Rare disease auxiliary diagnosis recommendation method and system

Through reinforcement learning, the integration of different diagnostic lists and the characteristics of simulated and real case libraries can be used to generate more accurate diagnostic lists of rare diseases, solving the problems of inconsistent diagnosis lists and high misdiagnosis rates in the existing technology, and improving the efficiency of diagnosis of rare diseases and accuracy.

CN118969250APending Publication Date: 2024-11-15ZHEJIANG UNIV
View PDF 0 Cites 5 Cited by

Patent Information

Application Number
CN202411050576.9
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-01
Publication Date
2024-11-15

AI Technical Summary

Technical Problem

During the diagnosis of rare diseases, the diagnostic list generated by the existing phenotype-based diagnostic recommendation algorithm is inconsistent and has a high misdiagnosis rate, resulting in a high initial misdiagnosis rate, wasted medical resources and delayed diagnosis.

Method used

Using reinforcement learning-based methods, by constructing DQN neural networks, fusing different diagnostic lists, using simulated rare disease case library and real phenotype to collect quality characteristics, train agents to generate more accurate diagnostic lists, and assist doctors in diagnosis.

Benefits of technology

It improves the accuracy and efficiency of diagnosis of rare diseases, reduces the rate of misdiagnosis, and improves the ability of clinicians to make correct diagnosis.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118969250A_ABST
    Figure CN118969250A_ABST
Patent Text Reader

Abstract

The invention discloses a rare disease auxiliary diagnosis recommendation method and system, and the method comprises the steps: (1) collecting a rare disease clinical case report, and obtaining the quantitative statistical information of phenotypic collection quality features; the phenotype acquisition quality characteristics comprise distribution of phenotype acquisition quantity and distribution of different quality phenotype types; (2) constructing a simulated rare disease case library conforming to epidemiological characteristics and phenotypic acquisition quality characteristics; (3) constructing a DQN neural network, and training the DQN neural network by adopting the simulated rare disease case library based on a reinforcement learning algorithm to obtain a fusion agent; and (4) obtaining a plurality of diagnosis recommendation lists of the target patient case according to a plurality of existing rare disease recommendation methods, and fusing the plurality of diagnosis recommendation lists through a fusion agent to obtain a fusion recommendation list of the target patient case. According to the fusion diagnosis list obtained by the method, the comprehensive ranking of real diagnosis in the list can be improved, so that clinical doctors can make correct diagnosis more favorably.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the field of intelligent auxiliary diagnosis technology, and in particular to a method and system for recommending auxiliary diagnosis of rare diseases based on reinforcement learning and fusion of multi-source diagnosis lists. Background Art

[0002] Rare diseases, also known as orphan diseases, refer to diseases with extremely low incidence rates, usually genetic diseases. The World Health Organization (WHO) defines rare diseases as diseases with a prevalence rate of 6.5 to 10 per 10,000 people, based on the identification of rare diseases in different regions. Although the prevalence and morbidity of rare diseases are very low, due to the wide variety of rare diseases, there are nearly 7,000 rare diseases confirmed internationally, accounting for about 10% of human diseases. Overall, rare disease patients are not rare, with more than 300 million rare disease patients worldwide, and a large number of undiagnosed rare patients.

[0003] About 80% of rare diseases are caused by genetic factors, and other causes include infection, allergic reaction, collective degeneration or hyperplasia, etc.; about 50% of rare diseases occur at birth or in childhood; most rare diseases are chronic, serious diseases that are often life-threatening and are currently the most important factor in death among children under 5 years old.

[0004] Diagnosis is the first and most important link in the entire rare disease rescue chain. However, due to the complexity of rare disease knowledge, low awareness of rare diseases among clinicians, serious imbalance in the distribution of superior medical resources, and low accessibility of molecular diagnostic technologies such as gene sequencing, these problems lead to nearly half of rare disease patients being misdiagnosed in the early stage, and it usually takes several years to confirm the diagnosis. According to a survey by the European Organization for Rare Diseases (EURORDIS), a quarter of patients with rare genetic diseases take 5 to 30 years to be diagnosed, and the initial misdiagnosis rate is as high as 40%. All these misdiagnoses will lead to a large number of ineffective medical treatments and even unnecessary surgeries, seriously endangering the health of patients, causing waste of medical resources, and delaying the treatment time of patients. Many infants with rare genetic diseases have died before obtaining laboratory-confirmed genetic diagnoses. Genetic analysis is a common method for diagnosing rare genetic diseases, but genetic diagnosis has the disadvantages of high cost and long time. A fast and effective diagnostic method is urgently needed in the field of rare diseases.

