WO2024060508A1 - 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法 - Google Patents

知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法 Download PDF

Info

Publication number
WO2024060508A1
WO2024060508A1 PCT/CN2023/077386 CN2023077386W WO2024060508A1 WO 2024060508 A1 WO2024060508 A1 WO 2024060508A1 CN 2023077386 W CN2023077386 W CN 2023077386W WO 2024060508 A1 WO2024060508 A1 WO 2024060508A1
Authority
WO
WIPO (PCT)
Prior art keywords
disease
phenotype
prevalence
dialogue
information
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/CN2023/077386
Other languages
English (en)
French (fr)
Inventor
李昊旻
杨建�
段会龙
舒强
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Zhejiang University
Original Assignee
Zhejiang University
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Zhejiang University filed Critical Zhejiang University
Publication of WO2024060508A1 publication Critical patent/WO2024060508A1/zh
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/20—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for computer-aided diagnosis, e.g. based on medical expert systems
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16H—HEALTHCARE INFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR THE HANDLING OR PROCESSING OF MEDICAL OR HEALTHCARE DATA
    • G16H50/00—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics
    • G16H50/70—ICT specially adapted for medical diagnosis, medical simulation or medical data mining; ICT specially adapted for detecting, monitoring or modelling epidemics or pandemics for mining of medical data, e.g. analysing previous cases of other patients
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
    • Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
    • Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation

