Traditional Chinese medicine intelligent diagnosis and treatment generation device based on prompt learning and application thereof

Abstract

The invention discloses a traditional Chinese medicine intelligent diagnosis and treatment generation device based on prompt learning and application thereof, and relates to the technical field of traditional Chinese medicine intelligent diagnosis and treatment, the core of the generation device is a TCM-BART model, the TCM-BART model comprises a BART model and prompt learning based on a thinking chain of treatment based on syndrome differentiation and addition and subtraction along with syndromes, the BART model is a basic framework, a clinical manifestation text is coded, and the coded clinical manifestation text is stored in the BART model. According to the method, feature vectors fully expressing text semantic information are generated, the input feature vectors are processed, traditional Chinese medicine diagnosis and prescription texts most conforming to input text features are predicted, and learning is prompted to take the traditional Chinese medicine syndrome differentiation and treatment thought as a thinking chain instruction to guide a model to generate a better result. According to the traditional Chinese medicine intelligent diagnosis and treatment generation device based on prompt learning and the application thereof, the traditional Chinese medicine diagnosis and prescription recommendation tasks are integrated by using the prompt instruction, so that sharing of model parameters is promoted, the synergistic effect between the tasks is enhanced, and the effectiveness of prescription recommendation is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of intelligent diagnosis and treatment of traditional Chinese medicine, and specifically to a device for generating intelligent diagnosis and treatment of traditional Chinese medicine based on prompt learning and its application. Background Art

[0002] TCM diagnosis focuses on identifying the patient's disharmonious patterns, known as "syndromes" or "signs". This involves a holistic observation of the patient's condition to understand the nature and progression of the disease. TCM practitioners use these insights to select medicinal herbs and formulate personalized prescriptions based on herbal compatibility rules. With the advancement of TCM clinical informatics, TCM-assisted diagnosis and treatment are increasingly valued, aiming to digitize clinical information and provide automated diagnostic and treatment support.

[0003] TCM clinical records contain a wealth of data, including symptom profiles, tongue and pulse diagnosis, TCM diagnosis and prescriptions. These records are not only a repository of TCM knowledge and practical experience, but also a basic resource for studying personalized TCM diagnosis and treatment.

[0004] In the field of TCM-assisted diagnosis, researchers usually transform diagnostic challenges into classification or generation tasks. In terms of classification, various methods are used to map diagnostic data from the four traditional pillars of TCM to corresponding labels. In terms of generation tasks, researchers use text generation techniques to convert diagnostic inputs into corresponding text outputs. For TCM diagnosis and treatment, a common approach is to first pre-train the model on a broad corpus and then fine-tune it to a specific TCM task, such as diagnosis or prescription recommendation. This technique uses extensive pre-training to improve the performance of these tasks. However, handling these interrelated tasks separately may not fully capture the complex knowledge embedded in TCM cases, limiting the effectiveness of the model. Summary of the invention

[0005] The purpose of the present invention is to provide a Chinese medicine intelligent diagnosis and treatment generation device and its application based on the "differentiation of syndromes and treatment, addition and subtraction according to symptoms" thinking chain prompt learning, so as to solve the problem raised by the above background technology that the interrelated tasks in the current market may not fully capture the complex knowledge embedded in Chinese medicine cases, limiting the effectiveness of the model.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a TCM intelligent diagnosis and treatment generation device based on prompt learning, the generation device includes a TCM-BART model, the TCM-BART model is a model formed by fine-tuning the BART model as a basic framework and in combination with the "differentiation of syndromes and treatment, addition and subtraction according to syndromes" thinking chain prompt learning technology, the BART model is used to encode clinical manifestation texts and process input feature vectors to predict TCM diagnosis and prescription texts that best match the input text features; the prompt learning technology modifies the BART model framework by introducing prompt-based fine-tuning to allow multi-tasking;

[0007] The construction process of the TCM-BART model includes the following steps:

[0008] (1) Based on the BART model, we use task-specific prompt instructions and combine them with the Seq2Seq model to integrate TCM diagnosis and prescription recommendation into a unified text generation task;

[0009] (2) Use real diagnoses as prompts to form a thinking chain of "differentiation and treatment, adding or subtracting according to symptoms" to form the TCM-BART model;

[0010] (3) Evaluate the TCM-BART model.

