Diagnosis and prescription assistance system

Abstract

[Problem] To allow a large-scale language model to be additionally trained with patient records of a large number of diagnosed cases, to assist the additionally trained large-scale language model in proposing an efficient examination plan for establishing a diagnosis from symptom findings obtained from an undiagnosed patient visiting the clinic, and in obtaining an established diagnosis by integrating obtained examination results. [Solution] An electronic medical chart system in which a large-scale language model is used in combination is provided with: a "possible diagnosis name, additional symptom, body finding, and examination answering means" for answering at least one of a list of possible diagnosis names and probabilities thereof, a recommended additional symptom, a body finding, and examination; and a sequential diagnosis improving means for repeating the "symptom, body finding, and examination result data providing means" and the "possible diagnosis name, additional symptom, body finding, and examination answering means" until the list of possible diagnosis names and probabilities thereof has sufficient accuracy as an established diagnosis.

Description

Diagnosis and prescription support system

[0001] The present invention relates to a diagnosis and prescription support system that uses a large-scale language model to create an examination plan based on a patient's symptoms and physical findings, supports the diagnosis of a disease based on the results of the examination, and further supports the creation of an appropriate treatment plan.

[0002] Computer systems (electronic medical records) for efficiently recording medical instructions (orders) and records have become widely used in medical settings. Furthermore, in recent years, there has been remarkable progress in artificial intelligence (AI), with particular attention being paid to large-scale language models (LLMs) or generative AI, which are trained on large volumes of literature and data to enable inquiries and responses in natural language. Prior art documents related to this application include the following:

[0003] Patent Publication No. 2023-523644 Patent Publication No. 2023-73095 Patent Publication No. 5484317 Patent Publication No. 7313757 Patent Publication No. 6792750 Patent Publication No. 7284970 Patent Publication No. 7145367

[0004] When providing medical care, examination plans are created based on the patient's symptoms and examination findings, and treatment plans are then created based on the test results, and treatment is then implemented. Because a wide variety of medical professionals, including doctors and nurses, are involved, the usefulness of electronic medical records for efficient medical activities goes without saying. Patients must be examined, their symptoms and physical findings narrowed down to a certain extent to identify the disease, and tests must be performed to confirm the diagnosis. Treatment plans must then be created based on this disease name. Here, medical orders for tests and treatments are created manually by doctors. This requires an enormous amount of work, resulting in long working hours for doctors. Furthermore, differences in the knowledge and experience levels of each doctor can lead to problems such as misdiagnoses, unnecessary tests, and inconsistent treatment outcomes.

[0005] The present invention has been made to solve these conventional problems, and its purpose is to provide a diagnosis and prescription support system that additionally trains a large-scale language model with patient records of a large number of diagnosed cases (additionally trained large-scale language model), and has the additionally trained large-scale language model propose an efficient test plan for confirming a diagnosis based on the symptoms and findings obtained from undiagnosed patients who visit the hospital, and synthesizes the obtained test results to help obtain a definitive diagnosis, and further proposes an appropriate treatment plan based on the definitive diagnosis.

[0006] As a means for achieving the above object, the diagnosis and prescription support system according to claim 1 comprises: (1) a diagnostic additional training data providing means for providing diagnostic additional training data to an electronic medical record system using a large-scale language model in combination, the electronic medical record system using the diagnosis name, symptoms, physical findings, and test result data of diagnosed cases extracted from the electronic medical record, and using an attention mechanism to perform additional training to obtain, as attention weights for each interrelationship, the interrelationships between individual symptoms, physical findings, test findings, etc. and the disease names of diagnosed cases that are teacher labels, and further the interrelationships between individual symptoms, physical findings, and test findings, and to generate an additionally trained large-scale language model (an additionally trained large-scale language model creating means using diagnosed case data); (2) a "symptom, physical findings, and test result data providing means" for providing the symptom, physical findings, and test result data of the electronic medical record obtained for undiagnosed cases to the additionally trained large-scale language model; (3) The additionally trained large-scale language model is provided with a "possible diagnosis name, additional symptoms, physical findings, and test answering means" that uses attention weights for each of the symptom, physical findings, and test result data groups obtained by the symptom, physical findings, and test result data providing means and a context vector obtained by accumulating them to calculate the probability of each disease name at the intermediate point when the symptoms, findings, test results, etc. are obtained, thereby answering at least one of a list of possible diagnosis names and their probabilities, and recommended additional symptoms, physical findings, and tests; and (4) is characterized by being provided with a sequential diagnosis improvement means that repeats the "symptom, physical findings, and test result data providing means" and the "possible diagnosis name, additional symptoms, physical findings, and test answering means."

