An artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method

Abstract

The present application relates to the technical field of traditional Chinese medicine syndrome type classification, and particularly relates to a traditional Chinese medicine syndrome type classification and probability prediction method based on artificial intelligence, which comprises: normalizing scattered data, natural language text data and image data related to patient syndromes to obtain normalized data, and the collection of each normalized data is a normalized data set; inputting the normalized data set into a trained syndrome type classification model based on Transform to obtain each syndrome probability corresponding to the syndromes; the model sequentially comprises: a patient information multi-modal embedding layer, a first self-attention layer, and a first linear mapping+Softmax function calculation layer; after the normalized data set in the patient information multi-modal embedding layer is input into the first self-attention layer for attention calculation, each syndrome probability is calculated through the function calculation layer; wherein the patient information multi-modal embedding layer is used for storing the normalized data set. The intelligent classification and probability prediction of traditional Chinese medicine syndrome types are realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of traditional Chinese medicine syndrome classification, and in particular to a traditional Chinese medicine syndrome classification and probability prediction method based on artificial intelligence. BACKGROUND

[0002] In the field of traditional Chinese medicine, when diagnosing the symptoms of a patient, a physician usually determines the syndrome type of the patient according to experience and then gives a diagnosis. Purely relying on manual judgment of current symptoms is highly subjective and difficult to quantify, standardize and unify the syndrome type, and the medical skills of physicians are not the same, so it is difficult to make consistent decisions even for the same patient with the same syndrome. Therefore, there is an urgent need for a method that can intelligently classify and predict the probability of traditional Chinese medicine syndromes. SUMMARY

[0003] (I) Technical problems to be solved

[0004] In view of the above-mentioned shortcomings and deficiencies of the prior art, the present application provides a traditional Chinese medicine syndrome classification and probability prediction method based on artificial intelligence, which solves the technical problem that purely relying on manual judgment of current symptoms is highly subjective and difficult to quantify, standardize and unify the syndrome type, and the medical skills of physicians are not the same, so it is difficult to make consistent decisions even for the same patient with the same syndrome when determining the syndrome type.

[0005] (II) Technical solutions

[0006] In order to achieve the above-mentioned purposes, the main technical solutions adopted by the present application include:

[0007] The present application provides a traditional Chinese medicine syndrome classification and probability prediction method based on artificial intelligence, characterized in that it comprises:

[0008] S1, normalizing scattered data, natural language text data and image data related to the patient's syndrome, to obtain normalized data, and the set of each normalized data is a normalized data set;

[0009] S2, inputting the normalized data set into a trained syndrome classification model based on Transformer to obtain the probability of each syndrome corresponding to the syndrome;

[0010] The syndrome classification model based on Transformer comprises, in sequence: a patient information multi-modal embedding layer, a first self-attention layer, and a first linear mapping + Softmax function calculation layer.

[0011] The normalized data set in the patient information multi-modal embedding layer is input into a first self-attention layer for attention calculation, and then the respective syndrome probability is calculated through a linear mapping + Softmax function calculation layer; wherein the patient information multi-modal embedding layer is used to store the normalized data set.

[0012] Preferably, before S1, further comprising:

[0013] S0, training the Transformer-based syndrome classification model by using historical normalized data to obtain a trained Transformer-based syndrome classification model.

[0014] Preferably, comprising:

[0015] The scattered data includes gender, age, examination indicators, and test indicators.

[0016] The natural language text data includes medical history text and syndrome description text.

[0017] The image data includes MRI image pictures, face diagnosis pictures, and tongue diagnosis pictures.

[0018] Preferably, S1 comprises:

[0019] S10, performing vectorization processing on the scattered data, the natural language text data, and the image data related to the syndrome to obtain a vectorized data set.

[0020] S11, normalizing each vectorized data in the vectorized data set by adding type encoding to each vectorized data according to different types to form the normalized data.

[0021] Preferably, S10 comprises:

[0022] The scattered data is linearly mapped to obtain first vectorized data.

[0023] The natural language text data is converted into word embedding data by using word embedding on the natural language text data, the word embedding data is positionally encoded by using a position encoding method based on sine and cosine in the Transformer, and second vectorized data is obtained through a second self-attention layer feature extraction.

[0024] The image data is subjected to convolution calculation through a convolution layer to obtain third vectorized data.

