An artificial intelligence-based genetic disease assisted decision system and method

Abstract

The application discloses a kind of genetic disease auxiliary decision system and method based on artificial intelligence domestication, system includes: data preprocessing module, for obtaining and standardization processing local clinical data, establish the structured association between clinical phenotype, gene variation, pathogenicity rating and submission item;AI core engine module, with the local clinical data after standardization as main training set, domestication obtains phenotype-gene association model, variation full information analysis model and submission item recommendation model;Interactive input module, for obtaining the clinical phenotype input of natural language format;Result output module, for outputting complete result containing gene variation information and submission item recommendation according to predetermined order.The application also provides corresponding method.The application realizes the full-process automation closed-loop output from clinical phenotype input to submission item recommendation by domesticating AI model with local data, with the advantages of complete output information, high recommendation adaptability and fast response speed.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and medical information, specifically to a genetic disease auxiliary decision-making system and method based on artificial intelligence training. Background Technology

[0002] In the field of genetic disease testing, recommending appropriate tests to clinicians is one of the core tasks of testing institutions' sales efforts. However, existing technologies have the following technical shortcomings in practical applications:

[0003] (1) The information output is incomplete and lacks clinical evidence.

[0004] Current sales-side recommendations typically only provide the name of the test item, failing to simultaneously display core clinical evidence such as gene mutation information associated with the patient's clinical phenotype, interpretation of mutation function, and pathogenicity rating. Without a complete chain of evidence, clinicians struggle to quickly determine the scientific rationale and clinical suitability of the test, resulting in unconvincing recommendations.

[0005] (2) The recommendation logic has gaps and a complete knowledge connection path has not been established.

[0006] Current technologies have failed to establish a complete logical link from "clinical phenotype input" to "gene mutation identification," then to "pathogenicity assessment," and finally to "test recommendation." The lack of a systematic matching mechanism between the recommendation results and individual patient characteristics, local patient population characteristics, and the actual capabilities of testing institutions leads to low positive test rates and low clinical satisfaction.

[0007] (3) Over-reliance on human experience leads to high communication costs and low response efficiency.

[0008] Novice sales staff or clinicians lacking a genetics background often cannot independently determine a patient's suitability for a treatment when faced with complex phenotypes, requiring repeated consultations with professional genetic counselors. This process involves time costs for the sales, clinical, and counseling teams, resulting in delayed responses and potential biases in recommendations due to information loss during communication, which can severely impact clinical collaborations.

[0009] (4) Existing AI applications are limited to a single link and have not formed an integrated closed loop.

[0010] Among the publicly disclosed patent technologies, CN121306249A only achieves the matching of phenotype and gene mutation, without involving variant interpretation and project recommendation; CN111063392A focuses on variant pathogenicity analysis, without integrating clinical phenotype and sales-end recommendation needs. None of the above technologies achieve a fully integrated closed loop from phenotype input to test item recommendation, failing to meet the comprehensive needs of medical testing sales for "data support, complete interpretation, and accurate recommendation."

[0011] In summary, there is an urgent need for an intelligent genetic disease auxiliary decision-making system and method that can integrate local clinical data, automate the entire process from phenotype analysis to project recommendation, and possess high accuracy and adaptability. Summary of the Invention

[0012] The purpose of this invention is to provide a genetic disease auxiliary decision-making system and method based on artificial intelligence domestication, so as to solve the technical problems of incomplete information output, broken recommendation logic, over-reliance on human experience, and the lack of an integrated closed loop in existing AI applications.

[0013] To achieve the above objectives, the present invention provides the following technical solution:

[0014] A genetic disease auxiliary decision-making system based on artificial intelligence domestication includes:

[0015] The data preprocessing module is configured to acquire and standardize local clinical data, and establish structured associations between clinical phenotypes, gene variations, pathogenicity ratings, and submitted test items.

[0016] The AI ​​core engine module is configured to use the standardized local clinical data as the main training set to train multiple collaborative functional models, wherein the functional models include at least:

[0017] Phenotype-gene association model for outputting candidate gene variants associated with input clinical phenotypes;

[0018] A full-information analysis model for variants is used to output the associated functional information of the candidate gene variants;

[0019] The test item recommendation model is used to output suitable test items based on preset multi-dimensional weighted rules;

[0020] An interactive input module, configured to acquire clinical phenotype input in natural language format;

[0021] The result output module is configured to output complete results generated by the AI ​​core engine module, including gene variation information and recommended test items, in a predetermined order.

[0022] As a further preferred technical solution, the AI ​​core engine module also includes:

[0023] The pathogenicity rating and comprehensive analysis model is configured to construct a three-source fusion rating system based on local ACMG rating data, public database rating data and AI prediction rating data, and output the pathogenicity level of the candidate gene variant and a comprehensive analysis text conforming to clinical format.

[0024] The pathogenicity rating and comprehensive analysis model adopts a combination architecture of nonlinear feature mining model and linear weighted fusion model to collaboratively process the three-source fusion rating data.

[0025] As a further preferred technical solution, the multi-dimensional weighted rules of the test item recommendation model include a combination of at least two of the following dimensions: phenotypic-variant association fit, variant pathogenicity weight, local historical positive rate, test item cycle, and test item cost;

[0026] The recommended model for the submitted items is configured to calculate a comprehensive score based on weighted scores of each dimension, and output at least one recommended item and its recommendation reason in descending order of the comprehensive score.

