `high_priority_warning` object field. Upon receiving this data structure, the front-end displays the primary diagnosis as usual and forcibly renders the "high-priority warning" module at the top of the interface using a highlighted pop-up or a prominent color block. For example, the rule directly outputs "eosinophilic cystitis" based on "elevated urinary eosinophil count," but the LLM strongly suspects "schistosomiasis cystitis" based on the cystoscopy description of "plasma-like plaques." In this case, the system adopts "eosinophilic cystitis" as the primary possibility, but includes the LLM's objection as a high-priority warning in the report, strongly recommending confirmation in conjunction with a history of contact with epidemic areas or biopsy pathology.

[0055] When the large language model inference module conflicts with the output of the typical case database, the system maintains the disease ranking of the large language model while using the differences found in the case database as a key identification prompt. For example, the top-ranked disease in LLM is "interstitial cystitis," but the case database search shows that the five most similar cases are all "glandular cystitis." The system maintains the LLM ranking but uses the case database finding as an "important prompt: glandular cystitis needs to be identified," and may suggest a biopsy for definitive differentiation.

[0056] When the outputs of the three components are inconsistent, the system calculates the internal confidence score of each component and performs a weighted fusion. The weighted fusion uses a confidence-weighted algorithm, with the rule engine having a weight of 0.6, the large language model inference module having a weight of 0.3, and the typical case database having a weight of 0.1. The total confidence score is calculated as: Total Confidence = (0.6 × Rule Score) + (0.3 × LLM Score) + (0.1 × Case Matching Degree). Simultaneously, the system follows a "safety first" circuit breaker mechanism: when all three are inconsistent and the large language model inference module or the typical case database points to a disease with serious consequences or urgent intervention characteristics (such as a risk of malignant tumors or schistosomiasis infection), the system ignores the regular weights and directly places the suspected disease at the top of the report as the highest-level warning item.

[0057] VI. Probability Ranking Formation Mechanism The probability ranking of the AI ​​inference unit adopts a mechanism that combines "dynamic weighted scoring" and "gold standard one-vote trigger".

[0058] Each candidate disease starts with a score of 0. Matching a supporting feature adds points according to a preset weight; matching an exclusion feature resets the score to zero or significantly deducts points. The system's underlying discrimination model calculates the weighted score and performs Softmax normalization. Specifically, the total disease score = Σ(points added for ordinary supporting features) + Σ(points added for core / gold standard features) - Σ(points heavily penalized for exclusion features). After obtaining the total score, it is converted into a probability percentage of 0% to 1%, and the final candidate diseases are ranked according to their probability.

[0059] When pathological or endoscopic gold standard features are captured, including positive Congo red staining (bladder amyloidosis), discovery of parasite eggs or histological evidence (schistosomiasis cystitis), formation of glandular structures on biopsy (glandular cystitis), and dense fibrosis of the entire bladder wall (leather cystitis), they are given the highest priority and directly become the first candidate diagnosis.

[0060] The quantitative scoring logic for interstitial cystitis and eosinophilic cystitis is as follows: Candidate Disease A: Interstitial Cystitis (IC). Supporting Factors: Long-term urinary frequency and pelvic pain (+2 points); ineffective antibiotics (+3 points). Gold Standard Trigger: If cystoscopy reveals "Hunner's ulcer," it triggers a highly weighted diagnosis (+ points, immediately pinned to the top). Exclusion Factors: No urinary frequency or significantly enlarged bladder capacity (-20 points, significantly reducing probability).

[0061] Candidate Disease B: Eosinophilic Cystitis (EC). Supporting Factors: History of allergies / asthma (+3 points); bladder wall thickening (+2 points). Key Clue Intervention: Significantly elevated blood / urine eosinophil count (+15 points, significantly differentiating it from other non-infectious diseases). Exclusion Factors: Repeatedly normal eosinophil counts (-50 points or zero, forcibly removing it from the primary candidate list).

[0062] This quantitative algorithm, which combines "accumulating subtle clues + highlighting core features + using the gold standard to veto / diagnose," ensures that computers can accurately execute complex clinical identification logic.

[0063] Specific implementation of the "Exclude Infection" one-click judgment module The system also includes a one-click "infection exclusion" judgment module, which uses a dynamic scoring system instead of a fixed set of multiple rules. The scoring rules are as follows: Negative urine culture: +3 points (strong evidence); ineffective standard course of antibiotic treatment: +2 points (strong evidence); normal / normal white blood cell count in urinalysis: +1 point (supporting evidence); presence of non-infectious precipitating factors (such as history of radiotherapy or chemotherapy): +1 point (supporting evidence); presence of microscopic hematuria but no pyuria: +0.5 points (auxiliary clue).

[0064] Specific implementation of the "Exclude Infection" one-click judgment module The system also includes a one-click "infection exclusion" judgment module, which uses a dynamic scoring system instead of a fixed set of multiple rules. The scoring rules are as follows: Negative urine culture: +3 points (strong evidence); ineffective standard course of antibiotic treatment: +2 points (strong evidence); normal / normal white blood cell count in urinalysis: +1 point (supporting evidence); presence of non-infectious precipitating factors (such as history of radiotherapy or chemotherapy): +1 point (supporting evidence); presence of microscopic hematuria but no pyuria: +0.5 points (auxiliary clue).

[0065] The system calculates the total score and compares it with a preset threshold: when the total score is ≥5, a high-risk warning is triggered, and the system clearly indicates "highly suggestive of non-infectious cystitis, and it is recommended to initiate a specific differential diagnosis"; when the total score is ≤3 and the total score is <5, a medium-risk warning is triggered, and the message is "there is a possibility of non-infectious cystitis, and it is recommended to combine it with further clinical evaluation"; when the total score is <3, the routine diagnosis and treatment process is maintained.

[0066] When some conditions are missing but strong evidence exists, the system dynamically adjusts the prompts. For example, even if the condition "not high leukocyte count in urine" is missing (mildly elevated leukocyte count), if "negative urine culture" (+3 points) and "ineffective antibiotics" (+2 points) are present simultaneously, and the total score reaches 5 points, the system will still output a high-risk prompt, but the content will be adjusted to "Although mild inflammatory markers are present, combined with aseptic technique and ineffective treatment, non-infectious causes are still highly suspected, and key differential diagnoses are recommended," and may advance the ranking of special inflammations such as "eosinophilic cystitis."

[0067] Specific Implementation of Misconception Early Warning Unit The error warning unit focuses on identifying the types of errors and triggering conditions, including: First, antibiotic abuse. Triggering conditions: The rule engine detects a negative urine culture, no systemic signs of infection such as high fever, and there is an antibiotic prescription in the prescription system. The system triggers an alert: "There is currently no laboratory evidence to support bacterial infection, and the use of antibiotics lacks indication. Please review the diagnosis." Second, overtreatment. Triggering condition: A comparison between the case database and guidelines reveals that the AI ​​diagnosis ranking shows a high probability of benign or precancerous lesions and recommends major surgery, while the case database shows that most similar cases have received conservative treatment. The system triggers an alert: "This lesion is usually best treated with monitoring or conservative treatment. The current treatment recommendation is too aggressive. We suggest referring to clinical guidelines or conducting a multidisciplinary discussion." Third, the risk of missed diagnosis. Triggering condition: The large model inference features identify keywords such as "sand-like plaque" and "Hunner's ulcer" in the cystoscopy description, but the doctor's initial diagnosis does not include the corresponding disease. The system triggers an alert: "The cystoscopy description suggests [disease name] is possible. This diagnosis is not currently under your consideration and is recommended for differential diagnosis." Fourth, delayed referral. Triggering conditions: The rule engine and diagnostic database determine that the AI's final candidate diagnoses include diseases such as "pelvic lipomatosis," "bladder endometriosis," and "refractory interstitial cystitis," which typically require in-depth specialist treatment or multidisciplinary collaboration, and there are no recent referral or consultation plans in the medical records. The system triggers an alert: "The diagnosis / management of this disease often requires multidisciplinary intervention from urology, gynecology, or pain management specialists; it is recommended to assess the necessity of referral." Fifth, insufficient examination. Triggering condition: The diagnosis-examination mapping rule determines that the AI ​​has a high probability of diagnosing "glandular cystitis," "amyloidosis," "eosinophilic cystitis," or other diseases that require biopsy for diagnosis, and the examination plan / record does not include a cystoscopic biopsy. The system triggers an alert: "A definitive diagnosis of [disease name] requires cystoscopic biopsy and pathological examination; the current examination plan may be insufficient for a clear diagnosis." The alerts are triggered in real time when doctors save diagnoses, prescribe medications, or develop treatment plans. They are presented as a non-blocking but sufficiently prominent sidebar or bottom information bar, listing the type of misconception, a brief explanation, and suggested corrections, avoiding frequent pop-ups that disrupt the workflow. All triggered alerts are logged, and doctors can choose to "accept" or "ignore with explanation." This feedback is used to continuously optimize the accuracy and usability of the misconception database.

[0068] Specific implementation of the diagnosis and treatment decision-making unit The diagnosis and treatment decision-making unit includes a primary care prompt generation module and a mandatory referral red line module.