[0025] The first vectorized data, the second vectorized data, and the third vectorized data are combined into a vectorized data set.

[0026] Preferably, after S2, further comprising:

[0027] S3, splicing the current or historical syndrome probability with the information vector level in the current or historical patient information multi-modal embedding layer as a patient information summary, inputting into the trained syndrome evolution prediction model, and obtaining the syndrome probability after or without specific treatment;

[0028] Wherein, after training the syndrome evolution prediction model with the historical syndrome probability and the information in the historical patient information multi-modal embedding layer, the trained syndrome evolution prediction model is obtained;

[0029] The syndrome evolution prediction model comprises in sequence: a patient information summary layer, a third self-attention layer, a full connection layer, and a second linear mapping + Softmax function calculation layer;

[0030] The patient information summary layer is used for storing the current and historical patient information;

[0031] After the information of the patient information summary layer is input into the third self-attention layer after time series encoding, and is processed by the full connection layer, the syndrome probability after or without specific treatment is calculated by the second linear mapping + Softmax function calculation layer;

[0032] The full connection layer comprises a first full connection layer and a second full connection layer; the first full connection layer is used for processing the historical patient information summary, and outputs the syndrome probability after specific treatment;

[0033] The second full connection layer is used for processing the current patient information summary, and outputs the syndrome probability without specific treatment;

[0034] The first full connection layer and the second full connection layer extract the corresponding patient information by setting different extraction parameters.

[0035] Preferably, the time series encoding comprises:

[0036] A time window T is established according to a time window T, and a historical timeline is established;

[0037] It is judged whether the historical patient information exists at the current time point;

[0038] If not, the mask mechanism is used to mask the current time point, and the information is not included in the historical patient information summary;

[0039] If yes, the current time is set to 0, and the last existing data time point is set to -T.

[0040] Preferably, the timing coding adopts a position coding mode based on sine and cosine of the Transformer.

[0041] Preferably, after the S3, further comprising:

[0042] S4, inputting the syndrome type probability after the specific treatment or without the specific treatment into a trained syndrome evolution prediction model to obtain a probability of a syndrome corresponding to the syndrome type;

[0043] Wherein, the syndrome evolution prediction model is trained by using historical syndrome type probabilities after the specific treatment or without the specific treatment to obtain the trained syndrome evolution prediction model;

[0044] The syndrome evolution prediction model comprises a third fully connected layer and at least one third linear mapping + Softmax function calculation layer.

[0045] After the syndrome type probability after the specific treatment or without the specific treatment is processed by the third fully connected layer, the probability of the syndrome corresponding to the syndrome type is calculated by inputting the third linear mapping + Softmax function calculation layer.

[0046] Preferably, comprising:

[0047] When the syndrome type probability after the specific treatment is input, the probability of the syndrome corresponding to the syndrome type after the specific treatment is output;

[0048] When the syndrome type probability without the specific treatment is input, the probability of the syndrome corresponding to the syndrome type without the specific treatment is output.

[0049] (Three) beneficial effects

[0050] The beneficial effects of the present application are: the TCM syndrome type classification and probability prediction method based on artificial intelligence can intelligently classify and predict the probability of TCM syndrome types by inputting the normalized data set into the trained syndrome type classification model based on the Transformer to obtain the probability of each syndrome type corresponding to the syndrome.

[0051] The trained syndrome type classification model based on the Transformer, the trained syndrome evolution prediction model and the trained syndrome evolution prediction model are adopted in the present application. When used, the information related to the patient, such as the syndrome, is input into the trained syndrome type classification model based on the Transformer to output the probability of each syndrome type. After the probability of each syndrome type is input into the trained syndrome evolution prediction model to output the probability of the syndrome type after treatment or without treatment, the probability of the syndrome type after treatment or without treatment is input into the trained syndrome evolution prediction model to output the corresponding syndrome probability. The intelligent and trained model can realize the evolution prediction of the syndrome and the syndrome type, get rid of the traditional Chinese medicine prediction mode of the syndrome and the syndrome type, reduce the subjective prediction error, and increase the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0052] Figure 1 The flowchart of the present application based on the artificial intelligence-based traditional Chinese medicine syndrome type classification and probability prediction method is shown in the figure.

[0053] Figure 2 The structure diagram of the syndrome type classification model of the present application is shown in the figure.