[0005] Phenotype-based diagnostic methods have the potential to become a fast and effective method for diagnosing rare diseases. However, since the complex relationship network composed of more than 7,000 rare diseases and tens of thousands of human phenotypes often exceeds the ability of human knowledge learning, various phenotype-based rare disease recommendation algorithms have been proposed to assist clinical personnel in completing diagnosis. These include several typical technical routes:

[0006] (1) The diagnosis recommendation algorithm based on phenotype matching is to match the patient's phenotype with the disease phenotype in the knowledge base or obtain a potential disease recommendation list based on similarity distance calculation method.

[0007] (2) The diagnosis recommendation algorithm based on Bayesian theory obtains various quantitative probabilistic knowledge between phenotypes and diseases based on Bayesian theory, and recommends a diagnosis list through a set of calculation processes. Compared with simple phenotype matching, it can obtain a better diagnosis list.

[0008] (3) Recommendation algorithms based on large language models. Large language models trained with massive literature, such as ChatGPT, and large language models optimized with professional corpus training in the medical field have also been shown to be able to recommend potential diagnosis lists.

[0009] These different technical routes can all generate diagnostic lists, which perform differently under different disease rarity, different phenotypic relationships, and different information collection qualities, and usually have different effects. Summary of the invention

[0010] In the process of diagnosing rare diseases, the diagnosis lists recommended by different phenotype-based diagnosis recommendation algorithms are inconsistent and have potential complementary characteristics. The present invention provides a method based on reinforcement learning to fuse related diagnosis lists to obtain a more accurate diagnosis list, thereby assisting doctors in improving the diagnostic efficiency of rare diseases.

[0011] The technical solution of the present invention is as follows:

[0012] A recommended method for auxiliary diagnosis of rare diseases includes the following steps:

[0013] (1) Collect clinical case reports of rare diseases and obtain quantitative statistical information on the quality characteristics of phenotypic collection; the quality characteristics of phenotypic collection include the distribution of the number of phenotypic collections and the distribution of phenotypic types of different quality;

[0014] (2) construct a simulated rare disease case library that meets the epidemiological characteristics and phenotypic collection quality characteristics;

[0015] (3) Construct a DQN neural network and train it using a simulated rare disease case library based on a reinforcement learning algorithm to obtain a fusion agent;

[0016] (4) According to several existing rare disease recommendation methods, several diagnostic recommendation lists of target patient cases are obtained respectively, and the phenotypic information of the target patient, the several diagnostic recommendation lists and related quantitative information are input into the fusion agent to obtain the fusion recommendation list of the target patient case.

[0017] The present invention utilizes reinforcement learning methods to perform training on a large number of rare disease case libraries with real phenotypic collection quality characteristics and epidemiological characteristics, so that more valuable diagnoses can be preferentially obtained from different diagnosis lists, and a method for fusing different diagnosis lists is obtained, providing a more accurate diagnosis list. The obtained more accurate diagnosis list can better assist doctors in diagnosing rare diseases.

[0018] Since real cases of rare diseases are relatively scarce, the present invention first obtains statistical information on the real phenotypic collection quality based on clinical cases of rare diseases, and then constructs simulated rare disease cases based on the phenotypic collection quality characteristics of real cases and the epidemiological characteristics of rare diseases to form a simulated rare disease case library for subsequent training of the DQN neural network.

[0019] Step (1) comprises:

[0020] (1-1) Collect clinical case reports of rare diseases, perform manual or automatic phenotype annotation on the phenotypes, and classify the phenotypes in combination with rare disease knowledge bases (such as Orphanet);

[0021] Phenotype types include precise phenotype, imprecise phenotype, and noise phenotype;

[0022] (1-2) Statistically analyze the distribution of the number of phenotypes and the distribution of phenotypes of different qualities in clinical cases of rare diseases.

[0023] Precise phenotype refers to a phenotype that is clearly and accurately annotated in the rare knowledge base, or a sub-concept that is a phenotype that is clearly and accurately annotated. For example, "ventricular septal defect" is a sub-concept of "congenital heart disease". For rare diseases with "congenital heart disease" phenotypic annotation information, "ventricular septal defect" and "congenital heart disease" are precise phenotypes.

