Method for automatically generating a pre-consultation medical report and associated intelligent medical assistant

Abstract

The invention relates to a method (500) for the automatic generation of a pre-consultation medical report, said method being implemented by computer and comprising: a step (510) of collecting data from a patient (P) via an evolving electronic questionnaire (Q), which offers questions adapted as the answers are provided; a step (520) of structuring the collected data into graphs (G); a step (530) of preprocessing the collected data to eliminate errors and inconsistencies; a step (540) of predictive analysis of the preprocessed data by means of machine learning models (40), making it possible to extract information useful for a medical consultation from the graphs (G); and a step (550) of automatically generating a specialized report (R), in medical terminology, for a healthcare practitioner (D). Figure 1 for the abstract

Description

Title of the invention: Method for automatically generating a pre-consultation medical report and associated intelligent medical assistant technical field

[0001] The present invention belongs to the general field of medical technologies (medtech), particularly medical assistance technologies. It relates more specifically to a method for automatically generating a pre-consultation medical report and an associated intelligent medical assistant. State of the art

[0002] In the healthcare field, various digital solutions have been developed to assist physicians in preparing for consultations and managing patient records. However, these solutions often have significant limitations, particularly in terms of autonomy, predictive analytics capabilities, and effective integration of machine learning technologies.

[0003] Among current solutions, the most common are electronic health record (EHR) management software. These systems, such as EPIC and CERNER (registered trademarks), centralize patients' medical information and facilitate access for healthcare professionals. While effective at organizing data, this software is not designed to perform advanced predictive analysis or to interpret responses to evolving questionnaires.

[0004] Another example of an existing solution is the use of medical assistants based on conversational agents (chatbots) or Clinical Decision Support Systems (CDSS). Tools such as the former IBM Watson Health use artificial intelligence algorithms to suggest diagnoses or treatment plans.

[0005] In 2015, a team from the University of Bordeaux created an animated chatbot capable of asking relevant questions to a patient in order to diagnose excessive daytime sleepiness.

[0006] However, these systems often require significant human supervision and are not entirely autonomous. They also lack the flexibility to adapt questionnaires based on patient responses.

[0007] Furthermore, existing systems are not capable of automatically generating a complete medical report for healthcare professionals. This report generation would require not only a conversion of the natural language used by the Patients use specialized medical jargon, but also a precise contextual interpretation of the information provided. The machine learning models used in these systems are generally limited to specific data analyses and are not configured to interpret and transform patients' natural language into appropriate medical terms for clinical use.

[0008] These limitations show that current solutions do not fully meet the needs of healthcare professionals for optimal consultation preparation based on predictive and intelligent analysis of patient data.

[0009] Document KR20210135406A describes a remote medical consultation system comprising a server and a program. The system generates a medical questionnaire by matching a list of questionnaires stored in a database with information about the patient's symptoms. It then analyzes measurement data received from a patient terminal and enters response data corresponding to the questionnaire items using this measurement data. Finally, the system provides the completed medical questionnaire to the medical staff. A patient uses a terminal, such as a smartphone, to access a dedicated application for booking and receiving treatment. Although booking treatment is not essential, it is recommended to expedite the process.The server receives data regarding the patient's symptoms and measurements, generates a treatment plan based on this information, and automatically completes the medical questionnaire sections of the treatment plan using the received measurement data. Any questionnaire sections that cannot be automatically completed are filled out directly by the patient. The server then provides this medical plan to the medical staff. After these steps, the patient has already completed the necessary formalities before going to the hospital. This document therefore presents a typical telemedicine solution.

[0010] Document CN111326221A describes a patient medical inquiry system comprising medical processing, personal account, login and registration, and storage modules. The login module allows patients to connect to the system and use a registration button to create an account. Personal information is stored in the storage module, and the medical processing module, which is a discussion platform between the patient and the physician, is activated after the personal information has been registered. This system allows patients to communicate with physicians via a network platform to complete the inquiry or consultation. Again, this is a classic telemedicine solution.

[0011] These documents show the limitations of current solutions which focus mainly on telemedicine functionalities without offering an analysis advanced predictive or advanced integration of artificial intelligence for the processing of patient data collected via an evolving questionnaire and the generation of automated pre-consultation reports.

