Creating and utilizing a conversational chatbots of a specific medical specialty or disease

Abstract

An AI-based chatbot diagnostic system configured to: receive data from at least one data source related to at least one medical field or disease; store the received data in a medical data repository, wherein the data is categorized according to at least one criterion; access a medical index engine that uses the categorized data and a medical reference database to create and update an index for individual medical fields; utilize interactive data for training a diagnostic model in conjunction with a medical index; select the trained diagnostic model most suitable for a given diagnostic scenario from a model storage database; and manage and interact with the selected diagnostic model using an AI model manager.

Description

CREATING AND UTILIZING A CONVERSATIONAL CHATBOTS OF A SPECIFIC MEDICAL SPECIALTY OR DISEASE

[0001] The present disclosure pertains to the field of artificial intelligence (AI), more particularly to an AI-based chatbot diagnostic system. The system and method specifically address the provision of diagnostic services, treatment planning, and the delivery of medical information.

[0002] Modern healthcare systems globally struggle with challenges such as staffing shortages, limited resources, and disparities in access to specialized care. There exists an urgent need to optimize access to healthcare services to meet these pressing issues. This patent application introduces a multimodal, artificial intelligence-based conversational assistant, aimed at enhancing the diagnostic process and providing patients with quicker access to appropriate medical care.

[0003] The AI-based conversational assistant employs technologies like natural language processing to comprehend presented medical cases and aid in disease diagnosis. By leveraging specially designed tools underpinned by artificial intelligence algorithms, patients can receive help remotely, negating the necessity of in-person doctor visits. Consequently, patients gain an alternative, bypassing lengthy wait times for specialist consultations or the high costs tied to obtaining immediate assistance.

[0004] This patent application outlines a system that incorporates the usage of large language models (LLMs) in medicine, specifically for diagnosis, treatment planning, and medical information delivery.

[0005] Large language models (LLMs) represent sophisticated artificial intelligence algorithms capable of generating, understanding, and processing text similarly to human cognitive capabilities. These models, founded on transformer architectures as described in the significant publication "Attention is All You Need", employ the "attention" mechanism. This mechanism allows them to concentrate on the most crucial information in the text and grasp the context wherein words are used. They find use in various applications, including chatbots, machine translation, automatic text summarization, and sentiment analysis.

[0006] Publication: Vaswani, A. et al.(2017). Attention is all you need. arXiv preprint arXiv:1706.03762.

[0007] LLMs are typically trained on vast text datasets encompassing information from diverse knowledge domains, including sciences, humanities, technical subjects, and others. The learning process involves adjusting its internal "rules", enabling it to comprehend and predict the sequence of generated words based on previously encountered instances. Practically, this implies that the model learns to recognize and generate texts that resemble those it has already processed.

[0008] Models learning from extensive numbers of "rules" exhibit capabilities in understanding the context of statements and discerning the emotional state of individuals participating in the dialogue, thereby demonstrating some competencies concerning the theory of mind. Some publications suggest that large models may already display certain characteristics of artificial general intelligence (AGI), possessing cognitive abilities comparable to human cognition.

[0009] Publication: Bubeck, S. et al. (2023). Sparks of Artificial General Intelligence: Early experiments with GPT-4. ArXiv preprint arXiv:2303.12712.

[0010] The aforementioned "rules" may also be referred to as "parameters." Therefore, the size of an LLM, that is, the number of parameters of the model, influences its capability to process and generate texts. Models with a larger number of parameters usually excel at tasks related to natural language processing; however, they demand substantially more computational power both during training and task execution. Due to these demands, a large number of parameters can be a limitation for the broad application of LLMs, especially in environments where access to computational resources is constrained - for instance, by financial and / or hardware limitations.

[0011] Recent breakthroughs in the field of large language models (LLMs) have ushered in new prospects in a medical context, marking a significant advancement in the application of artificial intelligence in healthcare. LLMs demonstrate capabilities in diagnosing diseases, planning therapies, and delivering information about diseases and medications. Most recent research outcomes are validating the escalating effectiveness of properly trained LLMs in answering questions posed in professional medical examinations.

[0012] Publication: Singhal, K. et al. (2022). Large Language Models Encode Clinical Knowledge. ArXiv preprint arXiv:2212.13138.

[0013] Despite promising achievements, the implementation of LLMs in medical practice encounters certain challenges and limitations. Specifically, these models still produce incorrect responses, which in the context of medical care could lead to hazardous situations and errors in diagnosis or treatment. A characteristic trait of LLMs is their capability to "hallucinate," i.e., generate persuasive yet entirely inaccurate medical advice, which poses a particular risk in the realm of medicine.

[0014] Another restriction associated with the utilization of LLMs in medical practice is the obligation to provide significant computational power to these models, both throughout the learning process and during their operation. As previously noted, the larger the model, the more resources it necessitates. Large language models, especially those with AGI capabilities, in particular require vast computational resources, which could constitute a barrier to the broad deployment of these technologies. Their extensive application in medicine necessitates further innovation to streamline processes and minimize computational power demands.

[0015] With the aim of reducing computational power requirements, the proposed invention suggests creating, deploying, and continually updating specialized models for individual medical fields or even specific ailments. These models will only be automatically activated when the AI-based conversational assistant (hereinafter referred to as the AI Assistant) recognizes a need for their application, based on the analysis of context and conversation content. Therefore, instead of deploying a single large, resource-demanding model, the system will utilize relatively smaller specialized models, which are considerably more efficient in terms of computational resource utilization.

[0016] In terms of the application of specialized models in medicine, it is worth referencing a pioneering study in the realm of finance. This study successfully devised a relatively small model specializing in the field of economics and finance. The model, developed based on data from publications accessible in the Bloomberg database, demonstrated superior effectiveness in executing certain tasks compared to other, larger and more resource-demanding models.

[0017] Publication: Wu, S. et al. (2023). BloombergGPT: A Large Language Model for Finance. ArXiv preprint arXiv:2303.17564.

[0018] Drawing from these outcomes, it can be speculated that the creation of smaller, specialized models for individual medical fields could yield similar results in these areas while substantially reducing computational demands. The research conducted by Lehman et al. (2023) supports these conjectures, demonstrating that relatively small, specialized clinical models can significantly outperform larger, generic language models, even when working with a limited amount of annotated data. Moreover, these smaller models proved to be more efficient in terms of parameters, indicating that they can match or even outperform their larger equivalents trained on general texts.

[0019] Publication: Lehman, E. et al. (2023). Do We Still Need Clinical Language Models? ArXiv preprint arXiv:2302.08091.

[0020] Furthermore, in this patent application, we discuss in detail the use of LLMs in the medical context and how our invention can address key issues associated with model errors in medical diagnosis, the accessibility of LLMs, and the automatic application of specialized models in specific fields of medicine.

[0021] Among its various features, the invention incorporates medical supervision over responses generated by large language models (LLMs). This implies that every response (e.g., diagnosis, recommendation) produced by the model can be scrutinized by a qualified medical professional prior to being communicated to the patient or other user. This approach allows for the elimination of erroneous responses, thereby enhancing safety and reliability of the information disseminated by LLMs.

[0022] The proposed solutions integrate both medical oversight of responses and the utilization of specialized models, leading to a notable reduction in system-generated errors. As a result, LLMs become not only more precise but also more accessible to a broader user base, even in locations where computational resources are scarce.

[0023] An additional feature of the invention is the continuous learning capability of models based on conversation histories and user interactions. This enables the models to improve their diagnosis of various diseases and adjust to societal norms, values, and expectations related to medical ethics across different cultures and communities.

[0024] This adaptive aspect allows for real-time consideration of the impact of cultural context on medical practices, potentially facilitating communication between LLMs, patients, and other users. Tailoring models to local norms and values enhances their utility and efficacy, while reducing the risk of misunderstandings or errors due to cultural differences.

[0025] Ultimately, the invention facilitates a deeper understanding of patient needs, enables adaptation to specific cultural contexts, and fosters the creation of more empathetic and understandable responses in conversations with patients.

[0026] While this disclosure provides a detailed account of the problems and challenges the invention aims to address, the scope of the solutions discussed is not confined strictly to these particular matters. The invention, leveraging specialized diagnostic AI models and medical oversight of responses generated by large AI-based language models, could have broad applications in diverse medical contexts and beyond. Therefore, despite focusing on specific issues related to optimizing access to medical services, healthcare system efficiency, and challenges related to employing large language models in medicine, the examples included in this application should not be constrained to solving only these identified issues, as delineated in the background of the invention or elsewhere in this disclosure.