[0069] The primary care intervention prompt generation module is configured to, when the primary disease falls within the scope of initial intervention at the primary level, invoke a fixed rule template to output standardized lifestyle guidance, symptomatic relief plans, and follow-up appointment points. For example, when the primary disease identified is chemical cystitis or mild leukoplakia of the bladder mucosa, the system invokes a fixed template to output standardized lifestyle guidance (such as hydration with a daily water intake of ≥0 mL, urine alkalization, and avoiding spicy and irritating foods) and routine recommendations for symptomatic relief (such as oral anticholinergic drugs to relieve urinary frequency and urgency), and specifies follow-up appointment points (such as a urinalysis to be repeated after 2 weeks and a cystoscopy to be repeated after 1 month).

[0070] The mandatory referral red line module is configured to immediately trigger the referral template and highlight referral suggestions when any of the following situations are detected: persistent gross hematuria; upper urinary tract hydronephrosis (e.g., ultrasound showing hydronephrosis or elevated renal creatinine / urea nitrogen); ultrasound suspecting malignancy; significantly reduced bladder capacity (e.g., bladder capacity <150mL); biopsy required for diagnosis but local hospitals cannot perform the corresponding examination (e.g., glandular cystitis requires biopsy to rule out tumors, interstitial cystitis requires dilation under anesthesia); unclear diagnosis but symptoms persist above a preset threshold (e.g., persistent symptoms >3 months and ineffective conventional treatment); or the system's recommended disease for which local hospitals cannot perform the critical examinations required for diagnosis.

[0071] The system has a built-in high-risk referral feature database. If the above red line indicators are identified in the input information, the system will immediately trigger the referral template and highlight the output: "This case has [specific high-risk / difficult-to-diagnose features], which is beyond the scope of routine treatment at the primary level. It is strongly recommended to refer the patient to a higher-level urology department / multidisciplinary department for further evaluation" to prevent risks in primary healthcare.

[0072] Specific implementation of the result output unit The output unit integrates the outputs of the AI ​​inference unit, the error warning unit, and the diagnosis and treatment decision unit to generate a structured decision report. The report includes at least the following modules: Candidate disease ranking module: Lists the top 3-5 possible diagnoses in descending order of probability, with each diagnosis accompanied by a probability percentage.

[0073] Specific implementation of the result output unit The output unit integrates the outputs of the AI ​​inference unit, the error warning unit, and the diagnosis and treatment decision unit to generate a structured decision report. The report includes at least the following modules: Candidate disease ranking module: Lists the top 3-5 possible diagnoses in descending order of probability, with each diagnosis accompanied by a probability percentage.

[0074] Diagnostic Criteria Module: For each candidate diagnosis, a clear list of matching points from medical history, laboratory tests to cystoscopy is provided as supporting evidence; at the same time, evidence for exclusion or weakening is listed, such as "Evidence against bacterial cystitis: negative urine culture".

[0075] Risk alert module: Displays high-priority warnings, key identification prompts, and misconception alerts generated during the collaborative adjudication process of the three components.

[0076] Primary care module: Provides standardized lifestyle guidance, symptom relief plans, and follow-up appointments.

[0077] Referral prompt module: When the mandatory referral red line is triggered, the referral suggestion and specific reasons are highlighted with a conspicuous color block.

[0078] Based on the diagnostic decision-making system described in this application, its application in specific implementation is as follows: Example 1: AI-assisted diagnosis of interstitial cystitis The patient, a 42-year-old female, presented with a 2-year history of recurrent urinary frequency and urgency accompanied by lower abdominal pain, which had worsened in the past month. The primary care physician entered basic information into the system: female, 42 years old, chronic disease duration of 2 years. The core symptoms selected were: urinary frequency, urgency, bladder area pain, worsening with urinary retention, and increased nocturia (5-6 times per night). Medical history text input: "The patient has experienced recurrent urinary frequency and urgency for 2 years. Multiple urinalysis tests showed a slight increase in white blood cells. He was treated for urinary tract infection with oral levofloxacin for 2 weeks and cefixime for 1 week, but the symptoms did not significantly improve. The pain is located in the suprapubic region, slightly relieved after urination, and worsened when holding urine." Auxiliary examinations imported: Urinalysis showed 3-5 white blood cells / HP, 2-4 red blood cells / HP, and negative nitrite; urine culture was negative; blood routine was normal; renal function was normal; Cystoscopy description text: "Cystoscopy revealed scattered punctate hemorrhages in the mucosa of the bladder dome and lateral walls, and a Huntner's ulcer, approximately 5 mm in diameter, covered with a white coating. The trigone mucosa was generally normal." The system operation process is as follows: The data acquisition unit obtains all the above information. The feature quantification unit performs NLP processing on the medical history text: it identifies symptom keywords such as "frequent urination," "urinary urgency," "bladder area pain," "worsening with urination," and "increased nocturia"; it identifies "levofloxacin for 2 weeks," "cefixime for 1 week," and "no significant symptom relief," activating the antibiotics_ineffective:True tag; it identifies "negative urine culture," activating the urine_culture_negative:True tag. The cystoscopy description is processed using NER: it identifies "Hunner ulcer" and "glomerulation," converting them into cystoscopy_hunner_ulcer:True and cystoscopy_glomerulation:True tags.

[0079] The data acquisition unit obtains all the above information. The feature quantification unit performs NLP processing on the medical history text: it identifies symptom keywords such as "frequent urination," "urinary urgency," "bladder area pain," "worsening with urination," and "increased nocturia"; it identifies "levofloxacin for 2 weeks," "cefixime for 1 week," and "no significant symptom relief," activating the antibiotics_ineffective:True tag; it identifies "negative urine culture," activating the urine_culture_negative:True tag. The cystoscopy description is processed using NER: it identifies "Hunner ulcer" and "glomerulation," converting them into cystoscopy_hunner_ulcer:True and cystoscopy_glomerulation:True tags.

[0080] The AI ​​inference unit is activated. The rule engine first performs the first layer of infection triage: urine culture is negative, routine urine leukocytes are only slightly elevated and there is no positive nitrite, combined with antibiotics_ineffective:True, acute bacterial cystitis is ruled out. The second layer is a rapid screening of medical history and predisposing factors: no specific predisposing factors such as radiotherapy history, chemotherapy history, history of epidemic areas, or allergy history are matched. The third layer is a precise matching of features and endoscopy: the system identifies cystoscopy_hunner_ulcer:True, according to the rule base, this feature is the gold standard feature of interstitial cystitis, triggering a top-priority mechanism, directly assigning interstitial cystitis the highest priority.

[0081] The large language model reasoning module receives all information and performs deep reasoning. LLM analyzes semantic clues in the medical history text such as "recurring for 2 years", "ineffective antibiotics", "worsening with holding urine", and "5-6 times of nocturia", which further support the diagnosis of interstitial cystitis. At the same time, it points out that it needs to be differentiated from glandular cystitis (because both can present as chronic urinary frequency), but glandular cystitis usually does not have Hunter's ulcers.

[0082] The large language model reasoning module receives all information and performs deep reasoning. LLM analyzes semantic clues in the medical history text such as "recurring for 2 years", "ineffective antibiotics", "worsening with holding urine", and "5-6 times of nocturia", which further support the diagnosis of interstitial cystitis. At the same time, it points out that it needs to be differentiated from glandular cystitis (because both can present as chronic urinary frequency), but glandular cystitis usually does not have Hunter's ulcers.

[0083] The typical case database was used to search for similar cases based on the key features of interstitial cystitis. Among the top-5 cases returned, 4 were diagnosed with interstitial cystitis and 1 was radiation cystitis (this case also had similar pain but a different medical history), providing empirical support for LLM inference.

[0084] The three-component collaborative adjudication module makes the decision: the rule engine and LLM output are consistent (both pointing to interstitial cystitis), the case database has majority support, and there are no conflicts among the three. The system calculates the probability using Softmax normalization: interstitial cystitis probability 88.5%, glandular cystitis probability 6.2%, radiation cystitis probability 3.1%, and other diseases have even lower probabilities.

[0085] The system detects a potential "chronic urinary tract infection" diagnosis by a doctor, but the system identifies a negative urine culture, no systemic infection signs, and a history of antibiotic use, triggering a "misuse of antibiotics" warning: "There is currently no laboratory evidence to support bacterial infection, and previous antibiotic treatment has been ineffective. Please verify the 'chronic urinary tract infection' diagnosis, and interstitial cystitis should be considered as a priority." Simultaneously, the system detects the keyword "Hunner's ulcer" in the cystoscopy description. If the doctor's initial diagnosis does not include interstitial cystitis, a "missed diagnosis risk" warning is triggered.

