Rare disease-oriented brain-like auxiliary diagnosis and treatment method

Abstract

A brain-like auxiliary diagnosis and treatment method for rare diseases is realized based on a GPU or NPU server, a system review agent is used as a sensing front end of the method, is connected with an external medical database through a standard API interface, executes automatic retrieval and quality preliminary screening of literatures, and outputs an original literature set; the document processing agent is used for receiving an original document set of the system review agent, performing structured extraction on an unstructured PDF document by utilizing a deep analysis tool, and generating a structured knowledge base containing a core clinical field; and the NHSC brain-like agent adopts a brain-like architecture based on a dual processing theory, receives the structured knowledge base and the patient data of the document processing agent, and outputs a final dynamic diagnosis and treatment suggestion through fast and slow dual-channel cooperative processing. Algorithm evaluation of evidence quality is realized, an evidence screening mechanism is enhanced from the source, risks caused by illusion of a large language model are effectively reduced, and stability and reliability of an evidence synthesis result are guaranteed.

Description

Technical Field

[0001] This invention relates to the fields of automation of natural history comparison (NHSC) methods and literature review technology, specifically designing a brain-like auxiliary diagnosis and treatment method for rare diseases. Background Technology

[0002] In the development of intelligent medical diagnostic systems, especially in the field of rare diseases where diagnostic data is scarce and pathological mechanisms are complex, building interpretable and reliable auxiliary diagnostic models faces two core challenges: first, in the clinical trial design phase, how to determine a feasible and scientifically rigorous sample size plan for rare diseases given a limited patient population; and second, how to efficiently and accurately integrate high-quality evidence from the massive and ever-growing body of medical literature. Existing research still has significant shortcomings in these two key areas.

[0003] First, traditional methods for estimating sample size in rare disease clinical trials have inherent limitations. Currently widely used sample size calculations primarily rely on power analysis, but this method faces two prominent problems: first, the calculation process often relies on point estimations of unknown efficacy parameters, while actual efficacy typically has significant uncertainty, resulting in insufficient stability and reliability of the obtained sample size results; second, this method fails to adequately incorporate realistic constraints such as the population size corresponding to the disease and the actual number of patients that can be recruited. For rare diseases with a limited patient base, the sample size calculated using this method often far exceeds the actual recruitment capacity, directly restricting the feasibility of clinical trials. To address these issues, the rare disease clinical trial sample size estimation method "CN202410356198" based on a decision theory model introduces a multi-decision-maker payoff function, transforming the sample size estimation problem into an optimization problem that simultaneously considers real-world population constraints. This achieves an algorithmic expression of the core idea of ​​Natural History Comparison (NHSC), representing a significant improvement over traditional methods. However, this method has not yet incorporated a crucial but often overlooked factor in medical practice during model building: the "time weight"—the diminishing impact of evidence and treatment benefits as technology evolves and time progresses. This deficiency limits the model's ability to characterize the long-term natural history of diseases and the dynamic medical environment. In precision-driven medical research, such biases can have a significant impact on final decisions.

[0004] Secondly, in the automated synthesis of literature evidence supporting medical decisions, existing technologies also face a trade-off between efficiency and quality. Before the widespread application of large language models, systematic medical literature reviews mainly relied on manual work by domain experts, which was not only costly and time-consuming, but also often resulted in delayed evidence updates. This timeliness is particularly prominent in the rapidly evolving field of medicine. The literature review generation method "CN202311491249" based on large language models significantly improves the efficiency of evidence synthesis through automated retrieval, information extraction, and content integration, providing a new technical path to solve the problem of delayed evidence updates. However, this type of method relies solely on a relatively simple scoring mechanism to screen the generated results, often making it difficult to deeply assess the inherent quality bias of the included literature, and thus failing to effectively suppress the inherent "illusion" risk of large language models. This uncertainty directly affects the reliability of the evidence conclusions, thereby limiting its direct application in demanding medical decision-making scenarios.

[0005] Based on the above analysis, there is an urgent need for an innovative solution that can deeply integrate cutting-edge artificial intelligence technology and realize the intelligent and dynamic operation of the entire process of the Natural History Comparison (NHSC) method. Summary of the Invention

[0006] To systematically address the aforementioned bottlenecks, this invention proposes a brain-inspired auxiliary diagnostic and treatment method for rare diseases. On one hand, by introducing the TWES (Time-Weighted Evidence Synthesis) algorithm, dynamic attenuation weighting is applied to multi-source, cross-period medical evidence, enabling it to more realistically reflect the natural history of diseases and the impact of the evolution of medical knowledge over time in intelligent NHSC models such as rare disease sample size estimation, thereby improving the timeliness and clinical relevance of the calculation results. On the other hand, during the automated literature review generation process, mature and rigorous quality bias assessment tools such as ROB2 (Risk Of Bias 2) are deeply integrated to achieve algorithmic evaluation of evidence quality, strengthening the evidence screening mechanism from the source, effectively reducing the risks brought by the "illusion" of large language models, and ensuring the stability and reliability of the evidence synthesis results.

[0007] To achieve the above objectives, the present invention employs the following technical solution: a brain-like assisted diagnosis and treatment method for rare diseases, implemented on a GPU or NPU server, comprising: a system review agent, used as the perception front end of the method, connecting to an external medical database through a standard API interface, performing automated literature retrieval and initial quality screening, and outputting a set of original literature; and a document processing agent, used to receive the set of original literature from the system review agent, and using deep parsing tools to extract structured data from unstructured PDF documents, generating a structured knowledge base containing core clinical fields. The NHSC brain-inspired intelligent agent adopts a brain-inspired architecture based on the dual-process theory. It receives a structured knowledge base of a document processing intelligent agent and patient data, and outputs the final dynamic diagnosis and treatment suggestions through fast and slow dual-pathway collaborative processing.