[0054] Figure 3 The structure diagram of the syndrome evolution prediction model of the present application is shown in the figure.

[0055] Figure 4 The structure diagram of the syndrome evolution prediction model of the present application is shown in the figure. DETAILED DESCRIPTION

[0056] In order to better explain the present application, the present application is described in detail in the specific embodiments below, combined with the accompanying drawings.

[0057] The present application based on the artificial intelligence-based traditional Chinese medicine syndrome type classification and probability prediction method proposed in the embodiment of the present application adopts the normalized data set input into the trained syndrome type classification model based on the Transformer to obtain the probability of each syndrome type corresponding to the syndrome. Compared with the prior art, the present application can intelligently classify and predict the probability of the traditional Chinese medicine syndrome type.

[0058] In order to better understand the above technical solution, the exemplary embodiments of the present application will be described in more detail below with reference to the accompanying drawings. Although the exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a clearer, more thorough understanding of the present application and to fully convey the scope of the present application to those skilled in the art.

[0059] Traditional Chinese medicine (TCM) is an important part of the Chinese traditional medical system, and its core theories include'syndrome type' and'syndrome', which are comprehensive descriptions of diseases and patient states. However, TCM diagnosis and treatment usually rely on experienced physicians, which is highly subjective and difficult to quantify and standardize. In recent years, with the development of electronic medical data and artificial intelligence technology, it has become possible to introduce data-driven methods into the field of TCM research.

[0060] Currently, artificial intelligence technology has been widely used in disease diagnosis and treatment decision-making in the field of modern medicine, such as using machine learning models to diagnose pathological conditions from image data and predict treatment outcomes. However, in the field of TCM, especially in the study of syndrome type classification and syndrome evolution, the application of artificial intelligence technology is still in its early stages. Traditional TCM syndrome type classification relies heavily on expert experience and lacks uniform objective standards. The evolution of syndrome types and syndromes has not been fully revealed. The embodiment inputs the normalized data set into the trained syndrome type classification model based on Transformer to obtain the probability of each syndrome type corresponding to the syndrome. Compared with the prior art, it can intelligently classify and predict the probability of TCM syndrome types.

[0061] Embodiment one,

[0062] As shown in Figure 1 , a TCM syndrome type classification and probability prediction method based on artificial intelligence, comprising:

[0063] S1, normalize the scattered data, natural language text data, and image data related to the patient's syndrome to obtain normalized data, and the collection of each normalized data is a normalized data set;

[0064] S2, input the normalized data set into the trained syndrome type classification model based on Transformer to obtain the probability of each syndrome type corresponding to the syndrome;

[0065] The syndrome type classification model based on Transformer comprises, in sequence: a patient information multi-modal embedding layer, a first self-attention layer, and a first linear mapping + Softmax function calculation layer.

[0066] After the normalized data set in the patient information multi-modal embedding layer is input into the first self-attention layer for attention calculation, the linear mapping + Softmax function calculation layer is used to calculate the probability of each syndrome type. The patient information multi-modal embedding layer is used to store the normalized data set.

[0067] In this embodiment, because TCM diagnosis and treatment involve multimodal data such as patient basic information, symptoms and signs, examination and test indicators, and medical images, and this data is heterogeneous and complex, the data related to the patient's syndrome are divided into three categories according to their structure: scattered data, natural language text data, and image data. In actual modeling, other data can be added, such as EEG signals and fNIRS near-infrared spectral signals collected from patients for the disease post-stroke cognitive impairment, which can also be used as input to the syndrome classification model. Users can input the syndrome classification model according to their actual situation; this embodiment does not impose any special limitations. Before inputting data, it is necessary to vectorize the data and then normalize it. Generally, this is done by encoding the vector data into a numerical form so that it can be processed by machine learning algorithms. Machine learning models usually require numerical input; therefore, categorical variables need to be encoded into numbers in some way.

[0068] Before inputting data into the syndrome classification model, it needs to be trained. In practical use, data is input into the trained syndrome classification model to obtain the target data. In this embodiment, the training method and usage process of the syndrome classification model are the same: the normalized dataset in the multimodal embedding layer of patient information is input into the first self-attention layer for attention calculation, and then the probability of each syndrome is calculated through the linear mapping + Softmax function calculation layer.

