Synthetic medical data generation system and method for predicting postoperative complications

Abstract

The present invention relates to a synthetic medical data generation system and method for predicting postoperative complications. According to the present invention, the synthetic medical data generation system is provided, the system comprising: a medical data acquisition unit for acquiring an input vector in a form in which structured medical data related to clinical information of a patient and unstructured medical data in a text form recorded for the patient are arranged according to configuration rules; an embedding unit for embedding, in consideration of data type, the structured medical data and the unstructured medical data included in the input vector; and a synthetic medical data generation unit, which integrally learns the embedded input vector through a pre-trained deep learning-based synthetic data generation model, thereby generating structured synthetic medical data having the same form as the structured medical data.

Description

System and method for generating synthetic medical data for predicting postoperative complications

[0001] The present invention relates to a system and method for generating synthetic medical data for predicting post-surgical complications, and more specifically, to a system and method for generating synthetic medical data capable of synthesizing high-quality structured medical data by utilizing structured data and unstructured data.

[0002] Postoperative complications reported include bleeding, respiratory complications including pneumonia, sepsis, and circulatory complications. While the incidence varies by country, it is reported to range from approximately 6-33%. When one or more postoperative complications occur, the patient mortality rate reaches approximately 12-20%, leading to increased hospital stay and medical costs.

[0003] Among postoperative complications, acute kidney injury is associated with the development of other postoperative complications and the progression of chronic kidney disease, which leads to increased mortality. Therefore, it is essential to predict high-risk groups for acute kidney injury before surgery, stratify risk for complications, and proactively allocate medical resources.

[0004] Additionally, in order to apply deep learning / machine learning techniques, which are widely used across the medical data analysis field, to the medical field, it is necessary to secure a significant amount of medical and patient data.

[0005] Recently, rapid progress has been made in computerization / digitization of hospital medical data, as well as in data collection / production / processing / security technologies. However, the connectivity and exchange necessary for utilizing these technologies remains somewhat inadequate. This is due to the sensitivity of personal information protection issues regarding hospital medical information and its structural complexity, which limits the connectivity and exchange of vast amounts of data.

[0006] Therefore, to utilize vast amounts of medical data, efforts are needed to standardize and structure medical information across hospitals, while also establishing reliable measures to protect sensitive personal information. While various research methodologies have been proposed to protect sensitive personal information, obtaining high-quality, large-scale data remains challenging due to issues such as deteriorating data quality, the risk of re-identification, and a lack of standardization.

[0007] Synthetic medical data is emerging as a solution to overcome these challenges. Synthetic medical data can replace real medical data containing sensitive information such as personal information. It also offers the advantage of being free from pseudonymization / anonymization and data linking restrictions, maximizing data utility.

[0008] Moreover, because researchers can conduct research only with synthetic data without handling actual patient medical data, they can use and process the data without legal restrictions and freely share the data with external researchers.

[0009] Synthetic medical data can be produced in any desired quantity based on original data that has undergone various preprocessing, standardization, and labeling, dramatically reducing the time and cost required to build datasets for artificial intelligence learning.

[0010] Therefore, there is a need for the development of synthetic medical data generation technology that can build large amounts of clinical data while protecting sensitive personal information and preventing personal information from being identified.

[0011] The technology underlying the present invention is disclosed in Korean Patent No. 10-2482262 (announced on December 28, 2022).

[0012] The present invention aims to provide a system and method for generating synthetic medical data capable of generating high-quality synthetic medical data that reflects the diversity and complexity of medical data by synthesizing structured medical data by referencing not only structured medical data related to clinical information of a patient but also unstructured medical data corresponding to text data recorded in relation to the patient in an artificial intelligence model.