[0011] Preferably, the BART model consists of an encoder and a decoder;

[0012] The encoder encodes the input consisting of the "differentiation and treatment, addition and subtraction according to the symptoms" thinking chain prompt instructions and clinical symptom descriptions to form a semantically consistent representation. The encoder consists of 6-12 transformer encoder layers, which process the text input through embedding and position encoding to preserve the relative position information of the text when processing in parallel in each layer;

[0013] The decoder sequentially processes the relative position information in the encoder at each time step, interprets the text and generates the most appropriate treatment solution. The decoder is based on the output of the encoder and has 6-12 transformer decoder layers combined with a masked self-attention mechanism to generate text and process the output of the encoder.

[0014] Preferably, in step (3), the TCM-BART model evaluation uses a BILSTM-seq2seq model, a CPT model, and a GPT-2 model to compare with the TCM-BART model to evaluate the BLEU, ROUGE1, and ROUGE2 indicators.

[0015] Preferably, in step (1), a BILSTM model is used to form a seq2seq architecture and perform a text generation task.

[0016] Preferably, the CPT model is a Chinese pre-trained Transformer, whose encoder and decoder adopt an unbalanced structure;

[0017] GPT-2 uses the Transformer encoder structure and is based on a large-scale pre-trained text generation model.

[0018] Preferably, in step (1), for a single task, prompt learning is performed by guiding the input of feature information of enriched data to improve the accuracy of the generated results.

[0019] The present invention further discloses the application of the above-mentioned prompt learning-based intelligent Chinese medicine diagnosis and treatment generation device in simulating Chinese medicine-assisted diagnosis and treatment based on natural language processing text generation technology.

[0020] The present invention further discloses the application of the above-mentioned prompt learning-based intelligent Chinese medicine diagnosis and treatment generation device in simulating Chinese medicine-assisted diagnosis and treatment based on natural language processing text generation technology.

[0021] Compared with the prior art, the present invention has the following beneficial effects:

[0022] 1. By integrating TCM diagnosis and prescription recommendation tasks through the use of prompt instructions, thereby promoting the sharing of model parameters, this approach enhances the synergy between tasks and thus improves the effectiveness of prescription recommendations.

[0023] 2. Incorporate the actual diagnosis as a prompt into the prescription task, forming a thinking chain of "differentiation and treatment, addition and subtraction according to symptoms", and use relevant indicators to prove that its method conforms to the principles of TCM differentiation and treatment, and effectively aligns with traditional practice. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] Figure 1 The structure and phase flow chart of the KEGT model of the present invention;

[0025] Figure 2 Generate a loss comparison chart during model training for the first prescription;

[0026] Figure 3 Generate a comparison chart of loss values ​​during model training for diagnosis;

[0027] Figure 4 Generate a loss comparison chart during model training for the second prescription;

[0028] Figure 5 Prescription recommendation graph for multi-task;

[0029] Figure 6 Diagnosis diagram of traditional Chinese medicine under multi-task;

[0030] Figure 7Prescription recommendation graph for a single task;

[0031] Figure 8 This is a TCM diagnosis diagram for a single task;

[0032] Fig. 9 Recommended charts for prescribing based on diagnosis;

[0033] Fig.10 for Figure 1 The enlarged view of position (1) in the figure;

[0034] Fig.11 for Figure 1 The enlarged image of position (2) in the figure;

[0035] Fig.12 for Figure 1 Enlarged view of point (3) in the figure. DETAILED DESCRIPTION

[0036] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0037] The present invention provides the following technical solutions:

[0038] 1. Introduction to TCM diagnosis generation device

[0039] The core and essence of the TCM diagnosis generation device of the present invention is the TCM-BART model:

[0040] First, specific prompt templates are developed for TCM diagnosis and prescription recommendation tasks. Specific prompt templates can help the model effectively identify and classify various tasks.

[0041] Secondly, in the prescription recommendation task, the doctor's diagnosis is used as a template to form a "differentiation and treatment, adding or subtracting according to the symptoms" thinking chain, and templates related to various diagnoses are set according to different symptoms to improve the performance of the model in generating prescriptions;

[0042] Finally, the BART model is used, which is a pre-trained transformer-based text generation system. Its sequence-to-sequence architecture is particularly suitable for diagnosis and treatment tasks in traditional Chinese medicine.

