Medical dataset development method and system based on large model construction

Abstract

This invention discloses a method and system for developing a medical dataset based on a large-scale model. The method includes: acquiring multiple medical data sets to be processed; performing privacy data identification and encryption on each medical data set based on a de-identification algorithm to obtain multiple de-identified medical data sets; labeling each de-identified medical data set with medical information according to multiple preset student models to obtain corresponding labeled data; sampling and verifying all labeled data according to the teacher model corresponding to the student model, and organizing the labeled data into a medical dataset based on the verification results; the medical dataset is a pre-training dataset, instruction fine-tuning dataset, and / or test dataset for large-scale model training. Therefore, this invention can effectively ensure privacy and security during large-scale medical information processing, providing high-quality and standardized data asset support for the robust training of subsequent large-scale medical models.

Description

Technical Field

[0001] This invention relates to the field of data processing technology, and in particular to a method and system for developing medical datasets based on large model construction. Background Technology

[0002] With the deepening application of artificial intelligence in the medical field, constructing high-quality, standardized medical datasets has become crucial for improving model robustness and clinical application effectiveness. Existing medical data processing technologies typically perform initial cleaning of medical data manually or through simple rule-based filtering, and attempt basic de-identification of some sensitive fields. However, these techniques lack support for automatic identification and encryption of medical privacy data, and lack a collaborative mechanism that utilizes student model annotation combined with teacher model sampling verification. This results in an inability to achieve a deep integration of privacy security and annotation efficiency when processing large-scale medical information, leading to high risks of data privacy leakage, insufficient accuracy of annotation results, and low data cleaning efficiency. This limits the standardization level and data asset quality of subsequent training of large-scale medical models. Clearly, existing technologies have shortcomings that urgently need to be addressed. Summary of the Invention

[0003] The technical problem to be solved by this invention is to provide a method and system for developing medical datasets based on large-scale model construction, which can effectively protect the privacy and security in the process of large-scale medical information processing, and significantly improve the accuracy of annotation results and data cleaning efficiency through a teacher-student model collaboration mechanism, thereby providing high-quality and standardized data asset support for the robust training of subsequent large-scale medical models.

[0004] To address the aforementioned technical problems, the first aspect of this invention discloses a method for developing medical datasets based on large model construction, the method comprising: Acquire multiple medical data sets to be processed; Based on the de-identification algorithm, privacy data identification and encryption are performed on each of the aforementioned medical data to obtain multiple de-identified medical data; Based on multiple preset student models, each of the desensitized medical data is annotated with medical information to obtain the corresponding annotated data; Based on the teacher model corresponding to the student model, all the labeled data are sampled and verified, and the labeled data are organized into the medical dataset based on the verification results; the medical dataset is a pre-training dataset, instruction fine-tuning dataset and / or test dataset used for large model training.

[0005] As an optional implementation, in the first aspect of the present invention, the method of performing privacy data identification and encryption on each piece of medical data based on the de-identification algorithm to obtain multiple de-identified medical data includes: For each piece of medical data, the modal dimension corresponding to the medical data is determined; the modal dimension includes at least one of video modality, audio modality, tabular data modality, text modality, and sensor data modality; Determine the privacy recognition model corresponding to the modality dimension; The medical data is input into the privacy recognition model to obtain multiple privacy data components as output; Each of the privacy data portions is encrypted to obtain the de-identified medical data corresponding to the medical data.

[0006] As an optional implementation, in the first aspect of the present invention, determining the privacy recognition model corresponding to the modality dimension includes: For each candidate privacy recognition model, calculate the set of input dimensions that the model can process; Determine whether the input dimension set includes all the modal dimensions to obtain a first determination result; If the first judgment result is negative, the fit of the candidate privacy recognition model is set to the preset minimum value; If the first judgment result is yes, calculate the average value of the data similarity between multiple historical identification data corresponding to the candidate privacy identification model and the medical data to obtain the fit of the candidate privacy identification model; The candidate privacy recognition model with the highest fit is determined as the corresponding privacy recognition model.

[0007] As an optional implementation, in the first aspect of the invention, encrypting each of the privacy data portions to obtain the de-identified medical data corresponding to the medical data includes: For each privacy data portion, the privacy data portion is input into the trained vectorization model to obtain the data vector corresponding to the privacy data portion; The data vector is encrypted based on a preset vector encryption algorithm to obtain an encrypted vector; Determine the overlay data template corresponding to the data modality corresponding to the privacy data portion, and assemble the encryption vector and the overlay data template to obtain the replacement data corresponding to the privacy data portion; Each of the privacy data portions of the medical data is replaced with the replacement data to obtain the de-identified medical data corresponding to the medical data.

[0008] As an optional implementation, in the first aspect of the present invention, the teacher model is trained using a training dataset that includes multiple training medical data and corresponding full-dimensional medical analysis annotations; the student model is obtained by distilling the teacher model based on different distillation processes.