[0024] An imprecise phenotype refers to a phenotype that does not appear in the disease phenotype annotation of the rare disease knowledge base but belongs to the broad concept of a precise phenotype. For example, "heart disease" is a generalized concept of "congenital heart disease", and "congenital heart disease" is the superordinate broad concept of "ventricular septal defect". Superordinate concepts such as "heart disease" and "congenital heart disease" are imprecise phenotypes for rare diseases annotated with their subordinate concepts such as "congenital heart disease" or "ventricular septal defect". At the same time, this hierarchy is not limited to the "father-child" level relationship, but can be pushed up to the higher "ancestor" level concept.

[0025] Noise phenotypes refer to phenotypes that are not related to the annotated phenotypes of the rare disease in the rare disease knowledge base, that is, they belong neither to the descendant-level phenotype concept nor to the ancestral-level concept.

[0026] Step (2) includes: constructing simulated rare disease cases and establishing a simulated rare disease case library based on the incidence rate and phenotypic probability annotated in the rare disease knowledge base, combined with the distribution of phenotypic quantity and the distribution of phenotypic types of different qualities in rare disease clinical cases.

[0027] Furthermore, step (2) includes:

[0028] (2-1) Randomly select a candidate rare disease from the rare disease knowledge base based on the incidence of the rare disease;

[0029] (2-2) Randomly selecting phenotypes based on the phenotypic annotations of the candidate rare disease in the rare disease knowledge base and the phenotype occurrence frequency, and the number of selected phenotypes is randomly determined based on the distribution of the number of phenotypes in the real case library;

[0030] (2-3) Based on the distribution of different quality phenotype types, the phenotype is randomly changed to its ancestral node to generate an imprecise phenotype, and the rare disease knowledge base is combined to find irrelevant phenotypes to generate noise phenotypes;

[0031] (2-4) Repeat steps (2-1)-(2-3) to generate a large number of simulated rare disease cases and build a simulated rare disease case library.

[0032] Existing rare disease recommendation methods include at least two of the following: a diagnosis recommendation algorithm based on phenotype matching, a diagnosis recommendation algorithm based on Bayesian theory, and a recommendation algorithm based on a large language model.

[0033] The diagnosis recommendation algorithm based on phenotype matching uses the patient's phenotype to match the disease phenotype in the knowledge base or obtains a potential disease recommendation list based on similarity distance calculation.

[0034] The diagnosis recommendation algorithm based on Bayesian theory obtains various quantitative probabilistic knowledge between phenotypes and diseases based on Bayesian theory, and recommends a diagnosis list through a set of calculation processes. Compared with simple phenotype matching, it can obtain a better diagnosis list.

[0035] Recommendation algorithms based on large language models, large language models trained with massive literature, such as ChatGPT, and large language models trained and optimized with professional corpus in the medical field, have also been proven to be able to recommend potential diagnosis lists.

[0036] The DQN neural network is a multilayer perceptron (MLP).

[0037] Furthermore, step (3) includes:

[0038] (3-1) Construct a multilayer perceptron that takes the patient's phenotypic information, a diagnosis recommendation list obtained by multiple existing rare disease recommendation methods, and related quantitative information as input. t , the output is to take action a in this state t The expected cumulative reward Q(s t ,a t );

[0039] (3-2) randomly obtaining a simulated rare disease case from a simulated rare disease case library; based on the phenotype, generating multiple diagnostic recommendation lists and related quantitative information using multiple existing rare disease recommendation methods;

[0040] (3-3) Inputting the phenotypic information of the simulated rare disease case, multiple diagnosis recommendation lists and related quantitative information into the multilayer perceptron, the multilayer perceptron selects a diagnosis recommendation list, and the selected diagnosis recommendation list pops up the highest-ranked diagnosis. If the popped-up diagnosis is not in the current fusion diagnosis list, it is placed at the end of the fusion diagnosis list. If the popped-up diagnosis is in the current fusion diagnosis list, it is not placed in the fusion diagnosis list, and then the status is updated; the reward is calculated according to the reward mechanism; the status and reward are recorded;

[0041] (3-4) Repeat step (3-3) until the fusion recommendation list is filled;

[0042] (3-5) Repeat steps (3-2)-(3-4), train the multilayer perceptron according to the recorded states and rewards, and obtain a fusion agent.

