Explainable and active mining artificial intelligence system to accelerate the diagnosis for rare disease

Abstract

In an embodiment, the present invention provides a computer-implemented, machine learning method for active mining for rare diseases. A patient disease profile is generated based on received input. Data of the patient disease profile is transformed into a representation in a feature space for comparison to representations of diseases in the feature space. One or more candidate diseases are identified based on distances in the feature space between the representation of the data and the representations of the diseases. One or more discriminative disease features are determined for one or more candidate diseases. An interview question is generated based on the one or more discriminative disease features. The method has applications including, but not limited to, use cases in medical Al / healthcare for optimization of predictions or to support decision-making.

Description

EXPLAINABLE AND ACTIVE MINING ARTIFICIAL INTELLIGENCE SYSTEM TO ACCELERATE THE DIAGNOSIS FOR RARE DISEASECROSS-REFERENCE TO PRIOR APPLICATION

[0001] Priority is claimed to U.S. Provisional Application Serial No. 63 / 545,762 filed on October 26, 2023, the entire contents of which is hereby incorporated by reference herein. FIELD

[0002] The present invention relates to Artificial Intelligence (Al) and machine learning (ML), and in particular to a method, system, data structure, computer program product and computer-readable medium for an explainable Al (XAI) system for active mining for rare diseases.BACKGROUND

[0003] People may suffer from rare or not well-understood diseases, as well as diseases that are difficult to diagnose or detect. Examples include myasthenia gravis (MG), seizures and paralysis. For patients with such rare diseases, diagnosis and subsequent treatment can often take many years. For doctors, it is impossible to know every rare disease and to stay up to date with the latest research.

[0004] Most of the recent biotechnology research to improve the diagnosis of rare disease focuses on specific modalities such as genetic signatures (see Chulalongkorn University, “Fast- tracking rare disease diagnosis and treatment,” <<https: / / www.nature.eom / articles / d42473-020- 00209-6» (accessed September 9, 2023)) or epigenetic signatures (see Beatriz Martinez- Delgado and Maria J. Barrero, “Epigenomic Approaches for the Diagnosis of Rare Diseases,” Epigenomes 2022, 6, 21 (2022), hereinafter Martinez-Delgado et al., each of the foregoing being hereby incorporated by reference herein). There is also increasing contributions leveraging Al in this domain. Most of the existing technological approaches with Al focus on a specific disease through a classification approach. To overcome the lack of patient data for rare diseases, such an approach, for instance, uses the deep learning technology of Generative Adversarial Networks (GANs) to generate more patient data for exocrine pancreatic insufficiency detection (see Kezi Yu, Yunlong Wang, Yong Cai, Cao Xiao, Emily Zhao, Lucas Glass, and Jimeng Sun, “Rare Disease Detection by Sequence Modeling with Generative Adversarial Networks,” arXiv: 1907.01022 (2019), hereinafter Yu et al., which is hereby incorporated by reference herein).SUMMARY

[0005] In an embodiment, the present invention provides a computer-implemented, machine learning method for active mining for rare diseases. A patient disease profile is generated basedon received input. Data of the patient disease profde is transformed into a representation in a feature space for comparison to representations of diseases in the feature space. One or more candidate diseases are identified based on distances in the feature space between the representation of the data and the representations of the diseases. One or more discriminative disease features are determined for one or more candidate diseases. An interview question is generated based on the one or more discriminative disease features. The method has applications including, but not limited to, use cases in medical Al / healthcare for optimization of predictions or to support decision-making.BRIEF DESCRIPTION OF THE DRAWINGS

[0006] Embodiments of the present invention will be described in even greater detail below based on the exemplary figures. The present invention is not limited to the exemplary embodiments. All features described and / or illustrated herein can be used alone or combined in different combinations in embodiments of the present invention. The features and advantages of various embodiments of the present invention will become apparent by reading the following detailed description with reference to the attached drawings which illustrate the following:

[0007] FIG. 1 schematically illustrates a method and system architecture for active rare disease mining, and an active rare disease mining module, according to an embodiment of the present invention;

[0008] FIG. 2 illustrates a first step of the active rare disease mining module according to an embodiment of the present invention;

[0009] FIG. 3 illustrates the whole process of the active rare disease mining module according to an embodiment of the present invention;

[0010] FIG. 4 illustrates a flow diagram of a method for active mining for rare diseases according to an embodiment of the present invention; and

[0011] FIG. 5 is a block diagram of an exemplary processing system, which can be configured to perform any and all operations disclosed herein.DETAILED DESCRIPTION

[0012] People may suffer from diseases that are rare or hard to diagnose or to differentiate due to similar symptoms. Examples include MG, third nerve palsy, Homer syndrome, seizures and paralysis, among many others, including some for which there is only very limited information available. For patients with such rare diseases, diagnosis and subsequent treatment can often take many years. For doctors, it is impossible to know every rare disease and to stay up to date with the latest research. Given a patient’s disease profile, embodiments of the present invention are able to provide doctors with a list of likely diseases and corresponding reasoning / explanations, thereby aiding them in speeding up the diagnosis for rare diseases.Embodiments of the present invention in particular provide an Al tool that has enhanced computer functionality to be able to provide predictions for rare diseases in an explainable manner, which results in enhanced trust, security and reliability of the Al tool.

