Rare disease auxiliary diagnosis system based on multi-paradigm statistical fusion and large model agent introspection

Abstract

The application discloses a rare disease auxiliary diagnosis system based on multi-paradigm statistical fusion and large model agent reflection, comprising: a data standardization extraction module, which acquires a standardized HPO phenotype query set of a patient to be diagnosed; a source-level multi-paradigm reasoning module, which receives the HPO phenotype query set and inputs the same into multiple heterogeneous diagnosis engines in parallel to generate a source-level diagnosis evidence set; a two-stage meta-fusion module, which performs two-stage fusion processing to generate a statistical consensus priority diagnosis list; a knowledge enhancement agent reflection module, which performs reasoning reflection within a constraint diagnosis boundary through a thinking chain reasoning engine of a large language model and outputs a diagnosis list containing candidate diagnoses; and a deduplication splicing module, which sequentially connects remaining items after eliminating the candidate diagnoses to generate a final diagnosis recommendation list. The rare disease auxiliary diagnosis system of the application takes into account both rigorous mathematical precision and heuristic flexibility.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and medical auxiliary diagnosis technology, and in particular to a rare disease auxiliary diagnosis system based on multi-paradigm statistical fusion and large model intelligent agent reflection. Background Technology

[0002] Rare diseases (RDs) refer to a group of special diseases with an extremely low prevalence (usually less than 1 / 2000). Due to their extremely low incidence, high phenotypic heterogeneity, incomplete penetrance, and complex comorbidity patterns, the clinical manifestations of rare diseases are often subtle and easily overlap with common diseases. This leads to a prolonged "diagnostic odyssey" for rare disease patients, with an average diagnosis time of over five years, during which they face repeated missed diagnoses and redundant screenings, placing a heavy burden on patients' families and the healthcare system. Because rare disease diagnosis is interdisciplinary and involves multiple systems, it relies heavily on highly specialized clinical experience, resulting in significant structural diagnostic gaps in primary care or non-specialized medical institutions. Therefore, developing an efficient, accurate, and robust computational model for auxiliary diagnostic decisions to assist clinicians in making efficient and reliable differential diagnoses within the vast and complex spectrum of rare diseases is an urgent need shared by the medical and computational science communities.

[0003] To address this diagnostic challenge, the field of computational-assisted diagnosis for rare diseases has undergone several generations of paradigm evolution. The first generation of paradigms was represented by symbolic reasoning models driven by ontology and knowledge bases, such as Bayesian networks based on Human Phenotype Ontology (HPO) and semantic similarity measurement algorithms. These methods rely on rigorous formal logic, solidifying clinical experience from medical literature and databases into probabilistic priors, resulting in strong interpretability. The second generation was represented by data-driven deep learning pattern recognition technologies (such as neural network classifiers), aiming to automatically extract and identify deep nonlinear relationships between phenotypes through high-dimensional feature representations. In recent years, with breakthroughs in generative artificial intelligence, the fourth generation has emerged, represented by generative reasoning and agentic architectures based on Large Language Models (LLMs). These architectures, through parameterized medical knowledge, demonstrate strong context integration capabilities and flexible heuristic medical logic. In addition, there are similarity retrieval methods based on case matching, which compare the current patient's characteristics with confirmed cases in a case database to achieve "disease inference from cases."

[0004] The aforementioned multi-generational diagnostic paradigms exhibit natural orthogonality and heterogeneous complementarity in the dimension of phenotypic reasoning: symbolic models can ensure a rigid logical lower bound, deep learning can extract implicit phenotypic combination features, case retrieval can provide real-world precedent support, and large language models offer heuristic thinking that transcends the limitations of static databases. However, existing fusion strategies all suffer from serious structural limitations in how to organically integrate these paradigms with their distinct advantages to generate credible clinical decisions. On the one hand, traditional downstream linear fusion or simple voting algorithms only perform numerical accumulation or averaging of the candidate disease lists generated by each heterogeneous sub-model at the output end. This method ignores the conflicts and uncertainties in the underlying mathematical assumptions of each model, and is prone to the phenomenon of "regression to the mean," causing highly specific and advantageous information to be trivialized by noise signals. On the other hand, when faced with medical records with complex phenotypes, the end-to-end reflection mechanism of large language models based on prompts is prone to cognitive drift and amplification of cascade errors when the agent performs uncontrolled multi-step retrieval, iterative planning and self-correction processes due to the lack of pre-reasoning constraints. Ultimately, the output results cannot be accepted by high-security clinical practice due to the illusion that it contradicts medical common sense.