[0008] The system review agent executes a cascaded processing logic of multi-level funnel screening and multi-dimensional quantitative evaluation, specifically including: a first-level screening unit, used to filter the document titles for keywords through regular expression matching; The second-level screening unit is used to perform relevance analysis on document abstracts using a semantic model; The third-level screening unit performs standard compliance checks on the full text of the literature. Using the ROB2 evaluation engine, it receives the literature after the three-level screening and performs in-depth reviews in five dimensions in parallel according to the Cochrane bias risk assessment criteria: (D1) Review of the randomization process to assess grouping randomness and sequence concealment; (D2) Deviance from expected intervention review, analyzing intervention deviations caused by improper implementation; (D3) Review missing results data to check the completeness of follow-up data; (D4) Results measurement review, assessing the objectivity of the measurement method and the bias of the measurer; (D5) Selective reporting review to verify cases of concealment of negative results; The ROB2 evaluation engine calculates an evidence quality score of 0-20 based on the above five dimensions, and automatically eliminates low-quality studies by setting a threshold.

[0009] The document processing agent is specifically configured as follows: Integrates the MinerU deep parsing tool to identify and parse multi-column layouts, cross-page tables, and non-standardized charts in medical literature PDFs; By calling the large language model, core clinical fields such as age of onset, drug dosage, survival rate and genetic pattern are automatically extracted for each parsed document; The extracted fields are converted into vector data and stored in a vector database for downstream use.

[0010] The NHSC brain-like intelligent agent adopts a two-layer architecture that simulates the S1 intuition pathway and the S2 reasoning pathway of the human brain. Specifically, it includes a BioBERT embedding layer, which is used to convert the input patient symptom descriptions, gene sequences and phenotypic data into high-dimensional vector representations. The S1 intuitive pathway, used for rapid response tasks, employs a spiking neural network and simulates nerve impulses using a LIF neuron model to achieve millisecond-level rapid identification of common disease patterns. Its neuronal membrane potential dynamics equation is: (1) Where τ is the membrane time constant, V(t) is the membrane potential, and V rest V(t) is the resting potential, and I(t) is the input current; when V(t) exceeds the threshold V... th Pulse is emitted at specific times; The S2 reasoning pathway is used to handle complex and difficult tasks. It adopts a hierarchical reasoning model, which includes a high-level H layer responsible for abstract diagnostic planning and a low-level L layer responsible for detailed symptom analysis. The two layers interact with each other through a cross-attention mechanism. The S2 reasoning pathway integrates six sub-agents: natural history comparison, historical comparison, time series analysis, causal assessment, quality assessment, and evidence synthesis. A dual-path arbiter is used for calculations based on entropy uncertainty, dynamically determining whether to enable the S1 intuitive path to output direct results or activate the S2 reasoning path for deep reasoning.

[0011] The evidence synthesis sub-agent in the S2 reasoning pathway employs a time-weighted evidence synthesis algorithm to calculate the final synthesis weight ω for each piece of evidence. final Its calculation logic is as follows: parallel calculation time weight ω time With quality weight ω quality， The calculation formula is: (2) (3) Where, ω time Represents the time weight, ω quality The quality weight is represented by Δt, the publication duration of the literature, and λ, which is the decay coefficient. GRADE The quality score is based on GRADE grading; the final weight is obtained through nonlinear coupling and normalization, and the calculation formula is as follows: (4) This enables dynamic evidence fusion based on both timeliness and authority.

[0012] The advantages of this invention are as follows: First, this invention achieves a highly efficient automated process in the SR Agent and DP Agent modules. By combining three-level funnel screening, ROB2 automated scoring, and complex PDF parsing technology based on deep learning, this invention can shorten the traditional manual review process, which requires several weeks, to minutes. It also successfully overcomes the challenges of multi-column layout and cross-page table extraction in medical literature, ensuring the high quality and high structure of the data source.

[0013] Secondly, this invention innovatively constructs a brain-like cognitive architecture based on a dual pathway of "S1 intuition-S2 reasoning," a core capability not found in traditional medical aids. This method does not merely mechanically aggregate data, but simulates the working mode of the brains of top human medical experts: utilizing the S1 pathway to rapidly process routine evidence, using the S2 pathway to deeply analyze complex and difficult evidence, and achieving self-consolidation and incremental learning of experience through the hippocampal memory mechanism. This architecture endows the method with "metacognitive" capabilities, enabling it to autonomously assess the uncertainty (entropy) of evidence and flexibly switch between intuitive judgment and deep reasoning. Therefore, when dealing with highly complex and heterogeneous problems such as rare diseases, it provides more robust and intelligent decision support than traditional rule engines.

[0014] Third, this invention proposes a theoretically sound NHSC natural history synthetic control and Bradford-Hill causal quantitative assessment system, fundamentally addressing the lack of "gold standard evidence" in the field of rare diseases. Addressing the difficulty of conducting large-scale RCTs for rare diseases, this invention utilizes mathematical models (PSM matching, exponential decay trajectory prediction) to reconstruct a scientific "virtual control group" from historical data and calculates key clinical indicators such as RRR, ARR, and NNT, elevating case reports, which would otherwise only serve as low-level evidence, to high-quality evidence with causal inference value. Combined with the Bradford-Hill nine-dimensional quantitative scoring system, this method can rigorously examine each chain of evidence like a judge, quantifying its causal strength, thereby providing clinicians with sound diagnostic and treatment recommendations and filling the gap in evidence-based medicine for rare diseases.