[0013] New research and disease cases grow up in the world every day, but it is impossible for doctors to know all the latest research and disease cases. None of the existing approaches provide a generic approach to fdl this knowledge gap. Embodiments of the present invention enable to leverage the wealth of information found in the clinical literature to iteratively provide doctors the most valuable medical interview question, as well as a list of likely diseases with corresponding reasoning / explanations, therefore aiding them in speeding up the diagnosis for rare diseases.

[0014] In a first aspect, the present invention provides a computer-implemented, machine learning method for active mining for rare diseases. A patient disease profile is generated based on received input. Data of the patient disease profile is transformed into a representation in a feature space for comparison to representations of diseases in the feature space. One or more candidate diseases are identified based on distances in the feature space between the representation of the data and the representations of the diseases. One or more discriminative disease features are determined for one or more candidate diseases. An interview question is generated based on the one or more discriminative disease features.

[0015] In a second aspect, the present invention provides the method according to the first aspect, wherein the one or more discriminative disease features are determined based on a data purity measurement.

[0016] In a third aspect, the present invention provides the method according to the first aspect or the second aspect, wherein the data purity measurement is made using a disease database containing features of the diseases to identify one or more of the features that are most discriminative among the one or more candidate diseases.

[0017] In a fourth aspect, the present invention provides the method according to any of the first to third aspects, wherein the data purity measurement is based on a Gini Index, Entropy, and / or Information Gain method.

[0018] In a fifth aspect, the present invention provides the method according to any of the first to fourth aspects, further comprising: (a) receiving updated input in response to providing the interview question; (b) updating the patient disease profile based on the updated input; (c) transforming updated data of the updated patient disease profile into a new representation in the feature space; (d) identifying one or more updated candidate diseases based on distances in the feature space between the new representation and the representations of the diseases; (e) determining, for the one or more updated candidate diseases, one or more updated discriminativedisease features; and (f) generating an updated interview question based on the one or more updated discriminative disease features, wherein steps (a)-(f) are iteratively repeated until a stopping criteria is met.

[0019] In a sixth aspect, the present invention provides the method according to any of the first to fifth aspects, further comprising outputting, to a user, a disease diagnosis associated with a final one or more of the discriminative disease features determined in step (e) before the stopping criteria is met, along with an explanation of the disease diagnosis.

[0020] In a seventh aspect, the present invention provides the method according to any of the first to sixth aspects, wherein the disease diagnosis includes a ranked list of diseases that are based on a final set of candidates determined in step (d) before the stopping criteria is met, each of the diseases including an explanation based on the discriminative disease features.

[0021] In an eighth aspect, the present invention provides the method according to any of the first to seventh aspects, further comprising: building a disease feature extractor for extracting features of disease information from an internal disease database that was generated using electronic health records for patients and an external medical resources database that documents diseases; and storing the extracted features of the disease information in a disease database with a predefined schema.

[0022] In a ninth aspect, the present invention provides the method according to any of the first to eighth aspects, wherein the predefined schema includes a relational schema for a database or a non-structured query language (NoSQL) schema, and wherein the disease feature extractor includes a pattern-based extract, transform, and load (ETL) tool, or a language algorithm.

[0023] In a tenth aspect, the present invention provides the method according to any of the first to ninth aspects, wherein a generative language model is used to receive the input and generate the interview question.

[0024] In an eleventh aspect, the present invention provides the method according to any of the first to tenth aspects, wherein a neural network-based encoder is used to transform the data into the representation in the feature space and disease information of the diseases into the representations in the feature space, and wherein the neural network-based encoder transforms from heterogeneous modalities into a homogenous representation.

[0025] In a twelfth aspect, the present invention provides the method according to any of the first to eleventh aspects, wherein the interview question includes a request for results from a medical test or procedure, biographical information, or family disease history.

[0026] In a thirteenth aspect, the present invention provides the method according to any of the first to twelfth aspects, wherein the input includes user profile data and user health records.

[0027] In a fourteenth aspect, the present invention provides a computer system for active mining for rare diseases comprising one or more processors, which, alone or in combination, are configured to perform a machine learning method for mining for rare diseases according to any of the first to thirteenth aspects.

[0028] In a fifteenth aspect, the present invention provides a tangible, non-transitory computer-readable medium for active mining for rare diseases which, upon being executed by one or more hardware processors, provide for execution of a machine learning method according to any of the first to thirteenth aspects.

[0029] FIG. 1 illustrates a comprehensive system architecture according to an embodiment of the present invention (system components are numbered, such as (1) and (2), inputs and (intermediate) outputs are tagged, such as (a) and (b)). The system 100 can provide the user 102 with a list of potential rare diseases and corresponding explanations. From the user perspective, the system 100 works as a question-answering agent such as a Chatbot so the user 102 will feel like he / she is talking to an experienced doctor, can easily understand the diagnosis process and make an informed decision how to deal with a certain diagnosis result (e.g., accept, not accept, or further manual verification). From the system perspective, the tool is backed by an inventive active rare disease mining module. In the following, Part 1 explains the system overview (see FIG. 1) in a general way and Part 2 explains how the inventive active rare disease mining module works in a detailed way with an example (see FIGS. 2 and 3).PART 1 - COMPREHENSIVE SYSTEM ARCHITECTURE1. Construct disease database at the backend (a)(b)(c)(l)

[0030] To enable the automatic diagnosis in production, a Disease Feature Extractor (1) 104 such as a transformer-based information extraction model, will first extract disease information from both an Internal Disease DB (a) 106 and External Medical Resources (b) 108 to construct a unified Disease DB (c) 110. The Disease DB (c) 110 will store all disease information (including common diseases) in a consistent way, such as in a relational schema or Non-SQL schema to enable efficient information retrieval.