[0005] In summary, how to break through the traditional diagnostic integration methods that only perform downstream aggregation through simple linear weighting or contextual hints, and effectively coordinate the "rigid deterministic logic" of discrete ontology and statistical computation with the "elastic heuristic reasoning" of generative models at a deep architectural level, is a key technical bottleneck that urgently needs to be solved in the field of computational medicine for rare diseases. Summary of the Invention

[0006] This invention provides a rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large model intelligent agent reflection, which balances rigorous mathematical accuracy with heuristic flexibility.

[0007] The technical solution of the present invention is as follows: A rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection includes: The data standardization and extraction module obtains a standardized HPO phenotype query set for the patient to be diagnosed. The source-level multi-paradigm inference module receives the HPO phenotypic query set and inputs it in parallel to multiple heterogeneous diagnostic engines to generate a source-level diagnostic evidence set. ; The two-stage meta-fusion module provides source-level diagnostic evidence sets. A two-stage fusion process is performed to generate a statistical consensus priority diagnostic list. ; The knowledge-enhanced intelligent agent's reflection module prioritizes statistical consensus diagnostic lists. The former Using a disease as a constraint on the diagnostic boundary, combined with the HPO phenotype query set of the patient to be diagnosed and the source-level diagnostic evidence set. Statistical consensus priority diagnostic list In addition to the features of the rare disease knowledge base, the reasoning engine of the large language model performs reasoning and reflection within the aforementioned constrained diagnostic boundaries, outputting information including... A list of candidate diagnoses ; The deduplication and concatenation module will merge the diagnostic list. Priority Diagnostic List Based on Statistical Consensus Remove from The remaining items after each candidate diagnosis are sequentially concatenated to generate the final diagnostic recommendation list. .

[0008] Preferably, the multiple heterogeneous diagnostic engines include a knowledge-driven likelihood ratio diagnostic engine, a data-driven deep learning diagnostic engine, a similarity-matching-based case retrieval engine, and a large language model zero-shot diagnostic engine.

[0009] Preferably, the likelihood ratio diagnostic engine is based on Bayes' theorem, combines the frequency of disease phenotype annotations in the rare disease knowledge base, calculates the likelihood ratio of candidate diseases according to the query phenotype features in the HPO phenotype query set, and outputs a first diagnostic list sorted in descending order of likelihood ratio score. ; The deep learning diagnostic engine utilizes a multilayer perceptron trained on a real patient database to extract high-dimensional patient-phenotype association features. Using the HPO phenotype query set as input and the disease Logits distribution as output, it outputs a second diagnostic list in descending order of the disease Logits distribution. ; The case retrieval engine calculates the cumulative information similarity between the HPO phenotype query set and historical cases in the real patient database, and outputs a third diagnostic list recommended by historical case matching. ; The large language model zero-shot diagnostic engine organizes the phenotypic terms in the HPO phenotypic query set into text form and constructs diagnostic prompt words. The large language model directly generates priority diagnostic recommendations, aligns the diagnostic entities to disease entities in the rare disease knowledge base, and outputs a fourth diagnostic list. ; The output results of the above engines constitute the source-level diagnostic evidence set. .

[0010] Preferably, the two-stage fusion process includes: In the first phase, the source-level diagnostic evidence set is processed based on multiple complementary basic aggregators. Intermediate consensus scores were calculated for each candidate disease in the study. In the second stage, based on the aforementioned intermediate consensus scores, a weighted meta-fusion score is calculated for each candidate disease, and the statistical consensus priority diagnosis list is generated based on these meta-fusion scores. .