[0009] As an optional implementation, in the first aspect of the present invention, the student model is a diagnosis-related prediction model, a treatment-related prediction model, a medication-related prediction model, or a prognosis-related prediction model; the medical information annotation includes at least one of diagnosis-related annotation, treatment-related annotation, medication-related annotation, or prognosis-related annotation.

[0010] As an optional implementation, in the first aspect of the present invention, the step of annotating each of the de-identified medical data with medical information according to a plurality of preset student models to obtain corresponding annotated data includes: For each de-identified medical data, the de-identified medical data is input into the pre-prediction model corresponding to each student model to obtain the labeling probability output by each pre-prediction model; the pre-prediction model is trained by a training dataset including multiple training medical data and the corresponding labeling probability of the student model for the training medical data. Multiple student models are obtained by filtering out the student models corresponding to multiple pre-prediction models whose labeled probability is greater than a preset probability threshold. The anonymized medical data is input into each of the target student models to obtain multiple medical information annotations corresponding to the anonymized medical data; The multiple medical information labels are deduplicated, and the deduplicated labeling results are assembled with the desensitized medical data to obtain the labeled data corresponding to the desensitized medical data.

[0011] As an optional implementation, in the first aspect of the present invention, the step of sampling and validating all the labeled data according to the teacher model corresponding to the student model, and organizing the labeled data into a medical dataset based on the validation results, includes: All the labeled data are sampled to obtain multiple sampled data combinations; each sampled data combination includes multiple labeled data. Each of the aforementioned sampling data combinations is input into the teacher model to obtain the corresponding labeling rationality. For each student model, the average of the reasonableness of the annotations corresponding to all the labeled data for that student model is calculated to obtain the credibility of that student model. All labeled data whose credibility corresponding to the student model is greater than a preset credibility threshold are identified as credible labeled data. All the aforementioned trusted labeled data and their corresponding trust levels are compiled into a medical dataset.

[0012] A second aspect of this invention discloses a medical dataset development system based on large model construction, the system comprising: The acquisition module is used to acquire multiple medical data sets to be processed. The desensitization module is used to identify and encrypt privacy data for each piece of medical data based on the desensitization algorithm, resulting in multiple desensitized medical data. The annotation module is used to annotate each of the desensitized medical data with medical information according to multiple preset student models, so as to obtain the corresponding annotated data; The verification module is used to sample and verify all the labeled data according to the teacher model corresponding to the student model, and organize the labeled data into the medical dataset based on the verification results; the medical dataset is a pre-training dataset, instruction fine-tuning dataset and / or test dataset used for large model training.

[0013] As an optional implementation, in the second aspect of the present invention, the de-identification module, based on a de-identification algorithm, performs privacy data identification and encryption on each piece of medical data to obtain multiple de-identified medical data in the following specific ways: For each piece of medical data, the modal dimension corresponding to the medical data is determined; the modal dimension includes at least one of video modality, audio modality, tabular data modality, text modality, and sensor data modality; Determine the privacy recognition model corresponding to the modality dimension; The medical data is input into the privacy recognition model to obtain multiple privacy data components as output; Each of the privacy data portions is encrypted to obtain the de-identified medical data corresponding to the medical data.

[0014] As an optional implementation, in a second aspect of the invention, the desensitization module determines the specific method of the privacy recognition model corresponding to the modality dimension, including: For each candidate privacy recognition model, calculate the set of input dimensions that the model can process; Determine whether the input dimension set includes all the modal dimensions to obtain a first determination result; If the first judgment result is negative, the fit of the candidate privacy recognition model is set to the preset minimum value; If the first judgment result is yes, calculate the average value of the data similarity between multiple historical identification data corresponding to the candidate privacy identification model and the medical data to obtain the fit of the candidate privacy identification model; The candidate privacy recognition model with the highest fit is determined as the corresponding privacy recognition model.

[0015] As an optional implementation, in a second aspect of the invention, the specific method by which the de-identification module encrypts each of the privacy data portions to obtain the de-identified medical data corresponding to the medical data includes: For each privacy data portion, the privacy data portion is input into the trained vectorization model to obtain the data vector corresponding to the privacy data portion; The data vector is encrypted based on a preset vector encryption algorithm to obtain an encrypted vector; Determine the overlay data template corresponding to the data modality corresponding to the privacy data portion, and assemble the encryption vector and the overlay data template to obtain the replacement data corresponding to the privacy data portion; Each of the privacy data portions of the medical data is replaced with the replacement data to obtain the de-identified medical data corresponding to the medical data.

[0016] As an optional implementation, in a second aspect of the invention, the teacher model is trained using a training dataset that includes multiple training medical data and corresponding full-dimensional medical analysis annotations; the student model is obtained by distilling the teacher model based on different distillation processes.

[0017] As an optional implementation, in the second aspect of the present invention, the student model is a diagnosis-related prediction model, a treatment-related prediction model, a medication-related prediction model, or a prognosis-related prediction model; the medical information annotation includes at least one of diagnosis-related annotation, treatment-related annotation, medication-related annotation, or prognosis-related annotation.

