Medical support systems, medical support programs, and medical support methods

Abstract

To provide a medical support system that supports accurate and quick entry of medical information into a paper form of a predetermined format.SOLUTION: Provided is a medical support system 100 for automatic entry into a medical form related to medical care, the system comprising a medical form generation module 112 trained through machine learning to extract medical information for the items to be entered into the medical form from text information including medical information related to medical care.SELECTED DRAWING: Figure 4

Description

Technical Field

[0001] The present invention relates to a medical support system, a medical support program, and a medical support method.

Background Art

[0002] When medical workers such as emergency team members, doctors, and nurses check the patient's condition, they may give oral explanations. On the other hand, it is necessary to fill in the patient's condition on a form in a predetermined format.

[0003] For example, an integrated emergency service support system is disclosed that transmits information on the injured and sick received at a fire command center from an emergency server to a terminal used by emergency team members via a communication line (Patent Documents 1, 2).

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0005] By the way, when performing the process of filling in medical information on a form in a predetermined format, it is necessary to accurately and quickly perform the process of extracting the medical information included in the explanation, embedding it in the correct location on the form, and outputting it.

[0006] Particularly, in the case of diseases that require emergency treatment such as stroke and myocardial infarction, it is necessary to perform the process of filling in a form in a predetermined format so that medical information such as the patient's condition verbally explained by medical staff can be transmitted quickly and accurately.

Means for Solving the Problems

[0007] One aspect of the present invention is a medical support system for automatically filling out medical forms related to medical care, characterized in that it comprises a natural language understanding module that has been trained to extract medical information for items to be filled in on the medical form from text information containing medical information related to medical care.

[0008] Another aspect of the present invention is a medical support program for automatically filling out medical forms relating to medical care, characterized in that the computer functions as a natural language understanding module that has been trained to extract medical information for items to be filled out in the medical forms from text information containing medical information relating to medical care.

[0009] Another aspect of the present invention is a medical support method for automatically filling out medical forms relating to medical care, characterized in that a computer, which functions as a natural language understanding module trained to extract medical information for items to be filled in on a medical form from text information containing medical information relating to medical care, automatically extracts medical information for items to be filled in on a medical form from text information containing medical information relating to medical care.

[0010] Here, it is preferable that the natural language understanding module is machine-trained to output the medical information that is the training data when the input data is received, using supervised learning data that combines the medical information to be extracted from the text information for the items to be filled in in the medical form as input data, and the medical information that is the training data as training data as supervised learning data.

[0011] Furthermore, it is preferable that the natural language understanding module is machine-trained to output the medical information that is the training data when the input data is received, using supervised learning data that combines this information, with the medical information to be extracted from the text as training data for the items to be filled in on the medical form.

[0012] Furthermore, it is preferable to include a form embedding module that embeds the medical information output from the natural language understanding module into the medical form and outputs it.

[0013] Furthermore, it is preferable that the form embedding module is machine-trained to output the medical form, which is the training data, when medical information is input, using supervised learning data that combines medical information, with the medical form in which the medical information has been appropriately filled in as training data.

[0014] Furthermore, the natural language understanding module and the form embedding module are integrated, and when text information is input, it is preferable to extract medical information for items to be entered in the medical form from the text information, embed the medical information into the medical form, and output it.

[0015] Furthermore, it is preferable to include a speech-to-text conversion module that generates the aforementioned text information from speech information containing medical information related to medical care.

[0016] Furthermore, the speech-to-text conversion module is preferably configured to anonymize specific sensitive information before outputting the text information.

[0017] Furthermore, it is preferable that the speech-to-text conversion module removes noise from the speech information using a noise reduction algorithm before processing.

[0018] Also, it is preferably used in medical emergencies.

Effects of the Invention

[0019] According to the present invention, it is possible to provide a medical support system, a medical support program, and a medical support method that assist in accurately and quickly embedding medical information in a predetermined medical form and outputting it. Other objects of the embodiments of the present invention will become apparent by referring to the entire specification.

Brief Description of the Drawings

[0020] [Figure 1] It is a diagram showing the configuration of a medical support system in an embodiment of the present invention. [Figure 2] It is a diagram showing the configuration of a server in an embodiment of the present invention. [Figure 3] It is a diagram showing the configuration of a client in an embodiment of the present invention. [Figure 4] It is a functional block diagram showing the configuration of a medical support system in an embodiment of the present invention. [Figure 5] It is a diagram showing a voice text conversion module in an embodiment of the present invention. [Figure 6] It is a diagram showing a voice text conversion module in an embodiment of the present invention. [Figure 7] It is a diagram showing a voice text conversion module in an embodiment of the present invention. [Figure 8] It is a diagram showing a voice text conversion module in an embodiment of the present invention. [Figure 9] It is a diagram showing a medical form generation module in an embodiment of the present invention. [Figure 10] It is a diagram showing an example of text information in an embodiment of the present invention. [Figure 11] It is a diagram showing an example of form information in an embodiment of the present invention. [Figure 12] It is a diagram showing an example of output data in an embodiment of the present invention. [Figure 13] This is a flowchart showing a method for learning a natural language understanding module in an embodiment of the present invention. [Figure 14] This figure shows an example of the configuration of training data for a natural language understanding module in an embodiment of the present invention. [Figure 15] This figure shows an example of the configuration of training data for a natural language understanding module in an embodiment of the present invention. [Figure 16] This figure shows a method for learning a natural language understanding module in an embodiment of the present invention. [Figure 17] This figure shows a method for learning a natural language understanding module in an embodiment of the present invention. [Figure 18] This flowchart shows a method for learning a form-embedded module in an embodiment of the present invention. [Figure 19] This figure shows an example of the configuration of training data for a form embedding module in an embodiment of the present invention. [Figure 20] This figure shows a method for learning a form-embedded module in an embodiment of the present invention. [Figure 21] This flowchart shows the learning method for a medical form generation module in an embodiment of the present invention. [Figure 22] This figure shows a method for improving a medical support system according to an embodiment of the present invention. [Modes for carrying out the invention]

[0021] [System Configuration] The medical support system 100 in the embodiment of the present invention is configured to include a server 102 and a client 104, as shown in Figure 1. The client 104 may be one or multiple. The server 102 and the client 104 are connected to each other via an information and communication network 106 such as the Internet, enabling them to exchange information.