[0086] The diagnostic and treatment decision-making unit operates as follows: The primary disease, interstitial cystitis, falls under the category of requiring in-depth specialist treatment. The system invokes the primary care management prompt template and outputs: lifestyle guidance (avoiding irritating foods, bladder training, and maintaining psychological relaxation); symptomatic relief plan (oral amitriptyline or pentosan sodium polysulfate is recommended, but primary care physicians should use it under the guidance of a senior physician); and a clear referral follow-up point (immediate referral to a higher-level hospital for hydrocephalus under anesthesia and potassium sensitivity testing for further diagnosis). Simultaneously, the mandatory referral red line module identifies that this case requires biopsy and hydrocephalus for diagnosis, which the local hospital may not be able to perform, triggering the referral template: "This case presents with a cystoscopic Hunner's ulcer. A definitive diagnosis of interstitial cystitis requires hydrocephalus under anesthesia and a potassium sensitivity test, which exceeds the scope of routine primary care management. Referral to a higher-level urology department for further evaluation is strongly recommended." The output unit generates a final decision report, showing interstitial cystitis as the primary candidate diagnosis (probability 88.5%), detailing supporting evidence (chronic course, ineffective antibiotics, cystoscopic Huntner's ulcer) and exclusionary evidence (negative urine culture excluding bacterial infection), and highlighting referral recommendations.

[0087] Example 2: AI-assisted diagnosis of schistosomiasis cystitis The patient, a 38-year-old male, presented with intermittent hematuria and urinary frequency for 3 months. Basic information: Male, 38 years old, 3-month course of illness. Core symptoms: Gross hematuria, urinary frequency, and urgency. Medical history: "The patient began experiencing intermittent terminal hematuria 3 months ago, accompanied by urinary frequency and urgency, without dysuria. He was treated at a local clinic with oral antibiotics for 'urinary tract infection' for 1 week without effect. Five years ago, the patient worked in Nigeria, Africa for 2 years and had a history of exposure to freshwater." Ancillary examinations: Urinalysis showed numerous red blood cells and 2-3 white blood cells / HP; urine culture was negative; blood routine showed eosinophils 8% (mildly elevated); renal function was normal; cystoscopy description: "Cystoscopy revealed multiple scattered sand-like yellow plaques on the posterior wall and trigone of the bladder, with a rough surface, calcifications in some areas, and mucosal congestion and edema." System operation process: The feature quantification unit processes medical history text: It identifies "Africa, Nigeria" and "history of freshwater exposure," activating the `endemic_exposure:True` tag; it identifies "antibiotics ineffective after 1 week," activating the `antibiotics_ineffective:True` tag. It processes cystoscopy descriptions: NER identifies "sand-like yellow plaques" and "calcifications," converting them into `cystoscopy_sandy_patches:True` and `cystoscopy_calcification:True` tags.

[0088] The rule engine operates as follows: The first layer involves infection triage; urine culture is negative, and white blood cell count is only slightly elevated, ruling out acute bacterial cystitis. The second layer involves rapid screening of medical history and predisposing factors, matching "Africa, Nigeria" and "history of freshwater contact," quickly identifying the schistosomiasis cystitis category. The third layer involves precise feature matching with endoscopy findings, identifying cystoscopy_sandy_patches:True, a core supporting feature for schistosomiasis cystitis, which is assigned a high weight.

[0089] The LLM inference module receives all information and performs in-depth analysis of semantic associations such as "terminal hematuria", "history of working in Africa", and "sand-like plaques", outputting the disease ranking: schistosomiasis cystitis (probability 72.3%), vesicular leukoplakia (probability 12.1%, due to the presence of plaque-like changes but different colors), and interstitial cystitis (probability 8.5%).

[0090] The LLM inference module receives all information and performs in-depth analysis of semantic associations such as "terminal hematuria", "history of working in Africa", and "sand-like plaques", outputting the disease ranking: schistosomiasis cystitis (probability 72.3%), vesicular leukoplakia (probability 12.1%, due to the presence of plaque-like changes but different colors), and interstitial cystitis (probability 8.5%).

[0091] Typical case database search: Using "sand-like plaques", "history of epidemic area", and "terminal hematuria" as search criteria, the top-5 similar cases returned included 3 cases diagnosed as schistosomiasis cystitis and 2 cases of late-stage schistosomiasis complicated with bladder cancer (the system will display this as a warning message).

[0092] The three components worked in synergy: the outputs were largely consistent, but the case database suggested the possibility of concurrent bladder cancer. The system maintained schistosomiasis cystitis as the primary diagnosis (probability 72.3%), while also prioritizing the finding in the case database that "advanced schistosomiasis cystitis is a high-risk factor for bladder cancer" as a high-priority warning: "Patients diagnosed with schistosomiasis cystitis are at high risk for bladder squamous cell carcinoma. Systemic screening and regular monitoring are recommended, and biopsy should be performed if necessary to rule out malignancy." Misconception Warning Unit: If a doctor prescribes antibiotics, an "Antibiotic Abuse" warning is triggered (negative urine culture, no signs of infection). If the disease is detected, finding worm eggs is necessary for diagnosis. If the testing plan does not include finding worm eggs in urine sediment or a biopsy, an "Insufficient Testing" warning is triggered: "A definitive diagnosis of schistosomiasis cystitis requires finding worm eggs in urine or biopsy tissue; the current testing plan may not be sufficient for a definitive diagnosis." Treatment decision-making unit: Providing a primary care treatment plan (recommending plenty of water, avoiding contact with fresh water, and that oral praziquantel should be administered under the guidance of a parasitology specialist at a higher-level hospital); Simultaneously triggering mandatory referral: "This case has a history of contact with an epidemic area and sand-like plaques under cystoscopy. Diagnosis requires finding worm eggs in urine or tissue biopsy. Furthermore, schistosomiasis cystitis falls under the joint management of a specialized infectious disease and the urology department. It is strongly recommended that the patient be referred to a higher-level hospital for further diagnosis and standardized deworming treatment." Example 3: Differential Diagnosis of Glandular Cystitis and Bladder Tumor The patient, a 55-year-old male, presented with a 2-week history of painless gross hematuria. Basic information: Male, 55 years old, 2-week course of illness. Core symptom: Painless gross hematuria. Medical history: "The patient suddenly developed painless gross hematuria throughout urination 2 weeks ago. The hematuria was bright red, without clots, and accompanied by mild urinary frequency. No history of radiotherapy, chemotherapy, or allergies." Ancillary examinations: Urinalysis showed full field of red blood cells, negative white blood cells; urine culture was negative; complete blood count was normal; renal function was normal; urinary tract ultrasound showed a "space-occupying lesion in the bladder trigone, approximately 1.2 cm, with an irregular surface"; cystoscopy description: "Multiple follicular and villous proliferative lesions were seen in the bladder trigone and around the right ureteral orifice. The surface was grayish-white, with some areas showing papillary elevations. The lesions bled easily upon touch, and the lesion area was approximately 2.0 × 1.5 cm." System operation process: The feature quantization unit processes the cystoscopy description: NER extracts entities such as "follicular", "villous", "papillary", "triangle", and "prone to bleeding" and converts them into tags such as cystoscopy_follicular_hyperplasia:True, cystoscopy_villous_hyperplasia:True, and cystoscopy_papillary_projection:True.

[0093] Rule Engine: First layer excludes infection (negative urine culture, negative white blood cells). Second layer rapid screening of medical history and precipitating factors: no specific high-weight precipitating factors were matched. Third layer precise matching of features and endoscopy: "follicular" and "villous" hyperplasia were identified, triggering a high-weight bonus for glandular cystitis; however, "papillary elevation," "prone bleeding," and "space-occupying lesions" simultaneously triggered warning rules for bladder tumors (especially urothelial carcinoma).

[0094] The LLM inference module performs deep semantic analysis: LLM notices semantic clues highly suggestive of malignancy, such as "55-year-old male," "painless gross hematuria," "papillary protrusion," and "space-occupying lesion." It also identifies "follicular" and "villous" as typical manifestations of glandular cystitis. LLM output ranking: glandular cystitis (probability 45.2%), bladder urothelial tumor (probability 38.7%), interstitial cystitis (probability 5.1%), etc. LLM generates a special warning: "Glandular cystitis and low-grade malignant potential papillary urothelial tumors are extremely similar in appearance. This case presents with papillary protrusion and a space-occupying lesion; the risk of malignancy cannot be ignored." The LLM inference module performs deep semantic analysis: LLM notices semantic clues highly suggestive of malignancy, such as "55-year-old male," "painless gross hematuria," "papillary protrusion," and "space-occupying lesion." It also identifies "follicular" and "villous" as typical manifestations of glandular cystitis. LLM output ranking: glandular cystitis (probability 45.2%), bladder urothelial tumor (probability 38.7%), interstitial cystitis (probability 5.1%), etc. LLM generates a special warning: "Glandular cystitis and low-grade malignant potential papillary urothelial tumors are extremely similar in appearance. This case presents with papillary protrusion and a space-occupying lesion; the risk of malignancy cannot be ignored." Typical case database search: Using "follicular," "papillary," "triangle," and "painless hematuria" as search criteria, the top-5 cases returned included 3 cases of glandular cystitis and 2 cases of bladder tumors. The case database supports a cautious approach to LLM.