[0027] AI-based conversational assistant, referred to from now on as the AI Assistant, is a conversational chatbot capable of deploying multiple diagnostic artificial intelligence (AI) models, hereafter referred to as diagnostic AI models or diagnostic models, for patient diagnosis and aiding physicians through precise and efficient medical interviews. The AI Assistant has the ability to adaptively switch between different diagnostic AI models.

[0028] This adaptive shift signifies that the AI Assistant can preliminarily identify the medical field associated with the patient's problem during the interview, determine the relevant medical specialty, and transition to a diagnostic AI model specifically trained in diagnosing diseases from that medical field. Moreover, once a diagnosis is established by the diagnostic model for a given medical specialty, the AI Assistant can transition it to another diagnostic model that is specifically designed for managing and understanding the diagnosed disease. This shift between various specialty diagnostic models can occur unnoticeably and smoothly at any point during the interview. Therefore, whenever the AI Assistant changes specialty or diagnoses a disease, it can switch to a more suitable diagnostic model, enhancing diagnostic precision and effectiveness.

[0029] Furthermore, for patients with prior diagnoses, the Assistant can recognize the existing diagnosis and adapt accordingly, initializing the conversation using the model associated with the known condition. This capability enables the Assistant to ask more targeted and relevant questions from the onset, while retaining the ability to transition to other specialty models during the conversation, should the patient's responses suggest a new potential health issue or concern.

[0030] Leveraging AI technology, the Assistant engages in a conversation with the patient in natural language, gathering essential data about symptoms, medical history, and other pertinent information to understand the patient's health issue. The Assistant dynamically toggles between specialized models throughout the conversation, tailoring itself to the patient's needs and reported symptoms.

[0031] The designed diagnostic system features a distinctive algorithm that differentiates two primary conversation types, each designed uniquely to meet their intended objectives. The first stage, referred to in this application as the Initial Consultation Phase, aims not to diagnose, but to identify the most suitable medical specialty for treating the patient, to determine the appropriate level of medical care, and to assess the urgency of the patient's case. At this stage, the AI Assistant minimizes the number of asked questions to precisely narrow down potential disease units or groups of conditions. It then identifies the relevant criteria based on the patient's responses and the AI model's medical knowledge, and dynamically adjusts the diagnostic model to the identified medical field or specialty.

[0032] In the second stage, the process aims to establish a specific diagnosis and evaluate it by posing a larger number of detailed questions regarding both the patient's current symptoms and additional broad information related to their health condition and environment. During this stage, the Assistant can adaptively change to a diagnostic model representing a specific disease or sub-disease unit, allowing access to a broader information range, as the dedicated diagnostic model is based on a large number of specific disease cases. This adaptive shift enhances diagnostic precision and recommendations while simultaneously reducing computational power requirements compared to a universally-trained model.

[0033] Upon reaching the end of each stage, the AI Assistant or any employed diagnostic AI model may generate recommendations. Within the context of this patent application, 'recommendations' are defined as informed advice or suggested courses of action provided by either the AI Assistant or the relevant diagnostic AI model. These recommendations can include a range of potential next steps, inclusive but not limited to prescribing pertinent treatments or therapies related to the diagnosed disease or condition, suggesting specific medications, medical devices, or over-the-counter remedies, and advising on lifestyle modifications such as dietary changes, implementing exercise regimens or stress management techniques. They might also suggest further diagnostic tests, medical imaging, or laboratory work to validate or refine the diagnosis, provide referrals to relevant medical specialists for additional consultation, examination, or treatment, propose preventive measures or regular screenings, particularly for patients with high risk factors, give guidelines for monitoring symptoms, self-care, and managing the disease or condition at home, and relay information about potential risks, complications, or side effects associated with the disease or its treatment. These recommendations may form part of the output of both the Initial and Extended Diagnostic Phases, the specifics of which will be further detailed in this patent application.

[0034] While the system and its functionalities are exemplified within the framework of two specific patient conversation scenarios involving an AI-based conversational assistant, it's crucial to underscore that the scope of this invention extends beyond these instances. Consider, for instance, a scenario where a patient inputs their symptoms into an AI chatbot prior to hospital admission. In this context, the chatbot could provide initial triage, ask for further details regarding symptoms, and even facilitate the patient's admission to a specific department based on their responses. This proactive approach not only enhances the patient's comfort by offering immediate assistance, but also streamlines the healthcare process by providing the medical team with preliminary information, thus potentially increasing overall system efficiency. The multimodal approach disclosed herein, utilizing medical diagnostic models, is not only versatile but also adaptable to a wide array of medical contexts. These can range from patient interactions with an AI chatbot, to situations where AI-based conversational assistant acts as a supportive tool during in-person or telephonic consultations with a doctor, during hospital stays, home care monitoring, telemedicine interactions, remote patient monitoring, and more. Importantly, the application of this system is not constrained by the patient's medical condition, treatment location, stage of the diagnostic process, or the communication method used. Therefore, it retains its utility and adaptability across diverse healthcare scenarios.

[0035] This could include situations like an initial visit to a general practitioner, specialist consultations, emergency room admissions, hospital admissions, monitoring a hospitalized patient's health status, post-operative follow-up visits, telemedicine, conversations with chronically ill patients requiring regular health updates, or patient communication in home healthcare. Thus, while the described scenarios exemplify the system's operation, they are illustrative, not exhaustive, underscoring the multimodal technology's versatility and adaptability across various medical contexts.

[0036] In the invention at hand, the crucial aspect of the AI Assistant is the adaptive adjustment of diagnostic models, along with the ability to train those models based on conversation history and recommendations related to specific medical specialties or disease entities. This allows for the ongoing enhancement of dedicated diagnostic models through patient interactions.

[0037] For instance, consider a conversation with the AI Assistant:

[0038] Patient: Hi, I haven't been feeling great recently. I've noticed my legs are swollen, and I occasionally experience shortness of breath.

[0039] AI Assistant: Have you had any chest pains or heart palpitations?

[0040] Patient: No, I haven't had any pain, but sometimes it feels like something's obstructing my breath.

[0041] The AI Assistant identifies a potential cardiac issue and switches to a cardiology-specialized diagnostic model.

[0042] Cardiology-specialized diagnostic model: Have you noticed your breathlessness intensifies when you're lying down? Do you often wake up in the middle of the night feeling breathless?

[0043] Patient: Yes, that's exactly how I've been feeling! Sometimes I need to sit up to get a breath.

[0044] Based on this conversation, the cardiology-specialized diagnostic model might suggest that the patient likely suffers from heart failure. The AI Assistant could advise the patient to consult a primary care physician or a cardiologist for further tests and treatment.

[0045] Given the cardiology-specialized diagnostic model's suspicion of heart failure, during the subsequent Extended Diagnostics Phase, AI Assistant could shift to a diagnostic model specifically dedicated to the diagnosis and treatment of heart failure.

[0046] Heart failure diagnostic model: Do you have any chronic conditions, such as diabetes, high blood pressure, or kidney disease? Does heart disease run in your family?

[0047] Patient: My mother has high blood pressure, and I have type 2 diabetes. [...]

[0048] The specialized diagnostic model can ask more questions and then forward the results of the analysis.

[0049] This scenario demonstrates a situation where the AI Assistant needs to identify non-obvious symptoms and switch to a model specialized in a particular medical field and disease entity to best assist the patient.

[0050] The AI-based chatbot diagnostic system, which runs an AI assistant according to the present invention, includes at least one processing unit and a memory coupled to the at least one processing unit. The memory may comprise computer-executable instructions that, when executed by the at least one processing unit, generate an AI engine for creating a diagnostic model and perform a method of interaction.

[0051] The following figures serve as illustrative and not exhaustive or restrictive examples of the concepts discussed.

[0052] FIG.1demonstrates a client-server architecture used in the context of a medical conversational AI system as proposed in this application.

[0053] FIG.2displays a detailed layout of the server's structure, specifically designed for the client-server system discussed herein.

[0054] FIG.3illustrates the exemplary architecture of the AI Engine system, detailing its core components.

[0055] FIG.4illustrates the process of creating and integrating diagnostic models for diagnosing and treating medical conditions

[0056] FIG.5presents a flowchart detailing the AI Assistant's role in patient interaction, diagnosis verification, physician review, and reinforcement learning for future performance enhancement.

[0057] FIG.6outlines The Initial Consultation Phase, exhibiting the primary patient-AI interaction and the critical steps taken during this stage.

[0058] FIG.7represents The Extended Diagnostics Phase, emphasizing AI's use of specialist models for comprehensive diagnosis.