[0015] Fourth, this invention revolutionizes the timeliness of evidence synthesis through the TWES time-weighted algorithm. Unlike traditional meta-analysis that treats old and new evidence equally, this invention introduces a time decay factor, assigning extremely high decision weight to the latest high-quality evidence through a non-linear coupling mechanism. This simulates the sensitivity of human experts to cutting-edge medical advancements, ensuring that the diagnostic and treatment recommendations output by this method are always synchronized with the latest global research, avoiding misdiagnosis caused by citing outdated conclusions. Attached Figure Description

[0016] The accompanying drawings, which form part of this application, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0017] Figure 1 This is a flowchart of the brain-like auxiliary diagnosis and treatment method for rare diseases of the present invention; Figure 2 This is a schematic diagram of the process of the intelligent agent in the system overview of the present invention; Figure 3This is a schematic diagram of the document processing intelligent agent of the present invention; Figure 4 This is a schematic diagram of the process of the NHSC brain-like intelligent agent of the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0019] A brain-inspired auxiliary diagnostic and treatment method for rare diseases is proposed. This method utilizes a multi-agent collaborative architecture to automatically perform information retrieval, data extraction, quality assessment, and time-weighted evidence synthesis from massive amounts of unstructured medical literature. This further enables the construction of systematic reviews and support for diagnostic and treatment decisions in the field of evidence-based medicine. The multi-agent collaborative architecture adopts a modular and decoupled design. Each agent communicates and collaborates through standard interfaces and adapters, forming a complete technical closed loop of "literature retrieval - document processing - evidence synthesis." Specifically, it includes the following three core modules and their internal hierarchical structure: The first core module is the System Review Agent (SR Agent), which adopts a three-level funnel screening architecture. Each level is connected in a serial pipeline, including: the first-level screening unit, which connects to the world's mainstream medical databases through API interface, executes the title exact matching strategy, and the matching rule is to perform a string exact match between the rare disease keywords entered by the user and the database title field or to calculate the cosine similarity based on BioBERT semantic embedding. The similarity threshold is set to 0.85, and the set of literature related to the title is selected as L1.

[0020] The secondary screening unit performs semantic relevance analysis on the abstracts of documents in L1. The BioBERT model is used to encode the abstract text into a 768-dimensional vector, and the cosine similarity between the query vector q and the abstract vector a is calculated using the following formula: (5) The retained documents with a sim value ≥ 0.75 form the abstract-related document set L2; Where sim(q, a) represents the cosine similarity between the query vector q and the abstract vector a, q represents the numerical vector converted from the user-input query keywords, a represents the numerical vector converted from the document abstract text, "·" represents the dot product operation of the vectors, and "||" represents the magnitude of the vectors, used for normalization.

[0021] The three-level screening unit performs full-text standard compliance verification on the literature in L2. Verification dimensions include: D1 study type determination (priority weights are RCT 1.0, cohort study 0.7, and case report 0.3 respectively); D2 sample size adequacy test; D3 statistical method compliance review; D4 ethical approval completeness check; and D5 data usability assessment. A weighted scoring mechanism is used for these five dimensions, with a comprehensive score threshold set at S≥0.65. The three-level screening unit also performs ROB2 automated quality scoring, strictly adhering to the Cochrane Risk of Bias 2 standard. Parallel quantitative assessment is conducted across five core dimensions: D1 randomization process assessment; D2 deviation from expected intervention assessment; D3 missing outcome data assessment; D4 outcome measurement assessment; and D5 selective reporting assessment. Each dimension is scored from 0 to 4. This method sets a quality threshold τ=12, where τ represents the quality threshold, automatically eliminating low-quality studies and outputting a high-quality evidence set. .

[0022] The second core module is the document processing intelligent agent (DP Agent). This module adopts a two-stage pipeline architecture, including: the first stage PDF complex format parsing unit, which integrates the MinerU tool to specifically parse complex PDF format medical documents such as multi-column layout, cross-page tables and non-standard charts, breaking through the format processing limitations of traditional natural language processing tools. The parsing process includes document layout detection, separation of text blocks / table blocks / image blocks, cross-page table splicing and restoration and OCR text recognition. The parsing result output is structured data containing content type, extracted text and bounding box coordinates. The second phase, the precise extraction unit of clinical fields, uses a large language model to accurately extract core clinical fields from the parsed literature. The extracted fields include age of onset (numerical range or median), drug dosage (drug name + dosage + frequency), survival rate (time point + percentage + confidence interval), and genetic pattern (inheritance mode + gene name + mutation site). The extraction process adopts a few-sample learning paradigm and undergoes numerical rationality checks and unit standardization. The extracted results are stored in a vector knowledge base for structured storage and management. Each document is stored as an independent object, containing metadata and clinical fields. BioBERT encoding is used to generate 768-dimensional vectors for semantic retrieval. The indexing strategy is constructed using the HNSW algorithm, which supports hybrid queries based on keywords and vector similarity.