[0095] The three components work together to make a collaborative decision: the rule engine assigns a higher weight to glandular cystitis due to "follicular hyperplasia"; LLM significantly increases the probability of bladder tumor due to factors such as age, painless hematuria, and space-occupying lesions; the case database shows that both are possible. At this point, there is some conflict among the three outputs. The system calculates the weighted confidence scores: the rule engine score is 0.75 (supporting glandular cystitis), the LLM score is 0.82 (but LLM gives higher attention to bladder tumor), and the case match is 0.6. Calculated by weight: total confidence (glandular cystitis) = 0.6 × 0.75 + 0.3 × 0.65 + 0.1 × 0.6 = 0.69; total confidence (bladder tumor) = 0.6 × 0.3 + 0.3 × 0.82 + 0.1 × 0.4 = 0.406. Although the total score for glandular cystitis is still slightly higher, LLM strongly suggests the risk of malignancy, and the consequences of this disease are serious. Based on the "safety first" circuit breaker mechanism, the system places the suspicion of bladder tumor as the highest level warning item at the top of the report, while keeping glandular cystitis as the primary candidate diagnosis.

[0096] Warning Unit for Misconceptions: Triggering the "Risk of Missed Diagnosis" warning (if the doctor only considers benign lesions). Triggering the "Insufficient Examination" warning: "A definitive diagnosis of glandular cystitis and the exclusion of bladder tumors both require cystoscopic biopsy and pathological examination. Currently, only a cystoscopic description is available, and no biopsy has been arranged, which is insufficient to make a definitive diagnosis." Triggering the "Delayed Referral" warning: "This case has a space-occupying lesion and papillary protrusion, with a high risk of malignancy. Further biopsy and CT staging evaluation by a specialist urologist are required. Immediate referral is recommended." Treatment Decision Unit: Although the primary disease is glandular cystitis, the system identified malignancy risk and space-occupying lesions, directly triggering the mandatory referral red line: "This case has a space-occupying lesion in the bladder, papillary bulges and painless gross hematuria, the risk of malignancy is extremely high, and the diagnosis must rely on biopsy pathology, which is beyond the scope of primary care. It is strongly recommended to immediately refer the patient to a higher-level hospital for cystoscopic biopsy and imaging evaluation, and any attempt at local treatment at the primary care level is prohibited." Treatment Decision Unit: Although the primary disease is glandular cystitis, the system identified malignancy risk and space-occupying lesions, directly triggering the mandatory referral red line: "This case has a space-occupying lesion in the bladder, papillary bulges and painless gross hematuria, the risk of malignancy is extremely high, and the diagnosis must rely on biopsy pathology, which is beyond the scope of primary care. It is strongly recommended to immediately refer the patient to a higher-level hospital for cystoscopic biopsy and imaging evaluation, and any attempt at local treatment at the primary care level is prohibited." Example 4: AI-assisted diagnosis of pelvic lipomatosis The patient, a 48-year-old male, presented with a 2-year history of progressive dysuria and urinary frequency. Basic information: Male, 48 years old, 2-year course of illness. Core symptoms: Urinary frequency, dysuria, weak urine stream, and increased nocturia. Medical history: "The patient has gradually developed difficulty urinating, a weak urine stream, and urinary frequency over the past 2 years, waking up 4-5 times per night. No hematuria, no dysuria. Previously healthy, no history of surgery, BMI 32 kg / m2." Ancillary examinations: Urinalysis normal; urine culture negative; blood routine normal; renal function tests showed creatinine 132 μmol / L (mildly elevated); urinary tract ultrasound showed "mild hydronephrosis in both kidneys, dilation of the upper ureter, elongated bladder, and thickened bladder wall"; cystoscopy description: "Cystoscopy was successful, revealing significant elevation and elongation of the bladder neck, compression and deformation of the trigone and posterior wall mucosa, showing external pressure changes; the mucosal surface was smooth, vascular patterns were clear, and no obvious neoplasms were observed." System operation process: The feature quantification unit processes the medical history text: it identifies "BMI 32", "difficulty urinating", and "urinary frequency" and activates relevant tags. It processes the cystoscopy description: NER extracts "elevated bladder neck", "elongated", "triangle compression", and "external pressure changes" and converts them into the tags cystoscopy_extrinsic_compression:True and cystoscopy_bladder_neck_elevation:True. It identifies elevated renal creatinine and activates the renal_function_elevated:True tag.

[0097] Rule Engine: First layer excludes infection (normal urinalysis, negative urine culture). Second layer rapid screening of medical history and precipitating factors: no match was found for radiotherapy, chemotherapy, epidemic areas, etc., but BMI 32 suggests metabolic factors. Third layer precise matching of features with endoscopy: "external pressure changes", "elevation of the bladder neck", and "elongation" were identified, highly suggestive of pelvic lipomatosis; at the same time, bilateral hydronephrosis and elevated renal function suggest upper urinary tract damage.

[0098] LLM inference module analysis: LLM makes comprehensive inferences based on clues such as "male", "obesity", "progressive dysuria", "external pressure changes of the bladder", and "hydronephrosis", and outputs the following order: pelvic lipomatosis (probability 68.4%), benign prostatic hyperplasia (probability 20.3%, due to similar symptoms but no prostatic hyperplasia on cystoscopy), and neurogenic bladder (probability 5.2%).

[0099] LLM inference module analysis: LLM makes comprehensive inferences based on clues such as "male", "obesity", "progressive dysuria", "external pressure changes of the bladder", and "hydronephrosis", and outputs the following order: pelvic lipomatosis (probability 68.4%), benign prostatic hyperplasia (probability 20.3%, due to similar symptoms but no prostatic hyperplasia on cystoscopy), and neurogenic bladder (probability 5.2%).

[0100] Typical case database search: Using "external pressure changes", "obesity", and "hydronephrosis" as criteria, the top-5 cases returned included 4 cases diagnosed with pelvic lipomatosis and 1 case diagnosed with retroperitoneal fibrosis.

[0101] The three components worked together to make the decision: their outputs were largely consistent. The system output that pelvic lipomatosis was the primary candidate (68.4% probability), followed by benign prostatic hyperplasia (20.3%).

[0102] Misconception Warning Unit: If a doctor initially diagnoses "benign prostatic hyperplasia" (BPH), the system triggers a "Missed Diagnosis Risk" warning: "The cystoscopy description suggests external pressure changes rather than BPH, and the patient's elevated and elongated bladder neck is more consistent with the characteristics of pelvic lipomatosis. It is recommended to include it in the differential diagnosis and confirm it with a pelvic CT scan." Misconception Warning Unit: If a doctor initially diagnoses "benign prostatic hyperplasia" (BPH), the system triggers a "Missed Diagnosis Risk" warning: "The cystoscopy description suggests external pressure changes rather than BPH, and the patient's elevated and elongated bladder neck is more consistent with the characteristics of pelvic lipomatosis. It is recommended to include it in the differential diagnosis and confirm it with a pelvic CT scan." Treatment Decision Unit: Primary care prompts and provides lifestyle guidance (weight loss, management of metabolic syndrome, regular monitoring of kidney function). The mandatory referral red line module identifies "bilateral hydronephrosis," "impaired kidney function," and "requires CT confirmation," immediately triggering a referral: "This case presents with upper urinary tract hydronephrosis and impaired kidney function. A diagnosis of pelvic lipomatosis requires pelvic CT / MRI showing adipose tissue hyperplasia, and hydronephrosis complications have already occurred. Referral to a higher-level hospital is strongly recommended to assess the necessity of surgical decompression." Example 5: AI-assisted diagnosis and three-component conflict resolution for eosinophilic cystitis The patient, a 35-year-old female, presented with hematuria, urinary frequency, and lower abdominal pain for one month. She had a history of allergic rhinitis and asthma. Basic information: Female, 35 years old, duration of illness: 1 month. Core symptoms: Gross hematuria, urinary frequency, and lower abdominal pain. Medical history: "The patient developed hematuria one month ago after a cold, accompanied by urinary frequency and dull lower abdominal pain. She has a 10-year history of allergic rhinitis and a 3-year history of asthma. No history of radiotherapy or chemotherapy." Auxiliary examinations: Urinalysis showed numerous red blood cells and 5-8 white blood cells / HP; urine culture was negative; blood routine showed eosinophils at 18% (significantly elevated, normal value <5%); microscopic examination of urine sediment revealed positive eosinophils; renal function was normal; cystoscopy description: "Cystoscopy revealed diffuse edema and congestion of the anterior and dorsal bladder mucosa, with broad-based polypoid elevations covered with a white coating, increased mucosal fragility, and a normal trigone." System operation process: Feature quantification unit: Identifies "allergic rhinitis" and "asthma", activates allergy_history:True; Identifies 18% eosinophils in blood and positive eosinophils in urine, activates eosinophil_elevated:True; Processes cystoscopy descriptions to extract "diffuse edema", "polypoid protrusions", and "increased mucosal fragility", and converts them into corresponding labels.

[0103] Rule Engine: First layer excludes infection (negative urine culture). Second layer rapid screening of medical history and triggers: matching "allergy history" and "asthma" advances the ranking of eosinophilic cystitis. Third layer precise matching of features and endoscopy: significantly elevated blood / urine eosinophil count is a core laboratory supporting feature for eosinophilic cystitis, triggering a high-weighted +15 score; cystoscopic "polypoid protrusions" also support this diagnosis. Rule Engine output: Eosinophilic cystitis is the primary candidate (rule score 0.85).