Definitions

  • the present invention relates to the technical field of intelligent auxiliary diagnosis, and in particular to a knowledge-driven visual question-and-answer auxiliary differential diagnosis system and method for rare diseases.
  • Rare diseases also known as orphan diseases, are those with extremely low incidence rates.
  • the World Health Organization (WHO) defines rare diseases as those with a prevalence of 6.5 to 10 per 10,000 people.
  • the "China Rare Disease Definition Research Report 2021” includes "diseases with an incidence rate of neonates less than 1/10,000, a prevalence rate less than 1/10,000, and the number of affected people less than 140,000" as rare diseases. There are nearly 7,000 internationally recognized rare diseases, accounting for approximately 10% of human diseases.
  • HPO Human Phenotype Ontology
  • the present invention provides a knowledge-driven visual question-and-answer auxiliary differential diagnosis system for rare diseases.
  • the system can collect additional symptoms beyond self-report by dialogue with patients/doctors and use visualization technology to display auxiliary information for differential diagnosis. Further, Alleviating the difficulty of diagnosing rare diseases.
  • a visual question-and-answer auxiliary differential diagnosis system for rare diseases including:
  • the front-end UI module is used for human-computer interaction, collecting basic patient information and phenotypic information to obtain the patient's initial self-report, and using multi-dimensional visualization views to display relevant information for clinical staff to make differential diagnoses;
  • the dialogue diagnosis service module provides phenotype inquiry and recommendation services, disease diagnosis recommendation services and knowledge base query services through API (Application Programming Interface); the phenotype inquiry and recommendation service recommends appropriate phenotypes based on the current dialogue status to achieve a higher level of efficiency.
  • Dialogue benefits Users can choose to answer "exist”, “non-exist” or “unsure” for the phenotype asked by the system; the disease diagnosis recommendation service is based on the patient's initial self-report collected by the front-end UI module and new information generated during the dialogue process Phenotypic information recommends appropriate diseases; the knowledge base query service is to query and call the knowledge stored in the knowledge storage module based on the needs of the phenotype query recommendation service and disease diagnosis recommendation service;
  • the knowledge storage module is used to store the human phenotype ontology knowledge base and disease-phenotype relationship knowledge base.
  • HPO Human Phenotype Ontology
  • HP:0000118-Phenotypic abnormality and its descendant phenotypes each abnormal phenotype has a unique HPO code.
  • the disease-phenotype relationship knowledge base is the authoritative rare disease knowledge base Orphanet.
  • the authoritative rare disease knowledge base Orphanet includes disease prevalence, average age of onset and death, disease phenotype frequency characteristics and disease-causing gene information.
  • the knowledge storage module provides knowledge drive for the dialogue diagnosis service module.
  • Front-end UI modules include:
  • Patient information collection submodule including patient basic information view and phenotype information entry view
  • the visual question-and-answer differential diagnosis sub-module, disease assessment view, conversation status view, patient phenotype status view, conversation history view and question and answer view uses multi-dimensional visual views to display relevant information for clinical personnel to perform differential diagnosis.
  • the patient information collection sub-module includes a patient basic information view and a phenotype information entry view.
  • users can enter information such as patient age, gender at birth, positive phenotype, and negative phenotype.
  • the question-and-answer differential diagnosis sub-module includes:
  • Disease assessment view used to display disease prevalence, posterior probabilities, and phenotype matching
  • the dialogue state view displays the posterior probability (labeled as P), phenotype matching (labeled as M), and the uncertainty measure of disease recommendation results (labeled as information entropy, E, of the disease with the greatest advantage) in the form of a radar chart.
  • the Gini index label is G);
  • the patient's phenotypic status view is a sunburst diagram that reflects the patient's phenotypic characteristics (positive, negative, uncertain) based on the human phenotype ontology and using annotation propagation rules;
  • the dialogue history view and question and answer view are used to display the phenotypic information asked by the system and provide users with feedback buttons (exist, non-exist, unsure).
  • the question-and-answer differential diagnosis sub-module displays relevant information in a multi-dimensional visual view for clinical personnel to make differential diagnoses.
  • the dialogue diagnosis service module provides phenotype inquiry and recommendation services, disease diagnosis recommendation services and knowledge base query services through API (Application Programming Interface).
  • Disease recommendation services include disease prior probability estimation and disease posterior probability calculation, and dialogue diagnosis
  • the service module recommends the disease with the highest posterior probability at the end of the conversation;
  • Methods for estimating the prior probability of disease include:
  • Prob pretest (D) represents the prior probability of disease D (Pretest Probability)
  • PointPrev D represents the point prevalence of disease D (Point Prevalence)
  • ⁇ PointPrev represents the sum of the point prevalence of all rare diseases
  • Methods for calculating the posterior probability of disease include:
  • P represents the collected phenotype set
  • (p 1 , p 2 ,..., p n ) ⁇ P represents all rare diseases except rare disease D
  • D) represents the frequency of occurrence of phenotype p i in disease D
  • the composite likelihood ratio of phenotype set P is calculated as the likelihood of a single phenotype p i product of ratios
  • Odds posterior (P,D) Odds pretest (D)*CompositeLR(P,D)#(4)
  • Odds pretest (D) represents the prior odds of winning rare disease D
  • step (a-1) the method for uniformly converting the disease epidemic type into point prevalence is:
  • the entry whose prevalence geographical area is "worldwide” is preferred, and the numerical point prevalence or category point prevalence of the entry is used. the median of ; if the prevalence geographic area does not have a "worldwide” entry but does have a country-based prevalence entry, the point prevalence average of the country-based prevalence entry is used as the point prevalence of the disease;
  • the epidemic type of the disease is annual incidence
  • the average age of onset and age of death of the disease are obtained from the Orphanet database, and the duration of the disease is taken from the age of onset to the age of death.
  • the point prevalence of the disease is estimated by multiplying the annual incidence rate by the duration of the disease;
  • the epidemic type of the disease is birth prevalence, and the disease onset age is in the neonatal period, the birth prevalence multiplied by the proportion of the population in the age group of the disease in the world's total population is used as the point prevalence of the disease;
  • the present invention applies annotation relay rules on the human phenotype ontology when calculating the phenotype likelihood ratio, making the analysis of disease-phenotype relationships more efficient.
  • the annotation relay rule is: a patient's positive phenotype can be inferred that the phenotype of his ancestors is also positive in the patient, and a patient's negative phenotype can be inferred that the phenotypes of his descendants are all negative in the patient.
  • the phenotype recommendation service selects the phenotype with the greatest expected benefit for query by the system based on the current disease posterior probability distribution and disease-phenotype annotation relationship.
  • Phenotype recommendation services include:
  • Phenotype information gain (Info Gain) calculation including:
  • W is a weight set
  • w i represents the weight of a single element in the weight set W
  • W sum represents the sum of the weights of all elements in W
  • Entropy cur (D) Entropy( ⁇ Prob posterior (d)
  • D represents the set of all rare diseases
  • Prob posterior (d) represents the posterior probability of disease d
  • the first item is the information entropy of the posterior probability of disease in the current dialogue state
  • the second item is the conditional entropy when phenotype p is positive
  • the third item is the conditional entropy when phenotype p is negative
  • D represents all A set of rare diseases
  • Prob posterior (d) represents the posterior probability of disease d
  • ⁇ is the positive phenotype branch weight, that is, the sum of ⁇ Prob posterior (d)*Pr(p
  • d) represents the frequency of phenotype p in disease d
  • Gini index calculation for phenotypes including:
  • W is a weight set
  • w i represents the weight of a single element in the weight set W
  • W sum represents the sum of the weights of all elements in W
  • Gini cur (D) Gini( ⁇ Prob posterior (d)
  • D represents the set of all rare diseases
  • Prob posterior (d) represents the posterior probability of disease d
  • the first item is the Gini index of the posterior probability of disease in the current dialogue state
  • the second item is the conditional Gini index when phenotype p is positive
  • the third item is the conditional Gini index when phenotype p is negative
  • the remaining parameters Define equivalent formula (8)
  • AIGGI Adaptive Information Gain and Gini Index
  • D represents the set of all rare diseases
  • p represents a single phenotype
  • ⁇ reflects the uncertainty or disorder of the disease posterior probability distribution in the current dialogue state
  • ⁇ is represented by Gini cur (D);
  • the system selects the phenotype with the largest fusion measure AIGGI for questioning based on the current conversation status, and the user selects an answer of "exist”, “does not exist” or “unsure” for this phenotype.
  • annotation broadcast rules are applied to filter out some phenotypes when asking for phenotype recommendations, including:
  • the default ancestor phenotypes are all positive, and the ancestor phenotypes will be filtered out and no further questions will be asked;
  • the present invention also provides a method for auxiliary differential diagnosis using a rare disease visual question-and-answer auxiliary differential diagnosis system, including:
  • the user enters patient information in the patient information collection sub-module of the front-end UI module, including patient age, gender at birth, positive phenotype and negative phenotype;
  • the dialogue diagnosis service module selects the phenotype with the largest fusion measure AIGGI according to the current dialogue status for inquiry. After the user selects the answer of "exist”, “does not exist” or “uncertain” for this phenotype, it then enters the next round. dialogue;
  • users can obtain a variety of visual information including disease priority assessment, current dialogue status, patient phenotypic status, and dialogue history in the visual question-and-answer differential diagnosis sub-module of the front-end UI module;
  • step (2) Repeat step (2) until the uncertainty of the disease recommendation result is lower than the threshold or reaches the maximum dialogue round, the system gives disease recommendations based on the disease posterior probability and ends the dialogue.
  • the core of this invention is a disease recommendation service.
  • the core of its algorithm is how to obtain the most effective phenotype question and answer among tens of thousands of phenotypes and approach the real diagnosis with the optimal path.
  • One of the important contributions is the application of a method for calculating disease posterior probability based on Bayesian medical testing to the disease recommendation service in the dialogue diagnosis service module, which includes two steps: disease prior probability estimation and disease posterior probability calculation. .
  • the present invention designs a conversion method for different epidemic types, homogenizes heterogeneous disease epidemic knowledge into point prevalence, and calculates the disease prior probability based on it.
  • the (Likelihood Ratio) indicator is a measure of the contribution of a single phenotypic condition to disease priority. It is multiplied to obtain the composite likelihood ratio of all phenotypic conditions, and then the posterior probability of the disease based on the phenotype is obtained.
  • Another core of the present invention is a phenotype recommendation method that adaptively fuses information gain and Gini index, which is applied to the phenotype inquiry and recommendation service in the dialogue diagnosis service module.
  • the purpose of this method is to select the phenotype with the greatest expected benefit to be interrogated by the system based on the current dialogue state, that is, the disease posterior probability distribution and the disease-phenotype annotation relationship.
  • This method includes three aspects: phenotypic information gain (Info Gain) calculation, Gini index gain (Gini index Gain) calculation, and adaptive fusion measure AIGGI (Adaptive Information Gain and Gini Index) calculation of the two.
  • This invention uses the disease posterior probability as the classification weight to calculate the information gain and Gini index gain of a single phenotype, adaptively fuses the two according to the uncertainty of the disease recommendation results, and the system recommends the phenotype with the largest AIGGI as the query object.
  • the present invention adopts the form of phenotypic question and answer to actively collect additional patient symptom information through human-computer interaction.
  • doctors can be inspired to notice previously missed patient phenotypes during the dialogue process, ultimately achieving more effective disease diagnosis.
  • the present invention designs a method for calculating the posterior probability of a disease based on Bayesian medical testing. This method applies annotation propagation rules on the human phenotype ontology to improve the robustness of disease recommendations.
  • Prevalence is a very important factor affecting disease diagnosis, but most disease recommendation methods do not consider prevalence. Aiming at the problem of heterogeneity of disease epidemic knowledge, this invention designs a set of epidemic type conversion calculation methods to unify disease epidemic types into point prevalence rates to achieve homogeneity, and calculate based on the point prevalence rates. Disease prior probabilities, which are applied to disease recommendation methods.
  • the present invention designs a phenotype recommendation method that adaptively fuses information gain and Gini index. According to the degree of uncertainty of disease recommendation results, the two measures of information gain and Gini index gain of phenotypic conditions are adaptively fused to obtain a more efficient phenotypic question and answer.
  • Figure 1 is a schematic diagram of the module structure of the visual question-and-answer auxiliary differential diagnosis system for rare diseases
  • Figure 2 is a radar chart of the status of visual question-and-answer auxiliary differential diagnosis of rare diseases
  • Figure 3 is a schematic diagram of the question and answer flow chart of the visual question and answer auxiliary differential diagnosis system for rare diseases.
  • the present invention designs a knowledge-driven conversational diagnosis system for rare diseases, which can collect additional phenotypes beyond their self-reports through dialogue with users (patients/family members/doctors), and automatically carry out assessments and give disease diagnoses based on this. .