[0069] like Figure 2 As shown, Figure 2 The patient's basic information includes: gender, age, and historical text; syndrome information such as symptoms and signs includes: syndrome description text; examination and test indicators; medical images and pictures include: MRI images, facial examination images, and tongue examination images. Gender, age, and examination and test indicators are vectorized through linear mapping to obtain the first vectorized data. Historical text and syndrome description text are linearly mapped through word embedding, and then vectorized through positional encoding in a self-attention layer (second self-attention layer) to obtain the second vectorized data. MRI images, facial examination images, and tongue examination images are vectorized after convolutional processing in a convolutional layer to obtain the third vectorized data. Each vectorized data is then type-encoded and input into the patient information multimodal embedding layer. After attention calculation in the first self-attention layer, the probability of each syndrome type is calculated through a linear mapping + Softmax function calculation layer.

[0070] As a preferred embodiment, before S1, it further includes:

[0071] S0, training the Transformer-based syndrome classification model using historical normalized data, to obtain a trained Transformer-based syndrome classification model.

[0072] In this embodiment, the syndrome classification model must be trained before it is actually used. Training is an essential process that gives the intelligent syndrome classification model the ability to understand and handle specific types of problems. An untrained syndrome classification model is like a blank sheet of paper, with no prior knowledge, and naturally cannot effectively complete the predetermined task. The training method of this embodiment is the same as the aforementioned syndrome classification model usage method. The input during training of the syndrome classification model is the historical normalized data accumulated through experience. The historical normalized data set in the patient information multi-modal embedding layer is input into the first self-attention layer for attention calculation, and then the linear mapping + Softmax function calculation layer is used to calculate the syndrome probability. At this time, one training of the syndrome classification model is completed.

[0073] Since the syndrome classification model is trained in advance before it is used, the accuracy when the syndrome classification model is actually used is improved.

[0074] As a preferred embodiment, it comprises:

[0075] The scattered data includes: gender, age, examination index, test index;

[0076] The natural language text data includes: medical history text, syndrome description text;

[0077] The image data includes: MRI image, face diagnosis picture, tongue diagnosis picture.

[0078] This embodiment classifies different types of data in order to subsequently perform vector processing according to the data type.

[0079] As a preferred embodiment, S1 comprises:

[0080] S10, vectorizing the scattered data, natural language text data, and image data related to the syndrome to obtain a vectorized data set;

[0081] S11, normalizing each vectorized data in the vectorized data set by adding type encoding to each vectorized data according to the different types, to form normalized data.

[0082] In this embodiment, different types of data need to be vectorized, and the process of converting non-numeric data (such as text, images, audio, etc.) into numeric vectors is necessary for the training and use of the syndrome classification model based on Transformer in this embodiment. The main purpose of vectorization is to enable the computer to effectively process and analyze these unstructured data, and also to enable the data to be input into various machine learning algorithms for training and prediction. At the same time, since the self-attention layer of Transformer cannot distinguish different data categories, we need to add type encoding to all types of data vectors. Type encoding is similar to position encoding, for example, gender type is regarded as the first position, and syndrome description is regarded as the fourth position, then the position encoding can be transformed into type encoding. We can use Transformer-based sine and cosine position encoding or other position encoding methods for transformation, but cannot have a large correlation with the position encoding function, otherwise it will pollute each other. After adding type encoding, all data is normalized to obtain the patient information multi-modal embedding at the current time.

[0083] According to the actual computer processing data requirements, different types of data are processed according to the type of data and the requirements of use, realizing the unification of data.

[0084] As a preferred embodiment, S10 comprises:

[0085] The scattered data is linearly mapped to obtain first vectorized data;

[0086] The natural language text data is converted into word embedding data by using word embedding, the word embedding data is position encoded by using the sine and cosine position encoding method in the Transformer, and the second self-attention layer feature extraction is performed to obtain the second vectorized data;

[0087] The image data is subjected to convolution calculation through the convolution layer to obtain third vectorized data;

[0088] The first vectorized data, the second vectorized data, and the third vectorized data are combined to form a vectorized data set.