[0022] The information and communication network 106 is not limited to the Internet, but can be any network that enables communication between the server 102 and the client 104. For example, it may be a dedicated line, a public line (telephone line, mobile communication line, etc.), a wired LAN (Local Area Network), a wireless LAN, or a combination of the Internet and these.

[0023] As shown in Figure 2, the server 102 is configured to include a processing unit 10, a storage unit 12, an input unit 14, an output unit 16, and a communication unit 18. The processing unit 10 includes means for performing arithmetic processing such as a CPU. The processing unit 10 executes a medical support server program stored in the storage unit 12 to realize the process of accurately and quickly filling in medical information on a predetermined form in the medical support system 100 in this embodiment. The storage unit 12 includes storage means such as a semiconductor memory or a memory card. The storage unit 12 is connected to the processing unit 10 in an accessible manner and stores the medical support server program and information necessary for its processing. The input unit 14 includes means for inputting information. The input unit 14 includes, for example, a keyboard, touch panel, buttons, etc., for receiving input from an administrator. The output unit 16 includes means for outputting the processing results of the server 102, such as a user interface screen (UI) for receiving input information from an administrator. The output unit 16 includes, for example, a display for presenting images to an administrator. The communication unit 18 is configured to include an interface for communicating information with the client 104 via the information communication network 106. Communications by the communications unit 18 may be wired or wireless.

[0024] Server 102 accepts access from each user's client 104 via the information and communication network 106 and provides each user with the service via the information and communication network 106.

[0025] As shown in Figure 3, client 104 is composed of a processing unit 20, a storage unit 22, an input unit 24, an output unit 26, and a communication unit 28. Client 104 is also referred to as a communication terminal. The processing unit 20 includes means for performing arithmetic processing such as a CPU. The processing unit 20 realizes the function of a client terminal in the medical support system 100 in this embodiment by executing the medical support client program stored in the storage unit 22. The storage unit 22 includes storage means such as semiconductor memory or a memory card. The storage unit 22 is connected to the processing unit 20 in an accessible manner and stores the medical support client program and information necessary for its processing. The input unit 24 includes means for inputting information. The input unit 24 includes, for example, a keyboard, touch panel, buttons, motion sensor, microphone, etc. for receiving input from the user. The output unit 26 includes means for outputting information necessary for processing in client 104, such as a display that shows image information such as a screen for receiving input information from the user or a user interface screen (UI). The communication unit 28 is composed of an interface that communicates information with server 102 via the information communication network 106. Communications by the communications unit 28 may be wired or wireless.

[0026] Client 104 can be any information processing device capable of executing a client program for providing medical support services. For example, Client 104 can be a stationary or portable personal computer (PC), a tablet computer, a smartphone, a mobile phone terminal, a PHS (Personal Handy-phone System) terminal, a personal digital assistant (PDA), or a multi-functional television receiver with information processing capabilities (so-called smart TV).

[0027] Furthermore, in order to ensure secure access from client 104 to server 102, user authentication or other processing may be required when client 104 accesses server 102. For example, the user authentication process could involve having client 104 input a user ID and password via the input unit 24, and authenticating access from client 104 to server 102 if the combination matches one of the user IDs and passwords pre-stored in the storage unit 12 of server 102.

[0028] [Medical support processing] Figure 4 shows a system configuration diagram of the medical support system 100 in this embodiment. As shown in Figure 4, the medical support system 100 is composed of a speech-to-text conversion module 110 and a medical form generation module 112.

[0029] The speech-to-text conversion module 110 is a module that processes speech information obtained from medical professionals, patients, etc., to extract medical text information. The medical form generation module 112 is a module that extracts and outputs medical information corresponding to a predetermined form related to medical information based on the medical text information.

[0030] As shown in Figure 4, the medical form generation module 112 can be configured to include a natural language understanding module 112a and a form embedding module 112b. The natural language understanding module 112a is a module that processes the extraction of medical information necessary to generate a medical form from medical text information. The form embedding module 112b is a module that processes the embedding of the medical information output from the natural language understanding module 112a into a predetermined medical form and outputs the medical form. The natural language understanding module 112a and the form embedding module 112b may be independent modules, or they may be integrated into a single module.

[0031] Medical forms are documents such as medical records used in medical institutions. However, medical forms are not limited to medical records; any form that contains medical information corresponding to specified items regarding the patient's condition and medical treatment is acceptable. Furthermore, medical forms may be electronic data or printed out on paper from the output unit 16 of the server 102 or the output unit 26 of the client 104.

[0032] The following describes the speech-to-text conversion module 110 and the medical form generation module 112, respectively.

[0033] [Speech-to-text conversion module] As shown in Figure 5, the speech-to-text conversion module 110 receives medical information such as the patient's medical status and treatment instructions as voice information from medical professionals involved in healthcare, such as doctors, nurses, clinic staff, and paramedics, and processes this voice information to convert it into text information.

[0034] The speech-to-text conversion module 110 can be fitted with existing speech-to-text conversion algorithms and AI modules.

[0035] For example, a speech-to-text conversion module 110 can be constructed by applying statistical speech recognition technology. Statistical speech recognition technology consists of a feature extractor that extracts features from speech, a decoder, and a speech recognition model (a model integrating an acoustic model, a pronunciation dictionary, and a language model). The feature extractor analyzes a sequence of speech signals (e.g., time-series speech information) and extracts feature vectors of several tens of dimensions, called logarithmic Mel-frequency spectra or Mel-frequency cepstrums. The decoder searches for and outputs the most likely word sequence based on the feature vectors. The speech recognition model is constructed by integrating an acoustic model, a pronunciation dictionary, and a language model. The acoustic model is a model that obtains the most likely phoneme sequence from a sequence of feature vectors. The acoustic model mainly employs a non-deterministic state transition model with a finite number of states called an HMM. The pronunciation dictionary is a probabilistic model that generates words from phoneme sequences. The language model is a model that obtains the most likely word sequence from the words obtained from the pronunciation dictionary. In other words, the language model is a model that defines the linguistic characteristics of any given sequence of symbols (sequence of words).