  • the system uses the Orphanet rare disease knowledge base and the human phenotype ontology as prior knowledge. After the initial patient information is entered, the system performs phenotypic inquiry. The system selects the phenotype with the greatest expected benefit based on the real-time dialogue status, and the user makes a judgment. Iterate the above process to achieve multiple rounds of question and answer, thereby enriching the collection of structured phenotypic data of patients and improving disease diagnosis.
  • the rare disease question-and-answer dialogue diagnosis system designed by the present invention includes a front-end UI (User Interface) module, a dialogue diagnosis service module and a knowledge storage module, as shown in FIG1 , wherein:
  • the front-end UI module is used for human-computer interaction, specifically including the patient information collection sub-module and the visual question-and-answer differential diagnosis sub-module.
  • the patient information collection sub-module includes a patient basic information view and a phenotype information entry view. In this module, users can enter information such as patient age, gender at birth, positive phenotype, and negative phenotype.
  • the question-and-answer differential diagnosis sub-module includes disease assessment view, conversation status view, patient phenotype status view, conversation history view, and question and answer view.
  • the disease assessment view displays disease prevalence, posterior probability, and phenotypic matching;
  • the conversation status view displays the posterior probability (labeled as P) and phenotypic matching (labeled as M) of the currently most advantageous disease in the form of a radar chart. ) and the uncertainty measure of disease recommendation results (the information entropy label is E, the Gini index label is G), as shown in Figure 2.
  • the status view is a sunburst diagram drawn based on the human phenotype ontology that reflects the patient's phenotypic characteristics (positive, negative, uncertain), and uses annotation propagation rules;
  • the question and answer view is the core interactive module of the dialogue diagnosis sub-module, displaying system queries Phenotypic information and provide users with feedback buttons (presence, absence, uncertainty).
  • the question-and-answer differential diagnosis sub-module displays relevant information in a multi-dimensional visual view for clinical personnel to make differential diagnoses.
  • the dialogue diagnosis service module mainly provides phenotype inquiry and recommendation services, disease diagnosis recommendation services and knowledge base query services through API (Application Programming Interface).
  • the phenotype query recommendation service recommends appropriate phenotypes based on the current conversation status to achieve higher Dialogue benefits, users can choose to answer "exist", “not present” or “unsure” for the phenotype asked by the system;
  • the disease diagnosis recommendation service is based on the patient's initial self-report and the new phenotype information generated during the dialogue process Recommend suitable diseases.
  • the knowledge storage module stores information such as disease prevalence, average age of onset and death, disease phenotype frequency characteristics, and disease-causing genes obtained from the authoritative rare disease knowledge base Orphanet. As of February 2022, a total of 4,257 rare diseases with phenotypic characteristics have been counted in the Orphanet knowledge base. These 4,257 rare diseases are the target recommended disease collection of this system. Orphanet provides a knowledge graph in XML file format, which is parsed and stored in the relational database MySQL.
  • HPO Human Phenotype Ontology
  • HPO is a standardized structured ontology of human abnormal phenotypes built based on ontological ideas. HPO is constructed as a directed acyclic graph with a hierarchical structure. HPO terms are connected through the "is_a" relationship, and children have more precise definitions than parents. As of February 2022, HPO has a total of 16,226 current phenotypic abnormality phenotypes (HP:0000118-Phenotypic abnormality and its descendant phenotypes), and each abnormal phenotype has a unique HPO code.
  • the human phenotype ontology applies the Annotation Propagation Rule, that is, a patient's positive phenotype can be inferred that his ancestor's phenotype is also positive in the patient, and a patient's negative phenotype can be inferred that his descendant's phenotype is also positive in the patient. is negative, this rule plays a huge role in disease posterior probability calculations and phenotype question and answer recommendations.
  • One of the important additional aspects of the present invention is a method for calculating the posterior probability of disease based on Bayesian medical testing, which method is applied to the disease recommendation service in the dialogue diagnosis service module, including disease prior probability estimation and disease posterior probability
  • LR Likelihood Ratio
  • the disease prevalence knowledge in the Orphanet knowledge base is obviously heterogeneous. It can be annotated as one or more of the five epidemic types (point prevalence, prevalence at birth, lifetime prevalence, annual incidence, and number of cases).
  • point prevalence prevalence at birth
  • lifetime prevalence prevalence at birth
  • annual incidence and number of cases.
  • the disease epidemic type is unified into point prevalence to achieve homogeneity. Then, the prior probability is calculated based on the disease point prevalence within the range of rare diseases:
  • Prob pretest (D) represents the prior probability of disease D (Pretest Probability)
  • PointPrev D represents the point prevalence of disease D (Point Prevalence)
  • ⁇ PointPrev represents the sum of the point prevalence of all rare diseases.
  • the specific rare disease prevalence information homogenization method is as follows. A specific value of representative point prevalence is calculated from the knowledge of five epidemic types:
  • the disease prevalence type is point prevalence
  • the entry whose prevalence geographical area is "worldwide” is preferred, and the numerical point prevalence or the median of the categorical point prevalence of the entry is used. If a disease does not have a "worldwide” entry but does have a prevalence entry by country or region, the average of the point prevalences of these entries is used as the point prevalence of the disease.
  • the epidemic type of the disease is annual incidence, we need to additionally obtain the average age of onset and age of death of the disease from Orphanet. Because rare diseases are difficult to cure, we took the age of onset to the age of death as the duration of the disease and estimated the point prevalence of the disease by multiplying the annual incidence rate by the years of disease duration.
  • the disease prevalence type is the number of cases, we directly use the number of cases divided by the total world population as the point prevalence of the disease.
  • the epidemic type of the disease is prevalence at birth and the age of disease onset is in the neonatal period, we use the prevalence at birth multiplied by the proportion of the population in the world's total population in the age group at which the disease persists as the point incidence of the disease. morbidity rate.
  • the prevalence type of the disease is lifetime prevalence
  • the lifetime prevalence is the proportion of the number of people who are sick in a certain period of time in the target population, it includes the number of people currently sick at the measurement time point and the number of people who were previously sick.
  • the number of people who have recovered from the disease, so this type is difficult to correspond to the point prevalence.
  • Bayes' theorem-based likelihood ratio calculation method to estimate the LR of a single phenotype as well as the disease priority.
  • posterior probability of disease First, the prior winning probability is calculated according to formula (2), the composite likelihood ratio of the phenotypic information is calculated according to formula (3) according to the disease phenotype frequency annotation and annotation propagation rules, then the post-test winning probability is calculated according to formula (4), and finally according to the formula (5) Obtain the posterior probability of the disease.
  • P represents the collected phenotype set
  • (p 1 , p 2 ,..., p n ) ⁇ P represents all rare diseases except rare disease D
  • D) represents the frequency of occurrence of phenotype p i in disease D
  • the composite likelihood ratio of phenotype set P is calculated as the likelihood of a single phenotype p i product of ratios.
  • Odds pretest (D) represents the prior winning probability of rare disease D.
  • this method will apply annotation broadcast rules on the human phenotype ontology when calculating the phenotype likelihood ratio, so as to facilitate the analysis of disease-phenotype relationships. More efficient. Specifically, for a disease-annotated related phenotype, it can be inferred that the disease is also related to the ancestral phenotype of that phenotype. If a positive phenotype information entered is the ancestor phenotype of a disease-related phenotype, then the likelihood ratio measure of the positive phenotype for the disease should be positive, that is, the likelihood ratio is greater than 1. If the entered negative phenotype information is the ancestor phenotype of a disease-related phenotype, then the likelihood ratio measure of the negative phenotype for the disease should be negative, that is, the likelihood ratio is less than 1.
  • Another important additional aspect of the present invention is an adaptive fusion of information gain and Gini index phenotype recommendation method, which method is applied to the phenotype recommendation service in the dialogue diagnosis service module.
  • the purpose of this method is to select the phenotype with the greatest expected benefit to be interrogated by the system based on the current dialogue state, that is, the disease posterior probability distribution and the disease-phenotype annotation relationship.
  • This method includes the table There are three aspects: Information Gain (Info Gain) calculation, Gini Index Gain (Gini Index Gain) calculation, and the adaptive fusion measure AIGGI (Adaptive Information Gain and Gini Index) calculation of the two.
  • the so-called classification weight of the disease posterior probability to calculate the information gain and Gini index gain of a single phenotype, adaptively fuse the two according to the uncertainty of the disease recommendation results, and recommend the phenotype with the largest AIGGI as the query object.
  • the system focuses more on asking for phenotypes with greater expected information gain to reduce the uncertainty of disease diagnosis; in the later stages of Q&A diagnosis or when the posterior probabilities of dominant diseases are more prominent, The system focuses more on interrogating phenotypes with a smaller expected Gini index to further improve the purity of disease diagnosis.
  • This invention uses information entropy (Information Entropy) to represent the uncertainty of disease recommendation results.
  • Information entropy is related to the information content of the system and can characterize the uncertainty of the sample set.
  • Information entropy focuses on information representation and has stronger expressive power for messy data.
  • the classic decision tree algorithm ID3 is designed with information entropy as the measurement. .
  • W is a weight set
  • w i represents the weight of a single element in the weight set W
  • W sum represents the sum of the weights of all elements in W.
  • Entropy cur (D) Entropy( ⁇ Prob posterior (d)
  • D represents the set of all rare diseases
  • Prob posterior (d) represents the posterior probability of disease d.
  • the information gain under a single phenotype condition is:
  • the first item is the disease posterior probability information entropy in the current dialogue state
  • the second item is the conditional entropy when phenotype p is positive
  • the third item is the conditional entropy when phenotype p is negative
  • D represents all rare
  • Prob posterior (d) represents the posterior probability of disease d
  • ⁇ is the weight of the positive phenotype branch, that is, the sum of the elements of the ⁇ Prob posterior (d)*Pr(p
  • d) represents the frequency of phenotype p in disease d
  • the present invention uses Gini Index to express the purity of disease recommendation results.
  • the Gini index is related to data complexity or purity.
  • the Gini index focuses on algebraic representation and is better at pure set classification.
  • the related CART algorithm is used to construct classification decision trees.
  • Gini(W) with categorical weights:
  • W is a weight set
  • w i represents the weight of a single element in the weight set W
  • W sum represents the sum of the weights of all elements in W.
  • Gini cur (D) Gini( ⁇ Prob posterior (d)
  • D represents the set of all rare diseases
  • Prob posterior (d) represents the posterior probability of disease d.
  • the Gini index gain under a single phenotype is:
  • the first item is the Gini index of the posterior probability of disease in the current dialogue state
  • the second item is the conditional Gini index when phenotype p is positive
  • the third item is the conditional Gini index when phenotype p is negative
  • the remaining parameters Define equivalent formula (8).
  • Information entropy and the Gini index serve similar uncertainty measurement functions but have different definitions and measurement focuses.
  • Information entropy focuses on information representation and is more discriminative for messy data; Gini index focuses on algebraic representation and is better at classifying pure sets. Fusion metrics can integrate the advantages of both.
  • the system In the early stages of question-and-answer diagnosis or when the posterior probabilities of dominant diseases are not prominent enough, the system focuses more on asking for phenotypes with greater expected information gain to reduce the uncertainty of disease diagnosis; in question-and-answer diagnosis When the posterior probability of dominant disease is more prominent, the system will focus more on querying phenotypes with smaller expected Gini index to further improve the purity of disease diagnosis.
  • D represents the set of all rare diseases
  • p represents a single phenotype
  • ⁇ reflects the uncertainty or disorder of the disease posterior probability distribution in the current dialogue state. Since the value range of the Gini index is [0,1], Therefore ⁇ can be expressed directly using Gini cur (D).
  • the dialogue diagnosis process of the rare disease dialogue diagnosis system is shown in Figure 3. After the initial patient self-report, human-computer interaction is conducted in the form of multiple rounds of question and answer to obtain more patient phenotype information. When the maximum dialogue round or disease is reached, When the uncertainty of the recommendation result is very low, the system gives disease diagnosis recommendations and ends the conversation.
  • the detailed steps of the dialogue diagnosis process are:
  • Phenotype question and answer The system selects the phenotype with the largest fusion measure AIGGI for questioning based on the current dialogue status. After the user selects the answer of "exist”, “does not exist” or “unsure” for this phenotype, the system enters the next round of dialogue.
  • users can obtain a variety of visual information including disease priority assessment, current dialogue status, patient phenotypic status, and dialogue history in the dialogue diagnosis sub-module of the front-end UI module.
  • step (2) Disease recommendations. Repeat step (2) until the uncertainty of the disease recommendation result is very low or the maximum dialogue round is reached, the system gives disease recommendations based on the disease posterior probability and ends the dialogue.