[0012] Furthermore, personal digital assistants (PDAs) are widely adopted by healthcare professionals. The article “Barrett, J., Strayer, S., & Schubart, J. (2004). Assessing medical residents' usage and perceived needs for personal digital assistants. International Journal of Medical Informatics, 73(1), 25–34.” describes how PDAs are used to access medical information and manage patient data. In addition, “Casey, K., & Pedersen, D. (2003). The Clinical Use of Personal Digital Assistants by Physician Assistants and Physician Assistant Students in the Primary Care Setting. The Journal of Physician Assistant Education, 14, 214–219.” reports that these tools are used for medical references, calendars, and specific medical calculations.

[0013] Virtual and intelligent assistants have also been developed. “Richard, A., Mayag, B., Talbot, F., Tsoukiâs, A., & Meinard, Y. (2021). A Virtual assistant dedicated to supporting day-to-day medical consultations. 2021 IEEE 9th International Conference on Healthcare Informatics (ICHI), 330-338.” describes a virtual assistant designed to help physicians during consultations by anticipating necessary information based on existing patient data. On the other hand, “Regli, S., Gashi, F., & Denecke, K. (2021). Digital Medical Interview Assistant AnCha for Obtaining the Medical History in General Medicine: Case study (Preprint).” presents AnCha, an assistant using conversational interfaces to collect patients' medical histories, thereby facilitating interactions and consultation preparation.

[0014] However, there are no fully autonomous solutions capable of collecting medical information from universal and scalable questionnaires. Available tools lack advanced capabilities to fully utilize machine learning (ML) models and predictive analytics for autonomous clinical decisions. Barrett et al. (2004) also raise concerns regarding data security and regulatory compliance, such as HIPAA, which must be addressed for wider adoption of PDAs and other digital devices.

[0015] In conclusion, digital medical assistants already provide significant value by supporting medical consultations, but the state of the art does not yet include advanced autonomous solutions based on universal and scalable questionnaires and intelligent predictive data analysis. Summary of the invention

[0016] The present invention aims to overcome all or part of the drawbacks of the prior art described above, by providing an "intelligent" digital medical assistant for physicians to help them prepare for consultations by processing data provided beforehand by patients. The medical assistant is based on the implementation of a predictive analysis process, using artificial intelligence in the form of machine learning models to explore data provided by patients and structured as graphs, in response to evolving questionnaires based on techniques derived from symbolic artificial intelligence.

[0017] Objectives of the invention are therefore to enable doctors to save time during consultations, and thus improve the quality of consultations, to ensure a reliable pre-diagnosis, to limit diagnostic errors by relying on conclusive massive (big data), to ensure therapeutic education of patients, and finally to offer better health for all.

[0018] To this end, the present invention relates to a method for automatically generating a pre-consultation medical report, said method being implemented by computer and comprising: • a step of collecting patient data via an evolving electronic questionnaire, offering questions adapted as answers are provided; • a step of structuring the collected data into graphs; • a preprocessing step of the collected data to eliminate errors and inconsistencies; • a predictive analysis step of the preprocessed data using machine learning models, allowing the extraction of information useful for a medical consultation from the graphs; and • a step of automatically generating a specialized report, in medical language, for a healthcare practitioner.

[0019] According to one aspect of the invention, the step of collecting data from a patient via the evolving electronic questionnaire includes the use of connected medical devices to collect additional biometric data, such as heart rate, blood pressure and glucose levels, which are then integrated into the graphs.

[0020] According to one aspect of the invention, the data structuring step into graphs uses bipartite semiotic graphs composed of nodes and arcs, where each node represents a specific medical information and each arc represents a conceptual relationship between this information.

[0021] According to one aspect of the invention, the data preprocessing step includes data cleaning techniques, such as value normalization, missing data handling, and duplicate or inconsistency detection.

[0022] According to one aspect of the invention, the predictive analysis step includes the use of machine learning models, such as deep neural networks or random forests, to identify patterns, trends and relevant correlations between the different medical information in the graphs.

[0023] According to one aspect of the invention, the predictive analysis step further includes disambiguating ambiguous patient responses based on the context provided by the graphs.