[0023] The third core module is the NHSC brain-inspired intelligent agent (NHSC Agent). This module is the core embodiment of the brain-inspired architecture and has a built-in "intuition-reasoning" dual-pathway cognitive architecture (S1-S2 Dual Pathway). It simulates the cognitive decision-making process of human medical experts when dealing with complex and rare disease evidence, and specifically includes the following deep logical architecture: The first layer involves dual cognitive processing using the S1 intuitive pathway and the S2 reasoning pathway. The S1 intuitive pathway simulates the rapid processing mechanism of the human brain's visual pathway, used to quickly filter common evidence with high confidence. Its core employs Poisson pulse coding to convert continuous evidence vectors into time-series pulse signals, and calculates the pulse firing probability using the Sigmoid function. The calculation formula is as follows: (6) Where `rate` represents the probability of pulse firing, `x` represents the input stimulus intensity or accumulated membrane potential value, and `exp` represents an exponential function with the natural constant `e` as the base. An integral firing mechanism is used based on the LIF neuron dynamics model; when the membrane potential exceeds a threshold, a pulse is triggered, and a preliminary judgment result is output. The S2 inference pathway simulates the deep thinking mechanism of the prefrontal cortex to handle complex evidence with high entropy and high uncertainty. It employs a Transformer self-attention mechanism to capture long-distance dependencies and uses a Cross-Attention mechanism to achieve bidirectional interaction between the "abstract planning layer" and the "detailed analysis layer" until the Q-Learning decision network's judgment and inference converge, outputting a deep analysis conclusion. This method introduces a dual-path arbitrator, fusing the outputs of the two pathways based on Bayesian posterior probabilities. The calculation formula is as follows: (7) Where P(D|X) represents the posterior probability of the final diagnosis conclusion D given the input X. s1 P represents the diagnostic confidence / probability of the S1 intuitive pathway output. s2 ω1 represents the diagnostic confidence / probability output by the S2 inference pathway, ω2 represents the weighting of the S1 pathway, and ω2 represents the weighting of the S2 pathway. ω1 and ω2 are dynamically allocated based on the confidence of each pathway to ensure accuracy without sacrificing efficiency.

[0024] The second layer is the NHSC natural history synthetic control construction layer. This layer addresses the challenge of "lack of contemporaneous controls" in rare diseases by introducing a Historical Control Agent to construct a virtual control group based on natural disease course data. First, natural disease course data is retrieved from the knowledge base, and a virtual cohort is constructed using propensity score matching (PSM). The propensity score calculation formula is as follows: (8) Where e(X) represents the propensity score (i.e., the probability that an individual is assigned to a particular group), exp represents an exponential function with the natural constant e as the base, X represents a vector of covariates (such as age, gender, baseline disease status, etc.), β0 represents the intercept term of the logistic regression model, and β TThis represents the transpose of the regression coefficient vector. Furthermore, the net effect of treatment is calculated based on the NHSC criteria, with specific indicators including: The formula for calculating the relative risk reduction rate (RRR) is: (9) The formula for calculating the absolute risk reduction rate (ARR) is: ARR = R control - R treatment (10) The formula for calculating the number of people requiring treatment (NNT) is as follows: (11) In the above formulas (9), (10), and (11), RRR represents the relative risk reduction rate, ARR represents the absolute risk reduction rate, NNT represents the number of patients requiring treatment (i.e., the number of patients requiring treatment to prevent one adverse event), and R control R represents the risk / incidence of events in the control group (or natural history virtual control group). treatment This indicates the risk / incidence of events in the treatment group; This method further introduces a natural history agent to track disease progression, employing an exponential decay model. The calculation formula for the model is as follows: Function(t) = F0·exp(-βt) (12) Predict the trajectory of functional decline in the untreated state and compare the actual clinical outcome after treatment with the predicted trajectory to quantify the deviation brought about by treatment, so as to scientifically evaluate the true efficacy of the intervention measures. In formula (12), Function(t) represents the patient's functional state or physiological indicators at time t, F0 represents the patient's initial functional state (baseline level), β represents the functional decline rate constant of natural disease progression, and t represents time variables (such as disease duration).

[0025] The third layer is the Bradford-Hill Causality Strength Quantification Layer, which performs rigorous tests across nine dimensions through a Causality Assessment Agent to calculate the overall causal score. The scoring formula is as follows: (13) Where Score represents the overall confidence score (out of 100) for the causal relationship, S i S represents the actual score of the i-th dimension (out of 9 dimensions). imax denoted as the maximum possible score for the i-th dimension, and "9" represents the nine evaluation dimensions of the Bradford-Hill criterion.

[0026] Specific dimensions include: (a) association strength, calculating the RR or OR value and converting it into a logarithmic score, the calculation formula is: (14) Where S i The score represents the association strength dimension, where RR represents relative risk, log represents the logarithmic function, and min represents the minimum value function, used to limit the maximum score to 10 points; (b) Consistency, based on heterogeneity statistics, is calculated using the following formula: (15) Adjust the rating, of which I 2 (c) Heterogeneity index, reflecting the proportion of variation among studies to the total variation; Q represents Cochran's Q statistic, used to test heterogeneity; k represents the number of studies included in the analysis; (d) Specificity, assessing the correspondence between "single exposure and single outcome"; (e) Temporality, verifying the necessary condition that exposure precedes outcome; if not met, S4=0 and the total score is reset to zero; (f) Biological gradient, fitting a linear regression model, calculated using the following formula: Y = β0 + β1·XY (16) To test the dose-response relationship, the coefficient of determination r² is calculated as the scoring criterion. In equation (16), Y represents the dependent variable (outcome variable), X represents the independent variable (exposure dose or intervention intensity), β0 represents the intercept of the regression equation, and β1 represents the regression coefficient (slope), reflecting the strength of the dose-response relationship; (f) biological rationality, assessing whether the mechanism is known or supported by animal experiments; (g) consistency, checking the consistency with existing epidemiological and laboratory knowledge; (h) experimental evidence, assigning different weights according to the evidence levels such as RCT, quasi-experiment, and animal experiments; (i) analogy, calculating the semantic similarity between the current evidence and known causal relationships, using the following formula: (17) Where sim(A,B) represents the semantic similarity between current evidence A and the known causal relationship B, v A Let v represent the feature vector of evidence A. B Represents the eigenvectors of a known relation B; The fourth layer is the TWES time-weighted evidence synthesis layer, which proposes a nonlinear "time-quality" coupling algorithm to address the lag in traditional meta-analysis. First, the time weight is calculated using the following formula: ω time (i) = e -λ·Δt (2) Where ω time(i) represents the time weight of the i-th document, exp represents an exponential function with the natural constant e as the base, λ is the decay coefficient to ensure that the weight of new evidence increases exponentially, and Δt represents the time difference (i.e., "current time - publication time"); next, the quality weight is calculated using the following formula: (3) Where ω quality (i) represents the quality weight of the i-th document, SCORE GRADE This indicates the actual quality score (based on the GRADE grading standard) obtained by the document, SCORE. max This represents the theoretical maximum score under this scoring system; finally, the comprehensive weight is calculated through multiplicative coupling, and the calculation formula is: (4) Where ω final (i) represents the final overall weight of the i-th document; the ω of the numerator... time (i)·ω quality (i) indicates multiplicative coupling, meaning that a document must simultaneously possess both "newness" (high time weight) and "quality" (high quality weight) for the product to be large. If a document is very new but of poor quality, or of good quality but too old, the product will be lowered; the denominator "ω" time (j)·ω quality (j) indicates normalization. The product of (time × quality) of all included literature is summed as the denominator, where j represents the loop variable in the summation symbol, and N represents the total number of pieces of evidence involved in the synthesis. This method ultimately uses the weighted effect size after Softmax normalization to generate a dynamic review report containing confidence intervals, heterogeneity indicators, and recommendation strengths. Furthermore, this module incorporates a built-in hippocampal memory system for long-term storage and consolidation of experience. This method stores diagnostic cases as episodic memory vectors, calculated using the following formula: m = (q, r, t, context) (18) Where m represents the contextual memory vector, q represents the question or input, r represents the response or result, t represents the timestamp, and context represents the contextual information. A time decay weight is introduced, calculated using the following formula: ω = exp(-λ·(t now -t m (19) The mechanism of memory forgetting is simulated, where ω represents the retention weight of memory, exp represents an exponential function with the natural constant e as the base, and t... now Indicates the current time, t mThis indicates that it is based on the time when it was initially generated.