[0007] The diagnosis and prescription support system of claim 2 is characterized in that in the diagnosis and prescription support system of claim 1, the "means for providing possible diagnoses, additional symptoms, physical findings, and test responses" includes a recommended test order creation means for creating an order for a recommended test in the electronic medical record.

[0008] The diagnostic and prescription support system of claim 3 is characterized in that in the diagnostic and prescription support system of claim 1 or 2, the diagnostic additional learning providing means includes not only "symptoms, physical findings, and test findings" to advance the diagnosis of diagnosed cases, but also treatment items such as administered medications, injections, rehabilitation, and procedures including surgery.

[0009] The diagnosis and prescription support system of claim 4 is characterized in that, in the diagnosis and prescription support system of claim 1 or 2, the additionally trained large-scale language model is provided with a treatment plan proposal means that proposes a treatment plan for procedures including administered medications, injections, rehabilitation, and surgery for the definitive diagnosis obtained by the sequential diagnosis improvement means.

[0010] The diagnosis and prescription support system of claim 5 is characterized in that the diagnosis and prescription support system of claim 4 is provided with a recommended treatment order creation means for creating individual treatment orders, such as administration of medications, injections, rehabilitation, and procedures including surgery, proposed by the treatment plan proposal means, in an electronic medical record.

[0011] The diagnosis and prescription support system according to claim 6 is the diagnosis and prescription support system according to claim 4, further comprising a progress log creation means for recording a progress log of the successive diagnosis improvement means and the treatment plan proposal means.

[0012] The diagnosis and prescription support system according to claim 1 includes a diagnostic additional training data providing means, which uses the diagnosis names, symptoms, physical findings, and test result data of diagnosed cases extracted from electronic medical records to perform additional training using an attention mechanism to determine the interrelationships between individual symptoms, physical findings, test findings, etc. and the disease names of diagnosed cases, which are teacher labels, as well as the interrelationships between individual symptoms, physical findings, and test findings, as attention weights for each interrelationship, resulting in an additionally trained large-scale language model (additional trained large-scale language model creation means using diagnosed case data). The system also includes a "symptom, physical findings, and test result data providing means," which provides the symptom, physical findings, and test result data of electronic medical records obtained for undiagnosed cases to the additionally trained large-scale language model. The system is provided with a "means for providing possible diagnoses, additional symptoms, physical findings, and test results," and uses the attention weights of each of the symptom, physical findings, and test result data groups obtained by the means for providing symptom, physical findings, and test result data, and the context vector obtained by integrating these, to calculate the probability of each disease name at the intermediate point when the symptoms, findings, test results, etc. are obtained, thereby providing at least one of a list of possible diagnoses and their probabilities, and recommended additional symptoms, physical findings, and tests.The system is provided with a means for sequentially improving diagnosis, and therefore repeats the "means for providing symptom, physical findings, and test result data" and the "means for providing possible diagnoses, additional symptoms, physical findings, and test results" until the list of diagnoses and their probabilities becomes accurate enough to be considered a definitive diagnosis.

[0013] The diagnosis and prescription support system according to claim 2 includes a recommended test order creating means, which creates an order for a recommended test for the electronic medical record.