[0018] As an optional implementation, in the second aspect of the present invention, the annotation module performs medical information annotation on each of the de-identified medical data according to a preset plurality of student models to obtain the corresponding annotated data in the following specific manner: For each de-identified medical data, the de-identified medical data is input into the pre-prediction model corresponding to each student model to obtain the labeling probability output by each pre-prediction model; the pre-prediction model is trained by a training dataset including multiple training medical data and the corresponding labeling probability of the student model for the training medical data. Multiple student models are obtained by filtering out the student models corresponding to multiple pre-prediction models whose labeled probability is greater than a preset probability threshold. The anonymized medical data is input into each of the target student models to obtain multiple medical information annotations corresponding to the anonymized medical data; The multiple medical information labels are deduplicated, and the deduplicated labeling results are assembled with the desensitized medical data to obtain the labeled data corresponding to the desensitized medical data.

[0019] As an optional implementation, in the second aspect of the present invention, the specific method by which the verification module performs sampling verification on all the labeled data according to the teacher model corresponding to the student model, and organizes the labeled data into the medical dataset based on the verification results includes: All the labeled data are sampled to obtain multiple sampled data combinations; each sampled data combination includes multiple labeled data. Each of the aforementioned sampling data combinations is input into the teacher model to obtain the corresponding labeling rationality. For each student model, the average of the reasonableness of the annotations corresponding to all the labeled data for that student model is calculated to obtain the credibility of that student model. All labeled data whose credibility corresponding to the student model is greater than a preset credibility threshold are identified as credible labeled data. All the aforementioned trusted labeled data and their corresponding trust levels are compiled into a medical dataset.

[0020] A third aspect of this invention discloses another medical dataset development system based on large model construction, the system comprising: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute some or all of the steps in the medical dataset development method based on large model construction disclosed in the first aspect of the present invention.

[0021] The fourth aspect of the present invention discloses a computer storage medium storing computer instructions, which, when invoked, are used to execute some or all of the steps in the medical dataset development method based on large model construction disclosed in the first aspect of the present invention.

[0022] Compared with the prior art, the embodiments of the present invention have the following beneficial effects: This invention acquires medical data and performs privacy data identification and encryption to obtain multiple de-identified medical data sets. These data sets are then labeled using student models and sampled and verified using teacher models to construct a large-scale medical dataset. This effectively ensures privacy and security during large-scale medical information processing. Furthermore, the collaborative mechanism between teacher and student models significantly improves the accuracy of labeling results and the efficiency of data cleaning, thus providing high-quality and standardized data assets to support the robust training of subsequent large-scale medical models. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating a method for developing a medical dataset based on large model construction, as disclosed in an embodiment of the present invention.

[0025] Figure 2 This is a schematic diagram of the structure of a medical dataset development system based on large model construction disclosed in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of another medical dataset development system based on large model construction disclosed in an embodiment of the present invention. Detailed Implementation

[0027] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0028] The terms "first," "second," etc., used in the specification, claims, and accompanying drawings of this invention are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or devices.

[0029] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0030] This invention discloses a method and system for developing medical datasets based on large-scale model construction. By acquiring medical data and performing privacy data identification and encryption, multiple de-identified medical datasets are obtained. These datasets are then labeled using student models and validated through sampling using teacher models to construct a large-scale medical dataset. This effectively ensures privacy and security during large-scale medical information processing. Furthermore, the teacher-student model collaboration mechanism significantly improves the accuracy of labeling results and data cleaning efficiency, thus providing high-quality and standardized data assets to support the robust training of subsequent large-scale medical models. Detailed explanations follow.

[0031] Example 1 Please see Figure 1 , Figure 1 This is a flowchart illustrating a method for developing a medical dataset based on large model construction, as disclosed in an embodiment of the present invention. Wherein, Figure 1 The described method for developing medical datasets based on large model construction can be applied to data processing systems / data processing equipment / data processing servers (including local processing servers or cloud processing servers). Figure 1 As shown, this method for developing medical datasets based on large model construction can include the following operations: 101. Acquire multiple medical data to be processed.

[0032] Optionally, the medical data may be electronic medical records (EMR), DICOM medical images, electrocardiogram (ECG) physiological electrical signals, biochemical test report text, surgical monitoring video, or pathological slide scan images; this invention does not limit the scope of the data.

[0033] 102. Based on the de-identification algorithm, privacy data identification and encryption are performed on each piece of medical data to obtain multiple de-identified medical data.

[0034] Optionally, the de-identification algorithm can be a k-anonymity algorithm, a l-diversity algorithm, a t-nearest neighbor algorithm, a differential privacy algorithm, or a de-identification algorithm based on data masking; this invention does not limit the specific algorithm.

[0035] 103. Based on multiple preset student models, medical information is labeled for each de-identified medical data to obtain the corresponding labeled data.

[0036] 104. Based on the teacher model corresponding to the student model, sample and verify all labeled data, and organize the labeled data into the medical dataset based on the verification results.

[0037] Optionally, the medical dataset can be a pre-trained dataset, an instruction fine-tuning dataset, and / or a test dataset for training large models.

[0038] Optionally, the teacher model is trained using a training dataset that includes multiple training medical data sets and corresponding full-dimensional medical analysis annotations.