[0027] As a further preferred technical solution, it also includes:

[0028] A multi-terminal collaboration and knowledge accumulation module, configured as follows:

[0029] Provides a real-time problem feedback and anomaly alert interface between the clinical, sales, and genetic counseling ends;

[0030] The system stores historical consultation question and answer data, and uses this data as a training set to train the AI ​​core engine module, enabling it to automatically answer common genetic counseling questions.

[0031] As a further preferred technical solution, the standardization processing of the data preprocessing module includes at least one of the following: mapping clinical phenotypes to standardized phenotype terms, unifying gene variation formats to HGVS format, unifying pathogenicity ratings to ACMG five-level classification, and associating disease information with OMIM numbers;

[0032] The system also includes a private data storage module, which is configured to store the local clinical data and AI training parameters on the institution's private server, so that the AI ​​calculation and recommendation process is executed locally in a closed loop.

[0033] A genetic disease decision-making assistance method based on artificial intelligence domestication, using the aforementioned genetic disease decision-making assistance system based on artificial intelligence domestication, includes the following steps:

[0034] S1. Data Standardization: Acquire local clinical data and standardize it to construct a structured association between clinical phenotypes, gene variations, pathogenicity ratings, and submitted test items;

[0035] S2. Model domestication: Using the standardized local clinical data as the main training set, multiple collaborative functional models are domesticated to obtain the model. The functional models include at least a phenotype-gene association model, a variant full information analysis model, a pathogenicity rating and comprehensive analysis model, and a test item recommendation model.

[0036] S3, Input Acquisition: Acquire clinical phenotype input in natural language format;

[0037] S4. Result Generation and Output: Call the domesticated functional model to generate and output complete results containing gene variation information and recommended test items in a predetermined order.

[0038] As a further preferred technical solution, the adaptation process of the pathogenicity rating and comprehensive analysis model includes:

[0039] Construct a three-source fusion dataset that includes local ACMG rating data, public database rating data, and AI-predicted rating data;

[0040] The three-source fusion dataset is trained using a combination of a nonlinear feature mining model and a linear weighted fusion model, enabling the trained model to output the pathogenicity level of candidate gene variants and a comprehensive analysis text conforming to clinical format.

[0041] As a further preferred technical solution, the training process of the recommended model for submitted items includes:

[0042] Set weighting rules that include at least two of the following dimensions: phenotypic-variable association fit, variant pathogenicity weight, local historical positivity rate, testing cycle, and testing cost;

[0043] The recommended model for the submitted items is trained to calculate the comprehensive score of each submitted item based on the weighted rules, and outputs at least one recommended item and its recommendation reason in descending order of the comprehensive score.

[0044] As a further preferred technical solution, it also includes:

[0045] S5: Multi-terminal collaboration and knowledge accumulation: Provides real-time problem feedback and anomaly alert interfaces between clinical, sales and genetic counseling terminals; stores historical counseling Q&A data, and uses the historical counseling Q&A data as a training set to train the AI ​​model, enabling it to automatically answer common genetic counseling questions.

[0046] As a further preferred technical solution, in the S1 data standardization step, the standardization process includes at least one of the following: mapping clinical phenotypes to standardized phenotype terms, unifying gene variation formats to HGVS format, unifying pathogenicity ratings to ACMG five-level classification, and associating disease information with OMIM numbers;

[0047] The method also includes S6, data privatization storage: storing the local clinical data and AI training parameters on the institution's private server, so that the AI ​​calculation and recommendation process is executed locally in a closed loop.

[0048] Compared with the prior art, the beneficial effects of the present invention are: (1) Output all-dimensional information at once to build a complete chain of clinical evidence.

[0049] This invention, through the collaborative work of the AI ​​core engine module, can output a list of gene mutations, full information on the mutations, pathogenicity rating, comprehensive analysis, and recommendations for test items in one go, without requiring additional queries from clinicians, thus solving the technical problems of incomplete information output and logical gaps in recommendations in existing technologies.

[0050] (2) Based on local data training, avoid the professional defects of general AI.

[0051] This invention uses local real clinical data as the main training set to train AI, avoiding the pain points of directly using general artificial intelligence technology, such as chaotic knowledge structure, cold and mechanical output content, and lack of humanistic care. The output results are in line with the institution's testing capabilities and the characteristics of local patients.

[0052] (3) Lower the professional threshold and improve communication efficiency and response speed

[0053] After the sales representative inputs the clinical phenotype, the system can output comprehensive results within 60 seconds, eliminating the need for manual data processing and significantly improving work efficiency. Simultaneously, the system can automatically answer common genetic counseling questions, reducing redundant communication costs between sales, clinicians, and counselors.

[0054] (4) Forming a data-driven positive closed loop

[0055] The system can be continuously optimized and iterated by adding new detection data locally, thus achieving continuous self-evolution. Attached Figure Description

[0056] Figure 1 This is a flowchart of the method of the AI-driven clinical phenotype-gene variation association and test item recommendation system of the present invention.

[0057] Figure 2 This is a data processing flowchart for the entire process from clinical information input to full information output, as described in this invention.

[0058] Figure 3 This is a schematic diagram of the process of the present invention to train AI automated rating based on multi-source data to achieve automatic output and iterative update of storage;

[0059] Figure 4This is a screenshot of the system interface of the present invention, which realizes automatic extraction of phenotypes and CHPO standardized conversion based on clinical information input using AI.

[0060] Figure 5 This invention is a schematic diagram illustrating the entire process of achieving AI-based automatic analysis of variant sites and matching with ACMG guidelines based on clinical information input, realizing standardized output of full variant information analysis and recommending appropriate test items.