[0031] For instance, if the system 100 is deployed by a hospital, it can be assumed there will be an internal disease database (a) 106 from the hospital itself, which stores all internal disease cases with patients, diseases and treatment data. Meanwhile, it can be assumed to have access to one or more public medical resources (b) 108 such as OncoAlert («oncoalert360.com»), or Orphanet (<<www.orphanet.net>>), where the worldwide rare disease patients and researchers share their cases, treatments, and latest research. As a result, the Disease Database (c) 110 will contain both common and rare diseases.2. Trigger the system in production (d) (2. 1)

[0032] A user (a) 102 who starts the communication with the Medical Interview Module (2.1) 112 by, for instance, speaking / typing a starting question, triggers the diagnosis process. Optionally, the user 102 can input related electronic heath record (EHR) 114 and personal profde data (user profile data) 116 to provide more initial data to speed up the diagnosis process. If the user 102 does not directly provide EHR 114 and profile data 116, the module 112 will collect the data via a question-answering approach.

[0033] The Medical Interview Module (2. 1) 112 works as a Chatbot, such as a large language model (LLM). It is responsible for understanding questions from frontend users and constructing follow-up questions or conclusions from the diagnosis output of the backend system back to the user 102. This interactive process can be implemented in an iterative manner so that the process will feel like a real medical interview to the user 102.3. Diagnose rare disease in an active manner (2. l)(2.2)(2.3)(c)(e)(f)

[0034] Once the Medical Interview Module (2.1) 112 gets the needed patient data including both the user profile 116 and disease data 114 (such as gender, age, symptom, historical health record), it collects all of them as a Patient Disease Profile (e) 118 and sends it to the Disease Projector (2.2) 120.

[0035] The Disease Projector (2.2) 120 is a component that aims at transforming heterogeneous disease data into a homogeneous format so that the data can be further analyzed by computational algorithms. For instance, each disease case in Disease Database (c) 110 might have different features, many of them record age, gender, body temperature, blood pressure, etc., and some of them record height, weight, genetic test, nerve stimulation, etc. The Disease Projector (2.2) 120 converts them into a homogeneous format such as in a high dimensional vector. For example, the vector can be simply a row of a large tabular data where each column represents a disease feature or can be computed by a neural network-based feature encoder where each element or each block of elements represents a disease feature.

[0036] When it comes to a new patient case, the Disease Projector (2.2) 120 will convert the patient disease profile data 118 in the same way so that the new case will be represented into a Distribution Space (f) 122 with all diseases recorded in known Disease Database (c) 110 so that it is possible to further measure the distance between the new case and all recorded diseases.

[0037] Given the new patient case and all recorded diseases, the Disease Discriminator Detector (2.3) 124 will find the best or top-k best disease feature(s), which can most efficiently identify the new patient case out of all recorded diseases. The details of how the Disease Discriminator Detector (2.3) 124 works is described in Part 2.4. End the system (2.3) (2. 1) (d)

[0038] Based on patient case, the Disease Discriminator Detector (2.3) 124 will detect the best disease differentiable feature(s) (e.g., the response to anticholinesterase medicine is a strong discriminative feature for myasthenia gravis disease). The Medical Interview Module (2.1) 112 then constructs follow-up questions based on detected features to ask the user 102 to provide further answers. Once the user provides further data, the Disease Discriminator Detector (2.3) 124 will update the features accordingly.

[0039] Until the result of Disease Discriminator Detector (2.3) 124 fulfils certain criteria such as the distance between patient case and recorded disease is lower than X, the Medical Interview Module (2.1) 112 returns the list of likely diseases and corresponding explanation back to user (d) 102. In embodiments, X can be seen as a hyperparameter that the engineer associated with the system or deploys the system according to an embodiment of the present invention can fine tune. X could be a single value for all diseases or it could vary for different diseases (e.g., lower for diseases that cluster more tightly in the cases feature space).PART 2 - ACTIVE RARE DISEASE MINING MODULE

[0040] This section explains how the Active Rare Disease Mining Module (2) 126 works in details with an example. The whole module is implemented in an iterative way, as shown in FIG. 2. It is first described how the first step works (how the module is initialized), and second, whole process is described with reference to FIG. 3.1. Initialization of the active rare disease mining module (FIG. 2) 126

[0041] Assume a doctor user (d) 102 wants to use the system according to an embodiment of the present invention as an Al tool to diagnose if a patient has normal eye disease or the rare disease of myasthenia gravis (MG) because the patient has symptoms of muscle weakness and droopy eyelids. First, the user 102 will trigger the system by talking to the Medical Interview Module (2.1) 202 and providing related data of the patient case.

[0042] The Medical Interview Module (2.1) 202 collects all received data into a Patient Disease Profile (e) 204 and sends it to the Disease Projector (2.2) 206. The Patient Disease Profile (e) 204 contains both personal profile data such as gender and age, and disease-related data such as disease history and current symptoms.

[0043] The Disease Projector (2.2) 206 transforms the given data into a predefined homogeneous format and projects it in the Cases Feature Space (f) 208 to enable to measure the distance between the new case with all recorded diseases. In this example of the Cases Feature Space (f) 208, each dot is an individual case and different patients having the same disease are considered separately. FIG. 2 shows only a 2-dimensional space to simplify the visualization, the space can be highly dimensional with real data such as the tabular example (g) 210. Also, it can be extremely highly dimensional if it’s in vector representation.