[0039] Optionally, the student model is obtained by distilling the teacher model based on different distillation processes.

[0040] Optionally, the student model can be a diagnosis-related prediction model, a treatment-related prediction model, a medication-related prediction model, or a prognosis-related prediction model.

[0041] Optionally, medical information labeling includes at least one of diagnosis-related labeling, treatment-related labeling, medication-related labeling, or prognosis-related labeling.

[0042] Optionally, the student model can be a lightweight model after knowledge distillation, specifically BERT-Base, TinyBERT, DistilBERT, MobileNetV3, or ShuffleNetV2, which is not limited in this invention.

[0043] Optionally, the large model can be a pre-trained language model based on the Transformer architecture, a multimodal large model, a large language model for the medical vertical field, or a visual language model (VLM), and this invention does not limit it.

[0044] As can be seen, the above-described embodiments of the invention obtain multiple de-identified medical data by acquiring medical data and performing privacy data identification and encryption. Then, student models are used for annotation, and teacher models are used for sampling and verification to construct a large-scale medical dataset. This effectively ensures privacy and security during large-scale medical information processing. Furthermore, the teacher-student model collaboration mechanism significantly improves the accuracy of annotation results and data cleaning efficiency, thereby providing high-quality and standardized data asset support for the robust training of subsequent large-scale medical models.

[0045] As an optional embodiment, in the above steps, based on the de-identification algorithm, privacy data identification and encryption are performed on each piece of medical data to obtain multiple de-identified medical data, including: For each piece of medical data, determine the modality dimension corresponding to that medical data; optionally, the modality dimension includes at least one of video modality, audio modality, tabular data modality, text modality, and sensor data modality; Determine the privacy recognition model corresponding to the modality dimension; The medical data is input into a privacy-identifying model to obtain multiple privacy-related data components in the output. Each piece of privacy data is encrypted to obtain the corresponding de-identified medical data.

[0046] Optionally, the sensor data modality may include blood oxygen saturation data, non-invasive blood pressure monitoring data, respiratory rate data, intracranial pressure data, or body surface temperature data, and the present invention does not limit it.

[0047] Optionally, the privacy recognition model can be a named entity recognition model based on Conditional Random Field (CRF), a text region detection model based on ResNet, a sequence labeling model based on attention mechanism, or a feature alignment model based on multimodal contrastive learning (CLIP). This invention does not limit the model.

[0048] As can be seen, through the above optional embodiments, by determining the modal dimension of medical data and matching the corresponding privacy recognition model to encrypt sensitive parts, a differentiated privacy protection strategy for cross-modal heterogeneous data such as video, audio and tables is realized. While achieving automated privacy recognition, the risk of sensitive information leakage in complex data environments is greatly reduced and the concurrent processing performance of the preprocessing system is optimized.

[0049] As an optional embodiment, the step of determining the privacy recognition model corresponding to the modality dimension in the above steps includes: For each candidate privacy recognition model, calculate the set of input dimensions that the model can process; Determine whether the input dimension set includes all modal dimensions to obtain the first determination result; If the first judgment result is negative, the fit of the candidate privacy recognition model is set to the preset minimum value; If the first judgment result is yes, calculate the average data similarity between multiple historical recognition data corresponding to the candidate privacy recognition model and the medical data to obtain the fit of the candidate privacy recognition model; The candidate privacy recognition model with the highest fit is determined as the corresponding privacy recognition model.

[0050] Optionally, the data similarity can be cosine similarity, Jaccard similarity coefficient, edit distance (Levenshtein Distance) similarity, or feature distribution similarity based on KL divergence; this invention does not impose any limitations.

[0051] As can be seen, through the above optional embodiments, the optimal privacy recognition model is determined by calculating the coverage of the modal dimension of the candidate model and the average similarity of historical data. The multi-dimensional fit evaluation mechanism ensures a high degree of fit between the recognition tool and the medical data to be processed, effectively avoiding the waste of computing resources caused by modal mismatch and significantly enhancing the recognition accuracy of the system when facing unknown medical application scenarios.

[0052] As an optional embodiment, the step of encrypting each privacy data portion to obtain the de-identified medical data corresponding to the medical data includes: For each privacy data part, the privacy data part is input into the trained vectorization model to obtain the data vector corresponding to that privacy data part; Based on a preset vector encryption algorithm, the data vector is encrypted to obtain an encrypted vector; Determine the overlay data template corresponding to the data modality corresponding to the privacy data portion, and assemble the encryption vector and the overlay data template to obtain the replacement data corresponding to the privacy data portion; Each privacy data portion of the medical data is replaced with replacement data to obtain the corresponding de-identified medical data.

[0053] Optionally, the vectorization model can be Word2Vec, FastText, GloVe, BERT embedding layer, ViT feature extraction layer, or Wave2Vec audio encoder; this invention does not limit the specific model.

[0054] Optionally, the vector encryption algorithm can be homomorphic encryption, orthogonal transformation encryption, chaotic mapping encryption, hash encoding encryption, or random noise injection algorithm; this invention does not limit the specific algorithm.