[0061] Figure 6 This is a schematic diagram of the real-time anomaly monitoring method for any stage from sample entry to result input according to the present invention;

[0062] Figure 7 This is a screenshot of the standardized data storage system for AI training and the AI ​​interpretation and call source interface of this invention;

[0063] Figure 8 This is a screenshot of the existing system's abnormal handling operation and the interface displaying the problems to be handled and the abnormalities in this invention. Detailed Implementation

[0064] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0065] like Figure 1 and Figure 2 As shown in the figure, this embodiment provides a genetic disease auxiliary decision-making system based on artificial intelligence domestication, including a data preprocessing module, an AI core engine module, a data storage module, an interactive input module, a result output module, and a clinical-sales-consultation docking module.

[0066] 1. System Deployment Environment

[0067] This system is deployed on a private server provided by the organization, with hardware configurations of at least 24 CPU cores, at least 128GB of memory, and at least 4TB of hard drive space to support concurrent access by multiple users and ensure the security of local data storage. The software environment runs on a Python 3.9 platform, integrating a GPT-4 fine-tuning model and a BERT association model. It includes a built-in HPO terminology database, an OMIM disease database, an HGMD database, and an ACMG guidelines rule base, supporting access from multiple terminals such as computers, mobile phones, and tablets.

[0068] 2. Data Preprocessing Module

[0069] The data preprocessing module is used to acquire local clinical data tables and perform standardized processing. The local clinical data tables include the following fields: clinical phenotype, CHPO phenotype, submitted test items, detected gene variants, variant description, gene function, related diseases and disease descriptions, variant interpretation, pathogenicity rating (ACMG standard), comprehensive analysis, and positive / negative test results.

[0070] The standardization process specifically includes: mapping clinical phenotypes to CHPO terms, using the HGVS format for gene variants, standardizing pathogenicity ratings to a five-level classification of P (pathogenicity) / LP (suspected pathogenicity) / VUS (undetermined significance) / LB (suspected benign) / B (benign), and associating disease information with OMIM numbers.

[0071] In a specific example, the data preprocessing module collects genetic disease testing data from the local database for the past three years, integrates authoritative databases such as OMIM, CHPO, ClinVar, HGMD, gnomAD, LIS system, domestic and international treatment guidelines (such as the "Guidelines for the Diagnosis and Treatment of Rare Diseases"), and expert consensus, and organizes them into standardized tables. AI training sets are then constructed for the gene-variation-variation description-corresponding disease-disease description section, ACMG rating section, comprehensive analysis section, and test item recommendation section.

[0072] 3. AI Core Engine Module

[0073] The AI ​​core engine module is the core of this system. It uses standardized local clinical data as the main training set and public big data such as OMIM, HPO, ClinVar, and PubMed as supplementary training sets to train multiple collaborative functional models.

[0074] 3.1 Phenotype-Gene Association Model

[0075] This model is used to learn the association rules of "clinical phenotype / CHPO phenotype → gene variant", outputs candidate variants that are highly correlated with the phenotype, sorts them by phenotype correlation score (0-10), and includes variants with correlation score ≥7 in the output list.

[0076] The key to model training lies in constructing standardized tabular samples. Specifically, all clinical phenotypes, corresponding CHPO terms, and gene variants from historical samples are compiled into a table, with standardized codes for both phenotypes and variants (e.g., International Disease Code ICD-10, and variant naming following the HGVS standard). Using machine learning algorithms (e.g., LightGBM or XGBoost), based on the standardized CSV table input, preprocessing, association rule mining, 0-10 score scoring, and ≥7 score filtering are performed.

[0077] In a specific example, 1000 standardized tabular samples (including data from the ClinVar database and local databases) are input for model training. The validation criteria are: input 100 new clinical phenotypes, and the AI ​​automatically outputs the CHPO phenotype, corresponding gene, standard transcript, variant description, related diseases, and disease information, with an association accuracy ≥95%. The following is an example format for the training samples:

[0078] 3.2 Full Information Analysis Model for Variation

[0079] This model is used to learn the mapping rules of "gene variation - variation description - gene function - related diseases and introduction - variation interpretation". Among them, the disease introduction includes clinical characteristics and pathogenesis; variation interpretation includes variation type, location, and potential impact.

[0080] The key to model training lies in organizing gene variants, variant descriptions, gene functions, related diseases and their descriptions, and variant interpretations from historical samples in the local database into standardized templates. A lightweight text generation model (such as BERT-base or T5-small) is used to automatically learn the template mapping rules. When a new gene variant is input, the AI ​​matches it with similar variants in the training set and automatically generates corresponding variant descriptions, gene functions, related diseases, and interpretations, ensuring that all interpretations strictly adhere to the table field specifications and avoiding the subjectivity and variability of manual interpretation. Through this model, the time for generating interpretations of new variants can be reduced from several hours to seconds.