[0014] The diagnosis and prescription support system described in claim 3 includes not only "symptoms, physical findings, and test findings" to advance the diagnosis of diagnosed cases, but also treatment items such as administered medications, injections, rehabilitation, and procedures including surgery.

[0015] In the diagnosis and prescription support system of claim 4, a treatment plan including medication, injections, rehabilitation, and surgery is proposed for the definitive diagnosis obtained by the sequential diagnosis improvement means.

[0016] In the diagnosis and prescription support system of claim 5, individual treatment orders such as medications, injections, rehabilitation, and procedures including surgery proposed by the treatment plan proposal means are created for the electronic medical record.

[0017] The diagnosis and prescription support system according to claim 6 further comprises a progress log creating means, which records progress logs of the diagnosis improving means and the treatment plan proposing means successively.

[0018] FIG. 1 is a hardware configuration diagram of the present invention. FIG. 2 is an explanatory diagram of performing additional learning on a large-scale language model (LLM) using a large amount of data on diagnosed cases to create an additionally trained large-scale language model. FIG. 3 shows the relationship between the basic learning operation of a large-scale language model and additional learning. FIG. 4 is an explanatory diagram of additional learning in the present invention. FIG. 5 is an explanatory diagram of creating a diagnosis and treatment plan while receiving suggestions by utilizing an additionally trained large-scale language model. FIG. 6 shows an example of the diagnosis and treatment planning process.

[0019] The diagnosis and prescription support system according to the present invention comprises a server device, a database, and a terminal. The server device is a known computer device and comprises an arithmetic unit, a main memory device, an auxiliary memory device, an input device, an output device, and a communication device. The arithmetic unit, the main memory device, the auxiliary memory device, the input device, the output device, and the communication device are connected to one another via a bus interface. The arithmetic unit comprises a known processor capable of executing an instruction set. The main memory device comprises volatile memory such as RAM capable of temporarily storing the instruction set. The auxiliary memory device comprises non-volatile data storage capable of recording an OS and programs. The data storage may be, for example, an HDD or an SSD. The input device is, for example, a keyboard. The output device is, for example, a display such as an LCD. The communication device comprises a network interface connectable to a network. The server device includes means for creating an additionally trained large-scale language model using diagnosed case data, means for providing additional diagnostic training data, means for providing symptom, physical findings, and test result data, means for providing possible diagnoses, additional symptoms, physical findings, and test response data, means for sequentially improving diagnoses, means for creating recommended test orders, means for proposing treatment plans, means for creating recommended treatment orders, and means for creating progress logs. The processor of the server device exerts the functions and effects of these means. The database of the present invention may be configured as an auxiliary storage device of the server device, or may be configured as a separate auxiliary storage device independent of the server device. The database stores information handled by the diagnosis and prescription support system. The terminal of the present invention, like the server device, has a known computer hardware configuration. The server device, database, and terminal of the present invention can communicate via a network.

[0020] FIG. 1 shows an example of a hardware configuration diagram of the present invention. Because large-scale, trained language models require huge amounts of data and computational resources, they are typically stored on the cloud and connected to a local area network (LAN) within a medical institution via an internet connection and a router. An electronic medical record server is located within the medical institution. Doctors, nurses, and other staff access the electronic medical records and large-scale language models via terminals connected to the LAN. Part or all of the electronic medical record system may be built on the cloud. Furthermore, a lightweight version of the system may be configured to install part or all of the large-scale, trained language models within the medical institution (edge ​​computing). This would speed up response times to inquiries and further reduce the risk of information leaks.