[0059] FIG.8presents a flowchart of The Integrated Diagnostic Process Under Doctor Supervision, showcasing how physicians are actively involved in the system's diagnostic process.

[0060] This invention pertains to a system grounded in client-server architecture (FIG.1), which can serve as a web application, but is not exclusively confined to this implementation. In actuality, the universality of the invention allows for a wide range of applications across various configurations and system architectures. For example, it can be an integral part of a hospital's website, providing a direct interface for patients. Additionally, the system can be integrated with diverse medical tools and systems, such as hospital software, patient information management systems, clinical data management systems, or telemedicine systems. Furthermore, even though the described implementation leverages a client-server architecture, the invention can also be implemented within other architectures, such as peer-to-peer architecture, multi-tier architecture, microservices architecture, or cloud-based architectures.

[0061] Additionally, the invention can be utilized not only as a web application but also as a mobile application, desktop application, embedded medical system, or even as a component of intelligent medical devices. Hence, while the detailed description concentrates on a particular example of implementation, the core of the invention is not restricted to this scenario and can encompass an array of diverse implementations and applications.

[0062] Fig.1illustrates a simple example of the client-server architecture. In this setup, a client device101is designated as any device with access to an internet or network connection, such as a computer, smartphone, tablet, or other hardware, which utilizes the client interface and connects through a computer network102, like the internet, to a server103by means of a dedicated application, web browser, or other software (e.g., a hospital system). The system permits multiple client devices101to connect to one or more servers103. During the conversational diagnostic process, participants (e.g., a patient, doctor, or other medical staff member) input information into the system via the client device101. The AI Assistant or diagnostic models partake in the diagnostic process, conducting a medical interview, posing questions, or acting as a doctor's assistant. The client device101may facilitate various forms of communication, such as text chat, voice calls, or video connections. The AI Assistant can either take an active role, conducting a medical interview and posing questions, or play a supportive role, assisting the doctor during the interview. For instance, during a virtual consultation, the AI Assistant might be employed to support the doctor in various ways. This could include real-time reference, providing relevant medical data or suggesting possible follow-up questions based on patient's symptoms or responses. At the same time, the client device101, through which the AI Assistant operates, can acquire authorization to store and access the patient's health data and / or gain access to their Electronic Health Records (EHR), which can be provided by authorized entities, such as hospitals or other healthcare organizational units. Both doctors and active AI diagnostic models participating in the diagnostic process can then analyze the EHR. The transfer of medical documents of the patient or doctor to the described system is also possible.

[0063] Fig2. displays the server's structure within the client-server system, including the role of the client210and the frontend website engine211. The server comprises various components that ensure the correct functionality and support for the system. Some of these components may reside on a single server or be distributed across multiple servers. In such cases, communication between the elements is established using appropriate communication protocols, e.g., API, through the computer network linking the individual servers, such as the Internet. The client210represents any device, like a computer or smartphone, accessing the system through an internet or network connection. This client interacts with the frontend application engine211, which serves as the user interface, enabling them to interact with the system and access its services. This engine could be part of a website, a standalone application, or any other form of software interface designed for user interaction. Within the Server220, there is a Server Engine221handling the logic of the access application, artificial intelligence models repositories, and communication with individual databases. The server also contains a User Database222, which stores essential information for managing system users, such as patients, doctors, and others. In addition, the server includes Healthcare Entity Databases223, gathering data on various healthcare entities, such as hospitals, clinics, or pharmacies. The server may also contain an Electronic Health Records (EHR) Database224, where individual patient medical information can be stored.

[0064] The AI Engine225constitutes a central element of the client-server system, coordinating the creation, modification, and functioning of various artificial intelligence (AI) models. Its fundamental role may involve conducting conversations with users, including patients and doctors, via an AI assistant or diagnostic models. It then analyzes the outcomes of these conversations to extract pertinent insights, such as medical diagnoses. The AI Engine can be deployed on a single server, across multiple servers, within a cloud-based structure, in a microservices architecture, or in any other compatible setup. It can interact with other components, like databases, server engines, or other system elements, with communication between the various system units facilitated through different protocols and interfaces, such as APIs.

[0065] The AI Engine225incorporates a range of AI models, including diagnostic models and the AI Assistant―a principal, universal large language AI model, potentially displaying AGI features. These models could be hosted in a variety of locations, such as locally on the user's device, on the server, or be accessible via an API interface on a different server. Specifically, the AI Engine can serve as a component of the IT system for a healthcare entity or other medical institutions.

[0066] The AI Assistant can facilitate the diagnostic process via interaction with the patient and / or the doctor involved in the diagnostic process. This interaction results in the creation of a chat index. This index encapsulates all information gathered during the conversation, including questions and responses, documents retrieved from the patient's record, documents submitted by the patient or doctor, recommendations and diagnoses, prior medical history, the patient's environmental and familial data, etc. Documents can exist in various forms―text, audio, image, video, and others―in a range of formats adapted to the patient's diagnostic needs. The chat index can also be constructed in an anonymized format, precluding identification of the patient subjected to the diagnostics. In addition, the chat index also comprises metadata such as the date and place of the chat, the participating doctor's information, and the AI models engaged in the diagnostic process.

[0067] The chat index can subsequently be utilized to establish orderly conversational repositories, which act as templates for data gathered in a specific field of medicine. The term medical field or field of medicine in this invention's context may refer to a broad spectrum of areas and sub-areas within medical knowledge, encompassing but not limited to, scientific research, diagnostics, treatment, and prevention of various ailments. A field of medicine could be associated with a medical specialty such as cardiology, oncology, or pediatrics, a particular disease, such as diabetes, Alzheimer's disease, or appendicitis, or a group of diseases sharing common characteristics, like cardiovascular or autoimmune diseases. The field of medicine could also include specific medical procedures, treatment methods, nursing care protocols, rehabilitation procedures, or any other relevant area within the scope of healthcare.

[0068] These conversational repositories can also be utilized for creating conversational algorithms and scenarios, tailoring them to the specifics of a particular field of medicine. Furthermore, the chat index can serve as a training data source for dedicated diagnostic models, the development and usage of which aim to enhance the precision and efficiency of diagnosing diseases within a specific field of medicine.

[0069] The AI Engine225can also access various types of data stored across multiple servers. These include, but are not limited to, medical personnel databases with various kinds of data such as information on their availability, patient profile data, patient health records, treatment plans, a directory of medical entities with data on their areas of expertise and referral levels, records of previous patient conversations (chat indexes), health recommendation repositories. There may also be various other data sources not shown in the diagrams, potentially including, but not limited to, medical entities' calendars with doctor appointment availability, AI Assistant recommendations, doctors' recommendations, drug indexes, prescription templates, referrals, etc.

[0070] This data could be stored on both local data storage devices and servers, including external servers owned by government institutions, hospitals, or other medical entities. The AI Assistant can simultaneously collaborate with multiple specialized diagnostic models, which provide support during diagnostic conversations. This collaboration entails tasks such as creating questions, determining diagnoses, issuing referrals and prescriptions, as well as summarizing conducted conversations.

[0071] FIG3. Depicts an illustrative structure of the AI Engine system, composed of various elements described herein. The User Interface301serves as a module for managing large language models, facilitating the generation of new models, modification of existing ones, and the removal of models. The interface has access to a diverse range of data sources, both generic and medical, encompassing one or more medical fields. These may include, but are not limited to, medical specialties (e.g., internal medicine), medical subspecialties (e.g., cardiology, pulmonology), disease entities (e.g., diabetes), disease sub-entities (e.g., Type II diabetes), among others.

[0072] These data sources can include, but are not limited to, internet search engines, content providers (journals, medical periodicals, websites, medical case studies, images, voice recordings), printed sources (books, journals, newspapers), transcripts from audio and video recordings, audio and video footage, summaries of audio and video content, doctor-patient interaction recordings, electronic patient records, conversational chat logs, among other materials. The User Interface collects data from one or multiple sources in response to a specific inquiry pertaining to a particular medical field.

[0073] The accumulated data is then consolidated and stored in a Medical Data Store302. This data store categorizes information based on various criteria, such as related diseases / specialties, data sources, the timing of data collection, or data type.

[0074] The Medical Indexing Engine303functions as a tool to create and update indexes for individual medical fields, in response to requests from the User Interface301. These indexes compile structured medical data from a range of sources like databases (e.g.,302), scholarly publications, or doctor-patient dialogues. Each index encompasses information about one or more medical fields, detailing diseases, their symptoms, diagnostic methods, treatments, and epidemiological data. An index can be dynamically augmented with supplementary information, including diseases exhibiting similar symptoms that can be used for differential diagnoses.