[0036] Furthermore, for example, the speech-to-text conversion module 110 can be constructed by applying an AI module based on deep learning. While various configurations such as CNN and RNN can be applied to the AI ​​module, it is common to train and utilize a deep neural network (DNN) that, when given a feature vector as input, outputs a vector whose elements represent the probability of belonging to an HMM state.

[0037] Specifically, for example, the speech-to-text conversion module 110 can be constructed by combining voice information from a single medical professional with text information transcribed from that voice information as training data, inputting the voice data into an AI module composed of a neural network, and training the AI ​​module to produce output that matches the training data.

[0038] Here, the speech-to-text conversion module 110 can be configured to convert speech information from a single healthcare professional into text information, as shown in Figure 5. Figure 5 shows an example where speech information related to medical care, spoken by Doctor 1, is input and converted into text information. Alternatively, as shown in Figure 6, the module may be configured to convert speech information in the form of a conversation between multiple healthcare professionals or between a healthcare professional and a patient into text information. Figure 6 shows an example where speech information related to medical care, spoken by Doctor 1, Doctor 2, and Nurse 1, is input and converted into text information in the form of a chat conversation.

[0039] Furthermore, as shown in Figure 7, in order to make medical terminology easier to recognize, the medical language model and the speech recognition model may be integrated and processed together. For example, the speech-to-text conversion module 110 can be constructed using existing methods such as Shallow Fusion, Cold Fusion, Deep Fusion, and Density Ratio Approach.

[0040] Furthermore, by relying on key-value pairs containing medical terms and abbreviations, which are used to train the algorithm incorporated into the medical form generation module 112 described later, free-text entries can be standardized into structured data mapped to predefined formats related to specific pathological conditions. In this case, free-text entries containing only medical terms and abbreviations can be transformed.

[0041] Furthermore, in environments with a high level of noise affecting speech, such as hospitals and clinics, it is preferable to apply a noise reduction algorithm, as shown in Figure 8. The noise reduction algorithm plays an important role in improving the accuracy of processing in the speech-to-text conversion module 110. In particular, in the conversion of medical speech to text, clarity of speech information is essential due to the importance and sensitivity of the content.

[0042] Useful noise reduction algorithms for the speech-to-text conversion module 110 include spectral subtraction, Wiener filtering, beamforming and microphone arrays, noise suppression using deep neural networks (DNNs), recurrent neural networks (RNNs) and long short-term memory (LSTMs), non-negative matrix factorization (NMFs), statistical model-based methods, binaural processing, and end-to-end noise reduction using autoencoders.

[0043] Spectral subtraction is a classical technique that subtracts an estimated noise spectrum from a noisy speech spectrum. It works by estimating the noise spectrum of the non-speech section and subtracting this estimated noise from the entire signal.

[0044] Wiener filtering is a statistically based approach. It minimizes the mean squared error between the estimated clean signal and the actual clean signal. It is particularly effective when the noise characteristics are known.

[0045] Beamforming and microphone arrays are technologies that reduce noise by capturing sound using multiple microphones. The signals from the multiple microphones are combined so that signals from certain directions are amplified and signals from other directions are suppressed. This allows for greater focus on the speaker's voice while minimizing ambient noise.

[0046] Noise suppression using deep neural networks (DNNs) is a method of reducing noise using deep learning models. By applying a DNN model, it can be trained to distinguish between clean speech and noise. Once trained, the model can suppress noise from new speech data. With sufficient training data, DNNs perform exceptionally well even in complex noisy environments.

[0047] Techniques that apply recurrent neural networks (RNNs) and long-term short-term memory (LSTM) have the ability to store past information and can be used for noise reduction. They are particularly effective for time-series information such as audio signals, where context and time-series information are useful for identifying and removing noise.

[0048] Non-negative matrix factorization (NMF) decomposes the spectrogram of a noise signal into a set of ground spectra and their activity levels. By assuming that speech signals and noise signals can be approximated by different sets of ground spectra, they can be separated.

[0049] Statistical model-based methods are techniques for estimating the statistical characteristics of speech and noise. Based on statistical models, noise is suppressed and speech is enhanced. Voice Activity Detection (VAD) is a technique for detecting the presence or absence of human speech. During periods when no speech is detected, the system captures samples of background noise to more effectively suppress noise in the speech segment. Binaural processing is a technique based on human auditory processing, using two or more microphones (mimicking ears) to identify and suppress noise sources. End-to-end denoising with autoencoders is a technique that uses autoencoders trained to directly map noisy speech to clean speech. The encoder compresses the noisy speech into a compact representation, and the decoder reconstructs the clean speech from this representation.

[0050] Furthermore, sensitive information such as personal information may be anonymized from text information representing medical information. Sensitive information can be defined as appropriate. For example, sensitive information can be defined as any type of information that is considered sensitive or capable of identifying an individual. Typically, in medical records, sensitive information includes the patient's name, address, telephone number, social security number, and other unique identifiers.

[0051] One method for anonymizing this sensitive information is data tokenization. Tokenization breaks down text into individual words or tokens, and then performs replacement processing to recognize and anonymize specific data points. Anonymization can also be achieved using Named Entity Recognition (NER). NER is a natural language processing technique that identifies named entities in text and classifies them into predefined categories such as names, organizations, and locations. Because conventional NER models may not capture certain medical terms or nuances, it is preferable to use a pre-trained NER model that has been specifically fine-tuned for medical records.

[0052] Several methods can be applied to anonymize sensitive information after it has been identified. One method of anonymization is to re-edit the text information to completely remove the sensitive information. Another method of anonymization is to use pseudonyms to represent the sensitive information contained in the text information. When using pseudonyms to represent personal names contained in text information, for example, if the word "Tanaka Taro" in the text information is recognized as a person's name, the process is to replace the word "Tanaka Taro" with information indicating that it is a person's name, such as "PERSON_NAME". This method is particularly effective when it is necessary to maintain the data structure of the text information. Yet another method of anonymization is to generalize the sensitive information contained in the text information. For example, this is a technique to replace information in a specific item with something more general, such as replacing a date in the text information with only the year.