[0021] Figure 2 is an explanatory diagram of how a large-scale language model (LLM) is trained using a large amount of diagnosed case data to create a trained large-scale language model. Case data such as symptoms, physical findings, and test results are extracted from electronic medical records (case data extraction means), and "diagnosed case data" is created along with the confirmed diagnosis. This data is then used for training the large-scale language model (trained large-scale language model creation means using diagnosed case data). Large-scale language models can sometimes return fictitious responses known as "hallucinations," which are problematic. "Hallucinations" occur when queries are made for content not included in the trained data. This problem can be prevented by limiting the search scope to the large amount of diagnosed case data used for training. If possible, improved accuracy can be expected if diagnosed case data from other hospitals can be used in addition to the hospital's own. Incidentally, even if a patient presents with the same symptom of "chest pain," a cardiovascular specialist may suspect "angina pectoris" or "myocardial infarction," while an orthopedic specialist may more likely suspect "rib fracture." For this reason, it may be useful to appropriately group diagnosed case data by hospital.

[0022] Large-scale language models are currently being developed rapidly, with numerous models being developed, including ChatGPT (a registered trademark of OpenAI), Bard, LaMDA (a registered trademark of Google), and LLaMA (a registered trademark of Meta). Large-scale language models train large amounts of data to create general-purpose language models (base learning models), but their knowledge of specific fields is not necessarily deep. For this reason, specialized models are constructed by intensively additionally learning knowledge about specific fields (this invention is part of this approach). Furthermore, large-scale language models based on general-purpose base learning models are known to be weak in numerical calculations and logical operations. To compensate for this weakness, there are specialized modules such as Wolfram Alpha (Wolfram Alpha, non-patent document 1), both of which respond to prompt queries from the LLM and return answers.

[0023] Another possible approach is to query a web search engine and retrieve the response. Function calling has been proposed as one method for solving this problem. This involves using the functions of other search engines, databases, or LLMs specialized for specific purposes from a large-scale language model via an API (Application Programming Interface). If dedicated large-scale language models can be used in parallel, response times can be reduced. Furthermore, as described below, if information can be referenced from and written to an electronic medical record system, the large-scale language model can be used to control the electronic medical record system directly.

[0024] Figure 3 illustrates the relationship between the basic learning process and incremental learning for large-scale language models. General-purpose large-scale language models collect as much document data as possible from sources such as the web. Any document elements (e.g., words) contained in the collected document data are masked, and the remaining document elements are used for prediction training. During this training process, the relationships between document elements are identified using a method called the attention mechanism. For document generation tasks, the system predicts the next document element based on highly related document elements and the document elements generated up to that point, and then adopts the document element with the highest probability. This is achieved by calculating a context vector by multiplying the vector representation of each document element by an attention weight for each of the document element sequences generated up to that point. When a document element is added, the system repeats this process of predicting the next document element, making it possible to generate long documents.

[0025] Conventional learning models required large volumes of documents (corpora) with manually assigned teacher labels. The costly labeling process limited the size of the corpus. Large-scale language models mask portions of a document and predict the masked document elements using the remaining portions of the document. Because the masked document elements serve as the correct answer (teacher labels), large corpora can now be used from unsupervised documents without manual intervention. Conventional methods, such as recurrent neural networks (RNNs), only use document elements near the desired document element and "forget" those further away. Therefore, accuracy rapidly declines once the number of words exceeds several dozen, even with careful consideration. LLMs use an attention mechanism to incorporate distant document elements into predictions, making it possible to generate long sentences.

[0026] Figure 4 is an explanatory diagram of the additional learning process used in the present invention. Large-scale language models are trained by collecting a large number of documents from a wide range of fields, so they can easily understand and generate general documents. However, their accuracy in specific fields is not necessarily high. Therefore, additional learning can be performed on documents related to the specific field, and the two can be combined to create an additionally trained large-scale language model suited to the specific field. In the reinforcement learning process of the present invention, supervised additional learning data is used, including the symptoms, physical findings, and test results of cases with a confirmed diagnosis, and the confirmed diagnosis as the teacher label. In this case, information extraction using tag information from electronic medical records, as described in Patent Document 7 (Japanese Patent Publication No. 7145367), would also be useful. Additional learning methods include "transfer learning," which adds only the output layer without changing the parameters of the substrate learning model, and fine-tuning, which retrains the parameters of the substrate learning model using additional learning data. While both are possible, fine-tuning is more preferable in the present invention.