[0027] When the frequency (freq(m)) and accuracy (accuracy(m)) of a certain diagnostic pattern exceed a set threshold, the memory consolidation mechanism is triggered, converting short-term episodic memory into long-term semantic memory (LTM). The update formula is as follows: LTM new = LTM old + ŋ·Δm (20) This enables the method to continuously learn and evolve autonomously in handling rare disease diagnosis and treatment tasks, where LTM... new Represents the updated long-term semantic memory, LTM old ŋ represents the long-term semantic memory before the update, ŋ represents the learning rate (which controls the speed of new knowledge integration), and Δm represents the increment of short-term episodic memory to be solidified. The hardware support module of this invention relies on servers equipped with high-performance GPUs or NPUs to meet the computing power requirements of parallel inference of large language models and semantic retrieval of vector databases, ensuring efficient operation of the process. The core intelligent agent module includes the aforementioned System Review Agent (SR Agent), Document Processing Agent (DPAgent), and NHSC Brain-Inspired Agent (NHSC Agent). Each module interacts with data through a RESTful standard interface. SR Agent is responsible for the initial screening and scoring of high-quality documents, DP Agent is responsible for complex format parsing and structure extraction, and NHSC Brain-Inspired Agent is responsible for deep causal reasoning and evidence synthesis based on a brain-inspired architecture.

[0028] refer to Figures 1 to 4 A brain-like auxiliary diagnostic and treatment method for rare diseases; Example 1: Application of a Brain-Inspired Auxiliary Diagnostic Method for Rare Diseases in the Diagnosis and Treatment of Hemophilia A Complicated with Intraperitoneal Hemorrhage. This example illustrates a brain-inspired auxiliary diagnostic method based on a multi-agent collaborative architecture, specifically applied to a complex clinical case of hemophilia A complicated by spontaneous massive intraperitoneal hematoma leading to hemorrhagic shock. This method, through the collaborative operation of the SR Agent, DP Agent, and NHSC Agent, achieves a complete closed loop from information retrieval and unstructured data parsing to fast and slow dual-pathway causal decision-making.

[0029] 1. System task startup and case input: This method first activates the sensing front end to receive clinical case documents as input. In this embodiment, the input subject is a 38-year-old male patient with a history of hemophilia A, presenting with abdominal pain for 4 hours. Physical examination upon admission revealed abdominal distension and a hard mass (approximately 15cm × 16cm) in the right abdomen. Laboratory tests showed a significantly prolonged activated partial thromboplastin time (APTT) of 74.1s. Emergency CT imaging revealed a large abdominal hematoma with active bleeding. The objective of this method is to automatically construct an evidence-based medicine review and generate a dynamic diagnostic and treatment decision support plan for this rare disease in an acute and critical situation, utilizing a brain-like architecture.

[0030] 2. Multi-agent collaborative processing flow: (1) Overview of SR Agent System Intelligent Agent: Multidimensional Quantitative Screening Based on ROB2 As the perception front-end of this method, the SR Agent first performs keyword extraction on the input cases, identifying "hemophilia A," "intra-abdominal hemorrhage," and "hemorrhagic shock" as the core search entities. It then connects to external databases such as PubMed and the Cochrane Library via standard API interfaces to perform the following cascading processing: First-level screening: A regular expression matching strategy is executed to quickly filter out documents whose titles contain "Hemophilia A" and "Retroperitoneal Hematoma". Second-level screening: A semantic model is used to perform vectorized relevance analysis on the abstracts, eliminating low-relevance documents that, although matching keywords, have the subject matter of "dental bleeding" or "joint bleeding". Third-level screening (ROB2 in-depth assessment): For the remaining documents, the SR Agent activates the ROB2 assessment engine, performing in-depth reviews in five dimensions in parallel: (D1) Reviewing the randomization process to confirm whether the control group allocation was concealed; (D2) Analyzing deviations from the expected intervention to exclude biases caused by improper use of hemostatic drugs; (D3) Detecting missing outcome data; (D4) Assessing the objectivity of outcome measurement; (D5) Verifying selective reporting. This method ultimately calculates the evidence quality score for each document, automatically eliminating low-quality studies with scores below the threshold, and outputting a high-quality original document set.