[0053] Specifically, regular expressions can be used to detect typical patterns in specific sensitive information such as birth dates and phone numbers contained in text data. Alternatively, a dictionary containing known names, place names, or specific terms used within an organization such as a medical institution can be created, and this dictionary can be used to extract known terms and replace them with other specific terms.

[0054] Furthermore, machine learning and deep learning may also be utilized in the anonymization process. By applying machine learning and deep learning models, text information related to medical information and anonymized information containing specific sensitive information such as birth dates and phone numbers contained in that text information are combined as training data, and machine learning is performed so that the training data is output when the text information is input. In this way, it is possible to construct an AI model for anonymization that outputs anonymized information containing specific sensitive information such as birth dates and phone numbers contained in the text information. By applying machine learning and deep learning, it becomes possible to detect and anonymize inconspicuous text patterns and relationships contained in the text information that may lead to non-anonymization.

[0055] Furthermore, it is preferable that consistency be maintained during the anonymization process. That is, in the process of pseudonymization, for example, if "John Doe" is replaced with "Patient A" in one record of text information, it is preferable that the same replacement be consistent across all records. Ideally, the anonymization process should also be verified by a human-monitored review process. This allows for the identification of overlooked cases and further improvement of the anonymization process. In addition, by continuously updating the model, it becomes possible to perform anonymization appropriately, especially when new data types or structures are introduced into the records.

[0056] [Medical Form Generation Module] As shown in Figure 9, the medical form generation module 112 takes text information containing medical information related to medical care and form information related to a medical form in which said medical information is embedded as input, and outputs a medical form in which the medical information contained in the text information is appropriately embedded.

[0057] The text information input to the medical form generation module 112 is text information containing medical information obtained from healthcare professionals. As shown in Figure 10, the text information can be in the form of chat-style text data from healthcare professionals. Alternatively, as described above, the text information can be text data extracted from audio information by the speech-to-text conversion module 110.

[0058] Furthermore, the form information input to the medical form generation module 112 is information indicating the structure of a medical form into which medical information extracted from text information is embedded. Medical forms can be, for example, medical records, documents that need to be filled out for each disease or symptom, etc. Medical forms can be, for example, diagnostic forms, treatment forms, patient follow-up forms, research data collection forms, etc. Therefore, the form information can be, for example, as shown in Figure 11, information that combines a question ID, the content of the question, and the type of answer to the question (type of information such as numerical or text) for each type of medical form. The medical form generation module 112 extracts the appropriate type of information corresponding to the questions included in the form information from the input text information and outputs it with that information embedded in the appropriate medical form.

[0059] Furthermore, the form information may be selected and entered by users such as healthcare professionals based on the patient's condition, medical procedures, etc. Alternatively, the medical form generation module 112 may automatically select the appropriate form information from a set of pre-configured form information based on the text information entered into the medical form generation module 112.

[0060] The output data of the medical form generation module 112 includes information such as the question ID corresponding to the input form information and the answer to each question. The format of the output medical form depends on the type of question involved. For example, if the question is multiple-choice, single-choice, input number, free text, etc., the output answer will be formatted accordingly. It is preferable that the output data be output in a predetermined format that is easy to process later, such as JSON format.

[0061] The natural language understanding module 112a included in the medical form generation module 112 is configured to extract and output medical information that should be included in the medical form based on the input text information and form information. Figure 12 shows an example of output data output from the natural language understanding module 112a. The natural language understanding module 112a receives text information and form information as input and extracts and outputs medical information from the text information for the medical information items required for the medical form identified in the form information. Figure 12 shows an example in which information necessary for embedding a medical form related to stroke, such as the patient's age "Age", gender "Gender", blood pressure "Blood Pressure", and pulse rate "Heart Rate", has been extracted.

[0062] <Natural Language Understanding Module> The natural language understanding module 112a included in the medical form generation module 112 can be constructed by applying a deep learning-based AI module. The natural language understanding module 112a can be an AI module that applies a large-scale language model (LLM) such as GPT-4 or LLaMA2. A large-scale language model (LLM) is a natural language processing model whose accuracy has been improved by machine learning using a large dataset.

[0063] There are several steps involved in training a language model to automatically generate medical forms based on medical information. The training process for building the natural language understanding module 112a is described below, referring to the flowchart in Figure 13.

[0064] Step S10 involves preparing training data to be used for training the natural language understanding module 112a. This training data requires collecting a large amount of text information representing medical records. Additionally, it is necessary to have a pre-filled medical form available as output for the text information input to the medical form generation module 112.

[0065] For the natural language understanding module 112a, supervised learning data is prepared by combining text information containing medical information and form information that identifies the medical information to be extracted from the text information in order to embed it in a desired medical form, as input data, and medical information that should be output when this text information and form information are input as training data.

[0066] For example, as shown in Figure 14, text information containing medical information and form information indicating the information (items) to be output by the natural language understanding module 112a are used as input data, and the medical information to be output by the natural language understanding module 112a is used as training data. These are combined to form training data. In the example in Figure 14, the form information is instruction information specifying the patient's medical history as an output item. The training data is "diabetes," which is the patient's medical history to be extracted from the text information included in the training data.

[0067] The instruction information can include, for example, information that instructs the output of items indicating the patient's condition, such as the patient's age, sex, medical history, blood pressure, pulse, oxygen saturation, presence or absence of dyspnea, level of consciousness, changes in the electrocardiogram, and blood test results. The instruction information can also include information that instructs the output of items indicating medical procedures to be performed on the patient, such as injections, imaging studies, and thrombolytic therapy, as well as diagnostic and treatment methods and test items. However, the instruction information is not limited to these examples; it should indicate the content of information that should be extracted and output from the text information input to the natural language understanding module 112a.

[0068] Furthermore, as shown in Figure 15, the training data may include form information that specifies multiple items included in a medical form at once. Figure 15 shows an example of form information that includes instruction information specifying that the items necessary to generate a medical form related to stroke, such as the patient's age "Age", gender "Gender", blood pressure "Blood Pressure", and heart rate "Heart Rate", should be output. The training data consists of the items pointed out in the form information and the information that should be extracted from the text information included in the training data corresponding to those items.

[0069] The training data can be divided into a training set, a validation set, and a test set. The training set is used to train the natural language understanding module 112a. The validation set is used to validate the trained natural language understanding module 112a. The test set is used to finally test the validated natural language understanding module 112a. This allows us to evaluate the performance of the natural language understanding module 112a on unknown data.