[0043] Preferably, the multilayer perceptron adopts a network structure with one hidden layer, uses a ReLU function as an activation function of the hidden layer, and uses a sigmoid function as an activation function of the output layer.

[0044] Preferably, the multilayer perceptron selects the diagnosis recommendation list based on a greedy strategy, selects a diagnosis recommendation list for exploration with a probability of ε, and then selects the current optimal diagnosis recommendation list for utilization with a probability of 1-ε.

[0045] The reward mechanism is:

[0046] (a) When the highest-ranked diagnosis popped up is the true target diagnosis, the reward is +2;

[0047] (b) In the diagnosis recommendation list after the status is updated, if the true target diagnosis improves in the highest ranking among all recommended diagnosis lists, a reward of +1 is given; if the true target diagnosis does not improve in the highest ranking among all recommended diagnosis lists, a penalty of -1 is given.

[0048] In step (3), the target network and experience replay are used during training;

[0049] The Q value can be calculated by the following equation:

[0050]

[0051] θ and θ′ represent the parameters of the training network and the target network respectively; Q(s t ,a t |θ) represents the state s of the training network at time t t Take action a t The maximum expected Q value; r t Indicates state s t Take action a t The reward value of state s t Take action a t Then enter the next state s t+1 ;Q * (s t+1 ,a t+1 |θ′) represents the target network in state s t+1 Take action a t+1 Q value; γ represents the discount rate;

[0052] The size of the experience replay buffer is set to 10,000, the batch size is 64, and the learning rate is 0.03;

[0053] The squared error loss function is adopted and the root mean square propagation is selected as the optimizer.

[0054] Preferably, an XGboost model is used to complete the filling of the experience replay buffer in the initial stage.

[0055] The present invention also provides a rare disease auxiliary diagnosis recommendation system, comprising a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor, wherein the computer processor implements the following steps when executing the computer program: inputting the phenotypic information of the target patient, a plurality of diagnosis recommendation lists, and related quantitative information into a fusion agent to obtain a fusion recommendation list of the target patient case;

[0056] Several diagnostic recommendation lists and related quantitative information are obtained from several existing rare disease recommendation methods based on phenotypic information of target patient cases.

[0057] Compared with the existing rare disease diagnosis recommendation technology, the beneficial effects of the present invention are:

[0058] (1) Recommendation algorithms of different technical routes have different advantages and disadvantages under different disease incidence, phenotype collection quality, knowledge base completeness, etc. Combining the advantages of these technologies can provide a more accurate diagnosis list;

[0059] (2) The present invention simulates the incidence of different diseases and the quality of phenotypes collected in actual clinical practice during the simulation case generation process. Through the reinforcement learning process, the fusion method can learn how to better adapt to different recommendation technology routes under different disease incidence, phenotype collection quality, and recommendation result credibility.

[0060] (3) The fusion diagnosis list obtained by the present invention will improve the comprehensive ranking of the real diagnosis in the list, thereby being more conducive to clinicians making correct diagnoses. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 Schematic diagram of the construction process of a simulated rare disease case library with epidemiological characteristics and phenotypic collection quality characteristics;

[0062] Figure 2 Schematic diagram of the training process of fusing multi-source diagnosis lists based on reinforcement learning. DETAILED DESCRIPTION

[0063] The present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be pointed out that the embodiments described below are intended to facilitate the understanding of the present invention and do not have any limiting effect on the present invention.

[0064] The present invention collects real clinical cases of rare diseases, and uses the annotations of the relevant human phenotype database (HPO, HumanPhenotype Ontology) and the most comprehensive rare disease knowledge base (Orphanet) of mankind to establish a statistical model for the quality level of the real collected clinical phenotypes. Then, based on the incidence of rare diseases and phenotype probability, a simulated rare disease case library with epidemiological characteristics and phenotype collection quality characteristics is constructed. A reinforcement learning strategy is used to design a reward mechanism. Through continuous reinforcement learning in the simulated case library, the intelligent agent can adaptively learn how to fuse and construct a more accurate recommendation list from the diagnosis lists recommended by different mechanisms according to disease characteristics, phenotype quality, recommendation list status, etc.