[0075] For instance, during the compilation of an index concerning respiratory diseases, such as asthma and chronic obstructive pulmonary disease (COPD), the training of an diagnostic model may lack precision due to the similarity of symptoms between these diseases and other conditions not strictly related to the respiratory system. To enhance the precision of diagnostic model training in differentiating between these diseases, the present invention anticipates the inclusion of additional information in the index about diseases with overlapping symptoms - in this example, conditions such as heart failure whose symptoms may mimic those of respiratory diseases.

[0076] Incorporating these additional data into the index enables the constructed diagnostic model to better comprehend various diagnostic possibilities and select the appropriate treatment, constituting a key innovative feature of the current invention.

[0077] During the creation or updating of indexes, the indexing engine utilizes a Medical References Database304supplemented with additional data sources such as web-based medical resources. This Medical References Database serves as a comprehensive repository that contains references to, and in certain cases copies of, a diverse range of medical knowledge resources. These resources may include, but is not limited to, scholarly articles, research papers, textbooks, medical guidelines, multimedia educational content, web-based medical resources, and other relevant publications or informational materials, available in the database or indexed for accessvia the internet. The use of these materials, whether as references or full-text copies, would be in accordance with the appropriate licensing agreements or authorized access where necessary, while other resources may be freely available. These references or copies pertain to various diseases, medical specialties, and broader medical topics, providing a rich source of information that can be utilized for the indexing process. In addition, data stored in the Chat Index Database305, which includes conversational indexes derived from doctor-patient dialogues, can also be utilized, enhancing the comprehensiveness of the medical indexes.

[0078] Following the stage of data collection and classification, a process of creating or updating selected diagnostic models takes place, employing various machine learning techniques. These techniques could include but are not limited to decision trees, naive Bayes classifiers, support vector machines, gradient boosting, neural networks, and reinforcement learning. Based on disease indexes303and the appropriate definition of specific disease entities within a particular medical field304, a fitting diagnostic model can be constructed. For instance, the foundation for creating diagnostic models can include one or more diseases from a specified medical field, which might closely resemble well-known medical specialties. The diagnostic model is then trained using detailed information about these diseases, ensuring it is capable of performing effective diagnostics through accurate differentiation, considering factors like disease progression, characteristic symptoms, age-specific incidence rates, and demographic, genetic, familial, and environmental data, among others.

[0079] Furthermore, diagnostic models can be trained using data relating to the characteristics of diseases in the context of public health, including statistics on disease prevalence in a specific group of people, from a particular region, and based on determined demographic data (age, gender, income, employment, education, etc.).

[0080] Within the scope of constructing a diagnostic model for a specific disease, the invention also identifies additional diseases that exhibit similarities to the primary one. For instance, if asthma serves as the main focus, the system not only creates a diagnostic model for asthma but also recognizes other conditions manifesting symptoms similar to asthma. This step is integral to improving the AI system's ability to distinguish between various diseases that share overlapping symptoms.

[0081] Upon recognizing these 'similar' diseases, the system proceeds to compute 'similarity coefficients' or 'measures' for each disease. These figures offer a comparative assessment of the likelihood of the presence of one disease unit in relation to another. They essentially provide a mathematical estimation of the degree of similarity across different diseases.

[0082] These similarity coefficients are then leveraged to create, organize, enrich, or update disease indexes for distinct medical fields. Picture these indexes as thorough references, each dedicated to a specific medical field such as respiratory or cardiovascular diseases. They include a comprehensive list of related diseases, each accompanied by its corresponding similarity coefficient.

[0083] The invention also uses conversation data for training diagnostic models. This data is derived from various dialogues and is included in the conversational indexes. It may encompass complex algorithms and instances from real-life conversations, such as doctor-patient dialogues about symptom descriptions and medical histories. The analysis of these conversations helps the system improve its ability to interpret and process natural language data within a medical context. Once these diagnostic models are trained and validated, they are stored in the Models Repository Database for future use.

[0084] The Models Repository Database306serves as a dedicated repository for all created and stored diagnostic models. It maintains a comprehensive inventory of these models, each with varying features, language versions, applications, and more. Each model stored in this repository has distinct attributes, thereby contributing uniquely to the diagnostic process. This may include aspects such as particular medical field focus, region-specific adaptations, and linguistic preferences. Therefore, the Models Repository Database facilitates efficient selection and utilization of the most appropriate model for any given diagnostic scenario, optimizing the accuracy and effectiveness of the diagnostic process.

[0085] Figure4provides a non-limiting example that illustrates an overview of the process for creating diagnostic models specialized in diagnosing and treating medical conditions. This depicted process includes, but is not limited to, the following steps:

[0086] Data from diverse sources, both generic and medical, is collected and prepared, encompassing a wide range of medical fields, specialties, disease entities, and sub-entities401. This consolidated data is then stored and organized in the Medical Data Store according to criteria such as related diseases / specialties, data sources, the timing of data collection, or data type402. Utilizing this stored data, the Medical Indexing Engine creates comprehensive indexes for individual medical fields. The indexes include detailed information about diseases, their symptoms, diagnostic methods, treatments, and epidemiological data403. An essential part of creating these indexes is the computation of 'similarity coefficients' or 'measures' for each disease, providing comparative assessments of disease similarities404. These similarity coefficients are crucial for improving differential diagnoses by offering a mathematical estimation of the degree of similarity across different diseases.

[0087] The compiled indexes and similarity coefficients are further enriched with resources from the Medical References Database, a comprehensive repository of medical knowledge resources405. Once the data is adequately organized and indexed, the process of creating or updating diagnostic models begins406. Leveraging machine learning techniques, these models are trained using a variety of information. The training data can consist of detailed information about diseases, various types of medical data, and even dialogue data. These dialogues can come from real doctor-patient interactions, providing insight into how symptoms are described and questions are asked in real-life medical scenarios. Alternatively, the dialogues can also be artificially generated, using AI techniques to simulate doctor-patient interactions, thereby increasing the diversity and volume of training data. Utilizing a broad range of data, the training process equips the models to diagnose various medical conditions, interpret diverse symptoms, and engage in diagnostic dialogue effectively, replicating medical professional expertise.

[0088] During the process of creating diagnostic models, the system offers a decision on what training approach to apply407. The first path involves building a diagnostic model from scratch408. In this process, the system starts with a blank slate, and the model is initially trained with general data to establish a foundation. This training aims to equip the model with necessary skills such as understanding the language structure, context interpretation, and other essential features. Following this, the model is further trained and constructed based on the accumulated and indexed medical data. This approach offers the advantage of customizability, allowing the system to create a model that is highly tailored to the specific medical field or disease entity. The second path, which often proves more efficient and cost-effective, involves fine-tuning existing models409. Rather than starting afresh, this process uses pre-existing models as a foundation and refines them based on the specific knowledge and data accumulated in the disease indexes. This approach enhances the accuracy of these models, tailoring them to the particular medical field or disease entity in focus. Once the models are fine-tuned, they are validated to ensure their reliability and effectiveness in diagnosing diseases. Whether a new model is created or an existing one is fine-tuned, the resulting diagnostic models are then integrated into the enriched disease indexes410. These indexes, now augmented with the diagnostic models, serve as comprehensive and specialized references for various medical fields.

[0089] The diagnostic models are then stored in the Models Repository Database, maintaining a comprehensive inventory of these models, each with varying features, language versions, applications, and more411. The system's flexibility in being able to create new models from scratch or fine-tune existing models ensures an effective and resource-efficient process of developing high-performing diagnostic tools.

[0090] To manage the use of the diagnostic models stored in the Models Repository Database306, the system utilizes an AI Models Manager307. This component is designed to interact with, manage, and utilize various artificial intelligence models. It may be configured to gain access to one or multiple AI models306. These models can be stored and managed on various servers both internally and externally to the organization, and can be accessed through different pathways, including licenses, partnership agreements, research collaborations, resource sharing arrangements, or other forms of cooperation between entities. The AI Models Manager communicates with these models, for instance, via an API interface, allowing secure bi-directional data transmission between the engine and the AI models, but may also include other methods of communication and data formats, depending on the specifics of each AI model.

[0091] During the execution and analysis of conversations between the patient and one or more diagnostic models, the AI Models Manager may obtain access to all necessary data for its correct functioning. This includes access to conversation data, the chat index, which encompasses doctor-patient, doctor-AI, and AI-to-AI interactions, as well as any relevant data generated or affected by these conversations. This could include, but is not limited to, diagnoses, recommendations, referrals, and test results. It's important to note that this model can consider conversation data from multiple users and across various conversations.