[0070] Furthermore, sensitive information contained in text information is preferably anonymized as described above in order to ensure privacy.

[0071] In step S12, data preprocessing is performed. The format of the text information input to the natural language understanding module 112a is preferably a consistent machine-readable format. Therefore, in this step, the text information is converted to a machine-readable format as necessary. For example, depending on the complexity of the medical form, the text information is structured as pairs of items and their values, or as a more complex schema.

[0072] Furthermore, medical terms and abbreviations are preferably converted as item-value pairs. For example, text information may contain various different words that refer to the same "myocardial infarction," such as "acute myocardial infarction," "AMI," "shinkin kousoku," and "Heart Attack." Therefore, these words are used as keys to associate with "myocardial infarction" and stored in a medical terminology database. Then, if these words are present in the text information, they are replaced with "myocardial infarction." A similar process of replacing words with common words that have the same meaning can be applied to other medical terms and abbreviations as well.

[0073] In step S14, the learning process for the natural language understanding module 112a is performed. The natural language understanding module 112a is trained using the training data prepared in steps S10 and S12. Specifically, the text information and form information contained in the training data are input to the AI ​​module of the natural language understanding module 112a, and machine learning is performed so that the output data, which is the training data contained in the training data, is output from the natural language understanding module 112a.

[0074] Here, the AI ​​modules constituting the natural language understanding module 112a can be modules containing pre-trained AI models such as GPT-3 or GPT-4. Such AI models can be fine-tuned to constitute the natural language understanding module 112a.

[0075] Figure 16 shows the learning method of the natural language understanding module 112a. By inputting supervised training data into the AI ​​module of the natural language understanding module 112a, output data corresponding to the input text information and form information (including instruction information) is output from the natural language understanding module 112a. This output data is compared with the training data in the supervised training data, and loss evaluation is performed to adjust the evaluation function and weights in the AI ​​module of the natural language understanding module 112a so that the output data matches the training data. A large amount of supervised training data is prepared, and by repeating this learning process on this large amount of supervised training data, the natural language understanding module 112a is constructed to output output data that matches the training data.

[0076] Figure 17 shows another example of how the natural language understanding module 112a is trained. In this example, the natural language understanding module 112a is trained to extract and output multiple items included in a medical form at once. In the example in Figure 17, the supervised training data includes form information that specifies that the module should output items necessary to generate a medical form related to stroke, such as the patient's age "Age", gender "Gender", blood pressure "Blood Pressure", and pulse rate "Heart Rate". By inputting the supervised training data into the AI ​​module of the natural language understanding module 112a, output data corresponding to the input text information and form information (including instruction information) is output from the natural language understanding module 112a. The output data is compared with the training data in the supervised training data, and a loss evaluation is performed to adjust the evaluation function and weights in the AI ​​module of the natural language understanding module 112a so that the output data for each item in the output data matches the data for each item in the training data. A large number of supervised training datasets are prepared, and the learning process is repeated on these datasets to construct a natural language understanding module 112a that outputs output data that matches the training data.

[0077] Furthermore, it is preferable to train the system to prioritize the extraction of items that are considered important in the output data. For example, the items included in the form information can be weighted for importance, and the natural language understanding module 112a can be trained so that the more important an item is, the more accurately medical information related to that item can be extracted.

[0078] Specifically, the loss used for evaluation when training the AI ​​module of the natural language understanding module 112a is weighted for each item. In other words, the higher the importance of an item, the greater the loss is valued when the natural language understanding module 112a produces an output different from the training data. Then, the natural language understanding module 112a is trained to minimize this loss.

[0079] For example, if the weights for the items gender, disease, time, and MRI are 1, 5, 3, and 3 respectively, and the output data for gender, disease, time, and MRI from the natural language understanding module 112a is correct, incorrect, incorrect, and correct relative to the training data, the weighted loss error is calculated as (5+3) / (1+5+3+3)=0.67. In contrast, if no weighting is applied to each item, the loss error is calculated as (1+1) / (1+1+1+1)=0.50. Therefore, by training the natural language understanding module 112a to minimize the weighted loss error, it is possible to train the module to output more accurate data for items with larger weights.

[0080] Furthermore, it is preferable to train the system to output special information such as "not included" if the input text information does not contain the necessary medical information. This is because, in medicine, when information is unknown, the fact that the information is unknown itself contains information.

[0081] In step S16, the natural language understanding module 112a is evaluated. The training data is divided into a training set, a validation set, and a test set. The natural language understanding module 112a, which has been trained using the training set, is evaluated using the validation set and the test set. This allows for evaluation of the AI ​​model performance of the natural language understanding module 112a on unknown data.

[0082] For example, the accuracy of output data entered into medical forms based on transcriptions can be evaluated. For instance, the natural language understanding module 112a can be evaluated based on metrics such as accuracy rate and mean absolute error, or domain-specific accuracy metrics such as perplexity and BLEU.

[0083] As described above, the natural language understanding module 112a included in the medical form generation module 112 can be constructed using machine learning.

[0084] <Form Embedding Module> In the medical form generation module 112, the medical information extracted from text information by the natural language understanding module 112a is embedded into a desired medical form by the form embedding module 112b and output.

[0085] The form embedding module 112b can be built as a machine learning-trained AI module. There are several steps involved in training the form embedding module 112b. The training process for building the form embedding module 112b is explained below, referring to the flowchart in Figure 18.

[0086] In step S20, a process is performed to prepare training data to be used for training the form embedding module 112b. As training data, supervised training data is prepared by combining medical information output from the natural language understanding module 112a and a medical form in which that medical information is appropriately embedded, as training data.

[0087] For example, as shown in Figure 19, medical information output from the natural language understanding module 112a is used as input data, a medical form with the medical information appropriately embedded is used as training data, and supervised learning data is prepared by combining this information.

[0088] The training data can be divided into a training set, a validation set, and a test set. The training set is used to train the form embedding module 112b. The validation set is used to validate the trained form embedding module 112b. The test set is used to finally test the validated form embedding module 112b. This allows us to evaluate the performance of the form embedding module 112b on unknown data.