[0104] LLM Inference Module: LLM analysis of all information largely supports eosinophilic cystitis. However, when semantically analyzing the cystoscopy description, LLM noted that "broad-based polypoid elevation" and "increased mucosal fragility" can also be seen in glandular cystitis and bladder tumors. Furthermore, the patient is a 35-year-old female with gross hematuria, requiring careful exclusion of malignancy. LLM output ranking: eosinophilic cystitis (probability 55.3%), glandular cystitis (probability 22.1%), bladder tumor (probability 12.4%).

[0105] Typical case database search: Using "elevated eosinophils", "polypoid" and "hematuria" as criteria, the top 5 cases returned included 3 cases of eosinophilic cystitis and 2 cases of glandular cystitis (one of the glandular cystitis patients also had mild eosinophil elevation and was a patient with a combined allergic constitution).

[0106] Typical case database search: Using "elevated eosinophils", "polypoid" and "hematuria" as criteria, the top 5 cases returned included 3 cases of eosinophilic cystitis and 2 cases of glandular cystitis (one of the glandular cystitis patients also had mild eosinophil elevation and was a patient with a combined allergic constitution).

[0107] The three-component collaborative decision-making system: The rule engine and LLM agree on the primary diagnosis (eosinophilic cystitis), but LLM assigns a higher probability to glandular cystitis and bladder tumors. The case database also indicates that glandular cystitis is possible. The system calculates the following using weighted fusion: Total confidence (eosinophilic cystitis) = 0.6 × 0.85 + 0.3 × 0.553 + 0.1 × 0.6 = 0.7359; Total confidence (glandular cystitis) = 0.6 × 0.1 + 0.3 × 0.221 + 0.1 × 0.4 = 0.1363. After Softmax normalization: the probability of eosinophilic cystitis is approximately 78.2%, the probability of glandular cystitis is approximately 14.5%, and the probability of bladder tumor is approximately 4.8%. The system maintains eosinophilic cystitis as the primary diagnosis, but highlights the need for differentiation from glandular cystitis and tumors (as suggested by LLM) and the case database's "glandular cystitis cases with allergic constitutions" as key differential diagnoses, recommending biopsy to clarify the nature of the polyps.

[0108] Misconception Warning Unit: If the doctor does not arrange a biopsy, it will trigger the "Insufficient Examination" warning: "A diagnosis of eosinophilic cystitis and the exclusion of glandular cystitis / tumor both require cystoscopic biopsy and pathological examination. The current examination plan may not be sufficient to make a clear diagnosis." Treatment decision-making unit: Outputs primary care treatment plan (suggests avoiding contact with known allergens, oral antihistamines, and glucocorticoid treatment should be carried out under the guidance of a senior physician); at the same time triggers referral recommendation: "This case has a broad-based polypoid lesion, and a biopsy is required to rule out glandular cystitis and tumors. It is recommended to refer the patient to a higher-level hospital for cystoscopic biopsy." Misconception Warning Unit: If the doctor does not arrange a biopsy, it will trigger the "Insufficient Examination" warning: "A diagnosis of eosinophilic cystitis and the exclusion of glandular cystitis / tumor both require cystoscopic biopsy and pathological examination. The current examination plan may not be sufficient to make a clear diagnosis." Treatment decision-making unit: Outputs primary care treatment plan (suggests avoiding contact with known allergens, oral antihistamines, and glucocorticoid treatment should be carried out under the guidance of a senior physician); at the same time triggers referral recommendation: "This case has a broad-based polypoid lesion, and a biopsy is required to rule out glandular cystitis and tumors. It is recommended to refer the patient to a higher-level hospital for cystoscopic biopsy." Example 6: Identification of the terminal manifestations of leptocystitis The patient, a 62-year-old male, presented with a 5-year history of severe urinary frequency, urgency, and progressive dysuria, followed by 3 days of urinary retention. Basic information: Male, 62 years old, 5-year disease duration. Core symptoms: Severe urinary frequency (>30 times during the day), urgency, dysuria, and acute urinary retention. Medical history: "Five years ago, the patient underwent partial cystectomy for bladder cancer and received pelvic radiotherapy. Post-operatively, he gradually developed urinary frequency and urgency, which progressively worsened. In the past six months, his urine stream became significantly thinner, and three days ago, he experienced complete urinary inability to urinate." Ancillary examinations: Urinalysis showed a small number of white blood cells; urine culture was negative; blood routine showed hemoglobin 85 g / L (anemia); renal function tests showed creatinine 280 μmol / L (significantly elevated); cystoscopy description: "Cystoscopy revealed a significantly reduced bladder cavity, pale and rigid mucosa, complete disappearance of blood vessels, extensive submucosal fibrosis, loss of bladder wall elasticity, difficulty in filling with water, and a capacity of approximately 80 mL." System operation process: Feature quantification unit: Recognizing "post-bladder cancer surgery" and "pelvic radiotherapy", activating radiation_history:True and bladder_cancer_surgery:True; recognizing "significantly reduced bladder cavity", "pale and rigid mucosa", "extensive fibrosis", and "volume approximately 80mL", converting them into cystoscopy_bladder_capacity_reduced:True, cystoscopy_fibrosis:True, and cystoscopy_mucosal_pallor:True tags.

[0109] Rule Engine: First layer excludes infection. Second layer rapid screening of medical history and predisposing factors: Matches "pelvic radiotherapy," prioritizing radiation cystitis. Third layer precise matching of features and endoscopy: Identifies "significantly reduced bladder cavity," "extensive fibrosis," and "rigid, leathery bladder wall," triggering the gold standard pathological morphological features of leathery cystitis (although no biopsy was performed, the endoscopic manifestations are highly characteristic), while also matching the pale mucosa and reduced vascularity characteristic of radiation cystitis.

[0110] LLM Inference Module: LLM comprehensive analysis of clues such as "history of radiotherapy", "post-bladder cancer surgery", "5-year progressive worsening", "bladder capacity 80mL", "extensive fibrosis", and "significantly elevated renal function" determined that the patient presented with end-stage radiation cystitis, complicated by upper urinary tract damage and renal failure. LLM output ranking: Leather cystitis (probability 52.1%), Radiation cystitis (probability 35.4%, as the underlying cause), Pelvic lipomatosis (probability 3.2%).

[0111] LLM Inference Module: The LLM comprehensively analyzes clues such as "radiation therapy history", "post-operative bladder cancer", "progressive aggravation over 5 years", "bladder capacity of 80 mL", "extensive fibrosis", and "significantly elevated renal function", and determines that the patient has the manifestation of end-stage leathery cystitis due to radiation cystitis, complicated with upper urinary tract damage and renal failure. The LLM output ranking: Leathery cystitis (probability 52.1%), Radiation cystitis (probability 35.4%, as the etiological basis), Pelvic lipomatosis (probability 3.2%).

[0112] Retrieval of Typical Case Library: With "radiation therapy history", "bladder capacity < mL", and "fibrosis" as conditions, the top-3 returned cases are all leathery cystitis (evolved from end-stage radiation cystitis), providing strong empirical support for the LLM inference.

[0113] Synergistic Ruling of Three Components: The outputs of the three are highly consistent. The system clearly designates leathery cystitis as the primary diagnosis (probability 52.1%), and radiation cystitis as the second candidate as the underlying etiology (35.4%).

[0114] Misconception Warning Unit: If the doctor attempts to indwelling a urinary catheter and only administers antibiotics afterwards, it triggers an "abuse of antibiotics" warning (without evidence of infection). It triggers a "delay in referral" warning: "This disease is an end-stage change, accompanied by severe reduction in bladder capacity and renal failure, and requires specialist treatment." Diagnosis and Treatment Decision-making Unit: Immediately triggers multiple mandatory referral red lines: "intractable dysuria / urinary retention", "significantly impaired renal function (creatinine 280 μmol / L)", "significantly reduced bladder capacity (about 80 mL)", "undiagnosed but symptoms persist for 5 years and end-stage changes". The system highlights the output: "This case has end-stage bladder fibrosis, severe reduction in capacity, and renal failure, which has definitely exceeded the scope of primary care and requires urgent referral to a superior urology department for evaluation of the possibility of bladder augmentation surgery or urinary diversion surgery." Example 7: Interdisciplinary Differentiation of Bladder Endometriosis A patient, female, 29 years old, presented for medical treatment due to "cyclic hematuria and frequent urination for 6 months". Basic information: Female, 29 years old, disease course of 6 months. Core symptoms: Hematuria (synchronous with menstruation), frequent urination, lower abdominal pain. Medical history text: "The patient has had gross hematuria every time menstruation occurs for 6 months, accompanied by frequent urination and stabbing pain in the lower abdomen. The symptoms resolve spontaneously after menstruation. There is a history of dysmenorrhea for 10 years." Auxiliary examinations: Urine routine shows full视野 of red blood cells during menstruation and normal during non-menstruation; urine culture is negative; blood routine is normal; renal function is normal; cystoscopy description (examination during menstruation): "There is a blue-purple raised nodule about 8 mm in diameter visible on the posterior wall of the bladder near the top. The surface mucosa is intact, and local bleeding can be seen. The volume of the nodule is larger during menstruation than during non-menstruation." System Operation Process: Feature quantization unit: Recognizes "appears during menstruation", "relieves after menstruation", and "dysmenorrhea", and activates menstrual_correlation:True; processes cystoscopy descriptions to extract "blue-purple raised nodules", "enlarges during menstruation", and "bleeding", and converts them into cystoscopy_blue_purple_nodule:True and cystoscopy_menstrual_bleeding:True labels.