[0044] From the distribution space (f) 208, it can be seen that the new patient case is in the middle among flu cases, eye disease cases and the MG case so it is possible to be any of them. The doctor needs to do further testing to reach the final diagnosis. The challenge is to find what test can help confirm the diagnosis in the fastest way. The Disease Discriminator Detector (2.3) 212 according to an embodiment of the present invention addresses this challenge by searching the most discriminative feature in a two-dimensional way. As shown in tabular Disease DB (g) 210, each row is a disease case, and each column is a feature of the disease case. Each disease case consists of user / patient features such as gender and age, and disease features such as symptoms, treatment response and side effects. First, the Disease Discriminator Detector (2.3) 212 searches all similar disease cases by measuring the distance between all recorded cases and the new case. For instance, the distance can be measured by a cosine similarity metric. In this example, each disease type has only one record in the database so that three diseases are easily seen (i.e., flu, Homer syndrome and MG) as similar to the new patient case. To identify whether the patient has rare MG disease or just common flu or droopy eyelid problem, the system continues the second detection step. Second, the Disease Discriminator Detector (2.3) 212 searches further for the most discriminative feature or a set of features from detected similar cases based on data purity. It is possible to discriminate between the diseases identified by the distance-based similarity metric (the most likely diagnoses for that case). As an example of searching for the most discriminative feature or set of features based on data purity, the system can consider decision rules such as “a case is disease A if feature 1 is greater than X or disease B otherwise.” In embodiments, for each these rules, using the history of diagnosed cases for diseases A and B, the system can evaluate how well the rule discriminates correctly between the two diseases - similar to a decision tree algorithm. A best rule will be the one giving rise to two groups that have highest purity (separates the best cases from disease A and B). The feature or features associated with this rule will be the most discriminative. Purity here means a measure of the extent to which clusters contain ideally a single class. It can be computed by multiple methods such as Gini Index, Entropy, and Information Gain. The goal of this step is to identify the feature, or set of features, which maximizes the purity level of segmented clusters. For example, given scenario for determining between a male or female gender using a data set included in a table where each column is a feature and each row is a person. The analysis starts with one feature from the table, such as height. The system could generate clusters by using a rule that if height > 150cm, the person goes to cluster A otherwise the person goes to cluster B. Once two clusters are formed the system can determine if the clusters are pure. In this scenario, if either cluster has both males and females than it is not pure. The system can determine that the height feature is not discriminative enough to distinguish males and females. This process isrepeated using different features until the system identifies which features generate the clusters with the highest purity. For the context of disease this usually includes identifying a set of features. For instance, if there are two diseases, Homer syndrome or MG that are similar to the patient case, the possible differentiable feature is the response to cold pack.

[0045] As a result, the Disease Discriminator Detector (2.3) 212 will return detected Discriminative Disease Feature(s) (h) 214 back to the Medical Interview Module (2.1) 202. The Medical Interview Module (2.1) 202 then constructs the follow-up interview question 216 from the given features and asks the user to provide further data. For instance, the follow-up question 261 can be “what’s the response of the patient after using cold pack on the eyelid for 5-10 minutes?” or “is the muscle weakness dramatically improved after taking Reminyl medicine (an anticholinesterase inhibitor medicine for MG)?”

[0046] At this stage, the first initialization step of the Active Rare Disease Mining Module (2) 126 ends.2. Whole iterative process of the active rare disease mining module 126 (FIG. 3)

[0047] FIG. 3 visualizes the whole iterative process of the active rare disease mining module 126 with the same MG example. The first initialization step shown in FIG. 2 can be mapped to FIG. 3 from step 1 to step 9.

[0048] After asking the follow-up question in the first initialization step, the user 300 will return to the system further data, such as required medical testing results (step 10). Then, the Medical Interview Module (2.1) 302 will send again the updated patient profile data to the Disease Projector (2.2) 304 to start the second step of the mining loop (step 11). In this step, with additional data, it can be seen from the Cases Feature Space 306 (State-2) that the patient case is now detected to be similar to only Homer Syndrome and MG rather than the flu - represented as the triangle shape of the new patient case being closer to Homer Syndrome set #1 and MG disease #3 in State-2 of the cases feature space 306(steps 12-13). Next, the Disease Discriminator Detector (2.3) 308 searches again the new discriminative feature(s) and sends them to the Medical Interview Module (2. 1) 302 to ask the user 300 (steps 14-16).