Landscapes

  • Engineering & Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Medical Informatics (AREA)
  • Public Health (AREA)
  • Biomedical Technology (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Pathology (AREA)
  • Epidemiology (AREA)
  • General Health & Medical Sciences (AREA)
  • Primary Health Care (AREA)
  • Medical Treatment And Welfare Office Work (AREA)

Abstract

一种罕见病可视化问答式辅助鉴别诊断系统及方法，系统包括：前端UI模块，采集患者基本信息和表型信息录入以获取患者的初始自我报告；对话诊断服务模块，通过API提供表型询问推荐服务、疾病诊断推荐服务以及知识库查询服务；表型询问推荐服务是根据当前的对话状态推荐合适的表型以达到较高的对话收益；疾病诊断推荐服务是根据患者初始自我报告和对话过程中产生的新的表型信息推荐合适的疾病；知识库查询服务是根据需要，查询并调用知识存储模块中存储的知识；知识存储模块，用于存储人类表型本体知识库以及疾病-表型关系知识库。该系统能够收集超出其自我报告的附加症状并利用可视化技术显示鉴别诊断的辅助信息。

Description

知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法 技术领域
本发明涉及智能辅助诊断技术领域，尤其涉及一种知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法。
背景技术
罕见病，又称孤儿病，是指那些发病率极低的疾病。世界卫生组织(WHO)将罕见病定义为每10000人中患病率为6.5到10人的疾病。《中国罕见病定义研究报告2021》将“新生儿发病率小于1/万、患病率小于1/万、患病人数小于14万的疾病”列入罕见病。国际确认的罕见病有近7000种，大约占人类疾病的10％。
罕见病诊疗困难，具体表现为确诊困难、治疗成本高和药物研发困难。罕见病种类繁多且通常具有复杂的表型特征和分子特征，这使得罕见病的诊断难度要高于其他种类的疾病。根据欧洲罕见病组织(EURORDIS)的一项调查，四分之一的罕见遗传病患者需要5到30年的时间才能被诊断出来，且初诊误诊率高达40％。这些误诊会导致医疗资源浪费和救治时间的延误，甚至直接危害患者生命。因此，如何实施快速有效的诊断成为罕见病领域的一项重要课题。
目前许多罕见病知识库被建立起来。Orphanet于1997年在法国成立，旨在收集稀缺的罕见病知识，提高对罕见病患者的诊断、护理和治疗水平。迄今为止，Orphanet已成为罕见病诊断的重要参考来源。人类表型本体(Human Phenotype Ontology，HPO)在2008年由Robinson教授等人提出，是人类疾病中已发现的表型异常的标准化词汇表，广泛应用罕见疾病表型基因型诊断分析。临床医生可以利用罕见病知识库，根据患者的表型特征去筛选患者潜在疾病，然而这一方法在临床实践中有颇多的局限性：一是在临床采集患者表型过程中经常会采集到不准确表型甚至引入噪声表型；二是因遗漏表型导致患者表型信息不完整；三是罕见病之间表型交叠较多导致鉴别诊断难度较大。这些问题在各种罕见病诊断场景中都会对诊断效果 造成不利影响。
当前，对话式聊天机器人或面向任务的对话系统引起越来越多的关注，尤其在医疗领域，以疾病诊断为目标的智能对话系统能够进行主动的患者额外症状信息收集以达到更好的疾病诊断效果。现有的医疗诊断对话系统大多依赖于数据驱动的机器学习方法。然而，在罕见病领域，由于疾病种类繁多且病例资源匮乏，使得真实数据驱动的学习难以开展。在这一背景下，整合知识库开展知识驱动的医疗诊断对话成为一种潜在的有效的方法。
发明内容
本发明提供了一种知识驱动的罕见病可视化问答式辅助鉴别诊断系统，该系统能够通过与患者/医生对话的方式收集超出其自我报告的附加症状并利用可视化技术显示鉴别诊断的辅助信息，进一步缓解罕见病诊断难的问题。
本发明技术方案如下：
一种罕见病可视化问答式辅助鉴别诊断系统，包括：
前端UI模块，用于进行人机交互，采集患者基本信息和表型信息录入以获取患者的初始自我报告，并利用多个维度可视化视图展现相关信息供临床人员进行鉴别诊断；
对话诊断服务模块，通过API(Application Programming Interface)提供表型询问推荐服务、疾病诊断推荐服务以及知识库查询服务；表型询问推荐服务是根据当前的对话状态推荐合适的表型以达到较高的对话收益，用户对系统询问的表型可以选择回答“存在”、“不存在”或者“不确定”；疾病诊断推荐服务是根据前端UI模块采集的患者初始自我报告和对话过程中产生的新的表型信息推荐合适的疾病；知识库查询服务是根据表型询问推荐服务和疾病诊断推荐服务的需要，查询调用知识存储模块中存储的知识；
知识存储模块，用于存储人类表型本体知识库以及疾病-表型关系知识库。
所述的人类表型本体知识库为人类表型本体(Human Phenotype Ontology，HPO)，HPO是基于本体论思想构建的人类异常表型的标准化结构化本体。HPO被构造为具有层级结构的有向无环图，HPO术语之间 通过“is_a”的关系进行连接，子项比父项拥有更为精确的定义。截至2022年2月份，HPO总共有了16226条现行的表型异常表型(HP:0000118-Phenotypic abnormality及其后代表型)，每条异常表型都有一个唯一的HPO编码。
所述的疾病-表型关系知识库为罕见病权威知识库Orphanet，罕见病权威知识库Orphanet中包含疾病流行性、平均发病死亡年龄、疾病表型频率特征以及致病基因信息。
知识存储模块为对话诊断服务模块提供知识驱动。
前端UI模块包括：
患者信息采集子模块，包括患者基本信息视图和表型信息录入视图；
可视化的问答式鉴别诊断子模块，疾病评估视图、对话状态视图、患者表型状态视图、对话历史视图以及问答视图，利用多个维度可视化视图展现相关信息供临床人员进行鉴别诊断。
患者信息采集子模块包括患者基本信息视图和表型信息录入视图，用户可以在该模块输入患者年龄、出生时性别、阳性表型以及阴性表型等信息。
问答式鉴别诊断子模块包括：
疾病评估视图，用于展示疾病患病率、后验概率以及表型匹配情况；
对话状态视图，以雷达图的形式展示当前优势最大的疾病的后验概率(标签为P)、表型匹配情况(标签为M)以及疾病推荐结果的不确定性度量(信息熵标签为E，基尼指数标签为G)；
患者表型状态视图，为基于人类表型本体并利用注释传播规则绘制的反应患者表型特征(阳性、阴性、不确定)的旭日图；
对话历史视图以及问答视图，用于展示系统询问的表型信息并为用户提供反馈按钮(存在、不存在、不确定)。
问答式鉴别诊断子模块以多个维度可视化视图展现相关信息供临床人员进行鉴别诊断。
对话诊断服务模块，通过API(Application Programming Interface)提供表型询问推荐服务、疾病诊断推荐服务以及知识库查询服务。
疾病推荐服务包括疾病先验概率估计和疾病后验概率计算，对话诊断 服务模块在对话结束时推荐后验概率最大的疾病；
疾病先验概率估计的方法包括：
(a-1)将Orphanet知识库中的疾病流行性类型统一转换为点患病率；
(a-2)在罕见病范围内，根据疾病的点患病率计算疾病先验概率，公式为：
其中，Probpretest(D)代表疾病D的先验概率(Pretest Probability)，PointPrevD代表疾病D的点患病率(Point Prevalence)，∑PointPrev表示所有罕见病点患病率之和；
疾病后验概率计算的方法包括：
(b-1)胜算(Odds)与概率(Prob)之间的换算关系为：
根据公式(2)将疾病先验概率转换为疾病先验胜算；
(b-2)根据疾病表型频率注释和注释传播规则依照公式(3)计算表型信息的复合似然比：
其中，P代表已采集的表型集合，(p1,p2,…,pn)∈P，代表除罕见病D以外的所有罕见病，Pr(pi|D)表示表型pi在疾病D中的出现频率，表型集合P的复合似然比被计算为单个表型pi似然比的乘积；
(b-3)根据公式(4)计算疾病的后验胜算Oddsposterior(P,D)：
Oddsposterior(P,D)＝Oddspretest(D)*CompositeLR(P,D)#(4)
其中，Oddspretest(D)代表罕见病D的先验胜算；
(b-4)根据公式(5)计算疾病后验概率Probposterior(D)：
步骤(a-1)中，将疾病流行性类型统一转换为点患病率的方法为：
①如果疾病的流行性类型为点患病率，优先选择患病率地理区域为“世界范围内”的条目，使用该条目的数值型点患病率或者类别型点患病率 的中值；如果患病率地理区域没有“世界范围内”的条目但有以国家或地区统计的患病率条目，则使用以国家或地区统计的患病率条目的点患病率平均值作为该疾病的点患病率；
②如果疾病的流行性类型为年发病率，则从Orphanet库获取该疾病的平均发病年龄以及死亡年龄信息，将发病年龄到死亡年龄作为疾病的持续时长，用年发病率乘以疾病持续年份估算该疾病的点患病率；
③如果疾病的流行性类型为病例数，则直接使用病例数除以世界总人口作为该疾病的点患病率；
④如果疾病的流行性类型为出生时患病率，且疾病发病年龄在新生儿期，则使用出生时患病率乘以该疾病持续年龄段的人口在世界总人口中的比例值作为疾病的点患病率；
⑤如果疾病的流行性类型为终生患病率，则使用默认的百万分之一作为该疾病的点患病率。
针对临床采集的患者表型不准确有噪声的问题，本发明在计算表型似然比时会在人类表型本体上应用注释转播规则，使疾病-表型关系的分析更加高效。
优选的，所述的注释转播规则为：患者的阳性表型可以推断出其祖先表型在患者处也为阳性，患者的阴性表型可以推断出其后代表型在患者处均为阴性。