[0011] For any heterogeneous diagnostic engine The generated list of diagnoses , A collection of heterogeneous diagnostic engines, defining candidate diseases. The normalized ranking score is ,in For disease In the diagnostic list The ranking in For the diagnostic list The total number of candidate diseases; if the disease Not detected by heterogeneous diagnostic engine A recall will result in punitive penalties. It is 1.0.

[0012] The basic aggregator includes: Minimum ranking aggregator, capturing the minimum ranking score of the best ranking of a candidate disease in a single heterogeneous diagnostic engine. , ; The geometric mean ranking aggregator calculates the geometric mean ranking score of candidates whose preferences are consistent across most heterogeneous diagnostic engines. , ; The number of heterogeneous diagnostic engines; A ranking aggregator based on the Bates distribution calculates Bates ranking scores that simulate the statistical significance of the average ranking using the standard normal cumulative distribution function. , ; The cumulative function of the standard normal distribution. ; A score-based evidence aggregator calculates a composite evidence score based on the standardized score predicted by heterogeneous diagnostic engines and the number of heterogeneous diagnostic engines supporting candidate diseases. , , This is the standard score of the original score within the current heterogeneous diagnostic engine output list.

[0013] The second stage includes: using a discretized weight matrix to perform a weighted geometric average on the intermediate consensus scores of the basic aggregator to obtain the meta-fusion score. : ; in, Candidate disease In the basic aggregator The normalized ranking score in the middle, A collection of basic aggregators; Basic Aggregator The weight.

[0014] according to The ranking yields a statistical consensus priority diagnostic list. .

[0015] Preferably, the knowledge-enhancing intelligent agent reflection module concatenates the following multi-source, multi-modal diagnostic features to construct structured text prompts for large language models: The patient's demographic characteristics and the HPO phenotype query set described above; Source-level diagnostic evidence set output by the source-level multi-paradigm reasoning module ; The statistical consensus priority diagnostic list output by the two-stage meta-fusion module ; Priority diagnostic list based on statistical consensus retrieved from a rare disease knowledge base. Center front Objective mapping and labeling frequency relationship between candidate diseases and the HPO phenotype query set .

[0016] Preferably, the reasoning reflection of the large language model in the knowledge-enhanced intelligent agent reflection module includes: Phenotypic matching assessment: Before examination The labeled phenotypes of each candidate disease provide the widest coverage of the most specific clinical symptoms for patients, with a greater focus on rare symptoms; Knowledge base consistency check: Calculate the degree of agreement between the necessarily occurring phenotypes and the patient phenotype set, and regard the absence of non-essential phenotypes as a secondary factor that does not affect the diagnosis; Paradigm conflict resolution: Identify and rationally interpret ranking conflicts between different heterogeneous diagnostic engines, and give priority to diseases supported by highly specific cases; The final output is the reordered front. A list of candidate diagnoses .

[0017] The large language model of the knowledge-enhanced intelligent agent reflection module is a large language model that has been fine-tuned in the medical field or optimized by prompting engineering, and is deployed on a local medical diagnostic workstation or called through a secure API interface.

[0018] Preferably, to prevent overfitting or inference bias in large language models from causing potential missed diagnoses of diseases, a lossless and secure splicing is performed through a deduplication and splicing module to obtain the final diagnostic recommendation list. : ; in, Indicates from Delete The diagnosis is made while maintaining the relative order of the remaining diagnoses; Indicates will The remaining diagnoses are pieced together. after.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) This invention changes the traditional approach of having large models perform zero-constraint inference directly from massive amounts of medical terminology. Instead, it uses statistical fusion to define a highly reliable consensus diagnostic boundary for the large model. The large model only plays the role of arbitration and fine-tuning based on this consensus boundary, eliminating the diagnostic cognitive limitations of the autonomous diagnostic agent and establishing a highly reliable statistical constraint boundary. This mechanism effectively suppresses the hallucination phenomenon in generative models, thereby ensuring the reliability and stability of diagnostic results in high-risk medical scenarios.

[0020] (2) This invention effectively overcomes the limitations of a single aggregator in different scenarios through a two-stage diagnostic aggregation, enhancing the robustness of the algorithm in real clinical applications. By performing two-stage fusion on the basic aggregator and three-level discretization and quantization on the fusion weights in the meta-fusion stage, the system can not only eliminate redundant and marginally contributing traditional integrators, but also more effectively enhance the numerical stability of the model. This ensures that this invention can resist the distribution migration caused by different hospitals' medical record writing habits, input noise, and specialty bias, and improves cross-center generalization ability.