[0043] The core of the TCM-BART model is the BART model and the "differentiation of syndromes and treatment, addition and subtraction according to the syndrome" thinking chain prompt learning technology. The role of the BART model is to encode the clinical manifestation text, generate a feature vector that fully expresses the semantic information of the text, and process the input feature vector to predict the TCM diagnosis and prescription text that best matches the input text characteristics. Prompt learning uses the TCM syndrome differentiation and treatment idea as an instruction to guide the model to generate better results.

[0044] 2. BART Model

[0045] The present invention uses the BART model as the basic framework for TCM-assisted diagnosis and treatment. The model is divided into two main components: an encoder, which encodes the input consisting of prompt instructions and clinical symptom descriptions to form a semantically consistent representation; a decoder, which sequentially processes this information at each time step, interprets the text and generates the most appropriate treatment solution, thereby promoting personalized medicine.

[0046] BART utilizes a pre-training strategy by corrupting the text and then learning to restore it, using a large-scale corpus to enhance its semantic understanding and feature representation capabilities, significantly improving its performance in natural language generation, translation, and understanding, which is beneficial to subsequent tasks.

[0047] The encoder consists of 6-12 transformer encoder layers, similar to the BERT model. It processes text input through embedding and position encoding to preserve the relative position information of the text while processing in parallel in each layer.

[0048] The decoder is based on the output of the encoder and has 6-12 transformer decoder layers. It combines a masked self-attention mechanism to generate text from left to right, process the output of the encoder, and sequentially generate relative position information in each time step. When performing generation tasks, the output of the model is generated sequentially according to the time step. The Decoder processes the feature vector output by the Encoder at each time step and undergoes a linear transformation, then obtains the output probability distribution through Softmax, and finally outputs the corresponding character with the highest probability as our predicted output through the vocab dictionary mapping.

[0049] 3. Prompt Instructions

[0050] For TCM diagnosis and treatment, the model is first pre-trained on a broad corpus and then fine-tuned to a specific TCM task, such as diagnosis or prescription recommendation. This technique leverages the extensive pre-training to improve the performance of these tasks. However, treating these interrelated tasks separately may not fully capture the complex knowledge embedded in TCM cases, which may limit the effectiveness of the model.

[0051] Prompt learning modifies this traditional framework by introducing prompt-based fine-tuning, allowing for multi-task processing. This approach involves creating specific prompt templates for different tasks, thereby guiding the model to recognize and perform each task with higher accuracy. This innovation allows pre-trained models to more effectively adapt to the requirements of different tasks in a specific domain. For a single task, prompt learning also improves the accuracy of generated results by guiding the feature information of the input rich data. TCM diagnosis and prescription recommendation are integrated into a unified text generation task using a shared Seq2Seq model structure. This configuration allows the model to synergize knowledge between diagnosis and prescription, improving the performance and effectiveness of both tasks.

[0052] In addition, using clinical diagnoses as prompts to form a "differentiation and treatment, addition and subtraction according to the symptoms" thinking chain improves the model's understanding of patient symptoms and differential diagnoses, helping to create a more precise treatment plan. This process begins with practitioners determining differential diagnoses based on clinical observations, such as tongue and pulse characteristics, which inform potential pathologies and syndromes. Treatment strategies are then formulated accordingly, enhancing the specificity and applicability of prescriptions. Through prompt learning, the model effectively utilizes the diagnostic principles of traditional Chinese medicine and improves the quality and interpretability of prescription generation.

[0053] IV. Model Evaluation

[0054] In order to better evaluate the superiority of the TCM-BART model in prescription generation tasks, this paper selects the BILSTM (seq2seq) model, CPT model and GPT-2 model for comparison. The relevant models are introduced as follows:

[0055] BILSTM (seq2seq): Use the BILSTM model to form a seq2seq architecture and perform text generation tasks.

[0056] CPT model: A Chinese pre-trained Transformer with an unbalanced structure for the encoder and decoder.

[0057] GPT-2: uses the Transformer encoder structure and is based on a large-scale pre-trained text generation model.