[0027] Learning using the attention mechanism obtains the relationship between individual symptoms, physical findings, and test findings and the teacher labels, or disease names. Furthermore, the relationship between individual symptoms, physical findings, and test findings is also learned. Using the attention weights for each initially obtained symptom and physical finding group and the context vector obtained by integrating them, the probability of each disease name can be calculated at an intermediate stage after several symptoms, findings, and test results have been obtained. That is, while a typical large-scale language model sequentially adds the most probable document element to the end of an already generated document element sequence, this invention uses the attention mechanism to calculate the probability of each disease name at a given time point based on the symptoms, physical findings, test findings, and other information observed in a patient at that time point. This corresponds to the posterior probability of each possible disease name after partial symptoms, physical findings, and test findings have been obtained.

[0028] In Figure 4, for ease of explanation, terms such as "fever" and "abdominal pain" are arranged in order. However, because the attention mechanism calculates individual document element vectors and their attention weights, the order in which individual symptoms, physical findings, and test results appear can be arbitrary. While it is desirable to have data for all items, it is not essential. Even incomplete data contains a certain amount of information and can contribute to output accuracy. Furthermore, variations in notation, for example, "+" in Figure 4, can also be "present" or "positive." Furthermore, LLM specializes in translation, and it is possible to mix multiple languages ​​such as Japanese, English, French, and German.

[0029] Symptoms, physical findings, and test findings are sufficient for simply proceeding with a diagnosis. However, to receive further treatment plan suggestions from the trained large-scale language model, treatment elements such as administered medications, injections, rehabilitation, and surgery must also be included in the additional training data (means for providing additional training for diagnosis and treatment). The attention mechanism has learned the relationship between each treatment element and disease name, making it possible for the trained large-scale language model to suggest each treatment element. While disease names and treatment patterns typically have a simple one-to-one correspondence, using the trained large-scale language model allows the attention mechanism to obtain not only the relationship between disease names and each treatment element, but also the relationship between individual symptoms, physical findings, and test findings, making it possible to suggest treatment plans tailored to the circumstances of each individual case.

[0030] Figure 5 is an explanatory diagram of how a diagnosis and treatment plan is created while receiving suggestions using a reinforcement-learned large-scale language model. For undiagnosed patients, symptoms and physical findings are obtained and entered into the electronic medical record. The symptoms and physical findings are then submitted to the additionally trained large-scale language model. Information extraction from the electronic medical record using tag information, as described in Patent Document 7 (Japanese Patent No. 7145367), may also be useful. The additionally trained large-scale language model suggests possible diagnoses, their probabilities, and additional symptoms, physical findings, and tests that are useful for increasing the accuracy of the diagnosis. The doctor inputs the suggested additional symptoms and physical findings, and then selects the necessary test suggestions from the proposed tests to create a test order for the electronic medical record.

[0031] Once additional symptoms, physical findings, and test results are obtained, they are presented to the additionally trained large-scale language model. The additionally trained large-scale language model again suggests possible diagnoses and their probabilities, as well as additional symptoms, physical findings, and tests that are useful for increasing the accuracy of the diagnosis. If a sufficiently high probability is obtained for the possible diagnoses and their probabilities and the doctor is satisfied with the diagnosis, the diagnosis is finalized and registered in the electronic medical record. If the probability is still insufficient and the doctor is not satisfied, the same procedure is repeated until the diagnosis is finalized (sequential diagnosis improvement method). Once the diagnosis is finalized, the additionally trained large-scale language model suggests appropriate medications, injections, rehabilitation, surgery, and other treatments, and the doctor selects the appropriate treatment and creates a treatment order in the electronic medical record. Here, the diagnosis and individual tests and treatment orders proposed by the additionally trained large-scale language model are merely suggestions, and final adoption is left to the doctor. This is because the doctor bears ultimate responsibility in the event of a medical accident or other problem. If there is a progress log creation means for recording the progress logs of the sequential diagnosis improvement means and treatment plan proposal means, it will be possible to verify on what basis the examination and testing was carried out and the diagnosis was obtained, as well as how the treatment plan was created, thereby providing the "explainability" that was a problem lacking in conventional AI.