[0065] Specifically include:

[0066] (1) The process of acquiring quality characteristics of phenotypic collection

[0067] In the collection of phenotypic characteristics of rare disease cases, there are often problems such as inaccurate phenotypic expression and mixing of other unrelated phenotypes. This is also the biggest challenge in conducting diagnoses based on phenotypes, and it is also the real state of using relevant tools in the real world. Therefore, we need to evaluate such quality levels to provide a basis for the construction of subsequent simulation cases. To this end, we need to collect real clinical cases. The most effective source of information is published rare disease cases. These rare disease case reports have clinical descriptions. Based on these descriptions, HPO terms can be annotated manually or with automated tools. The HPO terms annotated in these real cases can be divided into several situations through the relevant phenotypic annotations of the corresponding rare diseases in the Orphanet knowledge base:

[0068] The HPO phenotype term annotated in the case accurately matches the HPO phenotype annotation in the knowledge base, or belongs to the descendant concept of the HPO phenotype in the knowledge base, that is, the concept of the refinement of the concept. For example, "ventricular septal defect" belongs to the subconcept of "congenital heart disease". This phenotype is an "accurate" matching phenotype;

[0069] The HPO phenotype term annotated in the case does not appear in the phenotype annotation of the disease in the knowledge base, but the concept "heart disease" generalized from the knowledge concept is a generalized concept of "congenital heart disease". This phenotype is an "inaccurate" matching concept;

[0070] The HPO phenotype terms annotated in the case do not belong to the above situation and are completely irrelevant to the phenotypes annotated in the knowledge base. This phenotype is "noise" and can only play a negative role in phenotype-based diagnostic recommendations.

[0071] By collecting a certain number of real clinical cases, statistical models can be constructed for the "precision", "imprecision" and "noise" of the real collected phenotypes, as well as the number of phenotypes that may be collected in a real case. For example, in a database based on 283 real cases of rare diseases, the number of phenotypes collected in a case usually conforms to the normal distribution, with an average of 9.1. Further statistics of phenotypes of different quality levels show that there are 2.0 "precise" matching phenotypes, 3.1 "imprecise" phenotypes, and 3.9 "noise" phenotypes that have no relationship. These statistical models will serve to artificially introduce "imprecise" and "noise" phenotypes with the same statistical characteristics when constructing simulation cases in the later stage, so as to better simulate real cases.

[0072] (2) Process of obtaining epidemiological characteristics

[0073] Orphanet, an authoritative knowledge base for rare diseases, contains information on disease prevalence, average age of onset and death, disease phenotype frequency characteristics, and pathogenic genes. At the same time, Orphanet and HPO annotate the phenotype-disease relationship of rare diseases. At the same time, the frequency of phenotypes in rare diseases is also divided into several levels:

[0074] · Phenotypes that will inevitably appear 100%

[0075] Very common phenotype 80-99%

[0076] Common phenotypes 30-79%

[0077] Sometimes 5-29%

[0078] As of February 2022, a total of 4,257 rare diseases with phenotypic characteristics have been counted in the Orphanet knowledge base, which are the target recommended disease set of this system. Orphanet provides a knowledge graph in XML file format, which is parsed and stored in the relational database MySQL. These statistical information will also be used to simulate case generation.

[0079] (3) The process of building a simulated rare disease case library with epidemiological characteristics and phenotypic collection quality characteristics

[0080] In order to overcome the problem of scarcity of real cases of rare diseases, we use the incidence rate, phenotypic probability, etc. annotated in the rare disease knowledge base, the phenotypic collection quality characteristics obtained in (1) (2) and the epidemiological characteristics of rare diseases to construct a simulated case library. The process is as follows: Figure 1 As shown, the specific process is as follows:

[0081] (3-1) Based on the incidence of the disease, a candidate rare disease is randomly selected from 4257 diseases in the Orphanet database;

[0082] (3-2) Randomly select phenotypes based on the phenotypic annotations of the rare disease in Orphanet and their frequency of occurrence. The number of phenotypes selected is randomly determined based on the normal distribution model of the actual number of phenotypes collected;

[0083] (3-3) According to the distribution of "imprecision" and "noise" in the real phenotype collection, the "imprecise" phenotype is generated by randomly changing the generated phenotype as the ancestral node, and then the knowledge base is combined to find irrelevant phenotypes to generate "noise". This random process makes the quality of the relevant phenotypes in the final simulation case library consistent with the actual collection;

[0084] (3-4) By repeating the above process, a simulation case library of massive cases can be generated. In the embodiment, we generate 10,000 cases.