表型推荐服务根据当前的疾病后验概率分布情况和疾病-表型注释关系，选择期望收益最大的表型由系统进行询问。
表型推荐服务包括：
表型的信息增益(Info Gain)计算，包括：
(A-1)使用信息熵(Information Entropy)来表示疾病推荐结果的不确定性，以公式(6)定义有分类权重的信息熵Entropy(W)：
其中，W是一个权重集合，wi表示权重集合W中单一元素的权值，Wsum表示W中所有元素权值之和；
当前对话状态下的疾病后验概率的信息熵Entropycur(D)为：
Entropycur(D)＝Entropy({Probposterior(d)|d∈D})#(7)
其中，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验概率；
(A-2)单一表型条件下的信息增益为：
其中，第一项为当前对话状态下的疾病后验概率的信息熵，第二项为表型p为阳性时的条件熵，第三项为表型p为阴性时的条件熵，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验概率，γ为阳性表型分支权重，即{Probposterior(d)*Pr(p|d)|d∈D}集合元素之和，Pr(p|d)表示表型p在疾病d中的出现频率，表示表型p不在疾病d中出现的频率；
表型的基尼指数计算，包括：
(B-1)使用基尼指数(Gini Index)来表示疾病推荐结果的纯度，以公式(9)定义有分类权重的基尼指数Gini(W)：
其中，W是一个权重集合，wi表示权重集合W中单一元素的权值，Wsum表示W中所有元素权值之和；
当前对话状态下的疾病后验概率的基尼指数为：
Ginicur(D)＝Gini({Probposterior(d)|d∈D})#(10)
其中，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验概率；
(B-2)单一表型条件下的基尼指数增益为：
其中，第一项为当前对话状态下的疾病后验概率基尼指数，第二项为表型p为阳性时的条件基尼指数，第三项为表型p为阴性时的条件基尼指数，其余参数定义等同公式(8)；
将信息增益和基尼指数进行自适应融合，获得融合度量AIGGI(Adaptive Information Gain and Gini Index)：
AIGGI(p,D)＝α*InfoGain(p,D)+(1-α)*GiniGain(p,D)#(12)
其中，D表示所有罕见病的集合，p表示单一表型，α为反应当前对话状态下疾病后验概率分布的不确定程度或杂乱度，α使用Ginicur(D)来表示；
系统根据当前对话状态选择融合度量AIGGI最大的表型进行询问，用户针对该表型选择“存在”、“不存在”或者“不确定”的回答。
优选的，为了提高对话的效率，在表型询问推荐时应用注释转播规则来过滤掉一些表型，包括：
已汇报的表型不会在接下来的对话轮次中被询问；
对于已汇报的阳性表型，默认其祖先表型均为阳性，将其祖先表型过滤掉，不再询问；
对于已汇报的阴性表型，默认其后代表型均阴性，将其后代表型过滤掉，不再询问。
本发明还提供了采用罕见病可视化问答式辅助鉴别诊断系统进行辅助鉴别诊断的方法，包括：
(1)用户在前端UI模块的患者信息采集子模块中录入患者信息，包括患者年龄、出生时性别、阳性表型以及阴性表型；
(2)对话诊断服务模块根据当前对话状态选择融合度量AIGGI最大的表型进行询问，用户针对该表型选择“存在”、“不存在”或者“不确定”的回答后，之后进入下一轮对话；
问答过程中，用户可以在前端UI模块的可视化的问答式鉴别诊断子模块中获得包括疾病优先级评估、当前对话状态、患者表型状态以及对话历史等多种可视化信息；
(3)重复步骤(2)，直到疾病推荐结果的不确定性低于阈值或者达到最大对话轮次时，系统根据疾病后验概率给出疾病推荐并结束对话。
本发明的核心是疾病推荐服务，其算法核心是如何在数万种表型中获得最有效的表型问答，以最优路径逼近真实诊断。其中重要贡献之一是将一种基于贝叶斯医学检验计算疾病后验概率的方法，应用于对话诊断服务模块中的疾病推荐服务，包括疾病先验概率估计和疾病后验概率计算两个步骤。首先本发明设计了不同流行性类型的转换方法，将异构的疾病流行性知识同质化为点患病率，并根据其计算疾病先验概率。使用似然比 (Likelihood Ratio)指标作为单个表型条件对疾病优先级贡献程度的度量，累乘得到全部表型条件的复合似然比，进而得到疾病在表型基础上的后验概率。
本发明的另一个核心是自适应融合信息增益和基尼指数的表型推荐方法，应用于对话诊断服务模块中的表型询问推荐服务。该方法的目的是根据当前的对话状态，即疾病后验概率分布情况和疾病-表型注释关系，选择期望收益最大的表型由系统进行询问。该方法包括表型的信息增益(Info Gain)计算、基尼指数增益(Gini index Gain)计算以及二者的自适应融合度量AIGGI(Adaptive Information Gain and Gini Index)计算三个方面。本发明使用疾病后验概率作为分类权重计算单个表型的信息增益和基尼指数增益，根据疾病推荐结果的不确定程度将二者进行自适应融合，由系统推荐AIGGI最大的表型作为询问对象。
与现有技术相比，本发明的有益效果为：
(1)相较于基于知识库的表型关键词检索推荐疾病以及常见的基于患者相似性的疾病推荐方法，本发明采用表型问答的形式，通过人机交互进行主动的患者额外症状信息收集以丰富患者表型信息，对话过程中可以启发医生注意到之前漏采的患者表型，最终实现更有效的疾病诊断。
(2)针对临床采集的患者表型不准确有噪声的问题，本发明设计了一种基于贝叶斯医学检验计算疾病后验概率的方法，该方法在人类表型本体上应用注释转播规则来提高疾病推荐的鲁棒性。
(3)患病率是影响疾病诊断的一个非常重要的因素，而大多数疾病推荐方法没有考虑患病率。本发明针对疾病流行性知识的异构性问题，设计了一套流行性类型转换计算方法，将疾病流行性类型统一为点患病率，以实现同质性，并根据点患病率计算出疾病先验概率，将其应用到疾病推荐方法中。
(4)本发明设计了一种自适应融合信息增益和基尼指数的表型推荐方法。根据疾病推荐结果的不确定程度，将表型条件的信息增益和基尼指数增益两种度量进行自适应融合，得到更高效的表型问答。
附图说明
图1为罕见病可视化问答式辅助鉴别诊断系统的模块结构示意图；
图2为罕见病可视化问答式辅助鉴别诊断状态雷达图；
图3为罕见病可视化问答式辅助鉴别诊断系统的问答流程图示意。
具体实施方式
本发明设计了一种知识驱动的罕见病对话诊断系统，能够通过与用户(患者/家属/医生)对话的方式收集超出其自我报告的附加表型，并基于此自动开展评估并给出疾病诊断。系统以Orphanet罕见病知识库和人类表型本体作为先验知识，在初始患者信息录入后由系统进行表型询问，系统根据实时对话状态选择期望收益最大的表型，并由用户给出判断，迭代以上过程实现多轮问答，从而丰富患者的结构化表型数据采集并提高疾病诊断效果。
(一)罕见病对话诊断系统模块介绍
本发明设计的罕见病问答对话诊断系统包括前端UI(User Interface)模块、对话诊断服务模块和知识存储模块，如图1所示，其中：
前端UI模块，用以进行人机交互，具体包括患者信息采集子模块和可视化的问答式鉴别诊断子模块。患者信息采集子模块包括患者基本信息视图和表型信息录入视图，用户可以在该模块输入患者年龄、出生时性别、阳性表型以及阴性表型等信息。问答式鉴别诊断子模块包括疾病评估视图、对话状态视图、患者表型状态视图、对话历史视图以及问答视图。疾病评估视图展示疾病患病率、后验概率以及表型匹配情况；对话状态视图以雷达图的形式展示当前优势最大的疾病的后验概率(标签为P)、表型匹配情况(标签为M)以及疾病推荐结果的不确定性度量(信息熵标签为E，基尼指数标签为G)，如图2所示，雷达图阴影面积越大代表疾病推荐结果的不确定性越低；患者表型状态视图为一个基于人类表型本体绘制的反应患者表型特征(阳性、阴性、不确定)的旭日图，并利用了注释传播规则；问答视图为对话诊断子模块的核心交互模块，展示系统询问的表型信息并为用户提供反馈按钮(存在、不存在、不确定)。问答式鉴别诊断子模块以多个维度可视化视图展现相关信息供临床人员进行鉴别诊断。
对话诊断服务模块，主要是通过API(Application Programming Interface)提供表型询问推荐服务、疾病诊断推荐服务以及知识库查询服务。表型询问推荐服务是根据当前的对话状态推荐合适的表型以达到较高 的对话收益，用户对系统询问的表型可以选择回答“存在”、“不存在”或者“不确定”；疾病诊断推荐服务是根据患者初始的自我报告和对话过程中产生的新的表型信息推荐合适的疾病。我们为表型询问和疾病诊断设计了相应的算法，作为本发明的两个重要附加方面。
知识存储模块，存储从罕见病权威知识库Orphanet中获取的疾病流行性、平均发病死亡年龄、疾病表型频率特征以及致病基因等信息。截至2022年2月份，Orphanet知识库中总共统计了4257个带有表型特征的罕见病，这4257个罕见病为本系统的目标推荐疾病集合。Orphanet提供XML文件格式的知识图谱，将其解析后存储到关系型数据库MySQL中。
除Orphanet知识库之外，本系统还整合了人类表型本体(Human Phenotype Ontology，HPO)，HPO是基于本体论思想构建的人类异常表型的标准化结构化本体。HPO被构造为具有层级结构的有向无环图，HPO术语之间通过“is_a”的关系进行连接，子项比父项拥有更为精确的定义。截至2022年2月份，HPO总共有了16226条现行的表型异常表型(HP:0000118-Phenotypic abnormality及其后代表型)，每条异常表型都有一个唯一的HPO编码。
人类表型本体适用注释传播规则(Annotation Propagation Rule)，即患者的阳性表型可以推断出其祖先表型在患者处也为阳性，患者的阴性表型可以推断出其后代表型在患者处均为阴性，该规则在疾病后验概率计算和表型问答推荐中发挥了巨大作用。
(二)基于贝叶斯医学检验计算疾病后验概率的方法