[0055] Optionally, the overlay data template can be a DICOM-compliant security header template, an HL7-compliant anonymous text placeholder, a fixed noise pixel block, or a frequency domain masking template; this invention does not impose any limitations.

[0056] As can be seen, through the above optional embodiments, the privacy data is vectorized and encrypted, and then combined with the overlay data template to generate replacement data to complete the de-identification of medical data. Vector space encryption technology is used to achieve irreversible privacy masking at the physical level while maintaining the semantic logic of the data. Templated assembly ensures the integrity of the de-identified data structure, thereby providing high-fidelity data that balances security and feature availability for subsequent computing tasks.

[0057] As an optional embodiment, in the above steps, the medical information annotation of each de-identified medical data according to multiple preset student models to obtain the corresponding annotated data includes: For each de-identified medical data point, the de-identified medical data point is input into the pre-prediction model corresponding to each student model to obtain the labeling probability output by each pre-prediction model. The student models corresponding to multiple pre-prediction models whose labeling probability is greater than a preset probability threshold are selected to obtain multiple target student models. The anonymized medical data is input into each target student model to obtain multiple medical information annotations corresponding to the anonymized medical data; Multiple medical information labels are deduplicated, and the deduplicated labeling results are assembled with the de-identified medical data to obtain the labeled data corresponding to the de-identified medical data.

[0058] Optionally, the pre-prediction model is trained on a training dataset that includes multiple training medical data sets and corresponding student models indicating the probabilities of labeling the training medical data sets.

[0059] Optionally, the label probability can be represented as the probability distribution value output by the Softmax layer, the original Logits score, the confidence level after Sigmoid activation, or the variance estimate obtained by Monte Carlo sampling; this invention does not limit this.

[0060] Optionally, the probability threshold can be a fixed static threshold (such as 0.8 or 0.9), or an adaptive threshold based on Bayesian optimization, or a percentile threshold based on data quality distribution. This invention does not limit the threshold.

[0061] As can be seen, through the above optional embodiments, the annotation probability is calculated based on the pre-prediction model and the target student model is selected to perform de-identification annotation on the anonymized data. The calculation task is dynamically allocated to the most suitable sub-model through the pre-probability evaluation mechanism. This not only greatly reduces the hardware resource overhead caused by redundant calculation of low-relevance models, but also significantly improves the consistency and logical rigor of complex medical knowledge annotation by utilizing de-duplication assembly logic.

[0062] As an optional embodiment, the steps described above, including sampling and validating all labeled data based on the teacher model corresponding to the student model, and organizing the labeled data into the medical dataset based on the validation results, include: Sampling is performed on all labeled data to obtain multiple combinations of sampled data; optionally, the combination of sampled data includes multiple labeled data. Each sample data combination is input into the teacher model to obtain the corresponding labeling rationality. For each student model, the average of the labeling reasonableness of all labeled data corresponding to that student model is calculated to obtain the confidence level of that student model. All labeled data whose corresponding student model has a confidence level greater than a preset confidence level threshold are identified as reliable labeled data. All trusted labeled data and their corresponding trust levels are compiled into a medical dataset.

[0063] Optionally, the teacher model can be GPT-4, Claude 3, Llama 3 (70B / 405B), Med-PaLM 2, or a benchmark true value model manually verified by experts; this invention does not impose any limitations.

[0064] Optionally, the medical dataset can be stored in JSONL, TFRecord, Parquet, CSV, or a custom sparse tensor storage format; this invention does not impose any limitations on this.

[0065] Optionally, the medical information labeling may include ICD-10 diagnostic codes, SNOMED CT terminology labels, anatomical location labels, surgical grade labels, or drug dosage recommendations, which are not limited in this invention.

[0066] As can be seen, through the above optional embodiments, the sampling rationality of the labeled data of the student model is evaluated by the teacher model and the data is filtered into the database according to the mean confidence level. A closed-loop quality control system based on teacher supervision is established. By quantitatively evaluating the confidence level of the labeled data in multiple dimensions, the automatic filtering of low-quality samples is realized, thereby greatly improving the purity of the medical pre-training dataset and optimizing the convergence efficiency of subsequent model tuning.

[0067] Example 2 Please see Figure 2 , Figure 2 This is a schematic diagram of the structure of a medical dataset development system based on large model construction, as disclosed in an embodiment of the present invention. Figure 2 The described medical dataset development system based on a large model can be applied to data processing systems / data processing equipment / data processing servers (including local processing servers or cloud processing servers). For example... Figure 2 As shown, this medical dataset development system based on a large model can include: The acquisition module 201 is used to acquire multiple medical data to be processed.

[0068] The de-identification module 202 is used to identify and encrypt privacy data for each piece of medical data based on the de-identification algorithm, so as to obtain multiple de-identified medical data.

[0069] The annotation module 203 is used to annotate each desensitized medical data with medical information according to multiple preset student models, and obtain the corresponding annotated data.