[0089] In step S22, the form embedding module 112b is trained. The form embedding module 112b is trained using the training data prepared in step S20. Specifically, the input data included in the training data is input to the AI ​​module of the form embedding module 112b, and machine learning is performed so that the medical form, which is the training data included in the training data, is output from the form embedding module 112b.

[0090] Figure 20 shows the learning method of the form embedding module 112b. In the example in Figure 20, medical information related to stroke, such as patient age "Age", gender "Gender", blood pressure "Blood Pressure", and heart rate "Heart Rate", is used as input data in the supervised learning data, and a medical form with this information embedded is used as training data. By inputting the input data into the AI ​​module of the form embedding module 112b, the form embedding module 112b outputs a medical form with the input data embedded as output data. This output data is compared with the training data in the supervised learning data, and a loss evaluation is performed to adjust the evaluation function and weights in the AI ​​module of the form embedding module 112b so that the medical information of each item in the output data matches the medical information of each item in the medical form in the training data. A large number of supervised learning data are prepared, and by repeating this learning process on the large number of supervised learning data, the form embedding module 112b is constructed so that output data that matches the training data is output.

[0091] In step S24, the form embedding module 112b is evaluated. The training data is divided into a training set, a validation set, and a test set. The form embedding module 112b, which has been trained using the training set, is evaluated using the validation set and the test set. This allows for evaluation of the performance of the AI ​​model of the form embedding module 112b on unknown data.

[0092] As described above, the form embedding module 112b included in the medical form generation module 112 can be constructed using machine learning.

[0093] The medical forms actually generated by the form embedding module 112b are stored in the storage unit 12 of the server 102. The generated medical forms are also presented to the user (medical professionals, etc.) by the output unit 16 of the server 102 or the output unit 26 of the client 104, as needed. The generated medical forms may be displayed as electronic data on a display or the like, or printed out on paper and presented to the user.

[0094] Furthermore, the medical information output from the natural language understanding module 112a is output in association with each item of the medical form specified by the form information. Therefore, the form embedding module 112b may generate and output a medical form by embedding the medical information output from the natural language understanding module 112a in association with each item of the medical form.

[0095] Here, the form embedding module 112b may refer to a medical terminology database pre-stored in the storage unit 12 of the server 102 to determine whether the terms included in the medical information extracted by the natural language understanding module 112a match the terms included in the medical terminology database. For example, if the terms included in the medical information extracted by the natural language understanding module 112a do not match the terms included in the medical terminology database, an alert may be issued to the user, such as a medical professional, to warn them. Alternatively, for example, if the terms included in the medical information extracted by the natural language understanding module 112a do not match the terms included in the medical terminology database, the extracted terms may be replaced with terms from the medical terminology database that are expected to be correct, and then embedded and output in the medical form.

[0096] Furthermore, the system may be configured to output an alert (warning) if a specific performance evaluation value for a medical form output from the medical support system 100 falls below a predetermined standard value. For example, each of the above items may be weighted for importance, and a performance evaluation value may be calculated according to the accuracy of the medical information contained in the medical form output from the medical form generation module 112. If the performance evaluation value falls below the standard value, an alert (warning) may be output.

[0097] Specifically, for example, if the weights for the items gender, illness, time, and MRI are 1, 5, 3, and 3 respectively, and the output data for the items gender, illness, time, and MRI from the natural language understanding module 112a is correct, incorrect, incorrect, and correct respectively, the weighted loss error is calculated as (5+3) / (1+5+3+3)=0.67. If this weighted loss error exceeds a predetermined threshold value, an alert (warning) is output. Alternatively, for example, an alert (warning) may be output if medical information for a specific required item is missing in the output medical form.

[0098] Furthermore, the form embedding module 112b may detect potential contradictions or inconsistencies in medical information based on the relationships between the medical information output from the natural language understanding module 112a. For example, if specific medical information should be combined for multiple items, the combination of such specific medical information for those multiple items is pre-stored in the storage unit 12 of the server 102 as a medical information database, and an alert (warning) is output if a combination of medical information other than the specific combination of medical information for those multiple items is output from the natural language understanding module 112a.

[0099] Alternatively, as shown in Figure 21, the medical form generation module 112 may be constructed by combining the natural language understanding module 112a and the form embedding module 112b into a single AI module. In this case, text information and form information are used as input data, and the medical forms that should be output corresponding to this information are used as training data. This information is then combined and used as supervised learning data. The input data included in the learning data is input to the AI ​​module combining the natural language understanding module 112a and the form embedding module 112b, and machine learning is performed so that the medical forms, which are the training data included in the learning data, are output from the AI ​​module.

[0100] Furthermore, when training the natural language understanding module 112a and the form embedding module 112b with machine learning, they can be trained to predict the appropriate type of medical form based on the text information.

[0101] Furthermore, as shown in Figure 22, it is preferable to analyze errors and mistakes in the output data output from the natural language understanding module 112a. This provides insights into the need for improvement, and makes it possible to improve the AI ​​model and training data of the natural language understanding module 112a based on these insights. It is also preferable to continuously evaluate the natural language understanding module 112a when it is actually used. Feedback from users such as doctors and medical professionals is useful for understanding the problems and areas for improvement of the natural language understanding module 112a. It is also preferable to continuously train the AI ​​model of the natural language understanding module 112a (active learning) based on user feedback.

[0102] Alternatively, the medical support system 100 may be provided after pre-training the natural language understanding module 112a and the form embedding module 112b, or it may be possible to perform scratch learning to train and build the natural language understanding module 112a and the form embedding module 112b in-house using public data or the company's own service data.

[0103] Furthermore, medical forms should be categorized based on the type of disease, type of treatment, medical condition, surgical procedure, and other medically relevant factors. By associating metadata and tags with each form type, the medical form generation module 112 can quickly identify and select the appropriate form.

[0104] The medical support system 100 can identify patients upon arrival or login using patient-specific identifiers, biometric authentication, or near-field communication (NFC) technology, and output treatment plans and history for identified patients in appropriate medical forms. It can also synchronize with the electronic health record (EHR) and input patient data into the medical support system 100 in real time. This allows the medical support system 100 to predict and select appropriate medical forms based on recent diagnoses, medications, and test results, and then embed medical information into those forms for output.