[0058] The experiments in this study used a hardware configuration of CPUi713700KF, GPURTX4090 (24G), an experimental environment of python3.7.4, pytorch1.13.1, transformers4.18.0, and BART-large-Chinese as the BART model. The model fine-tuning optimizer was AdamW, the learning rate was 0.00002, the batch-size was 15, the epoch was 50, and the maximum generation length was 500 to ensure that the results of the model can tend to a stable value after training. At the same time, in order to comprehensively evaluate the efficiency of model generation, this paper selected BLEU, ROUGE, P@5, P@10 and P@15 indicators for evaluation.

[0059] 1) BLEU is currently recognized by the industry as an evaluation index for text generation models. It tends to evaluate the accuracy of text generation. Its essence is to calculate the similarity between the sentence generated by the model and the original sentence. First, count the number of times the n-grams of the two appear at the same time, and take the smaller value as the final number of matches, and then divide it by the total number of n-grams in the text to obtain the n-gram accuracy score. The calculation formula is as follows:

[0060]

[0061] Among them, Count is the number of times the n-gram appears in the generated results. is the maximum number of occurrences of n-grams in the reference document, Candidates refers to the characters in the candidate set, and pn is the precision of the n-gram. On this basis, the arithmetic mean of the logarithm is calculated and the length penalty factor BP is added to obtain its evaluation BLEU value, and its calculation formula is as follows:

[0062]

[0063] W n =1 / n

[0064] Where c is the actual length of the machine-generated document and r is the length of the reference document.

[0065] 2) ROUGE calculates the similarity between the sentence produced by the model and the original sentence, but is more inclined to evaluate the recall rate of the model. The ROUGE1 and ROUGE2 indicators are used to evaluate the generation performance of the model. The calculation formula of ROUGE-N is as follows:

[0066]

[0067] Among them, S is the reference document, ReferencesSentences is the reference document set, and gram n is an n-gram of length n

[0068] 3) In order to evaluate the influence of the front and back positions of the model generation sequence, this paper simultaneously adopts P@5, P@10, P@15, P@20, R@5, R@10, R@15 and R@20 indicators to evaluate the model generation performance, and compares the accuracy of the prescriptions generated by the first 5, 10, 15 and 20 respectively.

[0069] First, the BART model is compared with the untrained BILSTM (seq2seq) model, CPT model, GPT-2 model, etc. using the prescription generation task. The results are shown in Tables 1 and 2:

[0070] Table 1 Comparison of model results

[0071]

[0072] Table 2 Comparison of model results

[0073]

[0074] From the results in the table above, we can see that the TCM-BART model has an 8.92% improvement in BLEU, a 17.77% improvement in ROUGE1, and a 12.35% improvement in ROUGE2 compared to the BILSTM model. It has an improvement of 7.09-26.91 in P@5, P@10, and P@15. This fully proves that the use of the large-scale pre-trained BART model for prescription generation tasks can greatly improve the accuracy of model prediction compared to the BILSTM model, and proves that the BART model can fully understand the hidden semantic relationship in the clinical symptom text, and the complexity of its model can also meet the basic needs of achieving the prescription generation task. At the same time, since the CPT model adopts an unbalanced architecture for text generation, it is not suitable for the long text prescription generation task in this article, so it has a large gap with the BART model in various indicators. Unlike other seq2seq models, the GPT-2 model uses the Transformer decoder to perform one-way text generation tasks, and cannot efficiently understand the semantic relationship of the input text, so its generation length cannot be automatically controlled. Therefore, it has a gap of 19.76 with the BART model in the BLEU index, and a gap of 0.39-4.25 in the P@5, P@10 and P@15 indicators. From the above results, it can be seen that the structure and pre-training method of the BART model are more suitable for the task of generating clinical prescriptions for traditional Chinese medicine.

[0075] By setting corresponding prompts for TCM diagnosis and prescription tasks and sending them into the BART model for training, the model is compared with the TCM prescription task alone to generate prescriptions. The results are shown in Tables 3 and 4.