[0070] The verification module 204 is used to sample and verify all labeled data according to the teacher model corresponding to the student model, and organize the labeled data into the medical dataset based on the verification results.

[0071] Optionally, the medical dataset can be a pre-trained dataset, an instruction fine-tuning dataset, and / or a test dataset for training large models.

[0072] As can be seen, the above-described embodiments of the invention obtain multiple de-identified medical data by acquiring medical data and performing privacy data identification and encryption. Then, student models are used for annotation, and teacher models are used for sampling and verification to construct a large-scale medical dataset. This effectively ensures privacy and security during large-scale medical information processing. Furthermore, the teacher-student model collaboration mechanism significantly improves the accuracy of annotation results and data cleaning efficiency, thereby providing high-quality and standardized data asset support for the robust training of subsequent large-scale medical models.

[0073] As an optional implementation, the de-identification module, based on a de-identification algorithm, performs privacy data identification and encryption on each piece of medical data, resulting in multiple de-identified medical data in various ways, including: For each piece of medical data, determine the modality dimension corresponding to that medical data; optionally, the modality dimension includes at least one of video modality, audio modality, tabular data modality, text modality, and sensor data modality; Determine the privacy recognition model corresponding to the modality dimension; The medical data is input into a privacy-identifying model to obtain multiple privacy-related data components in the output. Each piece of privacy data is encrypted to obtain the corresponding de-identified medical data.

[0074] As can be seen, through the above optional embodiments, by determining the modal dimension of medical data and matching the corresponding privacy recognition model to encrypt sensitive parts, a differentiated privacy protection strategy for cross-modal heterogeneous data such as video, audio and tables is realized. While achieving automated privacy recognition, the risk of sensitive information leakage in complex data environments is greatly reduced and the concurrent processing performance of the preprocessing system is optimized.

[0075] As an optional embodiment, the desensitization module determines the specific method of the privacy recognition model corresponding to the modality dimension, including: For each candidate privacy recognition model, calculate the set of input dimensions that the model can process; Determine whether the input dimension set includes all modal dimensions to obtain the first determination result; If the first judgment result is negative, the fit of the candidate privacy recognition model is set to the preset minimum value; If the first judgment result is yes, calculate the average data similarity between multiple historical recognition data corresponding to the candidate privacy recognition model and the medical data to obtain the fit of the candidate privacy recognition model; The candidate privacy recognition model with the highest fit is determined as the corresponding privacy recognition model.

[0076] As can be seen, through the above optional embodiments, the optimal privacy recognition model is determined by calculating the coverage of the modal dimension of the candidate model and the average similarity of historical data. The multi-dimensional fit evaluation mechanism ensures a high degree of fit between the recognition tool and the medical data to be processed, effectively avoiding the waste of computing resources caused by modal mismatch and significantly enhancing the recognition accuracy of the system when facing unknown medical application scenarios.

[0077] As an optional embodiment, the specific method by which the de-identification module encrypts each privacy data portion to obtain the de-identified medical data corresponding to the medical data includes: For each privacy data part, the privacy data part is input into the trained vectorization model to obtain the data vector corresponding to that privacy data part; Based on a preset vector encryption algorithm, the data vector is encrypted to obtain an encrypted vector; Determine the overlay data template corresponding to the data modality corresponding to the privacy data portion, and assemble the encryption vector and the overlay data template to obtain the replacement data corresponding to the privacy data portion; Each privacy data portion of the medical data is replaced with replacement data to obtain the corresponding de-identified medical data.

[0078] As can be seen, through the above optional embodiments, the privacy data is vectorized and encrypted, and then combined with the overlay data template to generate replacement data to complete the de-identification of medical data. Vector space encryption technology is used to achieve irreversible privacy masking at the physical level while maintaining the semantic logic of the data. Templated assembly ensures the integrity of the de-identified data structure, thereby providing high-fidelity data that balances security and feature availability for subsequent computing tasks.

[0079] As an optional implementation, the teacher model is trained using a training dataset that includes multiple training medical data and corresponding full-dimensional medical analysis annotations; the student model is obtained by distilling the teacher model based on different distillation processes.

[0080] As can be seen, the above optional embodiments limit the model training details of the teacher model and the student model to help realize the teacher-student model collaboration mechanism, which can significantly improve the accuracy of annotation results and data cleaning efficiency.

[0081] As an optional embodiment, the student model is a diagnosis-related prediction model, a treatment-related prediction model, a medication-related prediction model, or a prognosis-related prediction model; the medical information annotation includes at least one of diagnosis-related annotation, treatment-related annotation, medication-related annotation, or prognosis-related annotation.

[0082] As can be seen, the above optional embodiments limit the model prediction details and annotation results of the student model to help realize the teacher-student model collaboration mechanism, which can significantly improve the accuracy of the annotation results and the efficiency of data cleaning.