[0032] Figure 6 shows an example of the diagnosis and treatment process shown in Figure 5. An undiagnosed patient visits the hospital with fever and abdominal pain. Because numerous diagnoses are possible based on these alone, confirmation of additional symptoms and physical findings, such as coughing and vomiting, is suggested. Tests such as blood tests and X-ray CT scans are also recommended. Test orders are issued for any tests deemed necessary by the physician. The additional symptoms, physical findings, and test results are returned to the additionally trained large-scale language model, which then replies that the diagnosis of cholelithiasis is highly likely. If the physician determines a definitive diagnosis at this point, the diagnosis is registered in the electronic medical record and the additionally trained large-scale language model is requested to suggest treatment for cholelithiasis. If the physician determines that the suggested treatment, such as painkillers or intravenous antibiotics, is appropriate, a corresponding treatment order is issued in the electronic medical record. Finally, a progress log is created.

[0033] When examining a patient, a physician first conducts a medical interview to determine the patient's symptoms, the cause of the symptoms, and the time course of the symptoms. Then, a physical examination is performed to determine a diagnosis and order blood tests and radiological examinations to confirm the diagnosis, or, if multiple diseases are suspected, to differentiate between them. Based on the test results, additional symptoms and physical findings are confirmed and additional tests are ordered, as necessary, until a diagnosis is confirmed. Once a specific disease is sufficiently likely, the disease is registered in the electronic medical record as a confirmed disease and necessary treatment orders are issued. When "determining a diagnosis," physicians often recall cases with similar symptom and physical findings from numerous previous cases. However, this approach makes diagnosis highly personal, and physician experience varies, making it difficult to improve the accuracy rate through objective diagnoses. Furthermore, even the most skilled physician loses their knowledge upon death. By continually accumulating knowledge, as in the present invention, diagnostic accuracy is expected to continue to improve.

[0034] To proceed with diagnosis based on objective criteria, it is necessary to collect a large amount of case data with confirmed diagnoses. While the collection method is optional, practically, it would be discharge summaries or outpatient summaries at the end of outpatient visits. Extracting information using tag information from electronic medical records, as described in Patent Document 7 (Patent Document 7: Patent Publication No. 7145367), would also be useful. In a situation with no prior information, such as when an unknown patient enters the examination room for their first visit, the probability of a disease name is the ratio of the frequency of each disease name (prior probability). When symptom or physical findings are added, the probability of a disease name with those symptoms or physical findings (posterior probability) increases, while the probability of a disease name without those symptoms or findings decreases. When multiple disease names remain as candidates, the correct diagnosis must be determined by integrating additional symptoms, physical findings, and test results (differential diagnosis). Needless to say, symptoms, physical findings, and tests with large differences in positive / negative frequency between diseases are effective as the basis for differential diagnosis. By collecting symptoms and physical findings with large differences in the frequency of positive / negative results between diseases, conducting tests, and providing these results to a large-scale, additionally trained language model, the posterior probability of each disease name can change significantly, bringing us closer to a definitive diagnosis. Taking into account the difficulty and cost of the test, we can then proceed with the next symptoms, physical findings, and tests. This process can be repeated until the probability of the disease name is sufficiently high (a method of successively improving diagnosis).