[0081] In a specific example, a standardized table was created by inputting 1000 sets of gene pathogenesis mechanism description data. The AI ​​was then trained to automatically output the gene pathogenesis mechanism in the format of "encoded protein - physiological function - pathway" when a new gene is input. The following is an example format for the training samples: Gene Pathogenic mechanisms of genes ZNFX1 The ZNFX1 gene encodes a member of the RNA helicase superfamily, a zinc finger protein that is a mitochondrial-localized double-stranded RNA (dsRNA) sensor that can initiate an immune response against RNA virus infection (PMID: 31685995). DOCK6 The DOCK6 gene encodes a member of the atypical guanine nucleotide exchange factor (DOCK) family of cytokinase-specific factors (CSDs). Guanine nucleotide exchange factors interact with small GTPases and are part of the intracellular signaling network (PMID: 17196961). AGRN The AGRN gene encodes agarin, a large and ubiquitous proteoglycan with multiple isoforms that function differently in different tissues. Agrin was initially thought to be an important neuromodulator that induces the aggregation of acetylcholine receptors (achr) and other postsynaptic proteins on muscle fibers, and is essential for the formation and maintenance of neuromuscular junctions (NMJs) (PMID:1659950, PMID:10402191, PMID:22205389). TANC2 The TANC2 gene encodes a synaptic scaffold protein that interacts with other proteins at postsynaptic density to regulate dendritic spines and excitatory synapse formation (PMID: 31616000). SERPINB8 The SERPINB8 gene encodes protease inhibitor-8, a member of the ovoserine serine subfamily, produced by platelets. It can bind to furin and inhibit its function, furin being a serine protease involved in platelet function. PDE6B The PDE6B gene encodes the β subunit of phosphodiesterase (PDE). PDE is a peripheral membrane heterotrimeric enzyme composed of α, β, and γ subunits. Photon absorption triggers a signal cascade in rod photoreceptors, activating cGMP phosphodiesterase (PDE), leading to rapid cGMP hydrolysis, closure of cGMP-gated cation channels, and cellular hyperpolarization.

[0082] To achieve automatic AI generation, this invention employs a "rule-constraint + sequence generation model" (such as GPT-4 fine-tuning + format validation layer), embedding field format templates during text generation (e.g., forcing gene function to be segmented according to "encoded protein → physiological function → pathway"). Simultaneously, a knowledge graph and multi-label classification model are used to construct a knowledge graph from UniProt / KEGG's "gene-protein-function-pathway" data, retrieving corresponding information from the graph when the AI ​​generates gene functions. Furthermore, a literature retrieval and semantic matching model is used. When the AI ​​needs to populate information for a new gene, it automatically connects to the PubMed API, retrieving literature by "gene name + function / disease," and selecting 1-3 of the most relevant articles based on trained priority rules, generating citations in a table format. In cases of abnormal scenarios, the AI ​​marks "unclear / no literature" according to rules, without fabricating information, and provides actionable suggestions in the "Remarks" field (such as family analysis, literature retrieval time). The validation criteria are: input 100 genes from the OMIM database that are not included in the training set, and the AI ​​automatically outputs the pathogenicity of the gene (including encoded protein, physiological function, and pathway), with a completeness of ≥98%.

[0083] 3.3 Pathogenicity rating and comprehensive analysis model

[0084] The model constructs a three-source rating system, which allows AI to learn local ACMG rating, ClinVar rating and AI intelligent rating rules, and combines evidence such as mutation frequency, functional prediction and literature support to assess pathogenicity level (P / LP / VUS / LB / B).

[0085] The key to model training lies in: extracting HGVS sequences, ACMG five-level ratings, rating evidence chains, and associated phenotypes from the local database as core training data; crawling submitter ratings, evidence levels, population frequencies, and conflict annotations of gene variants from the ClinVar database as external validation data; and connecting with existing institutional intelligent systems to obtain variant functional prediction results, automated ACMG rating suggestions, homologous gene evidence, and literature indexes as new variant inference data.

[0086] In a specific example, 1000 sets of variance data containing ACMG rating evidence (including ClinVar ratings and local database ratings) are input and formatted into a standardized table for model training. The following is an example format for the training samples:

[0087] This invention employs a dual-model collaborative rating system: (1) Rule engine model: Prioritize the official evidence rules of ClinVar and the clinically recognized variant ratings in the local database to make a "hard judgment" on variant sites to ensure that the ratings comply with clinical guidelines; (2) Machine learning model: The model is trained using the ClinVar labeled variant sites to learn the association between evidence and rating. Based on the ACMG guidelines, supplementary predictions are made for sites that the rule engine cannot determine (such as VUS with insufficient evidence).

[0088] Specifically, LightGBM is used as the primary prediction model to mine nonlinear feature associations in the three-source rating data (such as the nonlinear rule of "variant located in functional domain + existing institutional AI intelligent prediction of harmful → rating LP / P"), adapting to the complex feature relationships of gene variant rating. Logistic regression is used as the fusion decision model to linearly weight and fuse the prediction results of LightGBM, the consistency features of the three-source rating, and the core evidence features, calibrating the prediction results, improving model interpretability, and meeting the clinical need for traceable rating basis. The system automatically generates structured rating basis based on model feature weights, three-source data, and variant features. This combined model takes into account both nonlinear fitting and linear calibration, and the entire process from three-source data acquisition to feature generation, model inference, and result output requires no manual intervention. The validation standard is: inputting variant data from 200 local databases not included in the training set, the automatic rating accuracy is ≥90%.

[0089] like Figure 3 As shown, the AI ​​is trained using a large amount of data to automate four steps: feature extraction, evidence matching, rule verification, and rating calculation. This results in a new ACMG rating for each variant, forming a three-tiered rating system: local database, ClinVar, and AI prediction. The automated rating results are then manually reviewed. Review results and suggestions are collected, and the feedback data is standardized to further optimize and supplement the training set, fine-tuning the AI ​​to continuously improve its accuracy and practicality.