[0013] The present invention provides a synthetic medical data generation system, comprising: a medical data acquisition unit for acquiring an input vector in which structured medical data related to clinical information of a patient and unstructured medical data in the form of text recorded about the patient are listed according to a set rule; an embedding unit for embedding structured medical data and unstructured medical data included in the input vector in consideration of data type; and a synthetic medical data generation unit for generating structured synthetic medical data in the same format as the structured medical data by comprehensively learning the embedded input vector through a pre-trained deep learning-based synthetic data generation model.

[0014] In addition, the structured medical data may be composed of categorical data having an information type classified by category among the patient's clinical information and numerical data having an information type expressed numerically.

[0015] Additionally, the above unstructured medical data may be text data on a surgical record recorded during surgery.

[0016] In addition, the input vector may have a form in which a first vector consisting of categorical data of the setting type for the patient, a second vector consisting of numerical data of the setting type, and a third vector consisting of tokenized text data extracted from the surgical record are sequentially connected.

[0017] In addition, in the case of the structured medical data, the embedding unit may embed categorical features corresponding to the categorical data and linearly embed continuous numerical features corresponding to the numerical data, and in the case of the unstructured medical data, may convert individual tokens of tokenized text into fixed-size vectors using a text embedding method.

[0018] In addition, the embedding unit may perform token type embedding to enable the synthetic data generation model to identify whether medical data is structured, positional embedding to distinguish between each column in the structured medical data and each token in the unstructured medical data, and null type embedding to distinguish between whether each value in the input vector is a null value (blank value).

[0019] In addition, the synthetic data generation model may include a transformer that comprehensively learns the characteristics of structured medical data and unstructured medical data within the embedded input vector to generate output data, and a multi-head that generates structured synthetic medical data composed of categorical data and numerical data using the output data of the transformer and provides the structured synthetic medical data as an output vector.

[0020] In addition, the multi-head may include a classifier that generates categorical features constituting categorical data in the synthetic medical data from the output data of the transformer, a Gaussian head that generates continuous features constituting numerical data in the synthetic medical data, and a null classifier that determines the presence or absence of a null value in the output vector.

[0021] Additionally, the synthetic medical data generated above can be used as learning data for building a prediction model for predicting post-surgical complications.

[0022] And, the present invention provides a method for generating synthetic medical data performed by a synthetic medical data generation system, comprising the steps of: obtaining an input vector in which structured medical data related to clinical information of a patient and unstructured medical data in the form of text recorded about the patient are listed according to a set rule; embedding structured medical data and unstructured medical data included in the input vector in consideration of data type; and generating structured synthetic medical data in the same format as the structured medical data by comprehensively learning the embedded input vector through a pre-trained deep learning-based synthetic data generation model.

[0023] According to the present invention, a synthetic data generation model that understands the complexity and diversity of medical data can be provided by comprehensively considering structured data related to a patient's clinical information and unstructured text data such as surgical records.

[0024] The present invention can significantly reduce the time and cost required to build a dataset for artificial intelligence learning by efficiently and quickly generating the necessary amount of data based on raw data refined through various preprocessing, standardization, and labeling processes.

[0025] In this way, according to the present invention, high-quality synthetic medical data can be efficiently generated by synthesizing structured medical data while deeply understanding the complexity and diversity of data by referring to not only structured information but also unstructured text.

[0026] Additionally, synthetic high-quality medical data can be used as a dataset for learning a complication risk prediction model by replacing actual medical data, and can play an important role in various medical tasks.

[0027] FIG. 1 is a diagram showing the configuration of a synthetic medical data generation system according to an embodiment of the present invention.

[0028] FIG. 2 is a drawing specifically explaining the structure of a synthetic data generation model according to an embodiment of the present invention.

[0029] FIG. 3 is a drawing illustrating a method for generating synthetic medical data according to an embodiment of the present invention.

[0030] Figure 4 is a diagram showing a specific example of the use of synthetic medical data.

[0031] Hereinafter, embodiments of the present invention will be described in detail with reference to the attached drawings so that those skilled in the art can easily practice the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein. In addition, in the drawings, parts irrelevant to the description have been omitted to clearly explain the present invention, and similar parts have been designated with similar reference numerals throughout the specification.