[0085] (4) Reinforcement learning process based on a massive simulation case library

[0086] Reinforcement learning is a machine learning technique that emphasizes how to act based on the environment to maximize expected benefits. It mimics the trial-and-error learning process that humans take to achieve their goals. Actions that contribute to achieving the goal will be reinforced, while actions that deviate from the goal will be ignored. This relies on a set of learning strategies and reward mechanisms. In this process, we define an agent to prioritize a diagnosis from the top diagnosis in each diagnosis list to the fusion diagnosis list, and then each diagnosis list is in the form of a stack, popping the diagnosis at the head of the selected diagnosis list. If the popped-up diagnosis is not in the current fusion diagnosis list, it will be placed at the end of the fusion diagnosis list. If the popped-up diagnosis is in the current fusion diagnosis list, only the state calculation reward is recorded, and the popped-up diagnosis is not placed in the fusion diagnosis list. After the agent's selection, rewards are given based on whether the true diagnosis improves in the highest ranking of all candidate diagnosis lists. A reward of +2 is given for direct matches between the selected disease and the target disease. If the minimum distance between the target disease and the top of the stack in each candidate diagnosis list stack decreases, a reward of +1 is given. If the minimum distance between the target disease and the top of the stack in each candidate diagnosis list stack does not decrease, a penalty of -1 is given. This selection process is repeated until the fused diagnosis list is full.

[0087] The goal of reinforcement learning is to select a candidate diagnosis from multiple diagnosis lists generated by different technical routes and build a new diagnosis recommendation list. The reinforcement learning process needs to automatically optimize the selection strategy through rewards and penalties fed back by the environment in each selection. This training process is shown in Figure 2 As shown, the details are as follows:

[0088] (4-1) Randomly obtain a case from the simulation case library, which contains generated phenotypes with different quality levels and target diagnosis.

[0089] (4-2) Based on the phenotype, algorithms with different technical routes can generate multiple different diagnostic recommendation lists.

[0090] This technical route, in the embodiments, includes methods based on phenotype matching, methods based on Bayesian statistics, and methods based on GPT-4.0, etc. In addition to this diagnostic list, the quantitative outputs of different models, such as matching scores, posterior probabilities, and other outputs of different calculation route algorithms, are also used as features for subsequent reinforcement learning.

[0091] (4-3) The initial intelligent body will select a diagnosis list under the greedy strategy according to the input state. All diagnosis lists adopt the stack mode, and then only pop up the highest-ranking diagnosis in the selected diagnosis list. If the popped diagnosis is not in the current fusion diagnosis list, it will be placed at the end of the fusion diagnosis list. If the popped diagnosis is in the current fusion diagnosis list, it will not be placed in the fusion diagnosis list. The remaining diagnoses in the selected diagnosis list remain unchanged, and then the state is updated.

[0092] (4-4) This selection process rewards (+1) or penalizes (-1) by checking whether the minimum distance from the top of the stack to the true target diagnosis in each candidate diagnosis list stack has become smaller in the new state, especially when the selected diagnosis is the target diagnosis, an additional reward (+2) is given. For example, for two candidate diagnosis lists A and B, the true diagnosis in A ranks 3 and the true diagnosis in B ranks 4. At this time, the minimum distance from the top of the stack to the true diagnosis in the two list stacks is 3. At this time, if the disease at the top of list A is popped, the minimum distance from the top of the stack to the true diagnosis in the two list stacks is reduced to 2, and the reward is +1; if the disease at the top of list B is popped, the minimum distance from the top of the stack to the true diagnosis in the two list stacks is still 3, and the penalty is -1.

[0093] (4-5) Through reinforcement learning of a large number of simulated cases, the intelligent body will learn how to make choices with the highest reward in different states.

[0094] The present invention is implemented using Q-learning, which records the learned strategies and tells the agent what actions to take under what circumstances to get the maximum reward value. In the embodiment, the structure of the specific reinforcement learning model is as follows:

[0095] Construct a multilayer perceptron (MLP). The input of this MLP is the phenotypic information of the current patient, the list of various diagnostic recommendation algorithms, and the relevant quantitative output. Here, s t Indicates that its output is to take action a in this state t The expected cumulative reward Q(s t ,a t ), reinforcement learning uses a greedy strategy (ε-greedy strategy) to select a random action for exploration with a probability of ε (ε is 0.6 in this example), which may not be the current optimal action. Then, the current optimal action (with the highest cumulative reward or value) is selected for use with a probability of 1-ε.