[0024] According to one aspect of the invention, the automatic generation step of the specialized report uses natural language generation (NLG) techniques to transform the structured data and the results of the analysis into a comprehensible and coherent text in a specialized medical language.

[0025] According to one aspect of the invention, the predictive analysis step of the pre-processed data includes the integration of probabilistic graph models to evaluate and quantify the uncertainties associated with medical predictions, thus providing the practitioner with indications of the reliability of the results.

[0026] The invention also relates to: • a medical assistant in the form of a computer program product downloadable from a communication network and / or stored on a microprocessor-readable medium and executable by a microprocessor, comprising program code instructions for the execution of a procedure as described; and • a terminal-readable and non-transient storage medium, storing a computer program comprising a set of instructions executable by a computer or processor to implement such a process.

[0027] The fundamental concepts of the invention having been set out above in their most elementary form, other details and features will become clearer from the reading of the following description and with regard to the attached drawings, giving by way of non-limiting example an embodiment of a method for automatically generating a pre-consultation medical report, in accordance with the principles of the invention. Presentation of the drawings

[0028] The figures are given for illustrative purposes only to aid in understanding the invention without limiting its scope. The various elements may be represented schematically and are not necessarily to scale. Across all figures, identical or equivalent elements bear the same numerical reference.

[0029] It is thus illustrated in:

[0030] [Fig.l]: an flowchart of the main steps of a process for automatically generating a pre-consultation medical report according to the invention;

[0031] [Fig.2]: an elementary example of structuring patient responses in a graph;

[0032] [Fig.3]: an example of a graph with thousands of nodes and arcs;

[0033] [Fig.4]: a functional diagram of a medical assistant for the implementation of the process;

[0034] [Fig.5]: a computer architecture of the medical assistant. Detailed description of implementation methods

[0035] It should be noted that certain technical elements well known to those skilled in the art are recalled here to avoid any insufficiency or ambiguity in the understanding of the present invention.

[0036] The embodiment described below refers to a method for automatically generating a pre-consultation medical report, primarily intended for physicians. This non-limiting example is given for a better understanding of the invention and does not preclude adapting the method to other related applications.

[0037] In this description, and unless otherwise indicated, the following definitions shall apply: • A "model" refers to an abstract, data-driven concept with parameters, characterized by an output of interest obtained from input data. A typical example of a model is a machine learning model. • A "graph" refers to a data structure composed of nodes and edges, used to represent information and their relationships. Each node represents, for example, a concept, and the edges represent the relationships between these concepts.

[0038] All other technical and scientific terms used have the same meanings as those commonly accepted in the technical field of the invention.

[0039] Also, the terms “collect,” “structure,” “determine,” “process,” “analyze,” “generate,” “form,” and their derived forms, or more broadly, “executable operation” within the meaning of the invention, mean an action performed by a device or processor unless the context indicates otherwise. In this respect, operations refer to actions and / or processes of a data processing system, for example, a cloud computing system or a computing device. Autonomous electronics, which manipulate and transform data represented as physical (electronic) quantities in the memories of the computer system or other devices for storing, transmitting, or displaying information. These operations may be based on applications or software.

[0040] Figure 1 illustrates the main steps of a method 500 for the automatic generation of a pre-consultation medical report, said method comprising: • an initial step 510 of collecting patient data via an evolving electronic self-administered questionnaire Q, offering questions adapted as answers are provided; • a step 520 of structuring the collected data into graphs G; • a 530 preprocessing step of the collected data to eliminate the errors and inconsistencies; • a step 540 of predictive analysis of preprocessed data using machine learning models, allowing the extraction of information useful for a medical consultation from the graphs G; and • a final step 550 of automatic generation of a specialized R report, in medical language, intended for a health practitioner.

[0041] Of course, some of the above steps of the process can be carried out in any order, including in parallel. For example, pretreatment step 530 can be carried out in parallel with patient data collection step 510.

[0042] The 500 process thus makes it possible to efficiently structure complex medical data, to improve predictive analysis through machine learning and to automatically generate clear and relevant reports for doctors.

[0043] The graphs used, in particular semiotic graphs, ensure a structured and evolving representation of information, allowing dynamic adaptation of questionnaires and efficient integration of patient data into an interconnected knowledge base.