[0115] Rule Engine: First layer excludes infection. Second layer rapid screening of medical history and causes: matching "menstrual-related" and "dysmenorrhea" quickly identifies bladder endometriosis. Third layer precise matching of features and endoscopy: identifying "blue-purple nodules" and "enlargement during menstruation" are characteristic endoscopic manifestations of bladder endometriosis, assigning them high weight.

[0116] LLM reasoning module: Based on clues such as "periodic hematuria", "synchronous with menstruation", "blue-purple nodules", and "history of dysmenorrhea", LLM strongly supports bladder endometriosis, while also suggesting that it needs to be differentiated from bladder tumor (hemangioma type), but the latter does not have a menstrual cycle.

[0117] Typical case database search: Using "blue-purple nodules" and "menstrual-related hematuria" as criteria, the top-5 cases returned included 4 cases diagnosed as bladder endometriosis and 1 case of bladder hemangioma (this case was not related to menstruation).

[0118] Three-component synergistic assessment: all three consistently point to bladder endometriosis. Softmax normalized probability: bladder endometriosis 91.2%, bladder hemangioma 4.5%, and other diseases are even lower.

[0119] Misconception Warning Unit: If the doctor initially diagnoses "urinary tract infection" or "bladder tumor", triggering the "risk of missed diagnosis" warning: "The endoscopy description suggests possible bladder endometriosis. The core clue is that periodic hematuria is synchronized with menstruation. This diagnosis is not under your current consideration. It is recommended to include it in the differential diagnosis and seek joint diagnosis and treatment from a gynecologist." Treatment decision-making unit: Outputs primary care suggestions (suggests recording the relationship between symptoms and menstrual cycle, avoiding cystoscopy during menstruation, and symptomatic pain relief); at the same time triggers referral suggestions: "The diagnosis / management of this disease often requires multidisciplinary collaboration between urology and gynecology. It is recommended to evaluate and refer the patient to the gynecology-urology joint clinic of a higher-level hospital for further laparoscopic and cystoscopic evaluation." Example 8: Pathological-dependent diagnosis of bladder amyloidosis The patient, a 58-year-old male, presented with a 3-month history of painless hematuria accompanied by urinary discomfort. Basic information: Male, 58 years old, 3-month course of illness. Core symptoms: Painless gross hematuria, urinary discomfort. Medical history: "The patient has experienced recurrent painless gross hematuria for 3 months, accompanied by a feeling of difficulty urinating. No history of radiotherapy, chemotherapy, or travel to epidemic areas." Ancillary examinations: Urinalysis showed numerous red blood cells; urine culture was negative; blood routine was normal; renal function was normal; urinary tract ultrasound showed "localized thickening of the bladder wall, approximately 1.5 cm, hypoechoic, with relatively clear borders"; cystoscopy description: "A submucosal raised lesion was seen on the right lateral bladder wall, with normal mucosal color, a shallow ulcer visible in the center, firm to the touch, and no obvious bleeding." System operation process: Feature quantification unit: Processing cystoscopy descriptions to extract "submucosal elevation", "firm texture", and "superficial ulcer", and converting them into corresponding tags. No specific endoscopic keywords for 13 rare bladder diseases were matched (no Hunter's ulcer, no sand-like plaque, no white plaque, etc.).

[0120] Rule Engine: First layer excludes infection. Second layer rapid screening of medical history and precipitating factors: no specific precipitating factor was matched. Third layer precise matching of features and endoscopy: no specific endoscopic feature match, the rule engine outputs "to be identified", listing bladder amyloidosis, glandular cystitis, and bladder tumor as medium probability candidates.

[0121] LLM Inference Module: LLM analyzes clues such as "submucosal elevation," "firm texture," "superficial ulcer," and "painless hematuria," suggesting bladder amyloidosis, bladder tumor, and glandular cystitis as possibilities. LLM specifically points out: "Bladder amyloidosis can present microscopically as a painless submucosal mass or ulcer, making it difficult to distinguish from tumors and interstitial cystitis, and highly dependent on pathological diagnosis." LLM output ranking: Bladder tumor (probability 40.2%), Bladder amyloidosis (probability 28.5%), Glandular cystitis (probability 18.3%).

[0122] Typical case database search: Using "submucosal elevation," "firm texture," and "ulceration" as criteria, the top-5 returned cases included 2 cases of bladder amyloidosis, 2 cases of bladder urothelial carcinoma, and 1 case of glandular cystitis. The case database indicates a high degree of overlap in the microscopic manifestations of the three diseases.

[0123] Three-component collaborative adjudication: The rule engine showed no clear bias, while LLM ranked bladder tumor first (due to the high-risk signal of painless hematuria in a 58-year-old male), and the case database showed that all three were common. At this point, the outputs for the three were inconsistent. The system calculated the weighted confidence scores: total confidence score for bladder tumor = 0.6 × 0.35 + 0.3 × 0.402 + 0.1 × 0.4 = 0.3706; total confidence score for bladder amyloidosis = 0.6 × 0.2 + 0.3 × 0.285 + 0.1 × 0.2 = 0.2255. Although the total score for bladder tumor was slightly higher, both bladder amyloidosis and bladder tumor are diseases requiring pathological diagnosis and with serious consequences. Based on the "safety first" circuit breaker mechanism, the system placed both "bladder tumor" and "bladder amyloidosis" as the highest-level warning items at the top of the report, clearly stating: "The current endoscopic findings lack specificity; neither bladder tumor nor bladder amyloidosis can be ruled out, and pathological diagnosis via biopsy is necessary." Misconception Warning Unit: Strongly triggers the "Insufficient Examination" warning: "A definitive diagnosis of bladder amyloidosis requires a positive Congo red staining result on the bladder mucosa (apple green birefringence under polarized light), and a definitive diagnosis of bladder tumor also requires biopsy pathology. Currently, there are no biopsies scheduled, indicating a severe shortage of examinations." Triggers the "Risk of Missed Diagnosis" warning. Triggers the "Delayed Referral" warning.

[0124] Treatment decision unit: Directly triggered the mandatory referral red line: "This case has submucosal raised lesions and painless gross hematuria. Malignant lesions and rare diseases (amyloidosis) cannot be ruled out. The diagnosis must rely on biopsy pathology. It is beyond the scope of primary care. It is strongly recommended to immediately refer the patient to a higher-level hospital for cystoscopic biopsy and Congo red staining pathological examination." Example 9: Rapid Identification of Chemical Cystitis The patient, a 52-year-old female, presented with a 2-day history of sudden onset of severe hematuria and dysuria. Basic information: Female, 52 years old, acute illness duration of 2 days. Core symptoms: Gross hematuria, dysuria, urinary frequency, and urgency. Medical history: "The patient is taking cyclophosphamide (CTX) mg / day for rheumatoid arthritis for 3 weeks. Two days ago, she suddenly developed massive gross hematuria, accompanied by painful urination, without blood clots or fever." Ancillary examinations: Urinalysis showed numerous red blood cells and 2-4 white blood cells / HP; urine culture was negative; blood routine was normal; renal function was normal; cystoscopy description: "Cystoscopy revealed diffuse hyperemia and edema of the bladder mucosa, significantly increased mucosal fragility, easy bleeding upon touch, and no ulcers or neoplasms." System operation process: Feature quantification unit: Identifies "cyclophosphamide", "CTX", and "3 weeks of use", and activates cyclophosphamide_history:True; processes cystoscopy descriptions to extract "diffuse congestion", "edema", "increased mucosal fragility", and "easy bleeding", and converts them into corresponding labels.

[0125] Rule Engine: First layer excludes infection (negative urine culture, no fever, normal white blood cell count). Second layer rapid screening of medical history and precipitating factors: matched "cyclophosphamide," quickly anchoring chemical / hemorrhagic cystitis. Third layer precise matching of features with endoscopy: identified "diffuse congestion," "edema," and "increased mucosal fragility," highly supporting chemical cystitis. Rule Engine output: Chemical / hemorrhagic cystitis is the primary candidate (rule score 0.90).

[0126] LLM reasoning module: LLM confirms the diagnosis of chemical cystitis based on clues such as "history of cyclophosphamide use," "sudden onset of gross hematuria," and "diffuse congestion and edema," while also suggesting differentiation from radiation cystitis (but without a history of radiotherapy), and pointing out that hemorrhagic cystitis is a severe manifestation of chemical cystitis. LLM output ranking: chemical cystitis (probability 75.6%), hemorrhagic cystitis (probability 18.2%, as a stratification of severity).

[0127] Typical case database search: Using "cyclophosphamide", "diffuse congestion" and "sudden hematuria" as criteria, the top-5 cases returned included 4 cases of chemical cystitis and 1 case of hemorrhagic cystitis (with a higher dose of CTX).

[0128] Three-component collaborative adjudication: all three are highly consistent. Softmax normalization: chemical cystitis 76.8%, hemorrhagic cystitis 17.5%, radiation cystitis 3.1%.

[0129] Typical case database search: Using "cyclophosphamide", "diffuse congestion" and "sudden hematuria" as criteria, the top-5 cases returned included 4 cases of chemical cystitis and 1 case of hemorrhagic cystitis (with a higher dose of CTX).