[0031] (2) DP Agent document processing intelligent agent: deep structure extraction based on MinerU The DP Agent receives the aforementioned high-quality literature and the patient's original medical records (PDF format) and performs the following operations: Deep analysis: Integrates the MinerU deep analysis tool to accurately identify and analyze the emergency laboratory reports (multi-column layout) and abdominal CTA imaging reports (non-standard charts) in the patient's medical records.

[0032] Field Extraction: The large language model is invoked to automatically extract core clinical fields from the parsed text. In this case, key information was extracted: age of onset "38 years old", medication dosage "irregular treatment", survival-related indicators "blood pressure 95 / 53 mmHg (shock state)" and inheritance pattern "hemophilia A".

[0033] Vectorized storage: The extracted fields are converted into high-dimensional vector data and stored in the vector knowledge base of this method for downstream cognitive core to call.

[0034] (3) NHSC Agent: Dynamic Decision Making Based on S1 / S2 Dual Pathways The NHSC Agent serves as the cognitive core of this method. It employs a BioBERT embedding layer to transform patient symptoms and extracted knowledge into high-dimensional vectors, which are then processed using a dual-processing theoretical architecture. S1 Intuitive Pathway (Fast Response): This method utilizes a spiking neural network (SNN) to process high-confidence signals. For the combination of "APTT 74.1s" and "giant abdominal mass," the LIF neuron model simulates nerve impulses according to the formula... (1) Integrating, when the membrane potential u instantaneously exceeds the threshold V th The pulse is delivered in a timely manner. The S1 pathway identifies the common pathological pattern of "spontaneous bleeding due to coagulation factor deficiency" within milliseconds, quickly ruling out the intuitive misjudgment of "abdominal tumor". In formula (1), τ is the membrane time constant, V(t) is the membrane potential, and V rest V(t) is the resting potential, and I(t) is the input current; when V(t) exceeds the threshold V... th Pulse is emitted at specific times; S2 Reasoning Path (In-depth Analysis): Since the decision of "whether to perform surgical exploration" is a high-risk, high-uncertainty decision, the dual-pathway arbitrator dynamically activates the S2 inference pathway based on entropy calculation. The S2 pathway adopts a hierarchical inference model (HRM), which performs "shock correction planning" in the H layer and "hematoma compression detail analysis" in the L layer.

[0035] During this process, the evidence synthesis sub-agent integrated within the S2 pathway executes the Time-Weighted Evidence Synthesis (TWES) algorithm: Parallel computation time weights (2) With quality weight ω q .

[0036] Where, ω t Represents the time weight, ω qThe quality weight is represented by T, the publication duration of the document, and λ is the decay coefficient. This method assigns extremely high weight to the latest literature published in 2025 on "conservative treatment of intraperitoneal hemorrhage in hemophilia" and reduces the weight of literature advocating early and aggressive surgery.

[0037] Through nonlinear coupling ω final = Softmax(ω t ·ω q (4) Generate the final weights.

[0038] Where ω final Indicates the final overall weight of the literature; "ω" time ·ω quality The "" indicates multiplicative coupling, meaning that a document must possess both the characteristics of being "new" (high time weight) and "good" (high quality weight) for the product to be large. If a document is very new but of poor quality, or of good quality but too old, the product will be reduced.

[0039] ③ Combining natural history with the predictions of the intelligent agent, it was inferred that if surgery was performed, the patient's risk of death was significantly higher than that of conservative treatment.

[0040] 3. Method Output and Decision Support: This method ultimately outputs dynamic diagnostic and treatment suggestions, clearly stating: Diagnosis: Hemophilia A complicated by severe intra-abdominal hemorrhage and hemorrhagic shock.

[0041] Treatment recommendations: Based on the latest evidence weighted by the TWES algorithm, immediate factor VIII replacement therapy is strongly recommended, with the goal of correcting APTT to the normal range, combined with blood transfusion to combat shock; exploratory laparotomy is strongly discouraged before factor levels are corrected (due to the high mortality risk inferred from S2 pathway causal assessment).

[0042] Prognostic prediction: Based on natural history data, it is predicted that the hematoma will be gradually absorbed within 2-3 weeks with standard replacement therapy.

[0043] 4. Beneficial effects: This embodiment shows that by integrating the MinerU parsing tool and the ROB2 assessment engine, the present invention solves the problem of multi-source heterogeneity and inconsistent quality of rare disease data; in particular, the S1 / S2 brain-like dual-pathway architecture introduced by the NHSC Agent not only ensures intuitive and rapid early warning of critical values ​​(such as abnormal APTT), but also ensures deep causal reasoning for complex treatment decisions (such as surgery vs. conservative treatment), which significantly improves the scientificity and timeliness of rare disease diagnosis and treatment.

[0044] Example 2: Brain-like Auxiliary Diagnosis and Treatment for Rare Diseases This method is applied in the differential diagnosis of Huntington's Disease. This embodiment illustrates a brain-like auxiliary diagnostic and therapeutic method based on a multi-agent collaborative architecture, specifically applied to a 56-year-old female patient who presented with "involuntary tremors of the limbs accompanied by slowed reaction time for 2 years," with a typical family history of the condition, and was clinically diagnosed as a complex neurodegenerative disease with "dystonia to be ruled out." This method, through the collaborative operation of the SR Agent, DP Agent, and NHSCAgent, achieves a complete closed loop from differential diagnostic evidence retrieval and multimodal data analysis to the construction of a dynamic disease causal map.