[0021] (3) This invention improves diagnostic coverage through the complementarity of multiple diagnostic paradigms. This invention perfectly integrates the rigor of symbolic reasoning, the pattern recognition of deep learning, the accuracy of historical case retrieval, and the heuristic reasoning of large language models. Experiments show that this segmented "diagnostic information source" Statistical consensus The disease prioritization accuracy of the "agent reflection" architecture is significantly better than that of existing dedicated clinical-grade systems. Attached Figure Description

[0022] Figure 1 This is a flowchart for the auxiliary diagnosis of rare diseases based on multi-paradigm statistical fusion and large model reflection. Detailed Implementation

[0023] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the embodiments described below are intended to facilitate the understanding of the present invention and do not limit it in any way.

[0024] This embodiment provides a rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection. Specifically, it runs on a medical diagnostic workstation equipped with a high-performance computing processor, a deep learning accelerator card, and a pre-deployed large language model service interface, and includes: 1. Data Standardization and Extraction Module: This module converts and extracts the original clinical medical records and laboratory characteristics of the individuals to be diagnosed into a standard HPO phenotype query set. ; 2. Source-level multi-paradigm inference engine cluster: This includes a parallel-deployed knowledge-driven likelihood ratio scoring model, a data-driven deep learning classifier, a similar case retrieval engine, and a zero-shot large-model diagnostic engine. It is used to generate heterogeneous source-level diagnostic evidence sets. ; 3. Two-stage meta-fusion module: Contains a basic aggregation submodule for the first stage and a meta-fusion calculation submodule for the second stage, designed to calculate and output an initial statistical consensus priority diagnostic list. ; 4. Knowledge-Enhanced Intelligent Agent Reflection Module: This module integrates patient clinical phenotypes, multi-source model evidence, initial priority diagnosis lists, and features from the unified medical knowledge base. It then uses the CoT (Cooperation of Thought) reasoning engine of a large language model to generate the most reasonable top-ranking algorithm. Diagnostic list ; 5. Deduplication and concatenation module: This module uses set subtraction and sequential concatenation algorithms to deduplicate and concatenate sets. Priority list of original statistical consensus The remaining set is reassembled to output a final, finely ranked diagnostic recommendation list for clinical decision support. .

[0025] A rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection is used to perform the following steps, such as... Figure 1 As shown: Step S1: Generation of Multi-paradigm Source-Level Diagnostic Evidence The system receives a standardized human phenotype ontology (HPO) phenotypic query set of patients to be diagnosed. For example, patient phenotype set Include A standardized phenotypic term: .

[0026] The system will use this vector The inputs are fed in parallel to four independent and orthogonal source-level diagnostic engines, and each engine is run to obtain a source-level candidate disease list: 1. Knowledge-driven Likelihood Ratio (LR) Engine: Based on Bayes' theorem, and combining the frequency of disease phenotype annotations from the rare disease knowledge base OMIM and Orphanet, it outputs a list of diseases sorted in descending order of likelihood ratio score based on the likelihood ratio values ​​of the query phenotype features for candidate diseases. ; 2. Data-driven Deep Learning (DL) Classifier: Utilizing a Multilayer Perceptron (MLP) trained on real patient databases (such as RareArena and Phenopacket), high-dimensional patient-phenotype association features are extracted. Taking the query phenotype features as input and the disease Logits distribution as output, a list of corresponding diseases is output in descending order of the disease Logits distribution. ; 3. Similarity-based case retrieval engine: First, the negative logarithm of the phenotypic frequency in the rare disease knowledge bases OMIM and Orphanet is taken to obtain the information content (IC) score of all phenotypic types. Then, the query phenotypic features are calculated. The cumulative IC similarity of overlapping phenotypes with historical cases in the real patient database is used to output a list of diseases recommended by historical case matching, sorted in descending order of IC score. ; 4. Large Language Model Zero-Shot Diagnostic Engine (LLM): This engine organizes query phenotypic terms into text and constructs diagnostic suggestion words. A large language model (such as DeepSeek-R1) directly generates priority diagnostic recommendations. After aligning diagnostic entities to OMIM / ORPHA disease entities, the output includes... List of most likely diagnoses ; The output results of the above engines constitute the source-level diagnostic evidence set. .