[0032] Throughout the specification, when a part is said to be "connected" to another part, this includes not only the cases where the parts are "directly connected" but also the cases where the parts are "electrically connected" with other elements intervening. Furthermore, when a part is said to "include" a component, this does not exclude other components, but rather includes other components, unless otherwise stated.

[0033] FIG. 1 is a diagram showing the configuration of a synthetic medical data generation system according to an embodiment of the present invention, and FIG. 2 is a diagram specifically explaining the structure of a synthetic data generation model according to an embodiment of the present invention.

[0034] As illustrated in FIG. 1, a synthetic medical data generation system (100) according to an embodiment of the present invention includes a medical data acquisition unit (110), an embedding unit (120), and a medical data generation unit (130). Here, the operation of each unit (110 to 130) and the data flow between each unit can be controlled by a control unit (not shown).

[0035] This synthetic medical data generation system (100) may be implemented as a computer device that is physically configured and includes a processor, memory, a user interface input / output device and a storage device, a network input / output unit, etc., or may be implemented as an application program running on a computer device or a user terminal.

[0036] The synthetic medical data generation system (100) according to an embodiment of the present invention can generate synthetic medical data by inputting actual patient medical data into a pre-trained synthetic data generation model. The synthetic medical data can be generated in a desired quantity to replace actual medical data. For example, the medical data of 10,000 individuals can be input into the synthetic data generation model to generate 10,000 new synthetic medical data sets.

[0037] As shown in FIGS. 1 and 2, the medical data acquisition unit (110) can acquire an input vector (10) in the form of structured medical data related to the patient's clinical information and unstructured medical data in the form of text recorded about the patient, which are listed according to a set rule.

[0038] The medical data acquisition unit (110) can collect patient-specific medical data or input vectors through a user interface input / output device, a wired / wireless connected user terminal, or a storage device. The medical data acquisition unit (110) can receive patient-specific medical data from the outside and directly process it into the form of an input vector, or can directly receive and acquire pre-processed patient-specific input vectors from the outside.

[0039] As shown in Fig. 2, the input vector (10) is largely divided into structured medical data and unstructured medical data, and the structured data can be further divided into categorical data and numerical data.

[0040] Accordingly, it can be seen that the input vector (10) is composed of a total of three types of data (categorical data, numerical data, and text data).

[0041] Structured data may correspond to medical data related to a patient's clinical information (e.g., gender, age, blood pressure, blood type, heart rate, past surgical history, past medical history), and unstructured medical data may correspond to medical data in text form recorded in relation to a patient.

[0042] Among structured data, categorical data may correspond to data that has a type of information classified as 0 or 1 according to a category among the patient's clinical information (e.g., gender (male / female), presence / absence of surgical history (yes / no), etc.), and numerical data may correspond to data that has a type of information expressed numerically among the patient's clinical information (e.g., blood pressure, heart rate, blood sugar, age, etc.).

[0043] Among the three types of data, the only unstructured medical data is text data, such as the surgical records recorded during a patient's surgery. Surgical records are text data recorded by medical staff during surgery, and can comprehensively record various information, including the patient's history, unusual events, and physiological responses that occurred during the surgery.

[0044] An embodiment of the present invention proposes a technique for synthesizing structured medical data for predicting post-surgical complications, wherein text data among the patient's medical data may include text data on a surgical record.

[0045] Additionally, the input vector (10) may have a form in which the three types of data described above are listed according to a setting rule. The setting rule may include the type of data that constitutes the input vector, the type or number of detailed data within the type, the listing order of the data, etc.

[0046] For example, the input vector (10) may have a form in which a first vector consisting of categorical data of a setting type for a patient, a second vector consisting of numerical data of a setting type, and a third vector consisting of tokenized text data extracted from a surgical record are sequentially connected, as shown in FIG. 2.