[0090] The comprehensive analysis model generates comprehensive analyses according to a template of "variant detection → literature report → pathogenicity rating → clinical recommendations," conforming to clinical table formats. In a specific example, 800 sets of samples containing comprehensive analyses are input, and standardized tables (including phenotypic combinations, association results, and comprehensive analyses) are created for model training. The following is an example format of the training samples:

[0091] Clinical diagnosis Mutation description Comprehensive analysis Burkitt lymphoma; liver tumor; (head and face) metastatic tumor (possibly); liver dysfunction; coagulation disorders. The subject tested positive for the MMUT gene NM_000255.4:intron9:c.1677-1G>A variant. This variant, located in intron9, causes the nucleotide at position 1677 of the MMUT gene cDNA to change from guanine to adenine, which is a splicing variant. The subject also tested positive for the TSC1 gene NM_000368.5:exon6:c.416T>C:p.M139T variant. This variant, located in exon6, causes the nucleotide at position 416 of the TSC1 gene cDNA to change from thymine to cytosine, and the amino acid encoded at position 139 to change from methionine to threonine, which is a missense variant. The subject tested positive for a heterozygous variant in the MMUT gene NM_000255.4:intron9:c.1677-1G>A. This variant has been reported in multiple cases of methylmalonyl-CoA mutase deficiency-induced methylmalonic aciduria, where it is a complex heterozygote, clinically presenting with developmental arrest, hypotonia, hepatomegaly, and anemia (PMID: 16281286, PMID: 17113806), which shows some consistency with the subject's clinical presentation of liver tumors. According to the ACMG guidelines, this variant is tentatively classified as pathogenic. The subject has a heterozygous variant at this locus, which does not conform to the inheritance pattern of the corresponding disease. The subject is a carrier of this variant and theoretically will not develop the disease. The physician should make a comprehensive assessment based on the subject's clinical presentation, family history, and other test results. The subject also tested positive for a heterozygous variant in the TSC1 gene NM_000368.5:exon6:c.416T>C:p.M139T. No relevant literature reports have been published regarding this variant. The disease caused by this gene abnormality, lymphangioleiomyomatosis, shows some consistency with the clinical presentation of Burkitt lymphoma in the subject. According to the ACMG guidelines, this variant is currently classified as of uncertain significance. A heterozygous variant was found at this locus in the subject, consistent with the inheritance pattern of the corresponding disease. It is recommended to comprehensively evaluate the clinical phenotype of the subject's mother and to make a comprehensive assessment based on the subject's clinical phenotype, family history, and other test results. hemolytic anemia The subject was found to have the SPTB gene NM_001355436.2:exon22:c.4479G>A:p.W1493* variant. This variant is located at exon22 and causes the 4479th nucleotide of the SPTB gene cDNA to change from guanine to adenine, and the 1493rd amino acid it encodes to change from tryptophan to a stop codon. This is a nonsense variant. The subject tested positive for a heterozygous variant in the SPTB gene NM_001355436.2:exon22:c.4479G>A:p.W1493*. No relevant literature has been reported on this variant. Diseases caused by this gene abnormality, such as elliptic polycythemia type 3 / anemia, neonatal hemolytic disease, lethal or near-lethal hemolytic disease, and spherocytosis, show some consistency with the clinical presentation of hemolytic anemia in the subject. According to the ACMG guidelines, this variant is tentatively classified as suspected pathogenic. Clinical observation of this mutation is recommended, with a comprehensive evaluation based on the subject's clinical phenotype, family history, and other test results.

[0092] To achieve AI-generated comprehensive analysis, this invention employs a "template constraint + GPT-4 fine-tuning" approach to ensure that the analysis text includes four core elements: "phenotypic matching degree, pathogenicity of variants, literature evidence, and clinical recommendations." The validation standard is: inputting data from 100 local databases not included in the training set, the AI ​​automatically outputs a comprehensive analysis text with ≥92% completeness.

[0093] 3.4 Recommendation Model for Submitted Items

[0094] The model is used to learn multi-dimensional weighted scoring rules and recommend the top 3 test items in descending order of comprehensive score. Specifically, the weight settings include: (1) phenotypic-variable fit (35%); (2) pathogenicity weight (30%); (3) local positive rate (25%); and (4) testing cost-effectiveness (10%).

[0095] Project Name Specimen type Clinical diagnosis Gene Yin and Yang Genetic screening for bone marrow failure diseases swab Aplastic anemia? RPS19 Positive Comprehensive Hematologic Oncology Surgical Edition swab 1. Hemolytic anemia? 2. Acute bronchitis 3. G6PD deficiency (favism) 4. Liver damage G6PD Positive Trios - Professional Whole-Exome Analysis of Genetic Diseases swab Bloodline reduction RAP1B Positive Comprehensive Hematologic Oncology Surgical Edition swab Thrombocytopenia PTPN11 Positive Comprehensive Hematologic Oncology Surgical Edition swab 1. Immune thrombocytopenic purpura 2. Acute bronchitis 3. Respiratory syncytial virus infection 4. Patent foramen ovale WAS Positive Trios - Professional Whole-Exome Analysis of Genetic Diseases swab Enlarged lymph nodes, splenomegaly (cause unknown), lymphoma?, lymphoproliferative disorder?, autoimmune lymphoproliferative syndrome?, recovery period from pneumonia, rhinitis PIK3CD Positive Comprehensive Analysis of Hematologic Oncology swab 1. Neutropenia; 2. (Right-sided) Lower extremity edema. GATA2 Positive Trios - Professional Whole-Exome Analysis of Genetic Diseases swab 1. Acute myeloid leukemia (M2 type); 2. Bronchopneumonia; 3. Pulmonary consolidation SMC1A Positive Trios - Professional Whole-Exome Analysis of Genetic Diseases swab 1. Acute lymphoblastic leukemia (ordinary B, BCR::ABL1(p190), intermediate risk); 2. Pneumonia; 3. Severe influenza; 4. Renal impairment; 5. Hyperuricemia; 6. Electrolyte imbalance; 7. Secondary coagulation disorders; 8. Folic acid deficiency; 9. Vitamin D deficiency; 10. Epstein-Barr virus infection; PAX5 Positive

[0096] To achieve AI-automated recommendation of test items, this invention combines a biomedical pre-trained model with a recommendation algorithm, incorporating a weighted ranking module (integrating expert recommendation scores, positive rates, and item priority) to recommend a top 3 items (ranked by priority) based on the input clinical phenotype. The preferred weighting scheme is as follows: (1) Clinical priority (guideline recommendation / positive rate): 45%; (2) Cycle adaptability: accounting for 20% (3) Cost suitability: accounting for 25%; (4) Big data recommendation: accounting for 10%.