[0096] In the embodiment, the MLP adopts a network structure with one hidden layer, uses a linear rectified unit (ReLU) as the activation function of the hidden layer, and a sigmoid function as the activation function of the output layer.

[0097] Two DQN training techniques are used during the training process, namely using the target network and experience replay. Specifically, according to the Bellman equation, the expected Q value can be calculated by the following equation:

[0098]

[0099] θ and θ′ represent the parameters of the training network and the target network respectively. t Indicates state s t Take action a t The reward value. γ represents the discount rate (set to 0.9 in the example). In each training set, we will conduct 100 diagnostic trainings and convert all the experiences (the experience value of each time step is e(s t ,a t ,r t ,s t+1 )) is updated to the experience replay pool. After each training set, we will extract samples from the experience replay pool for batch training. The size of the experience replay buffer is set to 10,000, the batch size is 64, and the learning rate is 0.03. We use the topic squared error loss function and choose the root mean square propagation (RMSprop) as the optimizer. The loss function can be expressed as:

[0100]

[0101] At the end of each training, the parameters of the target network are updated. In addition, we trained a simple XGboost model in the embodiment to complete the filling of the experience replay pool in the initial stage. This training technique allows DQN to obtain effective learning at the beginning and reduce the number of training times required. Finally, the best model of reinforcement learning is output to serve the diagnosis list fusion in real cases.

[0102] (5) Fusion recommendation in real diagnostic scenarios

[0103] In real clinical scenarios, only the phenotypes collected by clinical users are used as input, and then algorithms of various technical routes will recommend their own diagnosis lists. The phenotypes and these diagnosis recommendation lists are then used as input to enter the trained intelligent agent and output a fused recommendation list.

[0104] In the embodiment, we used 10,000 simulated rare disease cases to train a model. This model made diagnostic recommendations for 238 real cases, covering 69.33% of the real diagnoses in the top ten fusion diagnosis lists, and 63.87% of the real diagnoses in the top five fusion diagnosis lists. The first-ranked fusion recommendation diagnostic accuracy rate reached 41.60%. These results far exceed the corresponding results of GPT-4.0 (45.80%, 42.02% and 26.05%), and are also better than the Bayesian diagnostic method (phenotype-driven likelihood ratio analysis, PheLR) with the best single recommendation effect (68.49%, 58.40% and 41.60%), especially the top five recommendations, where the effect advantage is most obvious.

[0105] The embodiments described above provide a detailed description of the technical solutions and beneficial effects of the present invention. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, supplements and equivalent substitutions made within the scope of the principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for recommending auxiliary diagnosis of rare diseases, characterized in that: The following steps are involved: (1) Collect clinical case reports of rare diseases and obtain quantitative statistical information on the quality characteristics of phenotypic collection; the quality characteristics of phenotypic collection include the distribution of the number of phenotypic collections and the distribution of phenotypic types of different quality; (2) construct a simulated rare disease case library that meets the epidemiological characteristics and phenotypic collection quality characteristics; (3) Construct a DQN neural network and train it using a simulated rare disease case library based on a reinforcement learning algorithm to obtain a fusion agent; (4) According to several existing rare disease recommendation methods, several diagnostic recommendation lists of target patient cases are obtained respectively, and the phenotypic information of the target patient, the several diagnostic recommendation lists and related quantitative information are input into the fusion agent to obtain the fusion recommendation list of the target patient case.

2. The method for recommending auxiliary diagnosis of rare diseases according to claim 1, characterized in that: Step (1) comprises: (1-1) Collect clinical case reports of rare diseases, perform manual or automatic phenotype annotation on the phenotypes, and classify the phenotypes based on the rare disease knowledge base; Phenotype types include precise phenotype, imprecise phenotype, and noise phenotype; (1-2) Statistically analyze the distribution of the number of phenotypes and the distribution of phenotypes of different qualities in clinical cases of rare diseases.