[0105] Furthermore, sensors such as those from wearable health devices may be integrated into the medical support system 100. Real-time health data of the patient obtained from the wearable health device may be input to the medical support system 100 as text information, and if the health data indicates a specific condition (for example, an abnormal condition) for the patient, the medical support system 100 may output a medical form with medical information related to that condition embedded within it.

[0106] Furthermore, the system may predict which medical form will be needed next based on past medical forms. For example, if there is a history of recording a patient's condition on a monitoring form each month, the system may input the form information into the medical support system 100 so that the monitoring form is generated and output at the patient's monthly consultation. Alternatively, the system may analyze the patient's medical history and input the form information into the medical support system 100 so that the medical form required for that patient is generated and output. For example, it is preferable to generate and output a medical form with a cardiology-specific format for patients with a history of heart disease.

[0107] Alternatively, geolocation and indoor positioning systems may be integrated into the medical support system 100, and location information indicating the location of patients and medical devices may be input into the medical support system 100. Based on this location information, the medical support system 100 may select the medical forms to be generated and output. For example, if a patient is in the cardiology department of a medical institution (such as a hospital), it is preferable to configure the form information so that medical forms related to cardiology are generated and output from the medical support system 100.

[0108] Furthermore, the medical support system 100 may be configured to select a medical form in an appropriate format depending on the time of day, date, day of the week, week, month, or season. For example, during an influenza epidemic, it is preferable to configure the form information and input it into the medical support system 100 so that medical forms for evaluating influenza are preferentially generated and output.

[0109] Furthermore, by analyzing the behavior and preferences of healthcare professionals, if a healthcare professional frequently uses a particular medical form, that form may be prioritized for selection. For example, if a physician specializing in a particular disease frequently uses medical forms related to that disease, it is preferable to configure the form information and input it into the medical support system 100 so that the medical form is preferentially generated and output.

[0110] Furthermore, external data sources such as disease occurrence databases and regional health alerts may be integrated into the medical support system 100. This allows for the configuration of form information so that, in the event of a sudden outbreak of a specific disease in a region, medical forms related to that disease are preferentially generated and output.

[0111] Furthermore, natural language processing (NLP) may be used to implement chatbots and voice assistants in the medical support system 100. The chatbot or voice assistant may be used to allow patients or healthcare professionals to input their current needs (e.g., "I need a post-operative follow-up form") via voice, and the system may configure the form information to generate and output the specified medical form, which is then input into the medical support system 100.

[0112] Furthermore, by integrating the medical support system 100 in this embodiment with an existing clinical decision support system, the details of the patient's treatment plan can be input to the medical support system 100 as text information. As a result, the medical support system 100 can generate and output a medical form with appropriate information embedded in the treatment plan based on the patient's treatment plan input as text information.

[0113] Furthermore, the medical support system 100 may provide feedback to users, such as medical professionals, regarding whether the medical forms generated and output by the system were correct. Based on this feedback, the medical form generation module 112 may learn to generate and output more accurate medical forms.

[0114] Furthermore, by having the medical form generation module 112 perform machine learning using training data expressed in multiple languages, the medical support system 100 can be made multilingual and applied to the generation of medical forms that support multiple languages.

[0115] As described above, the medical support system 100 in this embodiment can generate text information containing medical information from voice information containing medical information transmitted by medical professionals such as doctors, nurses, and paramedics. Furthermore, based on the text information containing medical information, it can extract medical information for items to be filled in a desired medical form. In addition, the medical information extracted from the text information containing medical information can be automatically embedded into the desired medical form, generating and outputting it as a completed medical form.

[0116] The medical support system 100 can be used in various medical settings. It is particularly effective in creating medical forms in emergency medical situations such as acute stroke and acute myocardial infarction (AMI).

[0117] For example, when a patient is brought to the emergency room with symptoms of acute ischemic stroke, the physician quickly initiates a stroke protocol. Using the medical support system 100, the physician verbally provides information on the patient's symptoms, time of onset, risk factors, and initial physical examination findings. The system extracts this medical information from the voice input, accurately transcribes it into the desired medical form, and automatically fills it in. This expedites processes such as diagnosis, testing, and treatment. Healthcare professionals can then focus on patient treatment, including imaging, testing, and the possibility of thrombolytic therapy or thrombectomy.

[0118] For example, if a patient presents to the emergency room with severe chest pain, the doctor will suspect acute myocardial infarction. The doctor will immediately verbally communicate the patient's medical history, symptoms, changes in the electrocardiogram, and the results of the initial blood test to the medical support system 100. The medical support system 100 will quickly extract the necessary medical information from this voice information and fill in that information on the desired medical form and output it. This will allow for the rapid indication of, for example, the possibility of percutaneous coronary intervention (PCI), saving precious time in situations where every second counts in preserving the patient's myocardium and increasing the chances of survival.

[0119] In this way, the medical support system 100 enables the rapid and accurate generation and output of medical forms during medical emergencies. This minimizes the time healthcare workers spend on administrative tasks, leading to rapid diagnosis, testing, and treatment in time-sensitive situations such as strokes and AMIs, and enabling the efficient provision of medical services.