[0089] In this embodiment, for scattered data values, linear mapping is used for embedding, and the parameters of linear mapping are learned during the training of the syndrome type classification model. For natural language text data, first, the text is converted into vector data using pre-trained word vectors, open-source embedded syndrome type classification models or other models; then position encoding is added to the vector data, which can use the position encoding method based on sine and cosine of Transformer or other position encoding methods; then pass through a self-attention layer of Transformer, and the parameters of the self-attention layer are learned during the training of the syndrome type classification model. For image data, after image preprocessing, feature extraction is performed through a CNN convolutional layer, and the parameters of the convolutional layer are learned during the training of the syndrome type classification model. For different image data, different convolutional layers can be set, and the hyperparameters of the convolutional layers can also be different, and the hyperparameters are set according to the best experimental results.

[0090] Different vectorization methods are adopted according to different data, realizing the uniformity on the data acquisition side.

[0091] As a preferred embodiment, after S2, it further includes:

[0092] S3, the current or historical syndrome probability and the information vector level splicing of the current or historical patient information multi-modal embedding layer as patient information summary are input into the trained syndrome evolution prediction model to obtain the syndrome probability after or without specific treatment;

[0093] Wherein, the trained syndrome evolution prediction model is obtained by training the syndrome evolution prediction model with the historical syndrome probability and the information in the historical patient information multi-modal embedding layer;

[0094] The syndrome evolution prediction model sequentially includes: a patient information summary layer, a third self-attention layer, a full connection layer, a second linear mapping + Softmax function calculation layer;

[0095] The patient information summary layer is used to store the current and historical patient information;

[0096] The information of the patient information summary layer is input into the third self-attention layer after time encoding, and then processed by the full connection layer, and then input into the second linear mapping + Softmax function calculation layer to obtain the syndrome probability after or without specific treatment;

[0097] The full connection layer includes: a first full connection layer and a second full connection layer; the first full connection layer is used to process the historical patient information summary to output the syndrome probability after specific treatment;

[0098] The second full connection layer is used to process the current patient information summary to output the syndrome probability without specific treatment;

[0099] The first fully connected layer and the second fully connected layer extract the corresponding patient information by setting different extraction parameters.

[0100] In this embodiment, as Figure 3 As shown, after obtaining the syndrome type probability, the historical / current patient information is aggregated, time-encoded, and then input into the trained syndrome type evolution prediction model to obtain the syndrome type probability of having or not having undergone treatment. It should be noted that the syndrome type evolution prediction model in this embodiment also needs to be trained before use, and the training and usage processes are the same, involving aggregating historical patient information and inputting it into... Figure 3 The probability of syndrome evolution prediction model with or without treatment is obtained to complete one model training. In actual use, after inputting the current and historical patient information, the patient information summary with time-series coding is input into an n1-layer self-attention layer and two n2-layer fully connected layers. The parameters of the self-attention layer and fully connected layers are learned during the training of the syndrome evolution prediction model, and the hyperparameters n1 and n2 are set according to the best experimental results. The two fully connected layers correspond to the probability prediction of syndrome with and without specific treatment, respectively. After the data passes through the two fully connected layers, it is calculated by linear mapping + Softmax to obtain the probability of each syndrome with and without specific treatment after time T. The parameters of the linear mapping are learned during the training of the syndrome classification model, and the syndrome type is set during training, the same as the syndrome classification type setting, and the sample category is labeled by experts during training.

[0101] As a preferred embodiment, timing coding includes:

[0102] A historical timeline is established based on a time window T, with a time period length of T.

[0103] Determine if there is any historical patient information at the current point in time;

[0104] If it does not exist, the current time point will be masked using a masking mechanism, and the information here will not be included in the historical patient information summary.

[0105] If it exists, set the current time to 0 and the time of the last existing data to -T.

[0106] In this embodiment, a historical timeline is established according to a time window T, the current time is 0, the last time point is -T, the time point before the last time point is -2T, and so on. Then, the historical information of the same patient is summarized and classified into the nearest time point of the historical information. If there is no historical patient information summary at a certain time point, the mask mechanism in the Transformer is used to mask it. For example, if the patient has historical patient information summary before March and before September, the historical patient information summary before March and before September is taken as the input at -T and -3T, and the -2T at June is masked.

[0107] The time information is converted into a format that can be understood by the syndrome evolution prediction model by time series encoding of the historical data. This is very important in prediction analysis, pattern recognition and other tasks that need to consider time factors.

[0108] As a preferred embodiment, the time series encoding adopts the position encoding method based on sine and cosine of the Transformer.