[0027] Step S2: Two-stage phenotypic diagnostic ranking meta-fusion To integrate the aforementioned heterogeneous source-level diagnostic lists and establish a robust and fault-tolerant statistical consensus boundary, the system performs the following two-stage meta-fusion: The system first ranks and normalizes the candidate diseases generated by each source diagnostic model. For the candidate diseases... Assuming it is in the source diagnostic model Output list In the middle of Bit( (If it is a set of source-level diagnostic models), then calculate its value in the model. The normalized ranking score is as follows: ; in, For the diagnostic list Size. If the disease... Undiagnosed model Recall, i.e. Not appearing in the model The diagnosis list includes punitive scoring. .

[0028] Next, the system triggers the following four different basic aggregation mechanisms in parallel to calculate the disease. Intermediate consensus scores under each basic aggregator: 1. Minimum ranking score This is used to capture and retain the best diagnostic clues in any model. ; 2. Geometric mean ranking score This is used to favor candidates that perform consistently in most models and to mitigate the interference of outliers. ; Indicates the number of source-level diagnostic models; 3. Ranking scores based on Bates distribution The statistical significance of the average rank is simulated using the standard normal cumulative distribution function: ; in, The cumulative function of the standard normal distribution. ; 4. Aggregated scores based on fractional evidence Using models to study diseases Original rating The standard score (Z-score) And the comprehensive evidence calculated from the number of models supporting the disease: ; in, The value of varies depending on the model: for logistic regression (LR) models, the original score is the likelihood ratio score of the disease; for deep learning (DL) models, the original score is the Logits score of the disease; for case retrieval engines, the original score is the IC similarity score; for large-scale zero-shot diagnostic engines, the original score is the normalized ranking score (the highest ranking diagnostic score is 1, and the lowest is 0). and Representing the model respectively The mean and standard deviation of the scores for all candidate diseases. It is a model For disease Original rating Standard score among all candidate diseases.

[0029] Subsequently, in order to eliminate the systematic bias caused by a single aggregation strategy, this invention further implements weighted meta-fusion based on the basic aggregator score: ; in, Candidate disease In the basic aggregator Normalized ranking score in (A collection of basic aggregators). On the independent validation set, the Optuna framework is used to adjust the weights. Optimize to maximize Top-5 / Top-10 recall rate, and improve generalization ability by optimizing the [review / implementation]. Mapping to discrete sets middle.

[0030] Afterwards, according to Candidate diseases are sorted by score from lowest to highest to generate a statistical consensus diagnostic list. The higher the ranking, the higher the risk of disease. This step provides the system with a definitive statistical diagnostic boundary.

[0031] Step S3: Reflection on the Large-Scale Intelligent Agent with Knowledge Enhancement The system extracts the statistical aggregated diagnostic list generated in step S2. The former (like ( ) diseases as alternative options.

[0032] The system concatenates the following multi-source, multi-modal diagnostic features to construct structured text prompts for large language models: 1. Patient demographic characteristics (e.g., gender, age) and query typology ; 2. The source-level diagnostic evidence sets output by the four independent engines in step S1. ; 3. The candidate list generated in step S2 ; 4. Retrieved from the rare disease knowledge base The front of the middle One alternative disease and query form Objective mapping and labeling frequency relationship between them .

[0033] The system constructs diagnostic prompts based on the structured data mentioned above and inputs them into a large language model (such as DeepSeek-R1). It then activates the built-in Chain of Thought (CoT) reasoning prompt template and performs clinical reasoning and reflection according to the following steps: 1. Phenotypic Coverage Assessment: Examine which candidate disease phenotypes can best cover the patient's most specific clinical symptoms and give greater attention to rare symptoms; 2. Knowledge base consistency check: Calculate the degree of agreement between "inevitable (100% occurrence)" phenotypes and the patient phenotype set, and regard the absence of non-essential phenotypes as a secondary factor that does not affect the diagnosis; 3. Paradigm Conflict Resolution: If a certain disease ranks low in the likelihood ratio model due to probability calculation limitations, but is strongly supported by highly specific cases in the case similarity model, the agent can identify this conflict, interpret it reasonably, and determine that the disease should be given higher weight.