[0044] Step 510 of collecting patient data consists of interviewing a patient prior to their medical consultation via an evolving questionnaire.

[0045] This electronic questionnaire is deployed on a user interface accessible via electronic devices such as smartphones, tablets, or computers. The questions are dynamically adapted using decision algorithms, ensuring that the information collected is relevant and specific to each patient. For example, if a patient reports experiencing chest pain, additional questions about the nature and duration of the pain, as well as relevant medical history, will be automatically asked.

[0046] In addition to the electronic questionnaire, data could be collected from connected medical devices (connected bracelets, watches or patches, etc.) or other digital sources such as EMR (Electronic Medical Records).

[0047] In addition, data can be collected directly by questioning the patient himself, or indirectly by questioning a trusted third party (another attending physician, relative, legal representative, etc.).

[0048] During patient data collection, each response is associated with a node in a graph, representing specific information, in order to allow the overall structuring of all data into one or more graphs.

[0049] The collection step 510 can therefore be concomitant with the graph structuring step 520.

[0050] Step 520 of structuring the collected data into graphs consists of preparing the data in a form that facilitates the analyses that will be carried out via machine learning models.

[0051] According to one embodiment, the graphs used are semiotic graphs.

[0052] Indeed, semiotic graphs are bipartite structures composed of Nodes and arcs allow for the structuring and representation of complex relationships between the various medical data collected via the electronic questionnaire. Each patient response is converted into a node representing specific information, such as a symptom, medical history, or lifestyle habit. Arcs, representing conceptual relationships, describe the interactions and dependencies between the different medical information. For example, a node representing "chest pain" can be connected by an arc to a node representing "heart disease."

[0053] The structuring of the patient's responses and data into a graph thus forms a complex semantic network which may naturally contain semantic ambiguities.

[0054] When they occur, semantic ambiguities and polysemy in patient responses can be managed by using an empty type in the conceptual type lattice, allowing the system to retain flexibility until additional information is available.

[0055] Furthermore, the data structuring also includes labeling the responses provided by the patient. These responses are integrated into the graph as assertions and textual annotations, which makes it possible to maintain a structured and evolving knowledge base.

[0056] For example, if the question relates to the patient's first name, this could be represented in the graph by a specific relationship between the question and the answer. Thus, an answer to the question "What is your name?" could be represented by the following triplet:

[0057] [Question: #NPN] - (Answer) -> [Answer: #01] - (Graph) -> [Person: #x] - (01. DESCR) -> [First Name: #x[

[0058] where [Person: #x] is the node representing the patient, and [First name: #x] is the attribute of this node corresponding to the given response.

[0059] Each response from the patient can lead to the addition of new nodes and relationships in the graph, thus enriching the knowledge base. For example, if the patient indicates a specific medical condition, a new node representing that condition is created and connected to the existing nodes by appropriate arcs. This could result in relationships such as:

[0060] [Person: #x] - (A. SUFFERES_FOR) -> [Condition: Chest Pain]

[0061] If the patient answers an additional question regarding the duration of the pain, another node could be added to represent this information, and a relationship could be established between chest pain and its duration, for example:

[0062] [Condition: Chest Pain] - (DURATION) -> [Duration: 2 weeks]

[0063] Figure 2 shows an elementary example of a semiotic graph used to represent chronic kidney disease. This semiotic graph is a data structure composed of nodes and arcs, where each node represents a medical concept and each arc represents a relationship between these concepts.

[0064] In this specific case, the semiotic graph represents the following information: • The central node of the graph is the concept of "Disease", which is annotated as a general entity with a generic character "*". • From this central node, two arcs emanate to represent the specific relationships concerning this disease. • The first arc, labeled “LOC” (location), points to a node representing the concept of “Kidney”. This indicates that the disease is localized in the kidneys. • The second arc, labeled “EVOL” (evolution), points to a node representing the concept of “chronicity”. This indicates that the disease has a chronic nature.

[0065] By enriching the knowledge base, each new piece of information added by the patient allows for the construction of an increasingly detailed and interconnected semiotic graph. This facilitates subsequent analysis by enabling machine learning algorithms to explore complex relationships between the data, which is essential for predictive analysis and the generation of the pre-consultation report.