[0047] The embedding unit (120) can receive an input vector (10) of this type, embed it, and provide it to the synthetic medical data generation unit (130).

[0048] This embedding unit (120) can embed structured medical data and unstructured medical data included in the input vector by considering the data type.

[0049] Specifically, in the case of structured medical data, the embedding unit (120) can embed categorical features corresponding to categorical data and linearly embed continuous numerical features corresponding to numerical data. In addition, in the case of unstructured medical data, the embedding unit (120) can convert individual tokens of tokenized text into fixed-size vectors using a commonly known text embedding method.

[0050] In addition, the embedding unit (120) can perform token type embedding to enable a synthetic data generation model to identify whether medical data is structured, positional embedding to distinguish between each column in structured medical data and each token in unstructured medical data, and null type embedding to distinguish between whether each value in an input vector is a null value (blank value).

[0051] Here, null values ​​may exist in different locations for each patient. For example, for a patient without a heart rate, the value corresponding to the heart rate in the input vector may be represented as a null value.

[0052] In this way, the embedding unit (120) can perform the embedding process in various ways to appropriately model structured and unstructured data, and can transmit the input data for which embedding has been completed to the synthetic medical data generation unit (130).

[0053] The synthetic medical data generation unit (130) can generate structured synthetic medical data (20) in the same format as the structured medical data by comprehensively learning the embedded input vector through a pre-trained deep learning-based synthetic data generation model.

[0054] At this time, the synthetic medical data generation unit (130) applies the embedded input vector to a pre-learned synthetic data generation model, thereby comprehensively learning the characteristics of structured medical data and unstructured medical data within the input vector, and can generate synthetic medical data (20) synthesized with the same structure as the structured medical data as the final result.

[0055] Synthetic medical data (20) is composed of categorical data and numerical data, similar to the structured medical data used as input to the model, as shown in the upper part of Fig. 2, and may have the same data size as the input data. In addition, null values ​​within the synthetic medical data (20) may also exist in the same location as the null values ​​of the structured medical data used as input to the model.

[0056] In this way, synthetic medical data (20) has null values ​​in the same location as the original data used as input, which is structured data, and some or all of the remaining values ​​excluding the null values ​​may have a form that is modified from the original.

[0057] In an embodiment of the present invention, a synthetic data generation model may be formed by including a transformer (131) that exhibits the best performance in various natural language processing tasks, as shown in FIG. 2, and a multi-head (132) connected thereto.

[0058] The transformer (131) receives an input vector for which embedding has been completed, and can comprehensively learn the characteristics of structured and unstructured medical data within the vector to generate output data. In this process, the transformer (131) can utilize a masking-based generation method among various learning methods. This method allows the model to mask (hid) a portion of the input data and predict the masked portion, thereby enabling the model to gain a deeper understanding of the internal structure of the data and learn high-level patterns.

[0059] In the embodiment of the present invention, since the range of self-attention performed by the transformer (131) spans both structured medical data and unstructured medical data included in the input vector, data of various modalities can be modeled simultaneously.

[0060] In an embodiment of the present invention, the transformer (131) may be implemented using only a transformer encoder. Of course, the transformer may also utilize an encoder-decoder model structure as needed.

[0061] The multi-head (132) can generate structured synthetic medical data consisting of categorical data and numerical data using the output data of the transformer (131) and provide it as an output vector.

[0062] More specifically, the multi-head (132) can generate structured medical data by mapping the high-dimensional output data of the transformer into categorical data or numerical data.

[0063] At this time, the multi-head (132) may include, as shown in FIG. 2, a classifier that generates categorical features constituting categorical data in the synthetic medical data from the output data of the transformer (131), a Gaussian head that generates continuous features constituting numerical data in the synthetic medical data, and a null classifier that determines the presence or absence of a null value in the output vector.