[0120] [Structure of the invention] [Configuration 1] A medical support system that automatically fills out medical forms related to medical care, A medical support system characterized by comprising a natural language understanding module trained on machine learning to extract medical information for items to be entered in a medical form from text information containing medical information related to medical care. [Configuration 2] The medical support system described in Configuration 1, The aforementioned natural language understanding module is machine-trained to output the medical information that is the training data when the input data is received, using supervised learning data that combines this information, with text information containing medical information related to medical care as input data, and medical information that should be extracted from the text information for items to be filled in on the medical form as training data. [Configuration 3] The medical support system described in Configuration 2, The aforementioned natural language understanding module is a medical support system characterized by taking text information containing medical information related to medical care and form information indicating medical information to be extracted from the text information as input data, and using supervised learning data that combines this information, with the medical information to be extracted from the text for each item to be filled in on the medical form as training data, so that when the input data is entered, the medical information which is the training data is output. [Structure 4] A medical support system described in any one of items 1 to 3, A medical support system characterized by comprising a form embedding module that embeds medical information output from the natural language understanding module into the medical form and outputs it. [Composition 5] The medical support system described in Configuration 4, The aforementioned form embedding module is a medical support system characterized by being machine-trained to output the medical form, which is the training data, when medical information is input, using medical information as input data and the medical form in which the medical information is appropriately filled in as training data, and combining this information to create supervised learning data. [Composition 6] A medical support system as described in configuration 4 or 5, The aforementioned natural language understanding module and the aforementioned form embedding module are integrated, and when text information is input, the medical support system is characterized by extracting medical information for items to be entered in the medical form from the text information, embedding the medical information in the medical form, and outputting the result. [Composition 7] A medical support system described in any one of items 1 to 6, A medical support system characterized by comprising a speech-to-text conversion module that generates the aforementioned text information from speech information containing medical information related to medical care. [Structure 8] The medical support system described in Configuration 7, The aforementioned speech-to-text conversion module is a medical support system characterized by anonymizing specific sensitive information and outputting the text information. [Composition 9] A medical support system as described in configuration 7 or 8, The aforementioned speech-to-text conversion module is a medical support system characterized by processing the speech information by removing noise using a noise reduction algorithm. [Configuration 10] A medical support system described in any one of items 1 to 9, A medical support system characterized by its use in medical emergencies. [Composition 11] A medical support program that automatically fills out medical forms related to medical care, A medical support program characterized by using a computer as a machine learning-trained natural language understanding module to extract medical information for items to be entered in a medical form from text information containing medical information related to medical care. [Composition 12] A medical support method that automatically fills out medical forms related to medical care, A medical support method characterized by having a computer, which functions as a natural language understanding module trained to extract medical information for items to be entered in a medical form from text information containing medical information related to medical care, automatically extract medical information for items to be entered in a medical form from text information containing medical information related to medical care. [Explanation of symbols]

[0121] 10 Processing unit, 12 Storage unit, 14 Input unit, 16 Output unit, 18 Communication unit, 20 Processing unit, 22 Storage unit, 24 Input unit, 26 Output unit, 28 Communication unit, 100 Medical support system, 102 Server, 104 Client, 106 Information and communication network, 110 Speech-to-text conversion module, 112 Medical form generation module, 112a Natural language understanding module, 112b Form embedding module.

Claims

1. A medical support system that automatically fills out medical forms related to medical care, The system includes a natural language understanding module trained on machine learning to extract medical information for items to be filled in on the medical form from text information containing medical information related to medical care, The aforementioned natural language understanding module takes text information containing medical information related to medical care as input data, and medical information to be extracted from the text information for each item to be entered in the medical form as training data. Using supervised learning data that combines this information, the module is machine-trained so that when the input data is received, the medical information that is the training data is output. A medical support system characterized by setting items that are considered important in the medical information output from the natural language understanding module, setting weights for each item, calculating a weighted loss error for each item based on the accuracy of the output data for each item, and performing machine learning to prioritize the extraction of items in order to minimize the weighted loss error.

2. A medical support system according to claim 1, The aforementioned natural language understanding module is a medical support system characterized by being machine-trained to output the medical information that is the training data when the input data is received, using supervised learning data that combines this information, with the medical information that is the training data being used as training data for the items to be filled in on the medical form, with the natural language understanding module taking text information containing medical information related to medical care and form information indicating the medical information to be extracted from the text information as input data, and the medical support system being characterized by being machine-trained to output the medical information that is the training data when the input data is received.

3. A medical support system according to claim 1 or 2, A medical support system characterized by comprising a form embedding module that embeds medical information output from the natural language understanding module into the medical form and outputs it.

4. A medical support system according to claim 3, The aforementioned form embedding module is a medical support system characterized by being machine-trained to output the medical form, which is the training data, when medical information is input, using medical information as input data and the medical form in which the medical information is appropriately filled in as training data, and combining this information to create supervised learning data.

5. A medical support system according to claim 3, The aforementioned natural language understanding module and the aforementioned form embedding module are integrated, and when text information is input, the medical support system is characterized by extracting medical information for items to be entered in the medical form from the text information, embedding the medical information in the medical form, and outputting the result.

6. A medical support system according to claim 1, A medical support system characterized by comprising a speech-to-text conversion module that generates the aforementioned text information from speech information containing medical information related to medical care.

7. A medical support system according to claim 6, The aforementioned speech-to-text conversion module is a medical support system characterized by anonymizing specific sensitive information and outputting the text information.

8. A medical support system according to claim 6 or 7, The aforementioned speech-to-text conversion module is a medical support system characterized by processing the speech information by removing noise using a noise reduction algorithm.

9. A medical support system according to claim 1, A medical support system characterized by its use in medical emergencies.

10. A medical support program that automatically fills out medical forms related to medical care, The computer is made to function as a machine learning-trained natural language understanding module that extracts medical information for items to be filled in on the medical form from text information containing medical information related to medical care. The aforementioned natural language understanding module takes text information containing medical information related to medical care as input data, and medical information to be extracted from the text information for each item to be entered in the medical form as training data. Using supervised learning data that combines this information, the module is machine-trained so that when the input data is received, the medical information that is the training data is output. The system sets important items for the medical information output from the natural language understanding module, assigns weights to each item, calculates a weighted loss error for each item based on the accuracy of the output data for each item, and performs machine learning to prioritize the extraction of items in order to minimize the weighted loss error. A medical support program characterized by the following features.

11. A medical support method that automatically fills out medical forms related to medical care, A computer, which functions as a natural language understanding module trained to extract medical information for items to be entered in the medical form from text information containing medical information related to medical care, is made to automatically extract medical information for items to be entered in the medical form from text information containing medical information related to medical care. The aforementioned natural language understanding module takes text information containing medical information related to medical care as input data, and medical information to be extracted from the text information for each item to be entered in the medical form as training data. Using supervised learning data that combines this information, the module is machine-trained so that when the input data is received, the medical information that is the training data is output. A medical support method characterized by setting items that are considered important in the medical information output from the natural language understanding module, setting weights for each item, calculating a weighted loss error for each item based on the accuracy of the output data for each item, and using machine learning to prioritize the extraction of items in order to minimize the weighted loss error.

Images