[0092] In cases where the information contained in models specific to various medical fields is insufficient, or when the data incorporated in the diagnosis or recommendations needs to be verified for alignment with official guidelines, the AI Models Manager can access the Medical References Database304resources. These sources may contain general information about diseases, established procedural standards, severity indices, and other guidelines for medical procedures. The AI Models Manager may also utilize conversational data, which may include conversation scenarios such as questions / answers that can be predefined or automatically pre-generated. Conversational data may not necessarily be associated (or possible to associate) with a specific disease entity or specialization but can serve as a knowledge base supporting the diagnostic process and communication with the patient.

[0093] In situations where the diagnostic model is unavailable or does not exist for a certain condition, the AI Models Manager307can be configured to automatically use the medical reference database related to the specific condition. As a result, even in the absence of a diagnostic model, the AI Models Manager can continue the diagnostic process using the AI Assistant, relying on available reference data sources. This mechanism allows doctors and patients to receive information that aligns with the current medical standards and best practices, even in the absence of a suitable diagnostic model.

[0094] The AI Safety Mechanism309is an integral component of the diagnostic system, which is designed to verify and safeguard the outputs of the artificial intelligence models. It includes various mechanisms to check and validate the diagnoses and recommendations provided by the system. One such mechanism is a countermeasure against confabulations308, which refers to false or unrealistic information generated by the AI model that is not rooted in the input data or actuality.

[0095] This confabulation countermeasure aims to safeguard the system from generating erroneous recommendations potentially arising from AI models mistakes. In the context of artificial intelligence, confabulations are a phenomenon where the AI model produces false or unrealistic information that isn't derived from the input data or grounded in reality. The term "hallucinations" is sometimes used synonymously with confabulations in reference to AI, but it should be noted that this term has its origins in psychiatry where it denotes false sensory impressions. In truth, the issue at hand pertains to situations where AI model generates data not based on reality, a situation more accurately aligned with the definition of confabulation.

[0096] Confabulations could precipitate hazardous situations, particularly in the medical field, where inaccurate recommendations could have ramifications for patient health. To combat confabulations, the system could utilize an AI Assistant or, for instance, another AI model specifically tailored for this purpose, such as a counter-confabulation model. The counter-confabulation model could juxtapose the recommendations made by diagnostic models with reference values tied to various diseases, scientifically-based origins, official treatment guidelines, procedural standards, or other sources deemed reliable.

[0097] This mechanism could involve comparing the recommendation generated by the AI assistant and / or diagnostic models with reference values, and checking if there is a substantial similarity between them. If a significant discrepancy is detected, the counter-confabulation model may identify a potential confabulation and initiate measures to rectify it. This might include generating a new recommendation based on reference values, notifying the doctor of a potential issue, and / or providing directions to additional information sources that allow for more thorough verification.

[0098] Implementing a countermeasure against confabulations, executed by an AI Assistant or a dedicated counter-confabulation model, in the described diagnostic system based on artificial intelligence, could considerably enhance the safety and reliability of the AI-generated recommendations. This, in turn, could contribute to the improvement of healthcare quality and a reduction in the risk of diagnostic errors.

[0099] The adoption of such a counter-confabulation mechanism, through the use of an AI Assistant or a dedicated counter-confabulation model, could also foster greater trust among users in AI-based systems. This could, in turn, lead to wider acceptance of such technologies in medical practice.

[0100] In conclusion, the counter-confabulation mechanism, executed by the AI Assistant or dedicated counter-confabulation model, is a vital component of the AI-based diagnostic system. It enhances the safety and reliability of generated recommendations while simultaneously contributing to improved healthcare quality and a reduced risk of diagnostic errors.

[0101] Fig.5presents a simplified dialogue diagram of the AI Assistant interacting with a patient. Upon initiating the conversation501, the AI Assistant may recognize the relevant medical field, specialty or disease and determine the need to engage an additional, specialized diagnostic model502for a more precise diagnosis. This diagnostic model can actively participate in the patient dialogue by asking detailed questions about symptoms, medical history, and other significant aspects503. Alternatively, the diagnostic model may aid the AI Assistant by directly communicating with it and supplying essential information, such as suggestions for questions, diagnoses, or therapeutic recommendations. Once the AI Assistant concludes that the patient conversation has finished and the recommendations are prepared for presentation to the physician504, the counter-confabulation model proceeds to verify them505. If any errors are discovered506, the counter-confabulation model generates pertinent information about these and communicates it to the physician507. The physician can then review the consultation findings508, and if they introduce modifications to the recommendations509, these changes can be used to retrospectively train the diagnostic model using a reinforcement learning technique510. Within this context, the physician's adjustments act as forms of rewards or penalties, assisting the model in better understanding which decisions were accurate and which required correction. As a result, the diagnostic model can learn to improve its future decisions, enhancing overall precision and effectiveness. Should the physician accept the diagnosis, the chat index and associated information can be added to the Chat Index Database305for the purpose of improving future iterations of the diagnostic model.

[0102] Both in employing the reinforcement learning technique and in incorporating data into the database, it should be emphasized that this process may specifically target the diagnostic model used during the diagnosis. This potential for focused improvement could lead to an optimally efficient learning process, progressively enhancing the precision and effectiveness of the individual diagnostic model, while not precluding the possibility of AI Assistant model training.

[0103] The two following examples illustrate how the system can learn from real-world instances and mistakes, continuously improving through reinforcement learning and the use of the Chat Index Database.

[0104] Example 1:

[0105] The pulmonology diagnostic model accurately diagnosed a lung disease in a patient. Nevertheless, the treatment recommendations, including medication dosages, didn't align with the standards recorded in the Medical References Database304. The counter-confabulation mechanism detected the discrepancies and alerted the physician. Upon reviewing the recommendations alongside the references from the Medical References Database, the physician amended the treatment, adjusting medication dosages accordingly. The data gathered from this diagnostic process (chat index) are appended to the Chat Index Database and utilized for training the same pulmonology diagnostic model. As a result, the system increasingly enhances its accuracy and efficiency in subsequent diagnoses.

[0106] Example 2:

[0107] A cardiology diagnostic model was employed to diagnose a patient's heart condition. The patient presented with symptoms such as chest pain, shortness of breath, and fatigue. The AI Assistant utilized the cardiology diagnostic model to pose detailed questions about the patient's medical history, lifestyle, and family background. Based on the patient's responses, the cardiology diagnostic model suggested an incorrect diagnosis. As the diagnostic model recommended what it deemed appropriate treatment for the misdiagnosed condition, the counter-confabulation mechanism, cross-referencing the Medical References Database304, did not register any anomalies. Upon reviewing the diagnostic model's recommendations, the doctor observed that the disease had been misdiagnosed, resulting in an incorrect treatment recommendation. The doctor revised the diagnosis, which prompted the diagnostic model to reassess the new diagnosis against the Medical References Database. This time, the diagnostic model proposed the correct treatment, aligned with the updated diagnosis. The doctor accepted these recommendations, and the data from this diagnostic process (chat index) were incorporated into the disease index to refine the cardiology diagnostic model for future use.

[0108] The previously mentioned description of the counter-confabulation mechanism represents just one of the possible uses of the AI safety mechanism308in the described diagnostic system. Various other verification mechanisms, not necessarily focusing on confabulations, can also be applied in this context. More generally, the diagnostic system can utilize any verification AI model that compares generated recommendations, information, or diagnoses with reference values obtained from different, reliable sources. Such a verification model could serve to confirm the accuracy of a diagnosis, detect discrepancies or errors, and enhance the overall quality and reliability of the output.

[0109] Consequently, the system continually improves in reliability and utility in diagnosing patients. The reinforcement learning mechanism, coupled with the use of chat indices, enables the ongoing refinement of diagnostic models. These improvements allow the models to better support doctors and patients in the diagnostic and treatment processes. Through this process, the system becomes increasingly reliable and useful to both physicians and patients, leading to superior medical services and improved patient health outcomes.

[0110] The objective of the AI Assistant-led medical interview is to efficiently and accurately gather information about the patient, their symptoms, and health needs from various medical perspectives. The AI Assistant employs advanced text analysis techniques and semantic understanding of the patient's utterances, facilitating precise symptom identification, narrowing down disease or disease group search criteria, and tailoring the interview process to the individual needs of the patient, factoring in elements such as age, gender, language preferences, and level of medical knowledge.