本发明的重要附加方面之一是一种基于贝叶斯医学检验计算疾病后验概率的方法，该方法应用于对话诊断服务模块中的疾病推荐服务，包括疾病先验概率估计和疾病后验概率计算两个步骤。我们通过设计流行性类型转换方法，将异构的疾病流行性知识同质化为点患病率，并根据其计算疾病先验概率。然后，我们使用似然比(Likelihood Ratio，LR)指标作为单个表型条件对疾病优先级贡献程度的度量，将其累乘得到全部表型条件的复合似然比，进而得到疾病的后验概率。详细实施方式如下：
(1)疾病先验概率估计。
Orphanet知识库中的疾病流行性知识有明显的异构性，一个罕见病可 能被注释为五种流行性类型(点患病率、出生时患病率、终生患病率、年发病率以及病例数)中的一个或几个。通过设计流行性类型转换计算方法，将疾病流行性类型统一为点患病率，以实现同质性。然后，在罕见病范围内根据疾病点患病率计算先验概率：
其中，Probpretest(D)代表疾病D的先验概率(Pretest Probability)，PointPrevD代表疾病D的点患病率(Point Prevalence)，∑PointPrev表示所有罕见病点患病率之和。
具体的罕见病患病率信息同质化方法如下，从五种流行性类型的知识中计算出一个代表点患病率的具体值：
如果疾病的流行性类型为点患病率，优先选择患病率地理区域为“世界范围内”的条目，使用该条目的数值型点患病率或者类别型点患病率的中值。如果疾病没有“世界范围内”的条目但有以国家或地区统计的患病率条目，则使用这些条目的点患病率平均值作为该疾病的点患病率。
如果疾病的流行性类型为年发病率，我们需要额外从Orphanet获取该疾病的平均发病年龄以及死亡年龄信息。由于罕见病难以治愈，因此我们将发病年龄到死亡年龄作为疾病的持续时长，用年发病率乘以疾病持续年份估算疾病的点患病率。
如果疾病的流行性类型为病例数，我们直接使用病例数除以世界总人口作为疾病的点患病率。
如果疾病的流行性类型为出生时患病率，且疾病发病年龄在新生儿期，我们使用出生时患病率乘以疾病持续年龄段的人口在世界总人口中的比例值作为疾病的点患病率。
最后，对于疾病的流行性类型为终生患病率的情况，由于终生患病率是在某个时间段里患病的人数在目标人群中的比例，包括测量时间点目前患病人数和以前患病但是现在痊愈的人数，因此该类型很难对应到点患病率上，我们使用默认的百万分之一作为该疾病的点患病率。
(2)疾病后验概率计算。
为了更好地提供单个表型观察结果对疾病优先级的贡献程度的度量，我们使用基于贝叶斯定理的似然比计算方法来估计单一表型的LR以及疾 病的后验概率。首先根据公式(2)计算先验胜算，根据疾病表型频率注释和注释传播规则依照公式(3)计算表型信息的复合似然比，然后根据公式(4)计算后测胜算，最后根据公式(5)得到疾病的后验概率。
胜算(Odds)与概率(Prob)之间的换算关系：
多表型的复合似然比(Composite Likelihood Ratio)计算公式：
其中，P代表已采集的表型集合，(p1,p2,…,pn)∈P，代表除罕见病D以外的所有罕见病，Pr(pi|D)表示表型pi在疾病D中的出现频率，表型集合P的复合似然比被计算为单个表型pi似然比的乘积。
后验胜算计算公式：
Oddsposterior(P,D)＝Oddspretest(D)*CompositeLR(P,D)#(4)
其中，Oddspretest(D)代表罕见病D的先验胜算。
最后是后验概率的计算：
值得一提的是，针对临床采集的患者表型不准确有噪声的问题，本方法在计算表型似然比时会在人类表型本体上应用注释转播规则，使疾病-表型关系的分析更加高效。具体来讲，对于疾病注释的相关表型，可以推断疾病也与该表型的祖先表型相关。如果录入的一个阳性的表型信息为疾病相关表型的祖先表型，那么该阳性表型对该疾病的似然比度量应该为积极的，即似然比大于1。如果录入的一个阴性的表型信息为疾病相关表型的祖先表型，那么该阴性表型对该疾病的似然比度量应该为消极的，即似然比小于1。
(三)自适应融合信息增益和基尼指数的表型推荐方法
本发明的另一个重要附加方面为一种自适应融合信息增益和基尼指数的表型推荐方法，该方法应用在对话诊断服务模块中的表型推荐服务。该方法的目的是根据当前的对话状态，即疾病后验概率分布情况和疾病-表型注释关系，选择期望收益最大的表型由系统进行询问。该方法包括表 型的信息增益(Info Gain)计算、基尼指数增益(Gini Index Gain)计算以及二者的自适应融合度量AIGGI(Adaptive Information Gain and Gini Index)计算三个方面。我们使用疾病后验概率所谓分类权重计算单个表型的信息增益和基尼指数增益，根据疾病推荐结果的不确定程度将二者进行自适应融合，推荐AIGGI最大的表型作为询问对象。在问答诊断前期或优势疾病后验概率不够突出时，系统更侧重于询问期望信息增益较大的表型来降低疾病诊断的不确定性；在问答诊断后期或优势疾病后验概率较为突出时，系统更侧重于询问期望基尼指数较小的表型来进一步提升疾病诊断的纯度。
(1)表型问答的信息增益计算。
本发明使用信息熵(Information Entropy)来表示疾病推荐结果的不确定性。信息熵关联于系统信息含量，能够表征样本集合的不确定性，信息熵侧重于信息表示，对较乱的数据具有更强的表现力，经典的决策树算法ID3就是以信息熵为度量设计的。我们将疾病后验概率作为分类的权重计算当前状态的信息熵，如公式(7)所示。如果疾病后验概率分布均匀，说明疾病诊断的不确定性较高，信息熵值较大。
首先定义基本的有分类权重的信息熵Entropy(W)：
其中，W是一个权重集合，wi表示权重集合W中单一元素的权值，Wsum表示W中所有元素权值之和。
当前对话状态下的疾病后验概率信息熵Entropycur(D)为：
Entropycur(D)＝Entropy({Probposterior(d)|d∈D})#(7)
其中，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验概率。
单一表型条件下的信息增益为：
其中，第一项为当前对话状态下的疾病后验概率信息熵，第二项为表型p为阳性时的条件熵，第三项为表型p为阴性时的条件熵，D表示所有罕 见病的集合，Probposterior(d)表示疾病d的后验概率，γ为阳性表型分支权重，即{Probposterior(d)*Pr(p|d)|d∈D}集合元素之和，Pr(p|d)表示表型p在疾病d中的出现频率，表示表型p不在疾病d中出现的频率。
(2)表型问答的基尼指数计算。
本发明使用基尼指数(Gini Index)来表示疾病推荐结果的纯度。基尼指数关联于数据杂乱度或纯度，基尼指数侧重于代数表示，更擅长较纯集合分类，相关的CART算法用于构造分类决策树。我们将疾病后验概率作为分类的权重计算当前状态的基尼值，如公式(10)所示。如果疾病后验概率分布集中在少数疾病上，说明疾病诊断结果的纯度较高，基尼值较小。首先定义基本的有分类权重的基尼指数Gini(W)：
其中，W是一个权重集合，wi表示权重集合W中单一元素的权值，Wsum表示W中所有元素权值之和。
当前对话状态下的疾病后验概率基尼指数为：
Ginicur(D)＝Gini({Probposterior(d)|d∈D})#(10)
其中，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验概率。
单一表型条件下的基尼指数增益为：
其中，第一项为当前对话状态下的疾病后验概率基尼指数，第二项为表型p为阳性时的条件基尼指数，第三项为表型p为阴性时的条件基尼指数，其余参数定义等同公式(8)。
(3)信息增益和基尼指数的自适应融合。
信息熵和基尼指数具有相似的不确定性度量功能，但具有不同的定义和测量重点。信息熵侧重于信息表示，对较乱数据具有更强区分性；基尼指数则侧重于代数表示，更擅长较纯集合分类。融合度量可以将二者的优势进行整合。在问答诊断前期或优势疾病后验概率不够突出时，系统更侧重询问期望信息增益较大的表型来降低疾病诊断的不确定性；在问答诊断 后期或优势疾病后验概率较为突出时，系统更侧重询问期望基尼指数较小的表型来进一步提升疾病诊断的纯度。本发明将信息增益和基尼指数进行自适应加权集成，产生具有更完备不确定性刻画的融合度量AIGGI(Adaptive Information Gain and Gini Index)：
AIGGI(p,D)＝α*InfoGain(p,D)+(1-α)*GiniGain(p,D)#(12)
其中，D表示所有罕见病的集合，p表示单一表型，α为反应当前对话状态下疾病后验概率分布的不确定程度或杂乱度，由于基尼指数的取值范围为[0,1],因此α可以直接使用Ginicur(D)来表示。
值得一提的是，为了提高对话的效率，我们在表型询问推荐时会应用注释转播规则来过滤掉一些表型。首先，已汇报的表型不会在接下来的对话轮次中被询问。除此之外，对于已汇报的阳性表型，默认其祖先表型均为阳性，将祖先表型过滤掉；对于已汇报的阴性表型，默认其后代表型均阴性，将其后代表型过滤掉，系统不再询问。
(四)罕见病对话诊断系统对话诊断流程介绍
罕见病对话诊断系统对话诊断流程如图3所示，在初始的患者自我报告后，以多轮问答的形式进行人机交互来获得更多的患者表型信息，当达到最大对话轮次或疾病推荐结果不确定性很低的时候，系统给出疾病诊断推荐并结束对话。对话诊断流程详细步骤为：
(1)初始自我报告。用户在前端UI模块的患者信息采集子模块中录入患者信息，包括患者年龄、出生时性别、阳性表型以及阴性表型。
(2)表型问答。系统根据当前对话状态选择融合度量AIGGI最大的表型进行询问，用户针对该表型选择“存在”、“不存在”或者“不确定”的回答后，进入下一轮对话。问答过程中，用户可以在前端UI模块的对话诊断子模块中获得包括疾病优先级评估、当前对话状态、患者表型状态以及对话历史等多种可视化信息。
(3)疾病推荐。重复步骤(2)直到疾病推荐结果的不确定性很低或者达到最大对话轮次时，系统根据疾病后验概率给出疾病推荐并结束对话。
以上所述的实施例对本发明的技术方案和有益效果进行了详细说明，应理解的是以上所述仅为本发明的具体实施例，并不用于限制本发明，凡在本发明的原则范围内所做的任何修改、补充和等同替换等，均应包含在本发明的保护范围之内。