[0049] By receiving further answers from the user 300 (step 17), the Active Rare Disease Mining Module (2) 126 starts the next mining step (steps 18-22). From the Cases Feature Space 306 (State-3), it can be seen that the system now identifies that the patient case is highly probable to be MG disease rather than Homer Syndrome - represented as the triangle of new patient case being closer to MG disease #3 in the case feature space 306. This determination can be made based on the distance between the patient case and the MG disease #3 being below a minimum threshold distance within feature space 306, for example. In embodiments a distance within a feature space criterion can also be used to mle out diseases (e.g., the top candidatediseases will be the ones for which the centroid of the associated clusters of past cases in the database will be the closest to the case at hand). Therefore, the Disease Discriminator Detector (2.3) 308 will output this time the final diagnosis result such as in the format of a list containing the most probable diagnoses. Meanwhile, it will output also the intermediate mining result as a diagnosis explanation (step 21-22). The Disease Discriminator Detector (2.3) 308 according to an embodiment of the present invention works based on data purity that can be completely computed by simple mathematical operations, so the intermediate diagnosis results are completely transparent, thereby increasing the trust, security and reliability of the Al system. The Medical Interview Module (2.1) 302 will then transform the diagnosis result and the explanation into human-readable style and send it back to user 300 (step 23). For instance, if the Disease Discriminator Detector (2.3) 308 generates “cold pack” as output (e.g., “cold pack” is determined to be the most discriminative feature), then Classification and Regression Trees may generate a question such as “Hello user, will your symptom be relieved by using a cold pack on your eyes?" In embodiments, the Medical Interview Module (2. 1) 302 may generate this output by using a large language model (LLM such as chatGPT). A system engineer can instruct an existing LLM to formulate a human-readable question with certain specific keywords (i.e., here the keywords are the discriminative features). The instruction may be realized as a so called “prompt.” As another example, the explanation could be “your disease is diagnosed as X, because you have A, B, C symptoms, according to existing patient cases, xxx% of cases with ABC symptoms are Z and X, xx% of X cases have recorded D symptoms, 0% of Z case recoded D symptoms, so your case is diagnosed as X.” In the above explanation, the values of X, Y, A, B, C, D, Z are generated by the Disease Discriminative Detector (2.3) 308, and the output in sentence form is formulated by the Medical Interview Module (2.1) 302 (supported by an LLM).

[0050] In this example, the mining process ends when the system finds the most likely disease which matches the patient case. It is also possible to stop the mining process with different criteria or thresholds, such as the number of detected likely diseases or to which extent the likely diseases match to the patient case.

[0051] Embodiments can be practically applied to use cases to effect further improvements in a number of technical fields including, but not limited to, medical (e.g., digital medicine, personalized healthcare, Al-assisted drug or vaccine development, diagnosis or treatment, etc.) and as an Al tool to support optimized decision-making by doctors and / or patients or other users.

[0052] In an exemplary embodiment, the present invention can be applied for an interactive diagnosis suggestion for a patient. Here, a use case is a patient has health issues with symptoms and looks for a diagnosis due to limited access to medical facilities. The user interacts with thesystem according to an embodiment of the present invention to do a digital medical interview for diagnosis. As a data source, the patient triggers the communication with the system with a starting question, for instance, “Hello system, my muscles recently feel sore, and my eyelid has been droopy for 2 days, could you help me diagnose what disease I have?” Then the user provides the data which the system needs, which could be personal biographical and demographic information, disease history or certain medical testing results. Application of the method according to an embodiment of the present invention represents the patient case in a cases feature space populated by all known medical records and searches for the similar cases and detects the best discriminative features. Based on detected features, the system constructs follow-up interview questions back to the user to ask for further data or results of medical testing until it finds a diagnosis with a sufficient level of confidence. If the disease database used by the system features the costs of medical tests (e.g., time, resources and monetary costs), the system can also decide the best feature while taking into account the testing cost. The system can output a list of ranked diseases and a corresponding diagnosis explanation for each of the diseases. The diseases may be ranked in the list based on distances in the feature space.Diagnosis Recommendation for Doctor to Patient

[0053] In an exemplary embodiment, the present invention can be applied in the healthcare domain for an interactive diagnosis recommendation for a doctor to a patient. Here, a use case is a patient has health issues with symptoms and goes to a doctor for a diagnosis. Due to the doctor, for example, being newly graduated and lacking practical experience, the doctor cannot simply diagnose the disease according to existing knowledge. Therefore, the doctor interacts with the system according to an embodiment of the present invention to receive advice about the best interviewing question to aid or speed up the diagnosis. As a data source, the doctor triggers the communication with the system with a starting question, for instance, “Hello system, my patient currently presents the following symptoms: sore muscle and ptosis, but according to the tests, it’s not a common eye disease, could you help me diagnose what disease it is?” Then, the doctor continuously provides data that the system requests, such as patient health records or existing medical test results. Application of the method according to an embodiment of the present invention represents the patient case in a cases feature space populated by all known medical records and searches for the similar cases and detects the best discriminative features. Based on the detected features, the system constructs follow-up interview questions, presents them to the doctor to obtain further data or results of new medical tests until it finds a diagnosis with a sufficient level of confidence. If the disease database used by the system also possesses such features, the system can also take into account other factors such as the costs of medical tests (e.g., time, resources and monetary costs), or the availability of the corresponding devicesor medical specialists, to provide the doctor with the optimal follow-up question. The system can output a list of ranked diseases and a corresponding diagnosis explanation for each of the diseases.