[0111] A key element of the invention is the novel methodology for conducting the medical interview by the AI Assistant. This methodology synergizes advanced machine learning techniques, semantic analysis, and adaptive question generation, optimized for effectiveness, accuracy, and patient comfort. The AI Assistant operates based on a hierarchical conversation structure, enabling progressive deepening of patient information, initiating with general health status and medical history queries, transitioning to more detailed inquiries about symptoms and other pertinent information.

[0112] Throughout the interview, the AI Assistant deploys advanced semantic and contextual analysis techniques to enhance understanding of the patient's statements and distill relevant information. Consequently, the AI Assistant can interpret the patient's responses more accurately and tailor the subsequent questions based on prior information gathered during the interview. Furthermore, the AI Assistant can proactively verify the information received from the patient, asking for additional details or confirmation of previous answers. This approach mitigates potential ambiguities and reduces the risk of errors emanating from imprecise or incomplete patient responses.

[0113] Conversation method:

[0114] The practical application of information verification is demonstrated in the following AI Assistant-patient conversation example.

[0115] AI Assistant: Hello! How may I assist you today?

[0116] Patient: I've been experiencing a headache and fever for a few days.

[0117] AI Assistant: I'm sorry to hear that. Could you provide more details about your headache? Is it throbbing, constant, or sharp pain?

[0118] Patient: It's more of a steady pain that intensifies when I move my head.

[0119] AI Assistant: I appreciate your input. About the fever, could you tell me the temperature you most recently recorded?

[0120] Patient: I last recorded a temperature of 37.8°C (100°F).

[0121] AI Assistant: Thank you. Have you noticed any other symptoms, such as a sore throat, cough, runny nose, or muscle aches?

[0122] Patient: Yes, I also have a sore throat and a runny nose.

[0123] AI Assistant: I apologize, I just want to verify that I understood you correctly. You've mentioned having a headache, a fever of 37.8°C (100°F), a sore throat, and a runny nose, correct?

[0124] Patient: Yes, that's correct.

[0125] AI Assistant: Thank you for your confirmation. Do you have any other concerns or pertinent health information that you have not disclosed yet?

[0126] Patient: No, I believe that covers everything.

[0127] AI Assistant: In that case, based on the presented symptoms and information verification, I can proceed with further analysis and suggest potential causes for your symptoms. Please bear with me for a moment.

[0128] In the example above, the AI Assistant actively verifies information received from the patient by seeking confirmation of the stated symptoms.

[0129] The AI Assistant has been designed to take into account various levels of healthcare referral, while assessing the urgency of a patient's needs in order to tailor its recommendations accordingly. It analyses information obtained during a conversation with the patient to determine the appropriate referral level of healthcare facility the patient should be directed to, and also evaluates how urgently medical help is required. It's worth noting that different countries may define individual referral levels of healthcare differently, and the examples given here do not limit the scope of the patent.

[0130] At the first referral level, the assistant may deem a visit to a primary care physician sufficient. Here, the physician would assess the overall health of the patient, and diagnose and treat basic conditions. For less urgent healthcare needs, the assistant might recommend scheduling an appointment at a suitable time.

[0131] If the assistant identifies that the patient requires specialized care, they could direct the patient to a specialist outpatient clinic. When the patient's health status necessitates immediate intervention, the assistant will suggest an urgent referral to a specialist. For needs of lesser urgency, the assistant might recommend scheduling a specialist visit in a planned manner.

[0132] In instances of health needs requiring advanced care, the AI Assistant could advise that the patient be referred to a suitable hospital. Hospitals can have varying levels of referral, for example, city, county, regional, university, academic, or specialized hospitals. Depending on the urgency of the patient's medical needs as assessed by the AI assistant, it may recommend either a planned hospital referral or immediate hospitalization.

[0133] The AI Assistant has the capability to adapt its recommendations to suit various healthcare systems, bearing in mind the distinct referral levels and the urgency of the patient's medical needs.

[0134] Fig.6illustrates The Initial Consultation Phase.

[0135] At the onset of a conversation with the patient (point601), the AI Assistant establishes contact, posing initial questions to understand the patient's needs, symptoms, and the history of their illness related to a specific disorder.

[0136] When the AI Assistant poses a question and receives a response from the patient (point602), it utilizes its capabilities to process general information, which also encompass medical knowledge. Its versatility and ability to process and learn from a broad range of data make it capable in general medical diagnosis. At this stage, depending on certain factors such as licensing limitations that might prevent the AI Assistant from being used for medical diagnostic purposes, a 'general medical diagnostic model' might be utilized in place of the AI Assistant. The term 'general medical diagnostic model' refers to a diagnostic model specifically designed and trained to diagnose a broad range of medical conditions based on symptom descriptions and other relevant patient information. However, if the specific context of the conversation suggests a need for more specialized analysis, the AI Assistant has the capability of switching diagnostic modes, and further patient symptom analysis may be conducted or facilitated by a diagnostic model specific to a certain medical field or specialty.

[0137] The patient may communicate responses through various means, including but not limited to text, voice, and visual aids such as photos or videos. Additional data beneficial in the diagnostic process, such as medical test results, may also be provided. The AI Assistant may inquire if the patient authorizes them to access their Electronic Medical Records and other health-related data linked to the patient. If authorization is granted, the assistant can integrate a suitable range of this data into the context of the ongoing conversation, thereby creating a comprehensive medical profile of the patient. This step is essential in gathering the necessary data for further analysis, evaluation, and recommendation issuance, aligned with the goals of the conversation - accurately determining the appropriate medical specialization, identifying the medical referral level, and effectively assessing the urgency of the case. The AI Assistant uses machine learning algorithms and adaptive diagnostic model switching to continually analyze the patient's responses, facilitating a better understanding of the patient's symptoms and needs. This enables prompt adaptation to the particular case.

[0138] The next step involves the analysis of the patient's most recent response (at point603). This analysis is conducted while considering any previous responses and the existing medical profile of the patient, if one has been created. This methodology allows the AI Assistant to better understand the context of the patient's symptoms, thereby guiding more precise inquiries and increasing the accuracy of the final recommendations.

[0139] At point604, the AI Assistant or a specialized diagnostic model dedicated to this task evaluates whether a life-threatening situation has been detected. If so, it proceeds to point605, where it advises the patient to seek immediate medical help or dial an emergency number. Furthermore, if feasible, the AI Assistant may initiate an immediate handover to a medical professional or an emergency dispatcher, concluding the diagnostic process. If no life-threatening situation is detected, the AI Assistant continues with the diagnosis, advancing to point606.

[0140] At point606, the AI Assistant evaluates whether, based on the analysis conducted by the currently applied diagnostic model, it is highly probable to categorize the case into one of the following: medical specialty, referral level, or case urgency. If affirmative, the algorithm moves to point607, where the respective category is updated or assigned based on the analysis. In case of ambiguity, the AI Assistant evaluates whether all defined categories have been determined (point614).

[0141] Subsequently, the Assistant moves to point608to check whether the newly assigned category falls under the medical specialization category. If not, the algorithm advances to point614.

[0142] Following positive condition of step608, if the newly defined category encompasses a medical specialty, the AI Assistant verifies whether there is a corresponding specialist diagnostic model dedicated to this specialty amongst all available diagnostic models (step609). If such a diagnostic model exists, it transitions to step610and selects the most suitable specialist diagnostic model based on previously analyzed data gathered during the ongoing conversation and the medical profile of the patient. This selection is followed by an adaptive modification of the diagnostic model throughout the conversation, aiming to enhance diagnostic precision and conversation procedure efficacy.

[0143] Subsequently, at step611, the AI Assistant compiles a summary of the conversation thus far and switches the diagnostic model to the chosen specialist diagnostic model (step612). This switch is dynamic and does not necessitate patient intervention. Depending on the application's settings, the patient might remain oblivious to this change. At this juncture, the AI Assistant can adopt one of two roles depending on the scenario. In the first scenario, the AI Assistant serves as a "conversation moderator," only intervening when necessary, while the communication primarily takes place between the patient and the specialist diagnostic model. In the second scenario, the AI Assistant remains the primary communicator with the patient, while the specialist diagnostic model supports by suggesting questions and providing insights. Regardless of the scenario, the AI Assistant, alongside the specialist diagnostic model, analyzes responses in the context of the entire conversation, aiming to identify instances where the specialist diagnostic model might have been incorrectly determined. In such situations, the AI Assistant may adjust the specialist diagnostic model (this is not detailed in the diagram).