Claims (8)

  1. 一种罕见病可视化问答式辅助鉴别诊断系统，其特征在于，包括：
    前端UI模块，用于进行人机交互，采集患者基本信息和表型信息录入以获取患者的初始自我报告，并利用多个维度可视化视图展现相关信息供临床人员进行鉴别诊断；
    对话诊断服务模块，通过API提供表型询问推荐服务、疾病诊断推荐服务以及知识库查询服务；表型询问推荐服务是根据当前的对话状态推荐合适的表型以达到较高的对话收益，用户对系统询问的表型可以选择回答“存在”、“不存在”或者“不确定”；疾病诊断推荐服务是根据前端UI模块采集的患者初始自我报告和对话过程中产生的新的表型信息推荐合适的疾病；知识库查询服务是根据表型询问推荐服务和疾病诊断推荐服务的需要，查询并调用知识存储模块中存储的知识；
    知识存储模块，用于存储人类表型本体知识库以及疾病-表型关系知识库。
  2. 根据权利要求1所述的罕见病可视化问答式辅助鉴别诊断系统，其特征在于，前端UI模块包括：
    患者信息采集子模块，包括患者基本信息视图和表型信息录入视图；
    可视化的问答式鉴别诊断子模块，疾病评估视图、对话状态视图、患者表型状态视图、对话历史视图以及问答视图，利用多个维度可视化视图展现相关信息供临床人员进行鉴别诊断。
  3. 根据权利要求2所述的罕见病可视化问答式辅助鉴别诊断系统，其特征在于，患者信息采集子模块包括患者基本信息视图和表型信息录入视图，用户可以在该模块输入信息包括患者年龄、出生时性别、阳性表型以及阴性表型。
    问答式鉴别诊断子模块包括：
    疾病评估视图，用于展示疾病患病率、后验概率以及表型匹配情况；
    对话状态视图，以雷达图的形式展示当前优势最大的疾病的后验概率、表型匹配情况以及疾病推荐结果的不确定性度量；
    患者表型状态视图，为基于人类表型本体并利用注释传播规则绘制的 反应患者表型特征的旭日图；
    对话历史视图以及问答视图，用于展示系统询问的表型信息并为用户提供反馈按钮。
  4. 根据权利要求1所述的罕见病可视化问答式辅助鉴别诊断系统，其特征在于，疾病推荐服务包括疾病先验概率估计和疾病后验概率计算，对话诊断服务模块在对话结束时推荐后验概率最大的疾病；
    疾病先验概率估计的方法包括：
    (a-1)将Orphanet知识库中的疾病流行性类型统一转换为点患病率；
    (a-2)在罕见病范围内，根据疾病的点患病率计算疾病先验概率，公式为：
    其中，Probpretest(D)代表疾病D的先验概率，PointPrevD代表疾病D的点患病率，∑PointPrev表示所有罕见病点患病率之和；
    疾病后验概率计算的方法包括：
    (b-1)胜算与概率之间的换算关系为：
    根据公式(2)将疾病先验概率转换为疾病先验胜算；
    (b-2)根据疾病表型频率注释和注释传播规则依照公式(3)计算表型信息的复合似然比：
    其中，P代表已采集的表型集合，(p1,p2,…,pn)∈P，代表除罕见病D以外的所有罕见病，Pr(pi|D)表示表型pi在疾病D中的出现频率，表型集合P的复合似然比被计算为单个表型pi似然比的乘积；
    (b-3)根据公式(4)计算疾病的后验胜算Oddsposterior(P,D)：
    Oddsposterior(P,D)＝Oddspretest(D)*CompositeLR(P,D)#(4)
    其中，Oddspretest(D)代表罕见病D的先验胜算；
    (b-4)根据公式(5)计算疾病后验概率Probposterior(D)：
  5. 根据权利要求4所述的罕见病可视化问答式辅助鉴别诊断系统，其特征在于，步骤(a-1)中，将疾病流行性类型统一转换为点患病率的方法为：
    ①如果疾病的流行性类型为点患病率，优先选择患病率地理区域为“世界范围内”的条目，使用该条目的数值型点患病率或者类别型点患病率的中值；如果患病率地理区域没有“世界范围内”的条目但有以国家或地区统计的患病率条目，则使用以国家或地区统计的患病率条目的点患病率平均值作为该疾病的点患病率；
    ②如果疾病的流行性类型为年发病率，则从Orphanet库获取该疾病的平均发病年龄以及死亡年龄信息，将发病年龄到死亡年龄作为疾病的持续时长，用年发病率乘以疾病持续年份估算该疾病的点患病率；
    ③如果疾病的流行性类型为病例数，则直接使用病例数除以世界总人口作为该疾病的点患病率；
    ④如果疾病的流行性类型为出生时患病率，且疾病发病年龄在新生儿期，则使用出生时患病率乘以该疾病持续年龄段的人口在世界总人口中的比例值作为疾病的点患病率；
    ⑤如果疾病的流行性类型为终生患病率，则使用默认的百万分之一作为该疾病的点患病率。
  6. 根据权利要求1所述的罕见病可视化问答式辅助鉴别诊断系统，其特征在于，表型推荐服务包括：
    表型的信息增益计算，包括：
    (A-1)使用信息熵来表示疾病推荐结果的不确定性，以公式(6)定义有分类权重的信息熵Entropy(W)：
    其中，W是一个权重集合，wi表示权重集合W中单一元素的权值，Wsum表示W中所有元素权值之和；
    当前对话状态下的疾病后验概率的信息熵Entropycur(D)为：
    Entropycur(D)＝Entropy({Probposterior(d)|d∈D})#(7)
    其中，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验 概率；
    (A-2)单一表型条件下的信息增益为：
    其中，第一项为当前对话状态下的疾病后验概率的信息熵，第二项为表型p为阳性时的条件熵，第三项为表型p为阴性时的条件熵，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验概率，γ为阳性表型分支权重，即{Probposterior(d)*Pr(p|d)|d∈D}集合元素之和，Pr(p|d)表示表型p在疾病d中的出现频率，表示表型p不在疾病d中出现的频率；
    表型的基尼指数计算，包括：
    (B-1)使用基尼指数来表示疾病推荐结果的纯度，以公式(9)定义有分类权重的基尼指数Gini(W)：
    其中，W是一个权重集合，wi表示权重集合W中单一元素的权值，Wsum表示W中所有元素权值之和；
    当前对话状态下的疾病后验概率的基尼指数为：
    Ginicur(D)＝Gini({Probposterior(d)|d∈D})#(10)
    其中，D表示所有罕见病的集合，Probposterior(d)表示疾病d的后验概率；
    (B-2)单一表型条件下的基尼指数增益为：
    其中，第一项为当前对话状态下的疾病后验概率基尼指数，第二项为表型p为阳性时的条件基尼指数，第三项为表型p为阴性时的条件基尼指数，其余参数定义等同公式(8)；
    将信息增益和基尼指数进行自适应融合，获得融合度量AIGGI：
    AIGGI(p,D)＝α*InfoGain(p,D)+(1-α)*GiniGain(p,D)#(12)
    其中，D表示所有罕见病的集合，p表示单一表型，α为反应当前对话 状态下疾病后验概率分布的不确定程度或杂乱度，α使用Ginicur(D)来表示；
    系统根据当前对话状态选择融合度量AIGGI最大的表型进行询问，用户针对该表型选择“存在”、“不存在”或者“不确定”的回答。
  7. 根据权利要求6所述的罕见病可视化问答式辅助鉴别诊断系统，其特征在于，在表型询问推荐时应用注释转播规则来过滤掉一些表型，包括：
    已汇报的表型不会在接下来的对话轮次中被询问；
    对于已汇报的阳性表型，默认其祖先表型均为阳性，将其祖先表型过滤掉，不再询问；
    对于已汇报的阴性表型，默认其后代表型均阴性，将其后代表型过滤掉，不再询问。
  8. 一种采用如权利要求1-7任一项所述的系统进行罕见病辅助鉴别诊断的方法，包括：
    (1)用户在前端UI模块的患者信息采集子模块中录入患者信息，包括患者年龄、出生时性别、阳性表型以及阴性表型；
    (2)对话诊断服务模块根据当前对话状态选择融合度量AIGGI最大的表型进行询问，用户针对该表型选择“存在”、“不存在”或者“不确定”的回答后，之后进入下一轮对话；
    问答过程中，用户可以在前端UI模块的可视化的问答式鉴别诊断子模块中获得多种可视化信息，包括疾病优先级评估、当前对话状态、患者表型状态以及对话历史；
    (3)重复步骤(2)，直到疾病推荐结果的不确定性低于阈值或者达到最大对话轮次时，系统根据疾病后验概率给出疾病推荐并结束对话。
PCT/CN2023/077386 2022-09-20 2023-02-21 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法 Ceased WO2024060508A1 (zh)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
CN202211146285.0 2022-09-20
CN202211146285.0A CN115482926B (zh) 2022-09-20 2022-09-20 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法