[0034] Ultimately, after the intelligent agent derives clinically validated conclusions through thought chains, it outputs a finely selected list of candidates. (like ) shortlist of candidates : .

[0035] Step S4: Assembling and Outputting Diagnostic Results To prevent overfitting or inference bias in the large model during step S3, which could lead to missed diagnoses of potential diseases, this embodiment of the invention performs a lossless and secure stitching: ; in, Indicates from Delete The diagnosis is made while maintaining the relative order of the remaining diagnoses; Indicates will The remaining diagnoses are pieced together. after.

[0036] That is, retain the first part of the large model's fine-ranking. The name was removed from the original statistical aggregation list. The remaining items following each disease are sequentially concatenated to form the final list of recommended diagnostic results. .

[0037] Experimental results show that by positioning the large language model as an arbitrator for fine-tuning based on statistical consensus boundaries, this invention improves the Top-1 accuracy of rare disease differential diagnosis to approximately 50%. Compared to the state-of-the-art rare disease differential diagnosis agent DeepRare, the Top-1 accuracy is improved by approximately 19%, while the error rate of rare disease code alignment is reduced to below 1%. This mechanism effectively suppresses hallucination phenomena in generative models, thereby ensuring the reliability and stability of diagnostic results in high-risk medical scenarios.

[0038] The embodiments described above provide a detailed explanation of the technical solutions and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, additions, and equivalent substitutions made within the scope of the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-scale model intelligent agent reflection, characterized in that, include: The data standardization and extraction module obtains a standardized HPO phenotype query set for the patient to be diagnosed. The source-level multi-paradigm inference module receives the HPO phenotypic query set and inputs it in parallel to multiple heterogeneous diagnostic engines to generate a source-level diagnostic evidence set. ; The two-stage meta-fusion module provides source-level diagnostic evidence sets. A two-stage fusion process is performed to generate a statistical consensus priority diagnostic list. ; The two-stage fusion process includes: In the first phase, the source-level diagnostic evidence set is processed based on multiple complementary basic aggregators. Intermediate consensus scores are calculated for each candidate disease in the dataset; for any heterogeneous diagnostic engine The generated list of diagnoses , A collection of heterogeneous diagnostic engines, defining candidate diseases. The normalized ranking score is ,in For disease In the diagnostic list The ranking in For the diagnostic list The total number of candidate diseases; if the disease Not detected by heterogeneous diagnostic engine A recall will result in punitive penalties. It is 1.0; The basic aggregator includes: Minimum ranking aggregator, capturing the minimum ranking score of the best ranking of a candidate disease in a single heterogeneous diagnostic engine. , ; The geometric mean ranking aggregator calculates the geometric mean ranking score of candidates whose preferences are consistent across most heterogeneous diagnostic engines. , ; The number of heterogeneous diagnostic engines; A ranking aggregator based on the Bates distribution calculates Bates ranking scores that simulate the statistical significance of the average ranking using the standard normal cumulative distribution function. , ; The cumulative function of the standard normal distribution. ; A score-based evidence aggregator calculates a composite evidence score based on the standardized score predicted by heterogeneous diagnostic engines and the number of heterogeneous diagnostic engines supporting candidate diseases. , , The standard score of the original score within the current heterogeneous diagnostic engine output list; In the second stage, based on the aforementioned intermediate consensus scores, a weighted meta-fusion score is calculated for each candidate disease, and the statistical consensus priority diagnosis list is generated based on the meta-fusion scores. This includes: using a discretized weight matrix to perform a weighted geometric average on the intermediate consensus scores of the basic aggregator to obtain the meta-fusion score. : in, Candidate disease In the basic aggregator The normalized ranking score in the middle, A collection of basic aggregators; Basic Aggregator The weights; The knowledge-enhanced intelligent agent's reflection module prioritizes statistical consensus diagnostic lists. The former Using a disease as a constraint on the diagnostic boundary, combined with the HPO phenotype query set of the patient to be diagnosed and the source-level diagnostic evidence set. Statistical consensus priority diagnostic list In addition to the features of the rare disease knowledge base, the reasoning engine of the large language model performs reasoning and reflection within the aforementioned constrained diagnostic boundaries, outputting information including... A list of candidate diagnoses ; The deduplication and concatenation module will merge the diagnostic list. Priority Diagnostic List Based on Statistical Consensus Remove from The remaining items after each candidate diagnosis are sequentially concatenated to generate the final diagnostic recommendation list. .