[0144] At step613, the AI Assistant provides the new diagnostic model with useful data which may include but is not limited to a summary of the current conversation, relevant parts of the conversation, and the medical profile of the patient (if available). This aids in better preparing the specialized diagnostic model for initiating and maintaining further conversation.

[0145] Subsequently, the Assistant evaluates whether all defined categories have been determined (point614). If affirmative, the AI Assistant moves to point615. If not, the algorithm circles back to point602to pose additional questions to aid in more accurately classifying the case.

[0146] At step615, the new specialist diagnostic model assesses the conversation's progress thus far, involving the evaluation of data and responses procured by the AI Assistant. If more information from the patient is required to meet the conversation's objectives, the set specialist diagnostic model returns to step602and proceeds with gathering additional information from the patient. The subsequent analysis is conducted by the designated specialist model. If this isn't the case, it advances to step617, where the AI Assistant generates and displays a summary along with recommendations.

[0147] Identifying distinct categories enables a more precise alignment of recommendations with the individual patient's needs, thereby facilitating improved case management and enhanced treatment efficacy. The preliminary algorithm ultimately provides recommendations for further medical procedures, including the identification of specific medical facilities or specialists the patient should consult next. Given the diversity of legal frameworks governing the use of AI in healthcare in different countries, this could encompass a broad spectrum of specific actions. These may include, but are not limited to, referring the patient to a specialist, advocating for additional diagnostic tests, suggesting subsequent steps in treatment, issuing valid referrals, prescribing medications, or implementing any other appropriate medical action, subject to the allowances of the specific legal and healthcare system in place. In practical terms, the AI Assistant can expedite patients' access to the appropriate specialists, thereby reducing the time and effort required for multiple doctor visits, a frequent necessity in traditional medical diagnostics. Consequently, patients may commence treatment more swiftly and achieve superior health outcomes, while the healthcare system can manage resources more effectively and optimize treatment pathways for diverse cases.

[0148] In summary, the Initial Consultation Phase presented herein centers on initiating patient contact, gathering requisite information, performing an initial analysis and evaluation of the situation, and subsequently employing a dedicated specialist diagnostic model for improved diagnostics. This entire process augments the enhancement of dedicated diagnostic models and heightens the precision of issued recommendations. Hence, the adaptive modification of diagnostic models during the conversation, as proposed in the process described, can contribute to enhancing medical care quality, boosting diagnostic efficiency, and promoting better resource management within the healthcare system.

[0149] At this point, the Initial Consultation Phase comes to an end. In the next phase, the AI Assistant has the ability to perform a detailed diagnosis, employing either a specialist diagnostic model or its inherent standard diagnostic capabilities. This aids in gathering more patient data and providing more precise recommendations. The AI Assistant may also pass the collected conversation data and its recommendations to a physician. This allows for verification of the results and physician oversight during the diagnostic process, potentially influencing the final decision regarding the diagnosis and subsequent medical procedures.

[0150] In the context of the Extended Diagnostics Phase, as illustrated on Figure7, the diagnosis process begins by first gathering and then aggregating data related to the patient's ailments, which are collected during the Initial Consultation Phase. This may encompass, but is not limited to, aspects such as: the conversation history, assigned categories, conversation summary, and provided recommendations (Point701). The collected data subsequently constitute the specialized medical profile of the patient. The incorporation of additional data into this profile is also feasible if the patient grants authorization to access other sources of medical data.

[0151] Subsequently, it is verified whether a specialist diagnostic model for the previously designated specialization category is available (Point702), an element critical for the adaptive modification of AI models during the conversation. If such a model exists, the system transition to the selection and adjustment of the most suitable specialist diagnostic model. The AI Assistant then advances with further diagnostic procedures utilizing an appropriate diagnostic model specialized in a specific medical field (Point703). If a specialist diagnostic model for a designated category is unavailable, the AI Assistant leverages its general knowledge, acquired from extensive prior training on diverse data, including medical, to proceed with the diagnostic procedures (Point704). If there are any constraints, such as licensing limitations or other factors, that prevent the AI Assistant from being used for medical diagnostics, a general medical diagnostic model can be engaged in place of the AI Assistant at this stage. Please note, the term 'model availability' refers not necessarily to the existence of a particular model, but rather its accessibility for use, such as whether it has been licensed for application.

[0152] Upon setting the suitable diagnostic model, it is supplied with the patient's data gathered in the specialized medical profile of the patient (Point705), thereby facilitating an evaluation of the information and initiation of the conversation with the patient via the AI Assistant (Point706). The current diagnostic model poses questions and receives responses from the patient (Point707). These responses may take various forms. For instance, this may include, but is not limited to, a text-based response, an uploaded image or video illustrating the patient's symptoms, a file with an illustrative image, laboratory test results, or even access to comprehensive patient records. Based on the analysis of these responses, accounting for prior responses and patient data, the AI Assistant endeavors to determine the patient's diagnosis using the diagnostic model (Point708).

[0153] The key element of the Extended Diagnostics Phase is the evaluation of the diagnostic model's capacity for medical assessment (point709). In cases where the diagnostic model or AI Assistant cannot provide a provisional diagnosis, they continue the conversation with the patient, asking more questions (point707).

[0154] Upon establishing a provisional diagnosis (point709), the system initiates a search for a dedicated disease model specifically designed for this diagnosed condition. This dedicated model, if assigned, provides focused and detailed diagnostic support for the particular ailment. If there is no assigned dedicated disease model available for the provisional diagnosis (point710), the system then seeks a specialized diagnostic model that covers the broader medical field or specialty relevant to the provisional diagnosis (point711).

[0155] This specialized diagnostic model, while not exclusively focused on the diagnosed condition, has the capacity to manage all diseases within its specific medical specialty. It serves as a more generalized support tool when a dedicated disease model is not available, ensuring that the diagnostic process can continue with expert assistance.

[0156] When a suitable model is identified, it is selected for the confirmed diagnosis (point712), a summary of the conversation is created (point713), and a transition is made to this chosen diagnostic model (point714). Just as in the Initial Consultation Phase, the AI Assistant can adopt one of two roles depending on the scenario. In the first scenario, the AI Assistant serves as a "conversation moderator," only intervening when necessary, while the interaction primarily occurs between the patient and the diagnostic model. In the second scenario, the AI Assistant remains the primary communicator with the patient, with the diagnostic model assisting by proposing questions and providing insights.

[0157] If no specialized AI model is available or one does not exist (point711, negative condition), the algorithm decides to persist with the diagnostic conversation employing either the current AI Assistant or diagnostic model (proceed to point716). Regardless of the scenario, both the AI Assistant and the diagnostic model analyze responses in the context of the entire conversation, aiming to identify instances where the diagnostic model may have been incorrectly determined. In such situations, adjustments may be made to the chosen diagnostic model.

[0158] Subsequently, the established diagnosis, the specialized medical profile of the patient, summary, and conversation history are all provided to the dedicated diagnostic model (point715). The data obtained in this way are then evaluated by the current diagnostic model (point716). Based on this analysis, if the current diagnostic model determines that additional questions are necessary to formulate a conclusive diagnosis for the patient, the AI Assistant manages the continuation of the interview (point717, positive condition). Depending on the cognitive capabilities of the diagnostic model, the AI Assistant may actively conduct the interview using the assistance of the diagnostic model, or the diagnostic model may guide the discussion with the AI Assistant providing a supportive role.

[0159] Finally, the AI Assistant recapitulates the conversation and generates the patient's examination results utilizing the implemented diagnostic model (point718).

[0160] The Extended Diagnostics Phase concludes by gathering the patient's comprehensive diagnostic results, which are then prepared for further utilization in accordance with the system's settings.

[0161] Integrated Diagnostic Process Under Doctor Supervision.

[0162] Highlighting the importance of existing legal constraints in various countries, structured to ensure heightened patient safety, the interaction with the AI Assistant has been programmed to permit real-time facilitation by a medical professional. This design allows the physician to supervise the whole diagnostic process, intervene whenever necessary, and formulate the final diagnosis based on the AI Assistant's findings. The physician's expertise, experience, and intuition can be employed during this interaction to pinpoint potential patient health concerns, which the AI Assistant might overlook or disregard. This collaboration enhances the physician's ability to leverage their knowledge and skills, fostering a deeper understanding of the patient's health conditions and augmenting diagnostic quality.

[0163] The Integrated Diagnostic Process Under Doctor Supervision, as illustrated in Figure8, is initiated upon the patient's decision to participate (point801), offering further refinement of diagnosis via specialized AI models. Should the patient opt out of this process, it is subsequently terminated (point801, negative condition).