[0064] Specifically, the classifier can determine which categorical class each output of the transformer (131) corresponding to the categorical features belongs to. The Gaussian head can model a normal distribution by predicting the mean and variance for each column using the output of the transformer (131) corresponding to the continuous features. By sampling values ​​from the distribution estimated in this way, probabilistic generation of numerical data can be performed. The null classifier can perform the function of distinguishing whether a column is a missing value or not using the output of the transformer (131) and synthesizing only for columns that are not missing values.

[0065] In this way, it can be seen that the multi-head (132) diversifies the generator located at the top of the synthetic data generation model to synthesize data by encompassing various data types within the structured data.

[0066] The right side of Figure 2 exemplarily shows a data set accumulating synthetic medical data generated using the aforementioned method. According to the present invention, N new synthetic medical data sets can be generated based on N actual medical data sets. These synthetic medical data sets thus obtained can replace actual data sets in various deep learning-based analysis / prediction models and related research.

[0067] According to this, for example, medical data on a thousand patients can be applied to an artificial intelligence model to generate a thousand new synthetic medical data as big data, and the synthetic medical data generated in this way can solve privacy issues in the process of big data-based medical data analysis.

[0068] FIG. 3 is a drawing illustrating a method for generating synthetic medical data according to an embodiment of the present invention.

[0069] First, the synthetic medical data generation system (100) obtains an input vector in the form of structured medical data related to the patient's clinical information and unstructured medical data in the form of text recorded about the patient, which are listed according to set rules (S310).

[0070] Next, the synthetic medical data generation system (100) embeds structured medical data and unstructured medical data included in the input vector by considering the data type (S320).

[0071] Then, the synthetic medical data generation system (100) inputs the embedded input vector into the synthetic data generation model (S330), and through this, comprehensively learns the characteristics of structured medical data and unstructured medical data within the input vector, thereby generating synthetic medical data in the same format as the structured medical data (S340).

[0072] According to the present invention, by synthesizing structured medical data by referring to unstructured text medical data recorded in relation to the patient as well as structured medical data related to the patient's clinical information, it is possible to generate synthetic medical data that understands the diversity and complexity of medical data.

[0073] In addition, while existing medical data synthesis models mainly focus on the synthesis of a single modality such as medical images, the present invention can provide a differentiated synthesis model capable of generating high-quality synthetic medical data that understands complexity and diversity by comprehensively considering structured medical information and unstructured medical text information.

[0074] High-quality data synthesized in this way can play a crucial role in various medical tasks, including predicting the risk of complications. Furthermore, these synthetic models can be widely applied not only in the medical field but also to various data synthesis tasks where structured and textual information coexist.

[0075] Figure 4 is a diagram illustrating a specific example of the use of synthetic medical data. As shown in Figure 4, the synthetic medical data (20) generated through the present invention can be utilized as a learning data set for building a post-surgical complication prediction model. For example, it can be utilized to build an artificial intelligence-based deep learning / machine learning model to predict whether or not a patient will develop acute kidney injury within one week of surgery, or the probability of such development.

[0076] In this way, introducing an algorithm that automatically predicts the risk of post-surgical complications based on synthetic medical data can enable efficient use of medical resources and costs based on pre- and post-surgical risk assessment.

[0077] Furthermore, synthetic medical data holds significant value as a substitute for personal and sensitive information contained in real medical data. Therefore, utilizing it can circumvent data pseudonymization and anonymization issues, significantly enhancing data usability and applicability. Researchers can utilize this synthetic medical data without legal restrictions and, if necessary, share it with other researchers.

[0078] In this way, the use of synthetic data can significantly help address privacy concerns surrounding medical big data, thereby fostering a social consensus on the increased use of medical big data. Furthermore, developing a complication prediction model using synthetic data can reduce medical costs and improve patient health outcomes. Furthermore, the introduction of this system, even in areas with a shortage of medical professionals, can enable the early detection of complications.

[0079] The following describes the results of an experiment on the classification performance of a prediction model (post-surgical complication prediction model) that has completed learning using synthetic medical data generated according to the technique of the present invention.