[0097] AI, based on the statistical patterns of standardized tables, calculates positive rates across multiple dimensions to meet the needs of various recommendation scenarios in sales. The core calculation formula is:

[0098] Positive rate = (Number of positive tests for this item under certain conditions / Total number of tests for this item under certain conditions) × 100%.

[0099] The calculation dimensions may include: (1) Disease dimension: the positive rate of a certain item submitted for testing for a specific disease; (2) Phenotypic dimension: the positive rate of a certain item submitted for testing for a certain clinical phenotype / phenotype combination; (3) CHPO standardization dimension: the positive rate of a certain item submitted for testing for a certain CHPO terminology combination.

[0100] 4. Data storage module

[0101] like Figure 7 As shown, all standardized forms used for AI training are stored in the corresponding storage module of this system for automated system retrieval and AI automatic retrieval. This module's data is regularly maintained and updated as the sample size and database are updated.

[0102] 5. Interactive Input Module

[0103] The interactive input module supports sales personnel to input clinical phenotypes in natural language format in various ways (voice, images, medical records, etc.), and is compatible with single-phenotype and multi-phenotype combination input. AI automatically converts the input content into standardized CHPO phenotypes.

[0104] 6. Result Output Module

[0105] like Figure 5 As shown, after completing the standardization and storage of all data and training the AI, when sales inputs new clinical phenotypes or genes of focus, the AI ​​can automatically extract and retrieve the data. For missing data in the storage module, the AI ​​can automatically output corresponding results based on the training model. The output results are presented in the order of "gene variant list → full variant information → pathogenicity rating → comprehensive analysis → recommended test items," with a format adapted to clinical documents and supporting one-click export.

[0106] 7. Module for multi-terminal adaptation (clinical, sales, and consultation), problem communication, information updates, and anomaly maintenance.

[0107] like Figure 6 and Figure 8 As shown, this module consists of two parts: a consultation module for connecting with and answering clinical or sales questions, and an exception handling module for raising and handling any exceptions during the process from sample submission to report issuance. All questions and exceptions are displayed on the notification dashboard for real-time monitoring. If a problem occurs in the corresponding module, a notification will be displayed on the reporting system login interface, showing the consultation question and the exception pending processing.

[0108] The consultation module stores previously processed genetic counseling questions. When a sales or clinical patient inputs a question, the AI ​​automatically retrieves the corresponding answer. If the answer cannot be retrieved, the AI ​​can automatically reply using big data and label it as "AI Automatic Reply Content." The counselor then reviews and modifies the answer, adding the revised content to the stored table for continuous iteration and updating. The anomaly handling module, from front-end to back-end, promptly detects and resolves any non-compliance with processing procedures at each stage (such as sample contamination, sample mixing, or discrepancies between sample type and submitted testing requirements). This prevents sample flow from being terminated due to problems at the back-end, thus saving manpower, shortening the cycle, and minimizing losses.

[0109] The following is a sample format for common genetic counseling questions: question answer Why are the results of detecting FECH:NM_000140.5:intron3:c.315-48T>C (homozygous or heterozygous) not interpreted or reported? Most FECH-associated EPP patients carry this submorphic c.333-48T>C variant, which has a rare loss-of-function allele, resulting in a >70% reduction in enzyme activity. In contrast, individuals carrying the c.333-48T>C variant in a homozygous state are usually unaffected because their enzyme activity levels are insufficient to cause disease. In summary, although this variant is common in the general population, it meets the criteria for pathogenicity of autosomal recessive EPP; however, it should be noted that this variant only causes disease in complex heterozygosity with a loss-of-function allele. Why are the detections of homozygous and heterozygous UGT1A1:NM_000463.3:exon1:c.211G>A:p.G71R not interpreted or reported? Literature reports that heterozygous and homozygous variants of this mutation possess 71% and 22% wild-type enzyme activity, respectively (PMID: 35415244). Furthermore, literature reports that Gilbert syndrome is a benign disease and therefore generally requires no specific treatment (PMID: 22160004). Individual phenotypes vary. Homozygous mutations do not result in a complete loss of UGT1A1 enzyme activity. Why is the detection of GJB2:NM_004004.6:exon2:c.109G>A:p.V37I heterozygote not interpreted and reported? GJB2 is generally reported in literature as being caused by AR, which does not conform to the mode of inheritance. This is listed in the appendix as a suggestion and will not be interpreted further. Are genetic diseases always passed on to the next generation? Not necessarily. Genetic diseases have diverse modes of inheritance, and some diseases can be caused by de novo mutations even without a family history. For example, most chromosomal disorders (such as Down syndrome) are caused by de novo mutations unrelated to parental inheritance; the proportion of de novo mutations is also relatively high in dominant single-gene disorders. Therefore, even without a family history, the risk of offspring inheriting a genetic disease cannot be completely ruled out.