[0054] In an embodiment, the present invention provides an Al tool for rare disease diagnoses or disease detection in general that implements a method comprising the steps of: Setup a new system (cold start):1) Prepare an Internal Disease Database (a) integrating all the past EHRs from the hospital or medical care unit. Each record contains both patient and disease related information.2) Gain access to one or more External Medical Resources (b) such as OncoAlert, or Orphanet, where rare diseases are documented.3) Build a Disease Feature Extractor (1) which extracts needed features for each disease from different resources and stores it into a Disease Database (c) with predefined schema (e.g. relational DB orNon-SQL DB). For instance, the Disease Feature Extractor (1) can be a pattern- based extract, transform and load (ETL) tool or a language algorithm such as a Transformer model. In embodiments the Disease Feature Extractor (1) can be a different algorithm depending on the type of data source. For example, if a data source is text-based data, the Disease Feature Extractor (1) may be transformer-based neural network model such as Bidirectional Encoder Representations from Transformers (BERT) model which converts text to word embeddings. For tabular data it depends on a value type as Boolean or categorical data can be analyzed using hot encoding. In the current described system all data sources will be converted into vector representations.4) Build a Medical Interview Module (2.1) which can interact with users in a questionanswering manner. For instance, it can be a generative language model such as an LLM. A pretrained model such as ChatGPT may be used to interact with users.5) Build a Disease Projector (2.2) which transforms both new patient cases and all recorded cases from heterogeneous modalities into a homogeneous representation which allows the application of computational analysis. For instance, the Disease Projector (2.2) can be a neural network-based encoder which can compute vector representations from input data.6) Build a Disease Discriminator Detector (2.3) which mines the best differentiable disease feature to construct follow-up medical interview questions and generates the final diagnosis with evidence. For instance, the Disease Discriminator Detector (2.3) can be implemented as a trainable algorithm, which decides the feature based on data purity measurements.When the new cases come:1) The system reads the input from the user and aggregates its data.2) The system generates the diagnosis of potential diseases via:a. The Medical Interview Module (2. 1) sends collected user data as patient disease profiles to the Disease Projector (2.2). b. The Disease Projector (2.2) transforms the user data into the required representation and projects it to the Cases Feature Space (f) where the new user data can be compared with all recorded diseases. c. Based on the transformed user data, the Disease Discriminator Detector (2.3) further detects the best discriminative disease feature(s) from all related recorded diseases and sends the detected feature(s) to the Medical Interview Module (2. 1). d. The Medical Interview Module constructs the follow-up interview question based on the detected feature.3) When the Disease Discriminator Detector (2.3) detects likely diseases which match to the new user case, the Medical Interview Module (2.1) returns the disease diagnoses with evidence as explanation back to the user (d).

[0055] Embodiments of the present invention provide for the following improvements and technical advantages over existing technology:1) The disease discriminator detector provides an inventive method to enable the system to determine and ask the most efficient questions to the patient, to confirm the rare disease diagnosis in the fastest and most efficient way by database mining technology. It firstly detects the similar diseases candidates to the new patient case. Second, it searches further the best discriminative disease feature(s) from all candidates based on purity measurements. Based on the detected feature, the system constructs the proper follow-up query back to the database.2) Providing a completely explainable method to provide use diagnosis of potential rare diseases. In particular: a. The disease discriminator detector searches disease features based on data purity, which is feasible for a human to deduce manually, by calculation, hence the system can explain why it formulated certain follow-up questions (e.g., ask to provide blood test results) and why it confirms certain diseases as potential diagnoses. b. The system is implemented in a question-answering manner so the reasoning process is transparent to the user and can naturally present evidence to explain how a given diagnosis can eventually be confirmed.

[0056] Existing Al systems are not able to provide interactive advice on guiding questions to speed up the diagnosis process in realistic scenarios. Most research on rare diseases focus on biomedical experimental techniques or a specific type of rare disease (see Martinez-Delgado et al., Yu et al., and Emily Alsentzer, Michelle M. Li, Shilpa N. Kobren, Undiagnosed Diseases Network, Isaac S. Kohane, and Marinka Zitnik, “Deep learning for diagnosing patients with raregenetic diseases,” medRxiv 2022.12.07.22283238 (2022), which is hereby incorporated by reference herein). In contrast, embodiments of the present invention provide to support diagnosis of rare disease by leveraging realistic clinical cases and external knowledge using completely different methodology and providing different outcomes.

[0057] Feichen Shen, Yiqing Zhao, Liwei Wang, Majid Rastegar Mojarad, Yanshan Wang, Sijia Liu, and Hongfang Liu, “Rare disease knowledge enrichment through a data-driven approach,” BMC Med Inform Decis Mak, 19:32 (2019), which is hereby incorporated by reference herein, describe a mining method to extract information from large disease data sources. However, this method focuses on the association mining between phenotype and disease to enrich the existing rare disease knowledge. This method could be used to assist in the creation of a disease database. Embodiments of the present invention provide for a number of improvements and new features which are completely absent from this method, such as the mining of the most discriminative feature between candidate diseases, and providing an interactive interface with the user.

[0058] Jinmeng Jia, Ruiyuan Wang, Zhongxin An, Yongli Guo, Xi Ni, and Tieliu Shi, “RDAD: A Machine Learning System to Support Phenotype-Based Rare Disease Diagnosis”, Front. Genet., Vol. 9 (2018), which is hereby incorporated by reference herein, describe an existing approach to support the diagnosis of rare disease by using machine learning technique, with a methodology and outcome that are completely different than in embodiments of the present invention. Given prepared rare disease data such as gene and phenotype information of the patient, the existing approach uses several predictive models for different data types to directly generate final diagnosis of rare diseases. In contrast, embodiments of the present invention start from only a few basic data inputs, and guides the user to do the most efficient action (e.g., choose which medical tests to take first) step by step and reach the final diagnosis, which is also more accurate and explainable, in a reduced time. Instead of generating final diagnosis at one step, the system according to an embodiment of the present invention provides the most efficient advice on each step, and can practically help doctor speed up the diagnostic process. Moreover, the iterative approach according to embodiments of the present invention is transparent and explainable, which can easily gain user trust in the Al system.