[0109] In this embodiment, after the current or historical syndrome probability and the information vector in the current or historical patient information multi-modal embedding layer are horizontally spliced, the input vector needs to be time series encoded. The time series encoding is also a position encoding, which can adopt the position encoding method based on sine and cosine of the Transformer or other position encoding methods. Then, the patient information summary with added time series encoding is input into an n1 layer self-attention layer and two n2 layer fully connected layers.

[0110] As a preferred embodiment, after S3, it further includes:

[0111] S4, inputting the syndrome probability after or without specific treatment into the trained syndrome evolution prediction model to obtain the probability of the syndrome corresponding to the syndrome type;

[0112] The historical syndrome probability after or without specific treatment is used to train the syndrome evolution prediction model to obtain the trained syndrome evolution prediction model;

[0113] The syndrome evolution prediction model includes a third fully connected layer and at least one third linear mapping + Softmax function calculation layer.

[0114] The syndrome probability after or without specific treatment is processed by the third fully connected layer and then input into the third linear mapping + Softmax function calculation layer to calculate the probability of the syndrome corresponding to the syndrome type.

[0115] By setting different linear mapping + Softmax function calculation layers, the syndrome probability after or without specific treatment is distinguished, so that the processing result is clearer.

[0116] In this embodiment, asFigure 4 As shown, the syndrome evolution prediction model of the embodiment also needs to be trained. First, the historical syndrome probability is input into the syndrome evolution prediction model Figure 4 , and the corresponding syndrome probability is output. At this time, the training of the syndrome evolution prediction model is completed. In actual use of the syndrome evolution prediction model, the syndrome type evolution prediction output is first input as an input. After passing through the full connection layer and multiple linear mappings + Softmax, the occurrence probability of each syndrome is obtained. The parameters of the third full connection layer and linear mapping are learned during the training of the syndrome evolution prediction model. The syndrome types are set during training, for example, hundreds of syndromes such as taciturnity, memory decline, red tongue, little moss, and slippery pulse can be set. The syndromes are obtained from the patient syndrome information, and are labeled by experts after the patient syndrome type is obtained during training. The syndrome type corresponds to the syndrome type. When predicting, the syndrome type probability after a specific treatment is input, and the probability of each syndrome after the specific treatment is output. When inputting the syndrome type probability without a specific treatment, the probability of each syndrome without a specific treatment is output.

[0117] As a preferred embodiment, it includes:

[0118] When the syndrome type probability after a specific treatment is input, the probability of the syndrome corresponding to the syndrome type after a specific treatment is output.

[0119] When the syndrome type probability without a specific treatment is input, the probability of the syndrome corresponding to the syndrome type without a specific treatment is output.

[0120] Embodiment two,

[0121] W0, the patient's basic information, syndrome information such as symptoms and signs, examination and inspection indexes, medical images and pictures in the historical text data are input into the syndrome classification model based on Transformer for training to obtain the trained syndrome classification model based on Transformer.

[0122] W1, the patient's basic information, syndrome information such as symptoms and signs, examination and inspection indexes, medical images and pictures obtained at the current time point are processed and input into the trained syndrome classification model based on Transformer for data processing to obtain the syndrome type probability corresponding to the current data.

[0123] W1 includes the following sub-steps:

[0124] W10, the scattered data is linearly mapped to obtain first vectorized data;

[0125] The natural language text data is converted into word embedding data by using word embedding, the word embedding data is positionally encoded by a position encoding method based on a sine and cosine in the Transformer, and a second self-attention layer feature extraction is performed to obtain second vectorized data;

[0126] The image data is subjected to convolution calculation by a convolution layer to obtain third vectorized data;

[0127] The first vectorized data, the second vectorized data and the third vectorized data are combined to form a vectorized data set.

[0128] The scattered data includes gender, age, examination index and test index.

[0129] The natural language text data includes medical history text and syndrome description text.

[0130] The image data includes MRI image, face diagnosis picture and tongue diagnosis picture.

[0131] For example, a 24-year-old male is subjected to blood fat examination, i.e. blood lipid examination; the history text and the syndrome description text correspond to the history blood lipid examination result of the male and the current syndrome description; generally, no MRI image is needed, and if needed, face diagnosis and tongue diagnosis image data can be added; after the above data is quantitatively processed, type coding is added and normalization processing is performed, the normalized data is input into the trained syndrome classification model based on the Transformer.