[0080] Acute kidney injury (AKI) was considered a postoperative complication. In Table 1 below, AKI=0 indicates data for the group with postoperative complications, and AIO=1 indicates data for the group without postoperative complications.

[0081] Statistical testing was performed as follows. The chi-square test was used for categorical data, and the t-test was used for numerical data. The number listed in each item indicates the number of columns where the distributions of the synthetic and actual data were judged to be identical. Furthermore, the actual data for model performance testing were divided into groups with AKI = 0 and with AKI = 1, and statistical tests were performed on each group.

[0082]

[0083] The modality items in Table 1 indicate the type of input data utilized for data synthesis. The lower row (Table+Text) in Table 1 is an example using multi-modality, representing a case of the present invention where synthetic medical data was generated by considering both tables (structured data) and text (unstructured data). The upper row (Table) is an example using a single modality, representing a case where synthetic medical data was generated using only table information for comparison with the present invention.

[0084] In Table 1, the sampling and mask scheduler items represent elements related to the synthesis level of the artificial intelligence model used.

[0085] The performance indicators of the postoperative complication prediction model (hereinafter, the prediction model) were the Area Under the Receiver Operating Characteristic Curve (AUROC), the ratio of the most frequent class, and the mean value.

[0086] When a predictive model (comparison target) trained on real data was validated against a test dataset of actual patients (patients with known postoperative complications), the AUROC value was 0.853. If the synthetic data were successfully generated using the present invention's technique, the AUROC value when applying the predictive model trained on the synthetic data to a test dataset of actual patients would also be close to 0.853.

[0087] As shown in the results in Table 1, the AUROC value of the prediction model learned through synthetic medical data according to the present invention, which takes into account both table and text information, ranges from 0.783 to 0.799, which is almost similar to the AUROC value of 0.853 of the comparison prediction model learned through actual patient data.

[0088] Furthermore, in all cases, synthetic data utilizing multi-modality (Table+Text) according to the present invention exhibited higher quality than synthetic data utilizing a single modality (Table). Consequently, predictive models trained with synthetic data generated by referencing unstructured text data in all cases exhibited higher performance than those generated without considering text, demonstrating the high quality of the synthetic data produced by the present invention.

[0089] Next, the ratio of the most frequent class is a measure of how much it differs from the actual data by dividing the number of the most frequent classes in each categorical column by the total number of samples. A lower value indicates a smaller difference and better performance.

[0090] Also, the mean value is the difference between the average values ​​of the numeric columns of the actual data and the generated data added up for the entire column, and the smaller the value, the higher the performance.

[0091] In all cases, the ratio and average values ​​for the case of utilizing the multi-modality (Table+Text) of the present invention were lower than those for the case of utilizing a single modality (Text), so it can be seen that the synthetic data produced by the present invention exhibits superior quality than the case of a single modality that does not refer to text, which is unstructured data.

[0092] In addition, we tested the possibility of personal information leakage of synthetic medical data generated by the proposed invention through a privacy assessment called membership inference, and confirmed that the possibility of leakage was low.

[0093] In this way, according to the present invention, a synthetic data generation model that understands the complexity and diversity of medical data can be provided by comprehensively considering structured data related to a patient's clinical information and unstructured text data such as surgical records.

[0094] Additionally, by synthesizing structured medical data while deeply understanding the complexity and diversity of data by referencing not only structured information but also unstructured text, high-quality synthetic medical data can be produced.

[0095] While the present invention has been described with reference to the embodiments illustrated in the drawings, these are merely exemplary, and those skilled in the art will understand that various modifications and equivalent alternative embodiments are possible. Therefore, the true scope of technical protection of the present invention should be determined by the technical spirit of the appended claims.