[0035] Patent Document 5: Japanese Patent No. 6792750 describes a diagnostic support system that collects a large amount of diagnostic information, symptoms, physical findings, and test results (prior probability distribution) and narrows down diagnostic information for undiagnosed patients through set operations on the symptoms, physical findings, and test results (posterior probability). Patent Document 6: Japanese Patent No. 7284970 describes a key-value store, a technology for recording and quickly searching large volumes of records. However, problems include cumbersome query procedures, the need to adjust for variations in notation when collecting large amounts of data, and the enormous computational resources required to record and reference large amounts of data, resulting in considerable costs. Large-scale language models enable queries in everyday language, significantly reducing the complexity of queries. The attention mechanism used in large-scale language models learns the relationships between document elements through inferential learning of masked document elements. Similar relationships are also learned from data provided by a diagnostic reinforcement learning data providing means, creating an additionally trained large-scale language model. Using this mechanism, it is possible to calculate the posterior probability of each disease name after the symptoms, physical findings, and test results of observed undiagnosed patients are obtained. Because a huge language system has already been built, by utilizing this language model, it is possible to record and process large amounts of case data, even for cases with confirmed diagnoses, at a relatively low cost.

[0036] Although the above describes an embodiment, the specific configuration of the present invention is not limited to the above embodiment, and design changes that do not deviate from the gist of the invention are also included in the present invention. For example, while the medical field has been used as an example, similar systems can be used in nursing care. While ChatGPT and other systems have been used as examples of LLM, the same applies to other systems, and even newly developed systems will be included in the present invention. In typical large-scale language models, a query statement called a prompt is often entered and a response is obtained from the large-scale language model. While this method is acceptable in the present invention, prompts and responses can also be executed via an API between the electronic medical record and the large-scale language model.

Claims

1. An electronic medical record system using a large-scale language model, (1) comprising a diagnostic additional training data providing means for performing additional training using diagnosis names, symptoms, physical findings, and test result data of diagnosed cases extracted from electronic medical records to obtain, by an attention mechanism, interrelationships between individual symptoms, physical findings, and test findings and the names of diseases of diagnosed cases, which are teacher labels, and further interrelationships between individual symptoms, physical findings, and test findings, as attention weights for each interrelation, to generate an additional trained large-scale language model (an additional trained large-scale language model creating means using diagnosed case data); (2) comprising a "symptom, physical findings, and test result data providing means" for providing the symptom, physical findings, and test result data of electronic medical records obtained for undiagnosed cases to the additional trained large-scale language model; (3) The additionally trained large-scale language model is provided with a "possible diagnosis name, additional symptoms, physical findings, and test answering means" that uses attention weights for each of the symptom, physical findings, and test result data groups obtained by the symptom, physical findings, and test result data providing means and a context vector obtained by accumulating them to calculate the probability of each disease name at the intermediate point when the symptoms, findings, test results, etc. are obtained, thereby answering a list of possible diagnosis names and their probabilities, and at least one of recommended additional symptoms, physical findings, and tests; (4) A diagnosis and prescription support system characterized by being provided with a sequential diagnosis improvement means that repeats the "symptom, physical findings, and test result data providing means" and the "possible diagnosis name, additional symptoms, physical findings, and test answering means." 2. The diagnosis and prescription support system according to claim 1, further comprising a recommended test order creation means for creating an order for a recommended test in the electronic medical record in the "possible diagnosis, additional symptoms, physical findings, and test response means." 3. A diagnosis and prescription support system as described in claim 1 or 2, characterized in that the diagnostic additional learning providing means includes not only "symptoms, physical findings, and test findings" for advancing the diagnosis of diagnosed cases, but also treatment items such as administered medications, injections, rehabilitation, and procedures including surgery.

4. A diagnosis and prescription support system as described in claim 1 or 2, characterized in that the additionally trained large-scale language model is provided with a treatment plan proposal means for proposing a treatment plan including administered medications, injections, rehabilitation, and surgery for the definitive diagnosis obtained by the sequential diagnosis improvement means.

5. A diagnosis and prescription support system as claimed in claim 4, further comprising a recommended treatment order creation means for creating individual treatment orders, such as administration of drugs, injections, rehabilitation and procedures including surgery, proposed by said treatment plan proposal means, in an electronic medical record.

6. A diagnosis and prescription support system according to claim 4, further comprising progress log creating means for recording a progress log of said successive diagnosis improving means and said treatment plan proposing means.

Images