[0045] 1. Method / Task Initiation and Case Input: This method first activates the sensing front end to receive clinical case documents as input. In this example, the input subject is a 56-year-old female patient with one son, who presented with a two-year history of "involuntary tremors in the limbs and slowed reaction time." Upon admission, physical examination revealed involuntary choreiform movements and gait instability. Cognitive assessment showed an MMSE score of 23, a MoCA score of 18 (indicating cognitive decline), and moderate anxiety (SAS score of 63). Family history showed that her father and brother both had similar symptoms and died around the age of 50.

[0046] The system's task objective is to automatically retrieve differential diagnostic evidence for this complex movement disorder case with a genetic predisposition, construct a causal knowledge graph that reveals the "gene-pathology-phenotype" mechanism, and generate a personalized management plan.

[0047] 2. Multi-agent collaborative processing flow: (1) Overview of SR Agent Systems: Multidimensional Quantitative Screening Based on ROB2 The SR Agent, serving as the perceptual front end of this method, first performs natural language understanding on the input case data, extracting core entities such as "choreiform movements," "progressive cognitive decline," and "family history of autosomal dominant inheritance," and then identifies "Huntington's Disease," "Chorea," and "Dementia" as core search keywords. This method connects to an external database via a standard API interface, performing the following cascaded processing: First-level filtering: Execute a regular expression matching strategy to quickly filter out literature collections whose titles contain "Huntington" or "Hereditary Chorea".

[0048] Second-level screening: Using semantic models, vectorized relevance analysis is performed on the abstracts to eliminate low-relevance documents that match the keywords but are actually "drug-induced tardive dyskinesia" or "vasochorea".

[0049] The third level of screening (ROB2 in-depth assessment): For the remaining literature, the SR Agent activates the ROB2 assessment engine to perform in-depth reviews across five dimensions in parallel: (D1) randomization process review; (D2) deviation from expected intervention review (with particular attention to dropout rates in drug clinical trials); (D3) missing outcome data review; (D4) outcome measurement review (assessing the standardization of the UHDRS scale); and (D5) selective reporting review. This method automatically eliminates low-quality studies with scores below a threshold, outputting a high-quality collection of original literature on the diagnostic criteria and treatment guidelines for Huntington's disease.

[0050] (2) DP Agent document processing intelligent agent: deep structure extraction based on MinerU The DP Agent receives the aforementioned high-quality literature and the patient's original medical records (PDF format) and performs the following operations: Deep Analysis: Integrates the MinerU deep analysis tool to perform layout recognition and accurate analysis of complex rating scales (MMSE / MoCA / SAS / SDS) and cranial MRI image reports in medical records.

[0051] Field extraction: The large language model is invoked to automatically extract core clinical fields from the parsed text. In this case, key information was extracted: age of onset "54 years old (2 years ago)", core symptoms "choreiform movements + cognitive impairment + mental and behavioral abnormalities", and family history pattern "father and elder brother have the disease (autosomal dominant inheritance characteristics)".

[0052] Vectorization storage: At the same time, key parameters such as the CAG repeat count threshold (≥36 times) of the HTT gene are extracted from the literature, transformed into high-dimensional vector data, and stored in the vector knowledge base of this method.

[0053] (3) NHSC Agent: Dynamic Decision Making and Atlas Construction Based on S1 / S2 Dual Pathways The NHSC Agent, serving as the cognitive core of this method, employs a BioBERT embedding layer to transform patient data into high-dimensional vectors and processes them using a brain-like architecture. S1 Intuitive Pathway (Rapid Response): This method utilizes spiking neural networks (SNNs) to process high-confidence patterns. For the characteristic combination of "mid-life onset," "choreiform movements," and "dominant family history," the LIF neuron model simulates nerve impulses, with the membrane potential instantaneously exceeding the threshold to fire pulses. The S1 pathway identifies the typical pattern of "Huntington's disease (HD)" within milliseconds, rapidly narrowing the differential diagnosis from the broad category of "dystonia."

[0054] S2 Reasoning Path (Deep Analysis and Graph Construction): Given the complexity of disease mechanisms and the sensitivity of genetic counseling, the dual-pathway arbitrator activates the S2 reasoning pathway. The S2 pathway employs a hierarchical reasoning model (HRM) and invokes a causal evaluation sub-agent to construct a dynamic causal knowledge graph. ①Entity and relation extraction: This method automatically extracts causal chains from the literature: HTT gene CAG repeat amplification, rightarrow mutant huntingtin protein (mHTT) aggregation, rightarrow striatal medium polyspinous neuron toxicity, rightarrow basal ganglia-thalamus-cortical circuit disorder, rightarrow choreiform movement (motor domain) / executive function impairment (cognitive domain).

[0055] ② Path reasoning and completion: This method connects non-directly adjacent nodes through multi-hop reasoning. For example, it can deduce that "involvement of the prefrontal-striatal circuit" is the pathological basis of the patient's "MoCA score of 18 (decreased executive function)" rather than simply Alzheimer's disease.

[0056] ③ TWES Evidence Synthesis: During the treatment recommendation generation phase, the evidence synthesis sub-agent uses the TWES algorithm to assign extremely high weights to the latest literature from 2023-2024 on "VMAT2 inhibitors (such as deuterated benazine)" and "antisense oligonucleotide therapy". (Large), reducing the weight of traditional antipsychotic drugs in the treatment of chorea.

[0057] 3. Method Output and Decision Support: This method completes the entire process within 30 minutes and outputs interactive treatment suggestions: Diagnosis: Highly suspected Huntington's disease, excluding essential tremor and Parkinson's disease.