[0066] Figure 3 illustrates an example of graph G with more than 3000 nodes and more than 2500 relations.

[0067] By using this formalism, the medical assistant can adapt the questions in real time according to previous answers, thus ensuring that the data collected is both relevant and specific to each patient.

[0068] This structuring process therefore makes it possible to create an evolving and interconnected knowledge base, facilitating subsequent analysis.

[0069] Step 530, which preprocesses the collected data to eliminate errors and inconsistencies, includes data cleaning techniques, such as normalizing values, handling missing data, and detecting duplicates or incorrect entries. For example, cleaning scripts can transform text responses like "two weeks" into "14 days," ensuring data consistency. Anomaly detection algorithms identify and correct or remove inconsistencies, such as ages that are inconsistent with other personal data of the patient.

[0070] Predictive analysis step 540 consists of extracting information useful for a medical consultation from the created semiotic graphs. Machine learning models, such as neural networks or random forests, are used to analyze the semiotic graphs. These algorithms explore the graphs to identify patterns, trends, and relevant correlations between the different pieces of information.

[0071] For example, they can detect that certain symptoms frequently co-occur with specific medical conditions, even if these relationships are not immediately apparent in the raw data. The results of the analysis are integrated into semiotic graphs, thus enriching the nodes and arcs with derived information, until the questionnaire is completed.

[0072] Predictive analysis step 540 also includes the use of machine learning models to disambiguate ambiguous patient responses based on the context provided by the semiotic graph.

[0073] In addition, various machine learning models, such as ensemble models (boosting, bagging) or models based on convolutional neural networks (CNNs) for specific data, can be used.

[0074] Step 550 of the automatic generation of a specialist report uses natural language generation (NLG) technologies to transform structured data and analysis results into understandable and coherent text in specialized medical language. The report includes a summary of symptoms, medical history, predictive analysis results, and other contextual information relevant to the consultation. Language processing models natural (NLP) ensure that the report is adapted to the medical context and meets the specific needs of the practitioner.

[0075] Graphs facilitate the translation of the patient's responses into specialized medical terms.

[0076] In addition, the report may include insights on the correlations identified between symptoms and medical history, graphs or tables to illustrate the relationships between the data, thus facilitating clinical decision-making.

[0077] Figure 4 schematically represents an intelligent medical assistant 100 for implementing the method 500 described above. This assistant acts as an intermediary between patient P and physician D. Patient P, either directly or through a connected device W or a third party T, provides data in response to the questionnaire Q, which constitutes the input for the medical assistant 100. Subsequently, the method 500 is implemented according to the steps described above, up to the generation of the report R, which is intended for physician D and constitutes the output of the medical assistant 100.

[0078] The Intelligent Medical Assistant 100 is a computer program product designed to assist physicians in preparing for medical consultations by processing data provided beforehand by patients. Although the medical assistant is software-based, it is structured as a precise computer system, composed of several interconnected functional units, each performing a step or part of a step in the process 500 described above.

[0079] Figure 5 illustrates an example of the general architecture of the medical assistant 100.

[0080] The medical assistant 100 is implemented on a computer device comprising a processor, memory, input devices (such as a keyboard and mouse), and output devices (such as a screen and printer). The device is connected to a network enabling communication with other healthcare systems and medical databases. The medical assistant 100 consists of the following functional units: a data collection unit 10, a data structuring unit 22, a data preprocessing unit 23, a predictive analytics unit 24, and a report generation unit 50.

[0081] Units 22, 23, and 24 can be grouped into a processing module 20 deployed in the cloud. Of course, the assistant 100 can be fully deployed in the cloud.

[0082] The data collection unit 10 is responsible for interacting with the patient via an evolving electronic questionnaire. This unit comprises a user interface, operational via a touchscreen or a web application allowing patients to answer questions, an adaptive questionnaire engine in the form of a software module integrated into the processor that dynamically adapts the questions based on the patient's previous answers, and optionally data collection devices biometric data such as connected sensors (like blood pressure monitors, glucometers) that provide additional biometric data automatically integrated into the system.