[0076] Table 3 Comparison of the first prescription generation results

[0077]

[0078] Table 4 Comparison of the first prescription generation results

[0079]

[0080] The results in the table show that for the task of generating TCM prescriptions, after the diagnosis and treatment multi-task learning based on prompt learning, the results of the model prescription generation have an improvement of 1.53% in the BLEU index, 0.98% and 1.52% in the ROUGE1 and ROUGE2 indicators, and 0.59-2.49 in the P@5, P@10 and P@15 indicators, respectively. The loss value results show that there are certain advantages in performing diagnosis and treatment tasks at the same time, which means that TCM diagnosis and prescriptions involve multiple aspects of information such as symptoms, signs, tongue and pulse of patients. Through prompt learning, the model can learn diagnosis and prescriptions at the same time, and realize the comprehensive use of this information, which helps the model better understand the overall situation of the patient and further understand the doctor's diagnosis and treatment preferences, so as to provide more accurate and personalized prescription suggestions.

[0081] The results of comparing the models for diagnosis generation are shown in Table 5.

[0082] Table 5 Comparison of diagnostic generation results

[0083]

[0084] It can be seen from the results in the table that for the TCM diagnosis task, after the prompt-based learning-based diagnosis and treatment multi-task learning, the results of the model diagnosis generation have a 0.94% decrease in the BLEU index, and a 0.41% and 0.87% decrease in the ROUGE1 and ROUGE2 indicators respectively. The loss value of the model is also relatively high when performing diagnosis and treatment tasks. The difference in results may indicate that the model tends to focus more on the prescription generation task in the process of diagnosis and treatment multi-task learning, and the prescription generation task may interfere with the diagnosis generation task to a certain extent, thereby affecting the results of the model diagnosis generation task to a certain extent.

[0085] To further compare the impact of TCM diagnosis on model prescription generation, this paper uses TCM diagnosis as a prompt for the prescription generation task and feeds it into the model for generation. Compared with only clinical symptom text, the results are shown in Tables 6 and 7:

[0086] Table 6 Comparison of the second prescription generation results

[0087]

[0088]

[0089] Table 7 Comparison of the second prescription generation results

[0090]

[0091] It can be seen from the results in the table that for the task of generating TCM prescriptions, after taking the TCM diagnosis text as a prompt, a thinking chain of "differentiation and treatment, addition and subtraction according to symptoms" is formed. The results generated by the model have an improvement of 2.13% in the BLEU index, 1.99% and 2.99% in the ROUGE1 and ROUGE2 indicators, and 0.61-3.08 in the P@5, P@10 and P@15 indicators, respectively. This shows that the model can better integrate the features of the two information sources. This feature fusion may provide a more comprehensive and integrated feature representation, which helps the model generate better prescriptions. Therefore, the increase in the complexity of the input data leads to an increase in the model loss value. At the same time, this also shows that in the task of prescription generation, the model can further improve the generation effect and reliability of the model by combining the experience and diagnosis of professional TCM doctors.

[0092] Example: The symptoms in the test set: hemoptysis after taking medicine, irritability, controlled by hemostatic needles, cough relief, dry mouth, tongue and pulse: yellow and greasy in the middle of the tongue coating, small and slippery pulse; syndrome: heat evil blocking the lungs, this medical case is sent to the model for prescription generation, and its real prescription, single prescription generation, prescription generation under multi-task prompt learning, and prescription generation with diagnosis as prompt and the cross attention value are shown in Table 8 and Table 9 respectively. Figure 5-9 As shown:

[0093] Table 8 Comparison of model generation results

[0094]

[0095]

[0096]