[0083] As an optional implementation, the annotation module annotates each de-identified medical data point with medical information based on multiple preset student models, and the specific method for obtaining the corresponding annotated data includes: For each de-identified medical data, the de-identified medical data is input into the pre-prediction model corresponding to each student model to obtain the label probability output by each pre-prediction model; optionally, the pre-prediction model is trained using a training dataset that includes multiple training medical data and the label probabilities of the corresponding student models for the training medical data. The student models corresponding to multiple pre-prediction models whose labeling probability is greater than a preset probability threshold are selected to obtain multiple target student models. The anonymized medical data is input into each target student model to obtain multiple medical information annotations corresponding to the anonymized medical data; Multiple medical information labels are deduplicated, and the deduplicated labeling results are assembled with the de-identified medical data to obtain the labeled data corresponding to the de-identified medical data.

[0084] As can be seen, through the above optional embodiments, the annotation probability is calculated based on the pre-prediction model and the target student model is selected to perform de-identification annotation on the anonymized data. The calculation task is dynamically allocated to the most suitable sub-model through the pre-probability evaluation mechanism. This not only greatly reduces the hardware resource overhead caused by redundant calculation of low-relevance models, but also significantly improves the consistency and logical rigor of complex medical knowledge annotation by utilizing de-duplication assembly logic.

[0085] As an optional implementation, the verification module performs sampling verification on all labeled data based on the teacher model corresponding to the student model, and organizes the labeled data into the medical dataset based on the verification results in the following specific ways: Sampling is performed on all labeled data to obtain multiple combinations of sampled data; optionally, the combination of sampled data includes multiple labeled data. Each sample data combination is input into the teacher model to obtain the corresponding labeling rationality. For each student model, the average of the labeling reasonableness of all labeled data corresponding to that student model is calculated to obtain the confidence level of that student model. All labeled data whose corresponding student model has a confidence level greater than a preset confidence level threshold are identified as reliable labeled data. All trusted labeled data and their corresponding trust levels are compiled into a medical dataset.

[0086] As can be seen, through the above optional embodiments, the sampling rationality of the labeled data of the student model is evaluated by the teacher model and the data is filtered into the database according to the mean confidence level. A closed-loop quality control system based on teacher supervision is established. By quantitatively evaluating the confidence level of the labeled data in multiple dimensions, the automatic filtering of low-quality samples is realized, thereby greatly improving the purity of the medical pre-training dataset and optimizing the convergence efficiency of subsequent model tuning.

[0087] Example 3 Please see Figure 3 , Figure 3 This is another medical dataset development system based on large model construction disclosed in the embodiments of the present invention. Figure 3 The described medical dataset development system based on a large model is applied to data processing systems / data processing equipment / data processing servers (including local processing servers or cloud processing servers). For example... Figure 3 As shown, this medical dataset development system based on a large model can include: Memory 301 storing executable program code; Processor 302 coupled to memory 301; The processor 302 calls the executable program code stored in the memory 301 to execute the steps of the medical dataset development method based on large model construction described in Embodiment 1.

[0088] Example 4 This invention discloses a computer read storage medium that stores a computer program for electronic data interchange, wherein the computer program causes a computer to execute the steps of the medical dataset development method based on large model construction described in Embodiment 1.

[0089] Example 5 This invention discloses a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program, and the computer program is operable to cause a computer to perform the steps of the medical dataset development method based on large model construction described in Embodiment 1.

[0090] The foregoing has described specific embodiments of this specification; other embodiments are within the scope of the appended claims. In some cases, the actions or steps described in the claims may be performed in a different order than those shown in the embodiments and may still achieve the desired result. Furthermore, the processes depicted in the drawings do not necessarily have to follow the specific or sequential order shown to achieve the desired result. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.

[0091] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions. A typical implementation device is a computer. Specifically, a computer can be, for example, a personal computer, laptop computer, cellular phone, camera phone, smartphone, personal digital assistant, media player, navigation device, email device, game console, tablet computer, wearable device, or any combination of these devices.

[0092] For ease of description, the above devices are described in terms of function, divided into various units. Of course, in implementing this specification, the functions of each unit can be implemented in one or more software and / or hardware components.

[0093] Those skilled in the art will understand that the embodiments of this specification can be provided as methods, systems, or computer program products. Therefore, the embodiments of this specification can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the embodiments of this specification can take the form of a computer program product implemented on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0094] This specification is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this specification. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0095] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0096] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0097] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0098] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0099] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0100] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0101] This specification can be described in the general context of computer-executable instructions that are executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform a specific task or implement a specific abstract data type. This specification can also be practiced in distributed computing environments, where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0102] The various embodiments in this specification are described in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

[0103] Finally, it should be noted that the medical dataset development method and system based on large model construction disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, not to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for developing medical datasets based on large model construction, characterized in that, The method includes: Acquire multiple medical data sets to be processed; Based on the de-identification algorithm, privacy data identification and encryption are performed on each of the aforementioned medical data to obtain multiple de-identified medical data; Based on multiple preset student models, each of the desensitized medical data is annotated with medical information to obtain the corresponding annotated data; Based on the teacher model corresponding to the student model, all the labeled data are sampled and verified, and the labeled data are organized into the medical dataset based on the verification results; the medical dataset is a pre-training dataset, instruction fine-tuning dataset and / or test dataset used for large model training.