[0058] Causal mapping visualization: It shows the complete pathological pathway from "CAG repeat amplification" to "striatal neuron loss" and then to "choreiform movement / cognitive impairment", helping doctors explain the condition to family members.

[0059] Personalized recommendations: HTT gene testing (CAG repetitive sequence analysis) is strongly recommended for diagnosis; based on the TWES weighted results, deuterated benzidine (VMAT2 inhibitor) is recommended for choreiform movements; for moderate anxiety (SAS score 63), combined psychological intervention or SSRIs are recommended; the patient's offspring have a 50% risk of inheritance, and referral to the Department of Medical Genetics for consultation is recommended.

[0060] 4. Beneficial Effects: This embodiment demonstrates that, through the brain-like dual-pathway architecture of the NHSC Agent, this invention not only achieves rapid and intuitive identification of typical characteristics of rare genetic diseases (S1 pathway), but also successfully constructs a visualized disease mechanism atlas through deep causal reasoning via the S2 pathway. This atlas logically links the patient's microscopic genetic changes with macroscopic clinical phenotypes (motor, cognitive, and mental), greatly improving the interpretability of the diagnosis and treatment process and providing clinicians with precise, scientific, and highly humanistic decision support.

Claims

1. A brain-inspired auxiliary diagnostic and treatment method for rare diseases, characterized in that, The server implementation based on GPU or NPU includes: a system review agent, which serves as the perception front end of this method, connects to external medical databases through standard API interfaces, performs automated literature retrieval and initial quality screening, and outputs a set of original literature; and a document processing agent, which receives the set of original literature from the system review agent, uses deep parsing tools to extract structured data from unstructured PDF documents, and generates a structured knowledge base containing core clinical fields. The NHSC brain-inspired intelligent agent adopts a brain-inspired architecture based on the dual-process theory. It receives a structured knowledge base of a document processing intelligent agent and patient data, and outputs the final dynamic diagnosis and treatment suggestions through fast and slow dual-pathway collaborative processing.

2. The brain-like assisted diagnosis and treatment method for rare diseases according to claim 1, characterized in that, The system review agent executes a cascaded processing logic of multi-level funnel screening and multi-dimensional quantitative evaluation, specifically including: a first-level screening unit, used to filter the document titles for keywords through regular expression matching; The second-level screening unit is used to perform relevance analysis on document abstracts using a semantic model; The third-level screening unit performs standard compliance checks on the full text of the literature. Using the ROB2 evaluation engine, it receives the literature after the three-level screening and performs in-depth reviews in five dimensions in parallel according to the Cochrane bias risk assessment criteria: (D1) Review of the randomization process to assess grouping randomness and sequence concealment; (D2) Deviance from expected intervention review, analyzing intervention deviations caused by improper implementation; (D3) Review missing results data to check the completeness of follow-up data; (D4) Results measurement review, assessing the objectivity of the measurement method and the bias of the measurer; (D5) Selective reporting review to verify cases of concealment of negative results; The ROB2 evaluation engine calculates an evidence quality score of 0-20 based on the above five dimensions, and automatically eliminates low-quality studies by setting a threshold.

3. The brain-like assisted diagnosis and treatment method for rare diseases according to claim 1, characterized in that, The document processing agent is specifically configured as follows: Integrates the MinerU deep parsing tool to identify and parse multi-column layouts, cross-page tables, and non-standardized charts in medical literature PDFs; By calling the large language model, core clinical fields such as age of onset, drug dosage, survival rate and genetic pattern are automatically extracted for each parsed document; The extracted fields are converted into vector data and stored in a vector database for downstream use.

4. The brain-like assisted diagnosis and treatment method for rare diseases according to claim 1, characterized in that, The NHSC brain-like intelligent agent adopts a two-layer architecture that simulates the S1 intuition pathway and the S2 reasoning pathway of the human brain. Specifically, it includes a BioBERT embedding layer, which is used to convert the input patient symptom descriptions, gene sequences and phenotypic data into high-dimensional vector representations. The S1 intuitive pathway, used for rapid response tasks, employs a spiking neural network and simulates nerve impulses using a LIF neuron model to achieve millisecond-level rapid identification of common disease patterns. Its neuronal membrane potential dynamics equation is: (1) Where τ is the membrane time constant, V(t) is the membrane potential, and V rest V(t) is the resting potential, and I(t) is the input current; when V(t) exceeds the threshold V... th Pulse is emitted at specific times; The S2 reasoning pathway is used to handle complex and difficult tasks. It adopts a hierarchical reasoning model, which includes a high-level H layer responsible for abstract diagnostic planning and a low-level L layer responsible for detailed symptom analysis. The two layers interact with each other through a cross-attention mechanism. The S2 reasoning pathway integrates six sub-agents: natural history comparison, historical comparison, time series analysis, causal assessment, quality assessment, and evidence synthesis. A dual-path arbiter is used for calculations based on entropy uncertainty, dynamically determining whether to enable the S1 intuitive path to output direct results or activate the S2 reasoning path for deep reasoning.

5. The brain-like assisted diagnosis and treatment method for rare diseases according to claim 4, characterized in that, The evidence synthesis sub-agent in the S2 reasoning pathway employs a time-weighted evidence synthesis algorithm to calculate the final synthesis weight ω for each piece of evidence. final Its calculation logic is as follows: parallel calculation time weight ω time With quality weight ω quality， The calculation formula is: (2) (3) Where, ω time Represents the time weight, ω quality The quality weight is represented by Δt, the publication duration of the literature, and λ, which is the decay coefficient. GRADE Quality scores based on GRADE ratings; The final weights are obtained through nonlinear coupling and normalization, and the calculation formula is as follows: (4) This enables dynamic evidence fusion based on both timeliness and authority.

Images