[0083] The data structuring unit 22 organizes the collected information into semiotic graphs. It includes a data conversion module in the form of an algorithm that transforms patient responses into graph nodes representing medical concepts (symptoms, history) and creates arcs to illustrate the relationships between these concepts, and a database stored in the memory of the implementing computer device, said database containing standardized medical references (ICD-10, SNOMED CT) to ensure the interoperability and consistency of medical data.

[0084] The data preprocessing unit 23 ensures the quality and consistency of the data before predictive analysis. It includes a data cleaning module in the form of a data processing algorithm that normalizes responses, manages missing data and detects anomalies or inconsistencies in the responses, and anomaly detection algorithms which are programs executed by the processor to automatically identify and mark suspicious data for manual review.

[0085] The predictive analysis unit 24 uses machine learning models to extract useful information from semiotic graphs. It includes a machine learning engine, which is a set of models, such as deep neural networks or random forests, analyzing the semiotic graphs to identify relevant patterns, trends, and correlations, and a results integration module in the form of an algorithm integrating the results of the analysis into the semiotic graphs, thus enriching the nodes and edges with derived information.

[0086] Finally, the report generation unit 50 is responsible for the automatic creation of a medical report in medical language for the physician. It includes a natural language processing (NLP) generation module and a report customization interface in the form of software that allows the physician to personalize the format and content of the report according to the specific needs of the consultation.

[0087] Thus, when a patient uses the intelligent medical assistant, data is first collected via the evolving questionnaire and biometric devices. This data is then structured into semiotic graphs, cleaned and preprocessed, and then analyzed by machine learning models. Finally, a specialized report is generated and presented to the physician, providing a clear and precise summary of the medical information relevant to the consultation.

Claims

Demands

1. A method (500) for automatically generating a pre-consultation medical report, said method being implemented by computer and characterized in that it comprises: a step (510) of collecting data from a patient (P) via an evolving electronic questionnaire (Q), proposing questions adapted as the answers are provided; a step (520) of structuring the collected data into graphs (G); a step (530) of preprocessing the collected data to eliminate errors and inconsistencies; a step (540) of predictive analysis of the preprocessed data by means of machine learning models (40), allowing the extraction of information useful for a medical consultation from the graphs (G); and a step (550) of automatically generating a specialized report (R), in medical language, for a health practitioner (D).

2. A method according to claim 1, wherein the step (510) of collecting data from a patient via the evolving electronic questionnaire (Q) includes the use of connected medical devices to collect additional biometric data, such as heart rate, blood pressure and glucose levels, which are then integrated into the graphs (G).

3. A method according to claim 1 or 2, wherein the step (520) of structuring the data into graphs (G) uses bipartite semiotic graphs composed of nodes and arcs, where each node represents a specific medical information and each arc represents a conceptual relationship between this information.

4. A method according to any one of the preceding claims, wherein the data preprocessing step (530) includes data cleaning techniques, such as value normalization, missing data handling, and duplicate or inconsistency detection.

5. A method according to any one of the preceding claims, wherein the predictive analysis step (540) includes the use of machine learning models, such as deep neural networks or random forests, to identify patterns, trends, and relevant correlations among the different medical information in the graphs (G).

6. A method according to any one of the preceding claims, wherein the predictive analysis step (540) further includes disambiguating ambiguous patient responses (P) based on the context provided by the graphs (G).

7. A method according to any one of the preceding claims, wherein the automatic generation step (550) of the specialized report (R) uses natural language generation (NLG) techniques to transform the structured data and the results of the analysis into a comprehensible and coherent text in a specialized medical language.

8. A method according to any one of the preceding claims, wherein the predictive analysis step (540) of preprocessed data includes the integration of probabilistic graph models to evaluate and quantify the uncertainties associated with medical predictions, thereby providing the practitioner with indications of the reliability of the results.

9. Medical assistant (100) in the form of a computer program product downloadable from a communication network and / or stored on a microprocessor-readable medium and executable by a microprocessor, characterized in that it includes program code instructions for the execution of a method (500) according to any one of the preceding claims.

10. Terminal-readable and non-transient storage medium storing a computer program comprising a set of instructions executable by a computer or processor to implement a method (500) according to any one of claims 1 to 8.

Images