2. The method for developing medical datasets based on large model construction according to claim 1, characterized in that, The aforementioned de-identification algorithm performs privacy data identification and encryption on each piece of medical data, resulting in multiple de-identified medical data sets, including: For each piece of medical data, the modal dimension corresponding to the medical data is determined; the modal dimension includes at least one of video modality, audio modality, tabular data modality, text modality, and sensor data modality; Determine the privacy recognition model corresponding to the modality dimension; The medical data is input into the privacy recognition model to obtain multiple privacy data components as output; Each of the privacy data portions is encrypted to obtain the de-identified medical data corresponding to the medical data.

3. The method for developing medical datasets based on large model construction according to claim 2, characterized in that, The step of determining the privacy recognition model corresponding to the modality dimension includes: For each candidate privacy recognition model, calculate the set of input dimensions that the model can process; Determine whether the input dimension set includes all the modal dimensions to obtain a first determination result; If the first judgment result is negative, the fit of the candidate privacy recognition model is set to the preset minimum value; If the first judgment result is yes, calculate the average value of the data similarity between multiple historical identification data corresponding to the candidate privacy identification model and the medical data to obtain the fit of the candidate privacy identification model; The candidate privacy recognition model with the highest fit is determined as the corresponding privacy recognition model.

4. The method for developing medical datasets based on large model construction according to claim 2, characterized in that, The process of encrypting each privacy data portion to obtain the de-identified medical data corresponding to the medical data includes: For each privacy data portion, the privacy data portion is input into the trained vectorization model to obtain the data vector corresponding to the privacy data portion; The data vector is encrypted based on a preset vector encryption algorithm to obtain an encrypted vector; Determine the overlay data template corresponding to the data modality corresponding to the privacy data portion, and assemble the encryption vector and the overlay data template to obtain the replacement data corresponding to the privacy data portion; Each of the privacy data portions of the medical data is replaced with the replacement data to obtain the de-identified medical data corresponding to the medical data.

5. The method for developing medical datasets based on large model construction according to claim 1, characterized in that, The teacher model is trained using a training dataset that includes multiple training medical data and corresponding full-dimensional medical analysis annotations; the student model is obtained by distilling the teacher model using different distillation processes.

6. The method for developing medical datasets based on large model construction according to claim 1, characterized in that, The student model is a diagnosis-related prediction model, a treatment-related prediction model, a medication-related prediction model, or a prognosis-related prediction model; the medical information annotation includes at least one of diagnosis-related annotation, treatment-related annotation, medication-related annotation, or prognosis-related annotation.

7. The method for developing medical datasets based on large model construction according to claim 1, characterized in that, The step involves labeling each de-identified medical data point with medical information based on multiple preset student models to obtain corresponding labeled data, including: For each de-identified medical data, the de-identified medical data is input into the pre-prediction model corresponding to each student model to obtain the labeling probability output by each pre-prediction model; the pre-prediction model is trained by a training dataset including multiple training medical data and the corresponding labeling probability of the student model for the training medical data. Multiple student models are obtained by filtering out the student models corresponding to multiple pre-prediction models whose labeled probability is greater than a preset probability threshold. The anonymized medical data is input into each of the target student models to obtain multiple medical information annotations corresponding to the anonymized medical data; The multiple medical information labels are deduplicated, and the deduplicated labeling results are assembled with the desensitized medical data to obtain the labeled data corresponding to the desensitized medical data.

8. The method for developing medical datasets based on large model construction according to claim 1, characterized in that, The step of sampling and validating all labeled data according to the teacher model corresponding to the student model, and organizing the labeled data into the medical dataset based on the validation results, includes: All the labeled data are sampled to obtain multiple sampled data combinations; each sampled data combination includes multiple labeled data. Each of the aforementioned sampling data combinations is input into the teacher model to obtain the corresponding labeling rationality. For each student model, the average of the reasonableness of the annotations corresponding to all the labeled data for that student model is calculated to obtain the credibility of that student model. All labeled data whose credibility corresponding to the student model is greater than a preset credibility threshold are identified as credible labeled data. All the aforementioned trusted labeled data and their corresponding trust levels are compiled into a medical dataset.

9. A medical dataset development system based on large model construction, characterized in that, The system includes: The acquisition module is used to acquire multiple medical data sets to be processed. The desensitization module is used to identify and encrypt privacy data for each piece of medical data based on the desensitization algorithm, resulting in multiple desensitized medical data. The annotation module is used to annotate each of the desensitized medical data with medical information according to multiple preset student models, so as to obtain the corresponding annotated data; The verification module is used to sample and verify all the labeled data according to the teacher model corresponding to the student model, and organize the labeled data into the medical dataset based on the verification results; the medical dataset is a pre-training dataset, instruction fine-tuning dataset and / or test dataset used for large model training.

10. A medical dataset development system based on large model construction, characterized in that, The system includes: Memory containing executable program code; A processor coupled to the memory; The processor calls the executable program code stored in the memory to execute the medical dataset development method based on large model construction as described in any one of claims 1-8.

Images