[0059] Feichen Shen, Sijia Liu, Yanshan Wan, Andrew Wen, Liwei Wang, and Hongfang Liu, “Utilization of Electronic Medical Records and Biomedical Literature to Support the Diagnosis of Rare Diseases Using Data Fusion and Collaborative Filtering Approaches,” JMIR Med Inform, 6(4):e 11301 , doi: 10.2196 / 11301 (2018), which is hereby incorporated by reference herein, describe an existing approach to support the diagnosis of rare disease by using EHRs and external disease resources such as the literature. However, the existing approachfocuses on the data fusion strategies to link the EHR with disease literature data and generate a final diagnosis based on the similarity between the case at hand and previously encountered and reported cases. In contrast, embodiments of the present invention provide diagnostic question advice step by step by mining the most discriminating disease feature to speed up diagnosis process in an explainable way.

[0060] Thus, in contrast to existing technology, embodiments of the present invention provide an Al tool with enhanced computer functionality to provide an efficient, interactive approach with the patient and / or the doctor, and to further offer support to choose the next steps towards reaching a final diagnosis in a more efficient, safe, reliable, transparent, trustworthy and accurate manner.

[0061] FIG. 4 illustrates a flow diagram 400 of a method for active mining for rare diseases according to an embodiment of the present invention. The flow diagram 400 includes at step 402 generating a patient disease profile based on received input (e.g., from a user). In embodiments, the input may include a user providing EHR information and personal profile data (user profile data). In accordance with at least one embodiment, the medical interview module may collect the data to generate the patient disease profile in a question-answering approach. The flow diagram 400 includes at step 404 transforming data of the patient disease profile into a representation in a feature space for comparison to representations of diseases in the feature space. In embodiments, distances in the feature space may be measured by a cosine similarity metric. The flow diagram 400 includes at step 406 determining one or more discriminative disease features of the diseases based on a distance in the feature space between the representation of the data and the representation of the diseases. Determining the one or more discriminative disease features may be performed by the disease discriminator detector. The flow diagram 400 includes at step 408 generating an interview question, e.g., as a follow-up based on the one or more discriminative disease features. In embodiments, the responses to the interview questions are used to update the patient disease profile and subsequent representations within the feature space using the updated patient disease profile for comparing to representations of diseases within the same feature space.

[0062] Referring to FIG. 5, a processing system 500 can include one or more processors 502, memory 504, one or more input / output devices 506, one or more sensors 508, one or more user interfaces 510, and one or more actuators 512. Processing system 500 can be representative of each computing system disclosed herein.

[0063] Processors 502 can include one or more distinct processors, each having one or more cores. Each of the distinct processors can have the same or different structure. Processors 502 can include one or more central processing units (CPUs), one or more graphics processing units(GPUs), circuitry (e.g., application specific integrated circuits (ASICs)), digital signal processors (DSPs), and the like. Processors 502 can be mounted to a common substrate or to multiple different substrates.

[0064] Processors 502 are configured to perform a certain function, method, or operation (e.g., are configured to provide for performance of a function, method, or operation) at least when one of the one or more of the distinct processors is capable of performing operations embodying the function, method, or operation. Processors 502 can perform operations embodying the function, method, or operation by, for example, executing code (e.g., interpreting scripts) stored on memory 504 and / or trafficking data through one or more ASICs. Processors 502, and thus processing system 500, can be configured to perform, automatically, any and all functions, methods, and operations disclosed herein. Therefore, processing system 500 can be configured to implement any of (e.g., all of) the protocols, devices, mechanisms, systems, and methods described herein.

[0065] For example, when the present disclosure states that a method or device performs task “X” (or that task “X” is performed), such a statement should be understood to disclose that processing system 500 can be configured to perform task “X”. Processing system 500 is configured to perform a function, method, or operation at least when processors 502 are configured to do the same.

[0066] Memory 504 can include volatile memory, non-volatile memory, and any other medium capable of storing data. Each of the volatile memory, non-volatile memory, and any other type of memory can include multiple different memory devices, located at multiple distinct locations and each having a different structure. Memory 504 can include remotely hosted (e.g., cloud) storage.

[0067] Examples of memory 504 include a non-transitory computer-readable media such as RAM, ROM, flash memory, EEPROM, any kind of optical storage disk such as a DVD, a Blu- Ray® disc, magnetic storage, holographic storage, a HDD, a SSD, any medium that can be used to store program code in the form of instructions or data structures, and the like. Any and all of the methods, functions, and operations described herein can be fully embodied in the form of tangible and / or non-transitory machine-readable code (e.g., interpretable scripts) saved in memory 504.

[0068] Input-output devices 506 can include any component for trafficking data such as ports, antennas (i.e., transceivers), printed conductive paths, and the like. Input-output devices 506 can enable wired communication via USB®, DisplayPort®, HDMI®, Ethernet, and the like. Input-output devices 506 can enable electronic, optical, magnetic, and holographic, communication with suitable memory 506. Input-output devices 506 can enable wirelesscommunication via WiFi®, Bluetooth®, cellular (e.g., LTE®, CDMA®, GSM®, WiMax®, NFC®), GPS, and the like. Input-output devices 506 can include wired and / or wireless communication pathways.

[0069] Sensors 508 can capture physical measurements of environment and report the same to processors 502. User interface 510 can include displays, physical buttons, speakers, microphones, keyboards, and the like. Actuators 512 can enable processors 502 to control mechanical forces.

[0070] Processing system 500 can be distributed. For example, some components of processing system 500 can reside in a remote hosted network service (e.g., a cloud computing environment) while other components of processing system 500 can reside in a local computing system. Processing system 500 can have a modular design where certain modules include a plurality of the feature s / functions shown in FIG. 5. For example, I / O modules can include volatile memory and one or more processors. As another example, individual processor modules can include read-only-memory and / or local caches.