[0164] Following a decision to proceed, the system begins by consolidating the results of the Initial Consultation Phase, the primary objective of which is to identify the predetermined categories, which include the medical specialty, level of referral, and urgency of the case (point802). These results may be supplemented by additional information such as patient-specific data, comprehensive responses with corresponding examination outcomes, and an overall summary with recommendations. Subsequently, the requirement for physician oversight during the Extended Diagnostics Phase is evaluated (point803), taking into account the country-specific regulations, legal mandates, and the patient's individual needs. If physician supervision is warranted, or at the patient's behest, the physician is engaged in subsequent diagnostic steps (point804). In the absence of these conditions, the process proceeds with the patient's Extended Diagnostics Phase without physician supervision (point807).

[0165] When a physician participates in the diagnostic process, their initial responsibility involves scrutinizing the collated patient data (point804). This is followed by a decision regarding the necessity of the Extended Diagnostic Phase for the patient (point805). If the physician recommends an alternative diagnostic pathway for the patient (point806), this concludes this process. If the decision is to proceed with the Extended Diagnostics Phase, a corresponding algorithm is initiated (point807) (detailed in Fig.6), utilizing specialized diagnostic models to yield more precise results. Upon algorithm completion (point808), the results undergo analysis. In jurisdictions where physician oversight is not legally required and is therefore deemed unnecessary (point809), the patient is endowed with the autonomy to decide on a potential consultation with a physician (point810).

[0166] If the patient opts for a consultation with a physician, or if the physician is engaged in the process from the outset, the physician receives and analyzes the results of the Extended Diagnostics Phase (point811). The physician may provide additional advice to the patient during a subsequent consultation (point812), and then decide whether to adjust the results derived from the Extended Diagnostics Phase (point813). If the physician opts not to make any adjustments, the results are presented to the patient (point816). If adjustments are needed, the physician modifies the results (point814) and presents the patient with the adjusted diagnostic results (point815). This facilitates greater physician involvement in the diagnostic process and incorporates their expertise and experience. If the patient decides against a consultation with a physician (point810, negative condition), the results from the Extended Diagnostics Phase are presented as unadjusted (point816).

[0167] Upon completion of the entire diagnostic process, the Integrated diagnostic process under doctor supervision is finalized (point817), offering the patient more comprehensive information about their health status. This could lead to quicker and more precise referrals to appropriate medical specialists and facilities.

[0168] In summary, the Integrated Diagnostic Process Under Doctor Supervision introduces additional features within the scope of this invention aimed at minimizing potential harm from errors in artificial intelligence algorithms. This includes the option for patient consultation with a physician, physician oversight when necessary, and the modification of diagnostic results by the physician. These features facilitate improved collaboration between the patient, the AI Assistant, and the physician, enhancing the precision and efficiency of the entire diagnostic process.

Claims

1.An AI-based chatbot diagnostic system for providing a suitable diagnosis for patients, comprising:at least one processing unit; anda memory coupled to said at least one of the above processing units; whereinsaid memory includes computer-executable instructions that, when executed by said at least one processing unit, generate an AI engine for creating a diagnostic model and perform a method of interacting, said method comprises:receiving data from at least one data source related to at least one medical field or disease;storing the received data in a medical data repository, wherein the data is categorized according to at least one criterion;accessing a medical index engine that uses the categorized data and a medical reference database to create and update an index for individual medical fields;utilizing interactive data for training a diagnostic model in conjunction with the medical index;selecting the trained diagnostic model most suitable for a given diagnostic scenario from a model storage database; andmanaging and interacting with the selected diagnostic model using an AI model manager; constituting the AI-based chatbot diagnostic system.2.The AI-based chatbot diagnostic system of Claim 1, further comprising:a database storing user information, medical institution information, for providing suitable conversation services to the patient; andan AI engine including at least one diagnostic model for conducting a conversation with the patient;wherein said AI engine is configured to:perform questions corresponding to a first model to receive a first set of responses from the patient,based on the first set of responses and the information from the database, adaptively transition the first model to a second model;performs questions corresponding to the second model to receive a second set of responses from the patient, andoutput recommendations based on the first and second set of responses, constituting the AI-based chatbot diagnostic system.3.In accordance with claim 2, wherein the said AI engine:selects a diagnostic model based on the patient's user information, electronic health record information, and the first through second sets of responses provided by the patient; andprovides the selected diagnostic model with the patient's user information, electronic health record information, and the first through second sets of responses to output a diagnostic result; constitutes the AI-based chatbot diagnostic system.4.In accordance with claim 3, wherein the AI engine:determines whether a suitable diagnostic model has been assigned to the patient,if a suitable diagnostic model has been assigned to the patient, outputs a diagnostic result based on the assigned diagnostic model, andif no suitable diagnostic model has been assigned to the patient, selects an available diagnostic model for the patient based on the patient's user information, electronic health record information, and the first through second sets of responses, and then outputs a diagnostic result based on the selected diagnostic model;constitutes the AI-based chatbot diagnostic system.5.In accordance with claim 3, wherein the AI engine:,upon receiving a request for a doctor's diagnostic involvement from the patient, or when it determines that a doctor's diagnostic involvement is necessary, changes the initial diagnostic phase to an extended diagnostic phase; andreflects the doctor's additional consultation or a revised diagnostic result in the diagnostic result for the extended diagnostic phase;constitutes the AI-based chatbot diagnostic system.6.In accordance with claim 2 or claim 3, wherein the AI engine:employs a counter-confabulation model mechanism to evaluate the recommendations or diagnoses that have been outputted;outputs an error to the doctor if such an error is detected by the counter-confabulation model mechanism; andupon the doctor's correction of the error, provides feedback to the model that outputted the recommendations or diagnosis,constitutes the AI-based chatbot diagnostic system.7.In accordance with claim 5, wherein the AI engine:uses reinforcement learning techniques to calibrate a specific diagnostic model by employing changes, additions, or modifications to the doctor's diagnoses or consultations as rewards or penalties; constitutes the AI-based chatbot diagnostic system.8.A method of diagnosis for an AI-based chatbot diagnostic system comprising at least one processing unit and a memory coupled to said at least one processing unit, the method comprising:generating an AI engine for creating diagnostic model when executed by at least one processing unit;receiving data within at least one scope from at least one data source related to at least one medical field;storing the received data in a medical data repository, wherein the data is categorized according to at least one criterion;accessing a medical index engine that uses the categorized data and a medical reference database to create and update an index for individual medical fields;utilizing interactive data to train the diagnostic model in conjunction with the medical index;selecting the most suitable trained diagnostic model for a given diagnostic scenario from a model storage database; andmanaging and interacting with the selected diagnostic model using an AI model manager;constitutes the diagnostic method.9.In accordance with Claim 8, the method further comprising:storing user information and medical institution information to provide a conversational service suitable for the patient;an AI engine, including at least one AI diagnostic model, performing questions corresponding to a first model to receive a first plurality of responses from the patient;based on the first plurality of responses and the information in the database, adaptively transitioning the first model to a second model;performing questions corresponding to the second model to receive a second plurality of responses from the patient; andoutputting recommendations based on the first and second plurality of responses; constitutes the diagnostic method.10.In accordance with Claim 9, the method further comprising:determining whether a suitable diagnostic model has been assigned to the patient;if a suitable diagnostic model is assigned to the patient, outputting a diagnostic result based on the suitable diagnostic model; andif a suitable diagnostic model is not assigned, selecting a diagnostic model available to the patient based on the patient's user information, electronic health record information, and the first through second plurality of responses, then outputting diagnostic results based on the selected diagnostic model;constitutes the AI-based chatbot diagnostic method.11.In accordance with Claim 10, the method further comprising:changing the diagnostic model to an expanded diagnostic model upon receiving a request for a doctor's diagnostic participation from the patient; andincorporating any additional consultation from the doctor or revised diagnostic result in the diagnostic outcome for the expanded diagnostic model;using reinforcement learning techniques to calibrate a specific diagnostic model by employing changes, additions, or corrections made by the doctor to the diagnosis or consultation as rewards or penalties;constitutes the AI-based chatbot diagnostic method.12.In accordance with Claim 11, the method comprising:A stage in which a counter-confabulation model evaluates the recommendations or diagnosis that have been outputted;outputting the error when an error is detected by the counter-confabulation model; andproviding feedback on the received error corrections to the model that outputted the recommendations or the diagnoses;constitutes the AI-based chatbot diagnostic method.

Images