[0130] Three-component collaborative adjudication: all three are highly consistent. Softmax normalization: chemical cystitis 76.8%, hemorrhagic cystitis 17.5%, radiation cystitis 3.1%.

[0131] Misconception Warning Unit: If the doctor only prescribes hemostatic drugs without addressing the CTX medication, a misconception warning is triggered: "Hemorrhagic cystitis is caused by cyclophosphamide. If only symptomatic hemostasis is given without clarifying the cause and adjusting the medication, it is recommended to immediately discontinue or change CTX and hydrate and alkalize the urine." Treatment decision-making unit: Providing primary care treatment plan (immediately recommending discontinuation of cyclophosphamide and contacting the prescribing physician to adjust the immunosuppressive regimen, large-volume hydration, urine alkalization, and preparation for bladder irrigation); while also noting: "Hemorrhagic cystitis can be caused by cyclophosphamide. Although hydration and alkalization can be performed first, severe bleeding related to CTX requires bladder irrigation and hyperbaric oxygen assessment at a higher-level hospital. Referral is recommended." Example 10: Identification of an extremely rare disease, toxoplasmosis-induced cystitis The patient, a 34-year-old male, presented with a 6-month history of chronic cystitis symptoms, unresponsive to multiple treatments. Basic information: Male, 34 years old, 6-month course of illness. Core symptoms: Urinary frequency, urgency, dysuria, and lower abdominal discomfort. Medical history: "The patient has experienced recurrent urinary frequency, urgency, and dysuria for 6 months, with multiple negative urine cultures. Multiple antibiotics (cephalosporins, quinolones, fosfomycin) have been used without effect. The patient is HIV-infected, with a CD4 count of 120 cells / μL." Ancillary examinations: Urinalysis showed 5-10 white blood cells / HP and a small number of red blood cells; multiple negative urine cultures; blood tests showed lymphopenia; renal function was normal; cystoscopy description: "Diffuse hyperemia and edema of the bladder mucosa, with scattered shallow ulcers, and no specific plaques or nodules." System operation process: Feature quantification unit: When "HIV", "CD4120", and "immunodeficiency" are detected, the immunocompromised:True tag is activated; when "multiple urine cultures are negative" and "multiple antibiotics are ineffective" are detected, the antibiotics_ineffective:True and multiple_cultures_negative:True tags are activated.

[0132] Rule Engine: First layer excludes infection (multiple negative urine cultures). Second layer rapid screening of medical history and precipitating factors: no common precipitating factors were matched. Third layer precise matching of features and endoscopy: no specific endoscopic features (no Hunter's ulcer, no sand-like plaques, etc.), the rule engine outputs "to be identified", listing toxoplasmosis cystitis as a very low probability candidate (very low priority, only retained in the context of immunodeficiency).

[0133] The LLM inference module analyzes clues such as "immunodeficiency (HIV, low CD4)," "refractory cystitis," "negative culture," and "antibiotic ineffectiveness." It concludes that after ruling out all common and other rare diseases mentioned above, refractory cystitis in immunodeficient patients should raise suspicion of rare infections, including toxoplasmosis, tuberculosis, and fungal infections. LLM lists toxoplasmic cystitis as the last candidate (probability 3.2%), but specifically generates a warning: "The patient has severe immunodeficiency. After ruling out all common and other rare bladder diseases mentioned above, rare pathogen infection (such as toxoplasmosis or tuberculosis) should be considered. Tissue PCR testing and special staining are recommended." The LLM inference module analyzes clues such as "immunodeficiency (HIV, low CD4)," "refractory cystitis," "negative culture," and "antibiotic ineffectiveness." It concludes that after ruling out all common and other rare diseases mentioned above, refractory cystitis in immunodeficient patients should raise suspicion of rare infections, including toxoplasmosis, tuberculosis, and fungal infections. LLM lists toxoplasmic cystitis as the last candidate (probability 3.2%), but specifically generates a warning: "The patient has severe immunodeficiency. After ruling out all common and other rare bladder diseases mentioned above, rare pathogen infection (such as toxoplasmosis or tuberculosis) should be considered. Tissue PCR testing and special staining are recommended." Typical case database search: Due to the extreme rarity of toxoplasmic cystitis, there may be no directly matching cases in the case database. Similar cases returned by the search include "immune deficiency-associated cystitis (fungal)," etc.

[0134] The three-component collaborative decision-making process: Both the rule engine and LLM listed toxoplasmic cystitis as a low-probability candidate, but LLM raised a rare disease warning in the context of immunodeficiency. Following the "rule priority" principle, the system did not include toxoplasmic cystitis in the top three candidate diagnoses, but added the LLM warning as a "special situation alert" to the end of the report: "The patient has severe immunodeficiency (HIV, CD4 <). After ruling out common infections and the above 13 rare bladder diseases, it is recommended to consider the possibility of infection by rare pathogens (Toxoplasma gondii, Mycobacterium tuberculosis, etc.), and tissue PCR and special etiological examinations are required." Misconception Warning Unit: Triggering the "Antibiotic Abuse" warning (long-term use of broad-spectrum antibiotics despite multiple negative cultures). Triggering the "Insufficient Examination" warning: "This case has an immunodeficiency background, and routine bacterial cultures are negative. Special etiological examinations (fungal smears, acid-fast staining, tissue PCR) are required to determine the cause." Treatment Decision Unit: Provides recommendations for primary care (discontinue unnecessary antibiotics, maintain hydration, and recommend adjustment of antiviral therapy by HIV specialists); Recommendation for mandatory referral: "This case has an immunodeficiency background and refractory cystitis, making diagnosis extremely difficult and far beyond the capacity of primary care. It is strongly recommended that the patient be immediately referred to the infectious disease department and urology department of a higher-level hospital for joint diagnosis and treatment, cystoscopic biopsy, and special etiological testing." Treatment Decision Unit: Provides recommendations for primary care (discontinue unnecessary antibiotics, maintain hydration, and recommend adjustment of antiviral therapy by HIV specialists); Recommendation for mandatory referral: "This case has an immunodeficiency background and refractory cystitis, making diagnosis extremely difficult and far beyond the capacity of primary care. It is strongly recommended that the patient be immediately referred to the infectious disease department and urology department of a higher-level hospital for joint diagnosis and treatment, cystoscopic biopsy, and special etiological testing." System operation process summary Based on the above embodiments, the standard operating procedure of the system of the present invention can be summarized as follows: Step S1: Acquire the patient's basic information, core symptoms, auxiliary examination results, and medical history text data through the data acquisition unit. The auxiliary examination results include at least urinalysis, urine culture, complete blood count, renal function tests, and cystoscopy descriptions.

[0135] Step S2: The feature quantization unit uses natural language processing and medical named entity recognition technology to transform medical history text data and unstructured cystoscopy descriptions into structured feature vectors, including medical history labels, laboratory indicator Boolean values, and standardized labels for endoscopic features.

[0136] Step S3: The AI ​​inference unit initiates inference. The rule engine first performs hard clinical logic triage based on a hierarchical decision tree: the first layer, infection triage layer, excludes acute bacterial cystitis based on urine culture, urine leukocytes / nitrite, and antibiotic efficacy; the second layer, rapid screening layer for medical history and causes, scans specific high-weight medical history and causes to anchor rare disease categories; the third layer, precise matching layer of features and endoscopy, performs detailed verification based on specific laboratory indicators and cystoscopic morphological features.

[0137] Step S4: The large language model reasoning module receives all patient information and clues output by the rule engine, performs deep discrimination reasoning, and outputs a list of diseases sorted by probability and supporting / excluding evidence.

[0138] Step S5: The typical case database performs cosine similarity retrieval based on the main candidate diagnoses and key features output by LLM, and returns Top-K similar cases to verify the rationality of the reasoning.

[0139] Step S5: The typical case database performs cosine similarity retrieval based on the main candidate diagnoses and key features output by LLM, and returns Top-K similar cases to verify the rationality of the reasoning.

[0140] Step S6: The three-component collaborative adjudication module performs collaborative adjudication according to the principle of "rule priority, intelligent supplementation, and empirical verification." When there is a conflict in the output between components, the conflict is resolved according to the priority principle of "rule as the basis, LLM as the main body, and case as the auxiliary body." If necessary, a confidence weighted algorithm (rule engine weight 0.6, LLM weight 0.3, and case database weight 0.1) is used to calculate the total confidence, or the "safety first" circuit breaker mechanism is triggered to place the suspected serious consequences disease at the top of the warning list.

[0141] Step S7: The AI ​​inference unit uses a mechanism combining "dynamic weighted scoring" and "gold standard one-vote triggering" to form the final probability ranking. Each candidate disease starts with a score of 0. Matching supporting features adds points, while matching excluding features resets the score to zero or deducts a significant amount. After Softmax normalization, the scores are converted into probability percentages.

[0142] Step S8: When doctors save diagnostic conclusions, prescribe prescriptions, or formulate treatment plans, the misconception warning unit compares the results with the preset misconception rule base in real time to identify misconception types such as antibiotic abuse, overtreatment, risk of missed diagnosis, delayed referral, and insufficient examination, and generates non-blocking warning prompts.