Claims

1. A medical data acquisition unit that acquires an input vector in the form of structured medical data related to a patient's clinical information and unstructured medical data in the form of text recorded about the patient, which are listed according to a set rule; An embedding unit that embeds structured medical data and unstructured medical data included in the input vector by considering the data type; and A synthetic medical data generation system including a synthetic medical data generation unit that comprehensively learns the embedded input vector through a pre-trained deep learning-based synthetic data generation model to generate structured synthetic medical data in the same format as the structured medical data.

2. In claim 1, The above structured medical data is, A synthetic medical data generation system comprising categorical data having information types classified by category among the clinical information of the above patient and numerical data having information types expressed numerically.

3. In claim 1, The above unstructured medical data is, A synthetic medical data generation system that generates text data from surgical records recorded during surgery.

4. In claim 2, The above input vector is, A synthetic medical data generation system having a first vector composed of categorical data of the above patient type, a second vector composed of numerical data of the above patient type, and a third vector composed of tokenized text data extracted from the above surgical records, which are sequentially connected.

5. In claim 2, The above embedding part, A synthetic medical data generation system that, in the case of the above structured medical data, embeds categorical features corresponding to the above categorical data and embeds continuous numerical features corresponding to the above numerical data using linear embedding, and, in the case of the above unstructured medical data, converts individual tokens of tokenized text into fixed-size vectors using a text embedding method.

6. In claim 2, The above embedding part, A synthetic medical data generation system that performs token type embedding to enable the synthetic data generation model to identify whether medical data is structured, positional embedding to distinguish between each column in the structured medical data and each token in the unstructured medical data, and null type embedding to distinguish between each value in the input vector and whether it is a null value (blank value).

7. In claim 2, The above synthetic data generation model is, A transformer that generates output data by comprehensively learning the features of structured medical data and unstructured medical data within the above-mentioned embedded input vector, A synthetic medical data generation system comprising a multi-head that generates structured synthetic medical data composed of categorical data and numerical data using the output data of the transformer and provides the structured synthetic medical data as an output vector.

8. In claim 7, The above multi-head, A synthetic medical data generation system comprising a classifier that generates categorical features constituting categorical data in the synthetic medical data from the output data of the transformer, a Gaussian head that generates continuous features constituting numerical data in the synthetic medical data, and a null classifier that determines the presence or absence of a null value in the output vector.

9. In claim 1, The synthetic medical data generated above is: A synthetic medical data generation system used as learning data for building a predictive model for predicting post-surgical complications.

10. In a synthetic medical data generation method performed by a synthetic medical data generation system, A step of obtaining an input vector in the form of structured medical data related to the patient's clinical information and unstructured medical data in the form of text recorded about the patient, which are listed according to a set rule; A step of embedding structured medical data and unstructured medical data included in the input vector by considering the data type; and A method for generating synthetic medical data, comprising the step of generating structured synthetic medical data in the same format as the structured medical data by comprehensively learning the embedded input vector through a pre-trained deep learning-based synthetic data generation model.

11. In claim 10, The above structured medical data is, A method for generating synthetic medical data consisting of categorical data having an information type classified by category among the clinical information of the above patient and numerical data having an information type expressed numerically.

12. In claim 10, The above unstructured medical data is, A method for generating synthetic medical data, which is text data from surgical records recorded during surgery.

13. In claim 11, The above embedding step is, A method for generating synthetic medical data, which performs token type embedding to enable the synthetic data generation model to identify whether medical data is structured, positional embedding to distinguish between each column in the structured medical data and each token in the unstructured medical data, and null type embedding to distinguish between each value in the input vector and whether it is a null value (blank value).

14. In claim 11, The step of generating the above synthetic medical data is: A step of generating output data by comprehensively learning the features of structured medical data and unstructured medical data within the embedded input vector in a transformer; and A method for generating synthetic medical data, comprising a step of generating structured synthetic medical data composed of categorical data and numerical data using output data of the transformer by a multi-head connected to the transformer and providing the structured synthetic medical data as an output vector.

Images