[0097] From the above results, we can see that for the diagnosis task, the real diagnosis and TCM diagnosis under single task are both heat evil obstructing lung syndrome and phlegm and blood stasis obstructing lung syndrome under multi-task are both diseases related to lung function and pathological state, corresponding to symptoms such as hemoptysis and cough. However, the diagnosis results under multi-task do not reflect the pathological factor of heat, indicating that the TCM diagnosis results under multi-task may be interfered by the prescription generation task. At the same time, for the prescription generation task, after sorting the training set prescriptions, the TCM-BART model can accurately predict the top-ranked Chinese medicines, which fully demonstrates that the model can recommend the more basic prescriptions in the prescription based on clinical symptoms. Then, for the Chinese medicine with a lower frequency of occurrence, the prediction effect of the model has a certain attenuation. Overall, it can be judged that the first half of the prescriptions generated by the model have a higher credibility. The prescription generation results under multi-tasks infer one more medicine compared with the single task. At the same time, after taking the diagnosis as a prompt, the model further predicted several Chinese medicines, including Beehive, Solanum nigrum, roasted Bombyx batryticatus, roasted mulberry bark and Panax notoginseng. Among them, Beehive has the effects of dispelling wind, attacking toxins and killing insects, Solanum nigrum has the effects of promoting blood circulation and reducing swelling, relieving fever and analgesia and antibacterial, roasted Bombyx batryticatus can resolve phlegm and disperse nodules, clear away heat and detoxify, and Panax notoginseng has the effects of dispersing blood stasis and stopping bleeding, reducing swelling and relieving pain. The relevant medicines are all related to the syndrome of heat evil blocking the lungs, which fully demonstrates that the model can make full use of syndrome information and thus improve the performance of the model prescription recommendation.

[0098] Although the present invention has been described in detail with reference to the aforementioned embodiments, it is still possible for those skilled in the art to modify the technical solutions described in the aforementioned embodiments, or to make equivalent substitutions for some of the technical features therein. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A TCM intelligent diagnosis and treatment generation device based on prompt learning, characterized in that: The generating device includes a TCM-BART model, which is a model formed by fine-tuning the BART model based on the BART model and the "differentiation and treatment, addition and subtraction according to the symptoms" thinking chain prompt learning technology. The BART model is used to encode the clinical manifestation text and process the input feature vector to predict the TCM diagnosis and prescription text that best meets the input text characteristics; The cue learning technique modifies the BART model framework by introducing cue-based fine-tuning, allowing for multi-tasking; The construction process of the TCM-BART model includes the following steps: (1) Based on the BART model, we use task-specific prompt instructions and combine them with the Seq2Seq model to integrate TCM diagnosis and prescription recommendation into a unified text generation task; (2) Use the real diagnosis as a prompt instruction to form a thinking chain of "differentiation and treatment, adding or subtracting according to the symptoms" to form the TCM-BART model; (3) Evaluate the TCM-BART model.

2. According to claim 1, a device for generating intelligent diagnosis and treatment of traditional Chinese medicine based on prompt learning, characterized in that: The BART model consists of an encoder and a decoder; The encoder encodes the input consisting of the "differentiation and treatment, addition and subtraction according to the symptoms" thinking chain prompt instructions and clinical symptom descriptions to form a semantically consistent representation. The encoder consists of 6-12 transformer encoder layers, which process the text input through embedding and position encoding to preserve the relative position information of the text when processing in parallel in each layer; The decoder sequentially processes the relative position information in the encoder at each time step, interprets the text and generates the most appropriate treatment solution. The decoder is based on the output of the encoder and has 6-12 transformer decoder layers combined with a masked self-attention mechanism to generate text and process the output of the encoder.

3. According to the prompt learning-based TCM intelligent diagnosis and treatment generation device of claim 1, it is characterized by: In step (3), the TCM-BART model evaluation uses the BILSTM-seq2seq model, CPT model, and GPT-2 model to compare with the TCM-BART model and evaluate the BLEU, ROUGE1, and ROUGE2 indicators.

4. According to the prompt learning-based TCM intelligent diagnosis and treatment generation device of claim 1, it is characterized by: In step (1), the BILSTM model is used to form a seq2seq architecture and perform text generation tasks.

5. According to claim 3, a device for generating intelligent diagnosis and treatment of traditional Chinese medicine based on prompt learning is characterized in that: The CPT model is a Chinese pre-trained Transformer, whose encoder and decoder adopt an unbalanced structure; GPT-2 uses the Transformer encoder structure and is based on a large-scale pre-trained text generation model.

6. According to claim 1, a device for generating intelligent diagnosis and treatment of traditional Chinese medicine based on prompt learning is characterized in that: In step (1), for a single task, prompt learning improves the accuracy of the generated results by guiding the feature information of the input enriched data.

7. The application of the TCM intelligent diagnosis and treatment generation device based on prompt learning as described in any one of claims 1 to 6 in simulating TCM-assisted diagnosis and treatment based on natural language processing text generation technology.

Images