[0110] This embodiment provides a genetic disease auxiliary decision-making method based on artificial intelligence domestication. The method uses the system described in Embodiment 1 and includes the following steps:

[0111] S1. Data standardization steps: Acquire local clinical data and perform standardization processing to construct structured associations between clinical phenotypes, gene variations, pathogenicity ratings, and submitted test items.

[0112] S2. Model domestication steps: Using standardized local clinical data as the main training set, multiple collaborative functional models are domesticated. The functional models include at least a phenotype-gene association model, a variant full information analysis model, a pathogenicity rating and comprehensive analysis model, and a test item recommendation model.

[0113] S3. Input Acquisition Steps: Acquire clinical phenotype input in natural language format.

[0114] S4. Result Generation and Output Steps: Call the domesticated functional model, generate and output complete results containing gene variation information and recommended test items in a predetermined order.

[0115] The workflow of this invention will be explained below with reference to a specific case (all steps prioritize retrieving data stored in the local system; if the local database does not have the data, the AI ​​will output the corresponding information based on the training model).

[0116] Step 1: Clinical Information Input

[0117] Sales staff are required to input clinical information in various ways (voice, images, medical records, etc.), such as "thrombocytopenia, splenomegaly".

[0118] Step 2: AI Standardization Processing

[0119] The AI ​​standardized the input as: thrombocytopenia (HP:0001873); splenomegaly (HP:0001744).

[0120] Step 3: Phenotype-gene variation association analysis

[0121] AI calls a phenotypic-gene association model and outputs candidate variants (sorted by phenotypic association degree):

[0122] KIF23:NM_001367805.3:exon15:c.1489T>C:p.C497R (Relevance score 9.2);

[0123] ANKRD26:NM_014915.3:exon28:c.3986A>C:p.K1329T (Relevance score 8.5).

[0124] Step 4: Analysis of complete mutation information

[0125] The AI ​​invokes a full-information mutation analysis model to output detailed information for each mutation. Taking the KIF23 gene as an example:

[0126] Variance description: Missense variant, located at exon 15, which causes the nucleotide at position 1489 of the KIF23 gene cDNA to change from thymine to cytosine, and the amino acid encoded at position 497 to change from cysteine ​​to arginine;

[0127] Gene function: The KIF23 gene encodes mitotic kinase-like protein 1 (MKLP1), which is crucial for the formation of the central spindle and mesosome during cytokinesis;

[0128] Related diseases and introduction: Congenital erythropoiesis-disordered anemia type IIIA (OMIM:105600), a rare autosomal dominant inherited blood disease characterized by non-progressive mild to moderate hemolytic anemia and peripheral blood macrocytosis.

[0129] Step 5: Pathogenicity rating

[0130] The AI ​​invoked the pathogenicity rating model and output the ACMG rating and basis for the variant: VUS (Various Unknown Significance), based on the extremely low frequency of the variant in databases such as ESP / 1000 Genomes (PM2_Supporting).

[0131] Step 6: Comprehensive Analysis and Generation

[0132] The AI-generated comprehensive analysis model generated a standardized comprehensive analysis text: "The sample tested positive for the KIF23 gene NM_001367805.3:exon15:c.1489T>C:p.C497R variant, which has no reported literature. The disease caused by this gene abnormality, congenital erythropoiesis type IIIA, is consistent with the patient's clinical manifestations of thrombocytopenia and splenomegaly. According to the ACMG guidelines, this variant is temporarily classified as VUS. Further family validation is recommended, along with a comprehensive assessment based on clinical phenotype and family history."

[0133] Step 7: Recommended Test Items

[0134] AI calls the submission item recommendation model and outputs a Top 3 recommendation:

[0135] Project 1: Based on the patient's symptoms of "thrombocytopenia + splenomegaly," and weighted calculations considering clinical priority, timeframe, and cost, AI prioritizes "Genetic Platelet Disorders NGSpanel" (overall score 89.3), recommending peripheral blood sample testing. This project is strongly recommended by the 2022 ASH guidelines, with a positive rate of 82.5% for similar symptoms, results available in 7 days, and a cost of 3200 yuan, which aligns with clinical standards and fits within a typical budget.

[0136] Project 2: If myeloproliferative disorders are suspected (especially in adult patients), the "Myeloproliferative Disorder Gene Panel" (overall score 75.6) can be selected, and it is recommended to send bone marrow fluid for testing. This project covers core genes such as JAK2, with a positive rate of 76.3%, results are available in 10 days, and the cost is 4500 yuan.

[0137] Project 3: If previous tests were negative and the phenotype is complex (such as combined developmental abnormalities), whole-exome sequencing (overall score 52.1) can be considered. It is recommended to submit a peripheral blood sample for testing. This project can screen for rare gene mutations, with a positive rate of 35.2%, and results are available in 14 days.

[0138] Step 8: Exception Handling and Multi-Terminal Collaboration

[0139] The sales team successfully recommended NGSpanel for hereditary platelet disorder testing to the clinicians. However, due to difficulties in sample collection or other circumstances, the clinicians did not submit peripheral blood for testing, instead submitting blank samples or bone marrow. Upon receiving the samples, the front-end team promptly reported the anomaly in the system. The sales team then reviewed the anomaly and communicated with the clinicians to determine whether to use the submitted sample for testing or reject the test and resubmit peripheral blood. After multiple anomaly alerts, both the sales and clinicians became more aware of the testing requirements when submitting samples, enabling them to confirm and complete the process at the front-end, successfully pushing the project forward.

[0140] This embodiment further ensures data security and scalability, specifically including:

[0141] Local data privatization: All core training data (local tables) are stored on the institution's private server. AI calculations and recommendations are all performed locally, avoiding the leakage of patient clinical data and complying with medical data security regulations.