2. The rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection as described in claim 1, characterized in that, Multiple heterogeneous diagnostic engines include a knowledge-driven likelihood ratio diagnostic engine, a data-driven deep learning diagnostic engine, a similarity-matching-based case retrieval engine, and a large language model zero-shot diagnostic engine.

3. The rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection as described in claim 2, characterized in that, The likelihood ratio diagnostic engine, based on Bayes' theorem and combined with the frequency of disease phenotype annotations in the rare disease knowledge base, calculates the likelihood ratio of candidate diseases according to the query phenotype features in the HPO phenotype query set, and outputs a first list of diagnoses sorted in descending order of likelihood ratio score. ; The deep learning diagnostic engine utilizes a multilayer perceptron trained on a real patient database to extract high-dimensional patient-phenotype association features. Using the HPO phenotype query set as input and the disease Logits distribution as output, it outputs a second diagnostic list in descending order of the disease Logits distribution. ; The case retrieval engine calculates the cumulative information similarity between the HPO phenotype query set and historical cases in the real patient database, and outputs a third diagnostic list recommended by historical case matching. ; The large language model zero-shot diagnostic engine organizes the phenotypic terms in the HPO phenotypic query set into text form and constructs diagnostic prompt words. The large language model directly generates priority diagnostic recommendations, aligns the diagnostic entities to disease entities in the rare disease knowledge base, and outputs a fourth diagnostic list. ; The output results of the above engines constitute the source-level diagnostic evidence set. .

4. The rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection as described in claim 1, characterized in that, The knowledge-enhancing agent reflection module concatenates the following multi-source, multi-modal diagnostic features to construct structured text prompts for large language models: The patient's demographic characteristics and the HPO phenotype query set described above; Source-level diagnostic evidence set output by the source-level multi-paradigm reasoning module ; The statistical consensus priority diagnostic list output by the two-stage meta-fusion module ; Priority diagnostic list based on statistical consensus retrieved from a rare disease knowledge base. Center front Objective mapping and labeling frequency relationship between candidate diseases and the HPO phenotype query set .

5. The rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection as described in claim 1 or 4, characterized in that, The reasoning reflection of the large language model in the knowledge-enhanced agent reflection module includes: Phenotypic matching assessment: Before examination The labeled phenotypes of each candidate disease provide the widest coverage of the most specific clinical symptoms for patients, with a greater focus on rare symptoms; Knowledge base consistency check: Calculate the degree of agreement between the necessarily occurring phenotypes and the patient phenotype set, and regard the absence of non-essential phenotypes as a secondary factor that does not affect the diagnosis; Paradigm conflict resolution: Identify and rationally interpret ranking conflicts between different heterogeneous diagnostic engines, and give priority to diseases supported by highly specific cases; The final output is the reordered front. A list of candidate diagnoses .

6. The rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection as described in claim 1, characterized in that, The large language model of the knowledge-enhanced intelligent agent reflection module is a large language model that has been fine-tuned in the medical field or optimized by prompting engineering, and is deployed on a local medical diagnostic workstation or called through a secure API interface.

7. The rare disease auxiliary diagnostic system based on multi-paradigm statistical fusion and large-model intelligent agent reflection as described in claim 1, characterized in that, The deduplication and splicing module performs lossless and secure splicing to obtain the final diagnostic recommendation list. : in, Indicates from Delete The diagnosis is made while maintaining the relative order of the remaining diagnoses; Indicates will The remaining diagnoses are pieced together. after.

Images