[0132] W11, the first, second and third vectorized data are normalized by adding type coding to each vectorized data according to different types, to form normalized data, and the normalized data is input into the trained syndrome classification model based on the Transformer.

[0133] After W1 obtains the syndrome probability corresponding to the current data.

[0134] W2, the syndrome probability of the historical data and the historical patient information are horizontally concatenated to form a historical patient information summary, which is input into the syndrome evolution prediction model, and the syndrome evolution prediction model is trained.

[0135] W21, the syndrome probability of the current historical data and the current patient information are horizontally concatenated to form a current patient information summary, and the current and historical patient information summaries are jointly input into the trained syndrome evolution prediction model to output the corresponding syndrome probability after or without treatment.

[0136] After W21 obtains the syndrome probability corresponding to the current data.

[0137] W3, input the historical syndrome type probability into the syndrome evolution prediction model, and train the syndrome evolution prediction model.

[0138] W31, input the current syndrome type probability into the trained syndrome evolution prediction model, and output the probability of the current syndrome corresponding to the current syndrome type.

[0139] In this embodiment, the trained syndrome type classification model based on the Transformer, the trained syndrome evolution prediction model, and the trained syndrome evolution prediction model are used. In use, the information related to the patient, such as the syndrome, is input into the trained syndrome type classification model based on the Transformer to output the probability of each syndrome type. After the probability of each syndrome type is input into the trained syndrome evolution prediction model to output the probability of the treated or untreated syndrome type, the probability of the treated or untreated syndrome type is input into the trained syndrome evolution prediction model to output the corresponding syndrome probability. The intelligent and trained model can realize the evolution prediction of the syndrome and the syndrome type, get rid of the traditional Chinese medicine prediction mode of the syndrome and the syndrome type, reduce the subjective prediction error, and increase the accuracy.

[0140] Those skilled in the art should understand that the embodiments of the present application can be provided as a method, a system or a computer program product. Therefore, the present application can adopt a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware aspects. Moreover, the present application can adopt the form of a computer program product implemented on one or more computer usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer usable program code.

[0141] The present application is described with reference to flowcharts and / or block diagrams of methods, devices (systems) and computer program products according to embodiments of the present application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be realized by computer program instructions.

[0142] It should be noted that in the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word 'comprising' does not exclude the presence of elements or steps not listed in a claim. The word 'a' or 'an' preceding an element does not exclude the presence of a plurality of such elements. The application can be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the claims, the word 'first','second', 'third' etc. does not imply any order. These words are used to distinguish between two entities having the same name. The word'step' does not imply any order. These words are used to distinguish between two entities having the same name.

[0143] Moreover, it should be noted that in the description of the application, the terms "one embodiment", "some embodiments", "an embodiment", "an example", "a specific example" or "some examples" are intended to mean that a particular feature, structure, material or characteristic is included in at least one embodiment or example of the present application. These terms do not necessarily indicate that the described particular feature, structure, material or characteristic is included in all embodiments or examples of the present application. Furthermore, such terms do not indicate that a particular feature, structure, material or characteristic is essential to the present application. In addition, in the description of the application, the terms "comprise" and "comprising" are used in a non-limiting sense to mean that the subsequent features, structures, materials or characteristics are included in the application. Therefore, the use of such terms in the description of the application should not be interpreted as excluding other features, structures, materials or characteristics from the application.

[0144] Although preferred embodiments of the application have been described, those skilled in the art will recognize that additional modifications and variations to the preferred embodiments are possible in light of the above teachings. It is therefore contemplated to be within the scope of the application to claim additional modifications and variations to the preferred embodiments as fall within the scope of the present application and its equivalents.

[0145] It will be apparent to those skilled in the art that various modifications and variations can be made to the present application without departing from the spirit or scope of the application. Thus, it is intended that the present application cover the modifications and variations of this application provided they come within the scope of the appended claims and their equivalents.