[0142] The functionality is expandable: if new fields are added to the table later (such as "testing cost" or "result turnaround time"), they can be directly added to the AI ​​training set to quickly implement new functions such as "sorting and recommending by cost / turnaround time".

[0143] Multi-terminal adaptation: Supports sales terminals to be used on multiple devices such as computers, mobile phones, and tablets, adapting to different work scenarios such as field visits and in-hospital communication.

[0144] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A genetic disease auxiliary decision-making system based on artificial intelligence domestication, characterized in that, include: The data preprocessing module is configured to acquire and standardize local clinical data, and establish structured associations between clinical phenotypes, gene variations, pathogenicity ratings, and submitted test items. The AI ​​core engine module is configured to use the standardized local clinical data as the main training set to train multiple collaborative functional models, wherein the functional models include at least: Phenotype-gene association model for outputting candidate gene variants associated with input clinical phenotype; A full-information analysis model for variants is used to output the associated functional information of the candidate gene variants; The test item recommendation model is used to output suitable test items based on preset multi-dimensional weighted rules; An interactive input module, configured to acquire clinical phenotype input in natural language format; The result output module is configured to output complete results generated by the AI ​​core engine module, including gene variation information and recommended test items, in a predetermined order.

2. The system according to claim 1, characterized in that, The AI ​​core engine module also includes: The pathogenicity rating and comprehensive analysis model is configured to construct a three-source fusion rating system based on local ACMG rating data, public database rating data and AI prediction rating data, and output the pathogenicity level of the candidate gene variant and a comprehensive analysis text conforming to clinical format. The pathogenicity rating and comprehensive analysis model adopts a combination architecture of nonlinear feature mining model and linear weighted fusion model to collaboratively process the three-source fusion rating data.

3. The system according to claim 1 or 2, characterized in that, The multi-dimensional weighting rules of the test item recommendation model include a combination of at least two of the following dimensions: phenotypic-variant association fit, variant pathogenicity weight, local historical positive rate, test item cycle, and test item cost; The recommended model for the submitted items is configured to calculate a comprehensive score based on weighted scores of each dimension, and output at least one recommended item and its recommendation reason in descending order of the comprehensive score.

4. The system according to claim 1, characterized in that, Also includes: A multi-terminal collaboration and knowledge accumulation module, configured as follows: Provides a real-time problem feedback and anomaly alert interface between the clinical, sales, and genetic counseling ends; The system stores historical consultation question and answer data, and uses this data as a training set to train the AI ​​core engine module, enabling it to automatically answer common genetic counseling questions.

5. The system according to claim 1, characterized in that, The standardization process of the data preprocessing module includes at least one of the following: mapping clinical phenotypes to standardized phenotype terms, unifying gene variation formats to HGVS format, unifying pathogenicity ratings to ACMG five-level classification, and associating disease information with OMIM numbers; The system also includes a private data storage module, which is configured to store the local clinical data and AI training parameters on the institution's private server, so that the AI ​​calculation and recommendation process is executed locally in a closed loop.

6. A genetic disease auxiliary decision-making method based on artificial intelligence domestication, characterized in that, The genetic disease auxiliary decision-making system based on artificial intelligence domestication as described in any one of claims 1 to 5 is used, and includes the following steps: S1. Data Standardization: Acquire local clinical data and standardize it to construct a structured association between clinical phenotypes, gene variations, pathogenicity ratings, and submitted test items; S2. Model domestication: Using the standardized local clinical data as the main training set, multiple collaborative functional models are domesticated to obtain the model. The functional models include at least a phenotype-gene association model, a variant full information analysis model, a pathogenicity rating and comprehensive analysis model, and a test item recommendation model. S3. Input Acquisition: Acquire clinical phenotype input in natural language format; S4. Result Generation and Output: Call the domesticated functional model to generate and output complete results containing gene variation information and recommended test items in a predetermined order.

7. The method according to claim 6, characterized in that, The adaptation process of the pathogenicity rating and comprehensive analysis model includes: Construct a three-source fusion dataset that includes local ACMG rating data, public database rating data, and AI-predicted rating data; The three-source fusion dataset is trained using a combination of a nonlinear feature mining model and a linear weighted fusion model, enabling the trained model to output the pathogenicity level of candidate gene variants and a comprehensive analysis text conforming to clinical format.

8. The method according to claim 6 or 7, characterized in that, The training process for the recommended model for the submitted items includes: Set weighting rules that include at least two of the following dimensions: phenotypic-variable association fit, variant pathogenicity weight, local historical positivity rate, testing cycle, and testing cost; The recommended model for the submitted items is trained to calculate the comprehensive score of each submitted item based on the weighted rules, and outputs at least one recommended item and its recommendation reason in descending order of the comprehensive score.

9. The method according to claim 6, characterized in that, Also includes: S5: Multi-terminal collaboration and knowledge accumulation: Provides real-time problem feedback and anomaly alert interfaces between clinical, sales and genetic counseling terminals; stores historical counseling Q&A data, and uses the historical counseling Q&A data as a training set to train the AI ​​model, enabling it to automatically answer common genetic counseling questions.

10. The method according to claim 6, characterized in that, In the S1 data standardization step, the standardization process includes at least one of the following: mapping clinical phenotypes to standardized phenotype terms, unifying gene variation formats to HGVS format, unifying pathogenicity ratings to ACMG five-level classification, and associating disease information with OMIM numbers; The method also includes S6, data privatization storage: storing the local clinical data and AI training parameters on the institution's private server, so that the AI ​​calculation and recommendation process is executed locally in a closed loop.

Images