[0143] Step S9: Based on the diagnostic results, the treatment decision unit outputs standardized lifestyle guidance, symptom relief plans, and follow-up visit nodes through the primary care prompt generation module; and immediately triggers the referral template by identifying situations such as persistent gross hematuria, upper urinary tract hydronephrosis, suspected malignant lesions on ultrasound, significantly reduced bladder capacity, need for biopsy for diagnosis but cannot be performed locally, and unclear diagnosis but symptoms persisting above the threshold.

[0144] Step S10: The result output unit integrates all the above outputs to generate and display a decision report that includes candidate disease ranking, diagnostic basis, risk warning, primary care treatment plan and referral suggestions.

[0145] Through the above specific implementation methods, this invention achieves accurate AI-assisted differential diagnosis of 13 rare bladder diseases. By combining the three components of a rule engine, a large language model, and a typical case database, along with dynamic weighted scoring, conflict resolution, misconception warning, and primary care decision support, it effectively solves the problem of rare diseases in primary healthcare institutions.

[0146] The above embodiments merely illustrate several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

[0147] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A rare bladder disease AI-assisted diagnosis decision system, characterized in that, include: The data acquisition unit is configured to acquire the patient's basic information, core symptoms, auxiliary examination results and medical history text data, wherein the auxiliary examination results include at least urinalysis, urine culture, blood routine, renal function and cystoscopy description; The feature quantization unit, connected to the data acquisition unit, is configured to convert the medical history text data and unstructured cystoscopy description into AI-recognizable structured feature vectors through natural language processing and medical named entity recognition technology. The structured feature vectors include medical history labels, laboratory indicator Boolean values, and standardized labels for endoscopic features. The AI ​​inference unit, connected to the feature quantization unit, includes a rule engine, a large language model inference module, a typical case database, and a three-component collaborative adjudication module. The rule engine is configured to perform hard clinical logic triage based on a preset rule base for the identification of rare bladder diseases, the rule base containing a hierarchical decision tree for rare bladder diseases; The large language model reasoning module is configured to perform semantic understanding, probabilistic reasoning, and differential diagnosis ranking on unstructured medical record texts. The typical case database is configured to store structured feature vectors of confirmed cases and provide empirical comparisons based on similarity calculations; The three-component collaborative adjudication module is configured to collaboratively adjudicate and resolve conflicts in the outputs of the rule engine, the large language model inference module, and the typical case database according to a preset priority principle, and output candidate disease probability ranking, supporting evidence, excluded evidence, and high-priority warnings. The error warning unit, connected to the AI ​​inference unit, is configured to identify common errors in clinical decision-making and generate non-blocking warning prompts in real time when doctors save diagnostic conclusions, prescribe prescriptions, or formulate treatment plans, based on a preset error rule base. The diagnosis and treatment decision unit, connected to the AI ​​reasoning unit and the error warning unit, is configured to output executable treatment plans and mandatory referral indications based on the diagnosis results and preset primary care treatment rule templates and referral high-risk feature database. The output unit is configured to integrate the outputs of the AI ​​inference unit, the error warning unit, and the diagnosis and treatment decision unit to generate and display a decision report that includes candidate disease ranking, diagnostic basis, risk warning, primary care treatment plan, and referral suggestions.

2. The system of claim 1, wherein, The rare bladder diseases mentioned include at least one of the following: radiation cystitis, chemical cystitis, hemorrhagic cystitis, toxoplasmosis cystitis, schistosomiasis cystitis, leathery cystitis, eosinophilic cystitis, glandular cystitis, interstitial cystitis, bladder leukoplakia, bladder endometriosis, bladder amyloidosis, and pelvic lipomatosis.

3. The system of claim 1, wherein, The calling relationship and workflow of the three-component collaborative adjudication module follow a sequential process of "rule priority, intelligent supplementation, and empirical verification": In the first round, the rule engine performs hard rule matching and triage based on structured data, and outputs preliminary diagnostic clues or instructions to be identified; In the next round, the large language model reasoning module receives all patient information and clues output by the rule engine, performs deep identification reasoning, and outputs a list of diseases sorted by probability. Finally, the typical case database is used to search for similar cases based on the main candidate diagnoses and their key features output by the large language model reasoning module, in order to verify the rationality of the reasoning or indicate the possibility of rarity.

4. The system of claim 3, wherein, The conflict resolution mechanism of the three-component collaborative adjudication module includes: When the rule engine and the large language model inference module output conflict, the rule engine result is written as the primary diagnosis in the main diagnosis field, and the high-risk objection raised by the large language model is encapsulated as a high-priority warning field. When the large language model inference module conflicts with the output of the typical case database, the disease ranking of the large language model is maintained, and the differences found in the case database are used as key identification prompts. When the outputs of the three components are inconsistent, the internal confidence of each component is calculated and weighted and fused. At the same time, following the "safety first" circuit breaker mechanism, the suspicion pointing to a disease with serious consequences or characteristics of emergency intervention is used as the highest level warning item.

5. The system according to claim 4, characterized in that, The weighted fusion employs a confidence-weighted algorithm, with the rule engine having a weight of 0.6, the large language model inference module having a weight of 0.3, and the typical case database having a weight of 0.1; the total confidence score is calculated using the following formula: Total confidence level = (0.6 × rule score) + (0.3 × LLM score) + (0.1 × case matching degree); When all three are inconsistent and the large language model inference module or typical case database points to a disease with serious consequences, the system ignores the regular weights and directly places the suspected disease at the top of the report as the highest level warning item.

6. The system according to claim 1, characterized in that, The hierarchical decision tree of the rule engine adopts a multi-level decision structure: The first layer is the infection triage layer, which prioritizes excluding acute bacterial cystitis based on urine culture results, urine leukocyte / nitrite index, and antibiotic efficacy data. The second layer is a rapid screening layer for medical history and triggers. Under the premise of non-infection, it scans for specific high-weight medical history and triggers, including history of pelvic radiotherapy, history of chemotherapy drug use, history of contact with epidemic areas, history of allergies, and menstrual-related history, in order to quickly anchor specific rare disease categories. The third layer is the feature and endoscopy precision matching layer, which is subdivided and verified based on specific laboratory indicators and cystoscopic morphological features. The cystoscopic morphological features include Huntner's ulcer, glomerulation, sand-like plaques, white plaques, follicular hyperplasia, villous hyperplasia, bluish-purple nodules, and external pressure changes.

7. The system according to claim 6, characterized in that, The probability ranking of the AI ​​inference unit adopts a mechanism combining "dynamic weighted scoring" and "gold standard one-vote triggering": Each candidate disease starts with a score of 0. If a supporting feature is matched, the score is added according to a preset weight. If an exclusion feature is matched, the score is reset to zero or significantly deducted. When pathological or endoscopic gold standard features are captured, including positive Congo red staining, discovery of parasite eggs, or histological evidence for diagnosis, they are given the highest priority and directly become the first candidate diagnosis. The total score for each disease is converted into a probability percentage after Softmax normalization, and the final candidate diseases are generated by ranking them according to their probability.

8. The system according to claim 1, characterized in that, The natural language processing in the feature quantization unit includes: By using a specialized fine-tuned large language model combined with medical named entity recognition technology, word segmentation and entity locking were performed on the cystoscopy description text, and site features and pathological modifiers were extracted. The extracted unstructured phrases are matched with a pre-defined standard dictionary of rare cystoscopy features using vector similarity. After identifying specific words, they are automatically converted into discrete, standardized Boolean labels, which are then used as calculation factors to input into the multi-level decision tree of the AI ​​inference unit.

9. The system according to claim 1, characterized in that, The error warning unit focuses on identifying the types of errors and triggering conditions, including: Antibiotic abuse: Triggered when a urine culture is negative, there are no signs of systemic infection, and there is a prescription for antibiotics; Overtreatment: Triggered when the AI ​​diagnosis ranking shows a high probability of benign or precancerous lesions and the treatment recommendation is major surgery, while most similar cases are treated conservatively; Risk of missed diagnosis: This is triggered when the system identifies specific keywords in the cystoscopy description but the doctor's initial diagnosis does not include the corresponding disease; Delayed referral: Triggered when the AI ​​candidate diagnosis includes a disease that requires in-depth specialist treatment or multidisciplinary collaboration, and there is no recent referral or consultation plan in the medical record; Insufficient examination: Triggered when the AI ​​has a high probability of diagnosing a disease that requires a biopsy for confirmation, and there is no corresponding biopsy scheduled in the examination plan.

10. The system according to claim 1, characterized in that, The diagnostic and treatment decision-making unit includes: The primary care intervention prompt generation module is configured to call a fixed rule template to output standardized lifestyle guidance, symptomatic relief plan and follow-up visit nodes when the primary disease falls within the scope of initial intervention at the primary level. The mandatory referral red line module is configured to immediately trigger the referral template and highlight the referral suggestion when the following conditions are identified: persistent gross hematuria, hydronephrosis of the upper urinary tract, suspected malignancy on ultrasound, significantly reduced bladder capacity, need for biopsy for diagnosis but local hospitals cannot perform the corresponding examination, or unclear diagnosis but symptoms continue to exceed the preset threshold.

Images