[0071] While subject matter of the present disclosure has been illustrated and described in detail in the drawings and foregoing description, such illustration and description are to be considered illustrative or exemplary and not restrictive. Any statement made herein characterizing the invention is also to be considered illustrative or exemplary and not restrictive as the invention is defined by the claims. It will be understood that changes and modifications may be made, by those of ordinary skill in the art, within the scope of the following claims, which may include any combination of features from different embodiments described above.

[0072] The terms used in the claims should be construed to have the broadest reasonable interpretation consistent with the foregoing description. For example, the use of the article “a” or “the” in introducing an element should not be interpreted as being exclusive of a plurality of elements. Likewise, the recitation of “or” should be interpreted as being inclusive, such that the recitation of “A or B” is not exclusive of “A and B,” unless it is clear from the context or the foregoing description that only one of A and B is intended. Further, the recitation of “at least one of A, B and C” should be interpreted as one or more of a group of elements consisting of A, B and C, and should not be interpreted as requiring at least one of each of the listed elements A, B and C, regardless of whether A, B and C are related as categories or otherwise. Moreover, the recitation of “A, B and / or C” or “at least one of A, B or C” should be interpreted as including any singular entity from the listed elements, e.g., A, any subset from the listed elements, e.g., A and B, or the entire list of elements A, B and C.

Claims

CLAIMSWhat is claimed is:

1. A computer-implemented method for active mining for rare diseases, the computer- implemented method comprising: generating a patient disease profde based on received input ; transforming data of the patient disease profde into a representation in a feature space for comparison to representations of diseases in the feature space; identifying one or more candidate diseases based on distances in the feature space between the representation of the data and the representations of the diseases; determining, for the one or more candidate diseases, one or more discriminative disease features; and generating an interview question based on the one or more discriminative disease features.

2. The computer-implemented method according to claim 1, wherein the one or more discriminative disease features are determined based on a data purity measurement.

3. The computer-implemented method according to claim 1 or claim 2, wherein the data purity measurement is made using a disease database containing features of the diseases to identify one or more of the features that are most discriminative among the one or more candidate diseases.

4. The computer-implemented method according to claim 2, wherein the data purity measurement is based on a Gini Index, Entropy, and / or Information Gain method.

5. The computer-implemented method according to any of the preceding claims, further comprising:(a) receiving updated input in response to providing the interview question;(b) updating the patient disease profde based on the updated input;(c) transforming updated data of the updated patient disease profde into a new representation in the feature space;(d) identifying one or more updated candidate diseases based on distances in the feature space between the new representation and the representations of the diseases;(e) determining, for the one or more updated candidate diseases, one or more updated discriminative disease features; and(f) generating an updated interview question based on the one or more updated discriminative disease features, wherein steps (a)-(f) are iteratively repeated until a stopping criteria is met.

6. The computer-implemented method according to claim 5, further comprising outputting, to a user, a disease diagnosis associated with a final one or more of the discriminative disease features determined in step (e) before the stopping criteria is met, along with an explanation of the disease diagnosis.

7. The computer-implemented method according to claim 6, wherein the disease diagnosis includes a ranked list of diseases that are based on a final set of candidates determined in step (d) before the stopping criteria is met, each of the diseases including an explanation based on the discriminative disease features.

8. The computer-implemented method according to any of the preceding claims, further comprising: building a disease feature extractor for extracting features of disease information from an internal disease database that was generated using electronic health records for patients and an external medical resources database that documents diseases; and storing the extracted features of the disease information in a disease database with a predefined schema.

9. The computer-implemented method according to claim 8, wherein the predefined schema includes a relational schema for a database or a non-structured query language (NoSQL) schema, and wherein the disease feature extractor includes a pattern-based extract, transform, and load (ETL) tool, or a language algorithm.

10. The computer-implemented method according to any of the preceding claims, wherein a generative language model is used to receive the input and generate the interview question.

11. The computer-implemented method according to any of the preceding claims, wherein a neural network-based encoder is used to transform the data into the representation in the feature space and disease information of the diseases into the representations in the feature space, and wherein the neural network-based encoder transforms from heterogeneous modalities into a homogenous representation.

12. The computer-implemented method according to any of the preceding claims, wherein the interview question includes a request for results from a medical test or procedure, biographical information, or family disease history.

13. The computer-implemented method according to any of the preceding claims, wherein the input includes user profile data and user health records.

14. A computer system for active mining for rare diseases, the computer system comprising one or more hardware processors which, alone or in combination, are configured to provide for execution of the following steps: generating a patient disease profile based on received input;transforming data of the patient disease profde into a representation in a feature space for comparison to representations of diseases in the feature space; identifying one or more candidate diseases based on distances in the feature space between the representation of the data and the representations of the diseases; determining, for the one or more candidate diseases, one or more discriminative disease features; and generating an interview question based on the one or more discriminative disease features.

15. A tangible, non-transitory computer-readable medium having instructions thereon which, upon being executed by one or more processors, provide for active mining for rare diseases by execution of the following steps: generating a patient disease profde based on received input; transforming data of the patient disease profde into a representation in a feature space for comparison to representations of diseases in the feature space; identifying one or more candidate diseases based on distances in the feature space between the representation of the data and the representations of the diseases; determining, for the one or more candidate diseases, one or more discriminative disease features; and generating an interview question based on the one or more discriminative disease features.

Images