Claims

1. An artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method, characterized in that, Comprise: S1, normalize the scattered data, natural language text data, image data related to the patient's syndrome, obtain normalized data, and the collection of each normalized data is a normalized data set; S1 includes: S10, vectorizing the scattered data, natural language text data, and image data related to the syndrome to obtain a vectorized data set; S11, normalize each vectorized data in the vectorized data set according to different types by adding type codes to each vectorized data, and form normalized data; S10 includes: The first vectorized data is obtained by linear mapping of the scattered data; The natural language text data is converted into word embedding data using word embedding, the word embedding data is positionally encoded through a position encoding method based on sine and cosine in the Transformer, and the second vectorized data is obtained through a second self-attention layer feature extraction; The third vectorized data is obtained by convolution calculation of the image data through a convolution layer; Wherein the first vectorized data, the second vectorized data and the third vectorized data are collectively combined into a vectorized data set; S2, input the normalized data set into the trained syndrome classification model based on Transformer to obtain each syndrome probability corresponding to the syndrome; The syndrome classification model based on Transformer comprises in turn: a patient information multi-modal embedding layer, a first self-attention layer, a first linear mapping + Softmax function calculation layer; The normalized data set in the patient information multi-modal embedding layer is input into the first self-attention layer for attention calculation, and then the first linear mapping + Softmax function calculation layer is calculated to obtain the syndrome probability; wherein the patient information multi-modal embedding layer is used to store the normalized data set.

2. The artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method according to claim 1, characterized in that, Before S1, it also includes: S0, train the syndrome classification model based on Transformer using historical normalized data to obtain a trained syndrome classification model based on Transformer. 3.The artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method according to claim 1, characterized in that, Comprise: The scattered data includes: gender, age, examination index, test index; The natural language text data includes: medical history text, syndrome description text; The image data includes: MRI image, face diagnosis picture, tongue diagnosis picture. 4.The artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method according to claim 1, characterized in that, After S2, it also includes: S3, splice the current or historical syndrome probability and the information vector in the current or historical patient information multi-modal embedding layer as patient information summary, input into the trained syndrome evolution prediction model, and obtain the syndrome probability after or without specific treatment; Wherein, the trained syndrome evolution prediction model is obtained by training the syndrome evolution prediction model using the historical syndrome probability and the information in the historical patient information multi-modal embedding layer; The syndrome evolution prediction model comprises in turn: a patient information summary layer, a third self-attention layer, a full connection layer, and a second linear mapping + Softmax function calculation layer; Wherein the patient information summary layer is used to store the current and historical patient information; The information of the patient information aggregation layer is input into the third self-attention layer after time sequence encoding, processed by a fully connected layer, and then input into the second linear mapping + Softmax function calculation layer to calculate the syndrome probability after treatment or without treatment; The fully connected layer includes a first fully connected layer and a second fully connected layer; the first fully connected layer is used to process the historical patient information aggregation to output the syndrome probability after treatment; The second fully connected layer is used to process the current patient information aggregation to output the syndrome probability without treatment; The first fully connected layer and the second fully connected layer extract the corresponding patient information by setting different extraction parameters.

5. The artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method according to claim 4, characterized in that, The time sequence encoding includes: According to the time window T, establish a historical timeline with T as a time period length; Determine whether the historical patient information exists at the current time point; If not, use the mask mechanism to mask the current time point, and the information here is not included in the historical patient information aggregation; If so, set the current time to 0 and the last existing data time point to -T. 6.The method of claim 5, wherein, It includes: The time sequence encoding adopts the position encoding method based on sine and cosine of the Transformer.

7. The artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method according to claim 6, characterized in that, After S3, it also includes: S4, input the syndrome probability after treatment or without treatment into the trained syndrome evolution prediction model to obtain the probability of the syndrome corresponding to the syndrome; The historical syndrome probability after treatment or without treatment is used to train the syndrome evolution prediction model to obtain the trained syndrome evolution prediction model; The syndrome evolution prediction model includes a third fully connected layer and at least one third linear mapping + Softmax function calculation layer; The syndrome probability after treatment or without treatment is input into the third linear mapping + Softmax function calculation layer to calculate the probability of the syndrome corresponding to the syndrome after processing by the third fully connected layer. 8.The artificial intelligence-based traditional Chinese medicine syndrome classification and probability prediction method according to claim 7, characterized in that, It includes: When the syndrome probability after treatment is input, the probability of the syndrome corresponding to the syndrome after treatment is output; When the syndrome probability without treatment is input, the probability of the syndrome corresponding to the syndrome without treatment is output.

Images