3. The method for recommending auxiliary diagnosis of rare diseases according to claim 1, characterized in that: Step (2) comprises: (2-1) Randomly select a candidate rare disease from the rare disease knowledge base based on the incidence of the rare disease; (2-2) Randomly selecting phenotypes based on the phenotypic annotations of the candidate rare disease in the rare disease knowledge base and the phenotype occurrence frequency, and the number of selected phenotypes is randomly determined based on the distribution of the number of phenotypes in the real case library; (2-3) Based on the distribution of different quality phenotype types, the phenotype is randomly changed to its ancestral node to generate an imprecise phenotype, and the rare disease knowledge base is combined to find irrelevant phenotypes to generate noise phenotypes; (2-4) Repeat steps (2-1)-(2-3) to generate a large number of simulated rare disease cases and build a simulated rare disease case library.

4. The method for recommending auxiliary diagnosis of rare diseases according to claim 1, characterized in that: Existing rare disease recommendation methods include at least two of the following: a diagnosis recommendation algorithm based on phenotype matching, a diagnosis recommendation algorithm based on Bayesian theory, and a recommendation algorithm based on a large language model.

5. The method for recommending auxiliary diagnosis of rare diseases according to claim 1, characterized in that: The DQN neural network is a multi-layer perceptron.

6. The method for recommending auxiliary diagnosis of rare diseases according to claim 5, characterized in that: Step (3) includes: (3-1) Construct a multilayer perceptron that takes the patient's phenotypic information, a diagnosis recommendation list obtained by multiple existing rare disease recommendation methods, and related quantitative information as input. t , the output is to take action a in this state t The expected cumulative reward Q(s t ,a t ); (3-2) randomly obtaining a simulated rare disease case from a simulated rare disease case library; based on the phenotype, generating multiple diagnostic recommendation lists and related quantitative information using multiple existing rare disease recommendation methods; (3-3) Inputting the phenotypic information of the simulated rare disease case, multiple diagnosis recommendation lists and related quantitative information into the multilayer perceptron, the multilayer perceptron selects a diagnosis recommendation list, and the selected diagnosis recommendation list pops up the highest-ranked diagnosis. If the popped-up diagnosis is not in the current fusion diagnosis list, it is placed at the end of the fusion diagnosis list. If the popped-up diagnosis is in the current fusion diagnosis list, it is not placed in the fusion diagnosis list, and then the status is updated; the reward is calculated according to the reward mechanism; the status and reward are recorded; (3-4) Repeat step (3-3) until the fusion recommendation list is filled; (3-5) Repeat steps (3-2)-(3-4), train the multilayer perceptron according to the recorded states and rewards, and obtain a fusion agent.

7. The method for recommending auxiliary diagnosis of rare diseases according to claim 5 or 6, characterized in that: The multi-layer perceptron adopts a network structure with one hidden layer, uses the ReLU function as the activation function of the hidden layer, and uses the sigmoid function as the activation function of the output layer.

8. The method for recommending auxiliary diagnosis of rare diseases according to claim 6, characterized in that: The multilayer perceptron selects the diagnosis recommendation list based on the greedy strategy, selects a diagnosis recommendation list for exploration with a probability of ε, and then selects the current optimal diagnosis recommendation list for utilization with a probability of 1-ε.

9. The method for recommending auxiliary diagnosis of rare diseases according to claim 6, characterized in that: The reward mechanism is: (a) When the highest-ranked diagnosis popped up is the true target diagnosis, the reward is +2; (b) In the diagnosis recommendation list after the status is updated, if the highest ranking of the true target diagnosis in all recommended diagnosis lists improves, a reward of +1 is given; if the highest ranking of the true target diagnosis in all recommended diagnosis lists does not improve, a penalty of -1 is given.

10. A rare disease auxiliary diagnosis recommendation system, characterized in that: The invention comprises a computer memory, a computer processor, and a computer program stored in the computer memory and executable on the computer processor. When the computer processor executes the computer program, the following steps are implemented: Inputting the phenotypic information of the target patient, several diagnostic recommendation lists and related quantitative information into the fusion agent constructed according to the method of claim 1 to obtain a fusion recommendation list of the target patient case; Several diagnostic recommendation lists and related quantitative information are obtained from several existing rare disease recommendation methods based on phenotypic information of target patient cases.

Citation Information

Cited By

  • Rare disease risk screening model training system and method, screening system and medium

    CN121075621A

  • Heart amyloidosis risk prediction method and device and electronic equipment

    CN121260479A

  • Rare disease scene interactive differential diagnosis method and device

    CN121565434A

  • An interactive differential diagnosis method and device for rare diseases

    CN121565434B

  • Rare disease phenotype auxiliary diagnosis system based on knowledge and data double-driven residual network

    CN122822303A