Publications (1)

Publication Number Publication Date
WO2024060508A1 true WO2024060508A1 (zh) 2024-03-28

Family

ID=84392464

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2023/077386 Ceased WO2024060508A1 (zh) 2022-09-20 2023-02-21 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法

Country Status (2)

Country Link
CN (1) CN115482926B (zh)
WO (1) WO2024060508A1 (zh)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118093252A (zh) * 2024-04-28 2024-05-28 浪潮云信息技术股份公司 一种云平台的数据库诊断方法及装置
CN119089370A (zh) * 2024-11-08 2024-12-06 中国电建集团华东勘测设计研究院有限公司 基于知识图谱和贝叶斯网络的大坝安全异常诊断方法
WO2026026689A1 (zh) * 2024-07-30 2026-02-05 四川大学华西医院 动态惊恐障碍治疗方案推荐系统、治疗系统及存储介质
CN121709289A (zh) * 2026-02-12 2026-03-20 湖南工商大学 基于多模态融合的罕见病诊疗资源协同调度方法及系统

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115482926B (zh) * 2022-09-20 2024-04-09 浙江大学 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107577907A (zh) * 2017-09-08 2018-01-12 成都奇恩生物科技有限公司 一种基于互联网的罕见病辅助诊断系统及使用方法
US20200286622A1 (en) * 2018-11-29 2020-09-10 Gachon University Of Industry-Academic Cooperation Foundation Data analysis methods and systems for diagnosis aids
CN112270988A (zh) * 2020-12-04 2021-01-26 厦门基源医疗科技有限公司 一种罕见病辅助诊断的方法及系统
CN113889265A (zh) * 2021-10-15 2022-01-04 浙江大学 一种基于表型可视化的罕见病辅助推理的方法及系统
CN113889259A (zh) * 2021-09-06 2022-01-04 浙江工业大学 一种知识图谱辅助下的自动诊断对话系统
CN115482926A (zh) * 2022-09-20 2022-12-16 浙江大学 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
DE102011016691A1 (de) * 2011-04-11 2012-10-11 medPad GmbH Verfahren und System zur Unterstützung bei der Auswahl wenigstens eines Objektes aus einer Gruppe von gespeicherten Objekten
ITMI20142149A1 (it) * 2014-12-16 2016-06-16 L I A Di Giuseppe Capasso Apparato di diagnosi differenziale in medicina adattato per determinare la sequenza di test ottimale atta ad identificare una patologia secondo criteri di appropriatezza diagnostica
US11404170B2 (en) * 2016-04-18 2022-08-02 Soap, Inc. Method and system for patients data collection and analysis
CN109147930A (zh) * 2017-06-28 2019-01-04 京东方科技集团股份有限公司 分诊对话方法、分诊对话设备及系统
WO2019225242A1 (ja) * 2018-05-23 2019-11-28 パナソニックＩｐマネジメント株式会社 摂食嚥下機能評価方法、プログラム、摂食嚥下機能評価装置および摂食嚥下機能評価システム
US20210343414A1 (en) * 2018-10-22 2021-11-04 The Jackson Laboratory Methods and apparatus for phenotype-driven clinical genomics using a likelihood ratio paradigm
WO2021211326A1 (en) * 2020-04-16 2021-10-21 Ix Layer Inc. Systems and methods for access management and clustering of genomic, phenotype, and diagnostic data
CN111899884A (zh) * 2020-06-23 2020-11-06 北京左医科技有限公司 智能辅助问诊方法、装置和存储介质
CN111984771A (zh) * 2020-07-17 2020-11-24 北京欧应信息技术有限公司 一种基于智能对话的自动问诊系统
CN113707299B (zh) * 2021-08-27 2024-09-27 平安科技（深圳）有限公司 基于问诊会话的辅助诊断方法、装置及计算机设备
CN114328864A (zh) * 2021-12-14 2022-04-12 合肥奥比斯科技有限公司 基于人工智能和知识图谱的眼科问答系统

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107577907A (zh) * 2017-09-08 2018-01-12 成都奇恩生物科技有限公司 一种基于互联网的罕见病辅助诊断系统及使用方法
US20200286622A1 (en) * 2018-11-29 2020-09-10 Gachon University Of Industry-Academic Cooperation Foundation Data analysis methods and systems for diagnosis aids
CN112270988A (zh) * 2020-12-04 2021-01-26 厦门基源医疗科技有限公司 一种罕见病辅助诊断的方法及系统
CN113889259A (zh) * 2021-09-06 2022-01-04 浙江工业大学 一种知识图谱辅助下的自动诊断对话系统
CN113889265A (zh) * 2021-10-15 2022-01-04 浙江大学 一种基于表型可视化的罕见病辅助推理的方法及系统
CN115482926A (zh) * 2022-09-20 2022-12-16 浙江大学 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN118093252A (zh) * 2024-04-28 2024-05-28 浪潮云信息技术股份公司 一种云平台的数据库诊断方法及装置
WO2026026689A1 (zh) * 2024-07-30 2026-02-05 四川大学华西医院 动态惊恐障碍治疗方案推荐系统、治疗系统及存储介质
CN119089370A (zh) * 2024-11-08 2024-12-06 中国电建集团华东勘测设计研究院有限公司 基于知识图谱和贝叶斯网络的大坝安全异常诊断方法
CN121709289A (zh) * 2026-02-12 2026-03-20 湖南工商大学 基于多模态融合的罕见病诊疗资源协同调度方法及系统

Also Published As

Publication number Publication date
CN115482926B (zh) 2024-04-09
CN115482926A (zh) 2022-12-16

Similar Documents

Publication Publication Date Title
CN115482926A (zh) 知识驱动的罕见病可视化问答式辅助鉴别诊断系统及方法
CN116775911B (zh) 基于问卷和大模型的医学队列随访对话辅助方法及系统
CN111191048B (zh) 基于知识图谱的急诊问答系统构建方法
CN111899884A (zh) 智能辅助问诊方法、装置和存储介质
CN110489566A (zh) 一种智能导诊服务机器人的导诊方法
CN112102937A (zh) 一种慢性病辅助决策的患者数据可视化方法及系统
US20180322954A1 (en) Method and device for constructing medical knowledge graph and assistant diagnosis method
CN111813957A (zh) 基于知识图谱的医疗导诊方法和可读存储介质
CN111984771A (zh) 一种基于智能对话的自动问诊系统
CN103559637A (zh) 一种为就诊患者推荐医生的方法及其系统
WO2022141925A1 (zh) 一种智能医学服务系统、方法及存储介质
CN107408143A (zh) 适于采用诊断适当性标准来确定用于识别病变的诊断试验的最佳顺序的医疗鉴别诊断装置
CN120032782A (zh) 基于大模型和专家策略的病史识别方法及系统
CN112927782B (zh) 一种基于文本情感分析的心身健康状态预警系统
CN120012896B (zh) 基于人工智能的主要诊断合理性判断方法及系统
CN120296135B (zh) 一种基于小专家模型的咨询分流对话系统
CN116993421A (zh) 一种基于大语言模型的患者评价系统
CN110827935A (zh) 一种中西医医疗大数据的数据质控方法及系统
CN120600353A (zh) 一种基于大模型智能体的ai护理白板系统及方法
CN119227784A (zh) 一种基于大语言模型的决策树生成方法及相关设备
CN119964827A (zh) 医疗检索方法及相关方法、装置、设备和存储介质
CN118888156A (zh) 基于患者对话的临床试验项目匹配方法、装置及电子设备
CN120280176B (zh) 基于互联网技术的跟随式慢病管理系统
CN117457180A (zh) 一种新生儿遗传代谢病筛查辅助系统
CN120913816A (zh) 一种患者-专家智能匹配系统及基于诊疗路径的分诊方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 23866812

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 23866812

Country of ref document: EP

Kind code of ref document: A1