Cerebrovascular disease risk prediction method and device based on distributed biological sample library

Abstract

The embodiment of the invention discloses a cerebrovascular disease risk prediction method and device based on a distributed biological sample library. According to the embodiment of the invention, at least one target prediction end matched with a clinical feature vector can be determined from a plurality of candidate prediction ends based on a first collaborative prediction request sent by a prediction request end and local sample data features of the plurality of candidate prediction ends; sending a second collaborative prediction request to each target prediction end; receiving a second initial cerebrovascular disease risk probability returned by each target prediction end and a second prediction confidence corresponding to the second initial cerebrovascular disease risk probability; performing weighted fusion on the first prediction confidence coefficient and each second prediction confidence coefficient to obtain a comprehensive prediction confidence coefficient, and performing weighted fusion on the first initial cerebrovascular disease risk probability and each second initial cerebrovascular disease risk probability to obtain a target cerebrovascular disease risk probability; and sending the target cerebrovascular disease risk probability and the corresponding comprehensive prediction confidence to a prediction request end. Therefore, the accuracy and reliability of cerebrovascular disease risk prediction are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of medical artificial intelligence, in particular to a cerebrovascular disease risk prediction method and device based on a distributed biological sample library. BACKGROUND

[0002] Cerebrovascular diseases (such as ischemic stroke, cerebral hemorrhage, etc.) have the characteristics of high incidence, high disability rate and high mortality. Early risk prediction is of great significance for disease prevention and intervention. At present, the commonly used cerebrovascular disease risk prediction models (such as Framingham stroke risk score, QRISK, etc.) are mostly developed based on the data of patients treated in a hospital or research institution in the past. These data are usually collected retrospectively after the occurrence of the disease, only reflecting the health characteristics of a specific region, a specific population (such as a certain gender, age group or race), and lack sufficient diversity and representativeness.

[0003] With the development of artificial intelligence technology, some medical institutions have begun to use the electronic health records (such as physical examination results, medical history, test indicators, etc.) accumulated in the hospital to train localized machine learning models for individual risk assessment of patients. However, due to the obvious limitations of the patient samples that a single medical institution can obtain in terms of population coverage, clinical variable dimension and long-term follow-up integrity, the trained model is often difficult to adapt to external patients from different regions and different health backgrounds. When facing individuals with large differences in clinical characteristics and local training data, the model is prone to bias, resulting in unstable prediction results, decreased accuracy, and even misleading risk judgments. SUMMARY

[0004] The embodiments of the present application provide a cerebrovascular disease risk prediction method and device based on a distributed biological sample library, which overcomes the problem of insufficient generalization ability of single-center models due to sample limitations, and improves the accuracy and reliability of cerebrovascular disease risk prediction by safely cooperating with external prediction resources.

[0005] The embodiments of the present application provide a cerebrovascular disease risk prediction method based on a distributed biological sample library, the method comprising:

[0006] receiving a first collaborative prediction request sent by a prediction request end, the first collaborative prediction request carrying a clinical feature vector, a first initial cerebrovascular disease risk probability and its corresponding first prediction confidence, the clinical feature vector being extracted by the prediction request end based on the clinical feature data of a to-be-diagnosed object, and the first initial cerebrovascular disease risk probability and its corresponding first prediction confidence being generated by the prediction request end based on the clinical feature vector using a cerebrovascular disease risk prediction model deployed locally by the prediction request end;

[0007] Based on the first collaborative prediction request and the local sample data features of multiple candidate prediction ends, at least one target prediction end that matches the clinical feature vector is determined from the multiple candidate prediction ends. The local sample data features are used to characterize the data attributes of the local biobank used by the corresponding candidate prediction end when training its locally deployed cerebrovascular disease risk prediction model.

[0008] A second collaborative prediction request is sent to each target prediction end. The second collaborative prediction request carries a clinical feature vector, so that each target prediction end can generate a corresponding second initial cerebrovascular disease risk probability and its corresponding second prediction confidence based on the clinical feature vector and its locally deployed cerebrovascular disease risk prediction model.

[0009] Receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal;

[0010] The first prediction confidence and each of the second prediction confidences are weighted and fused to obtain the comprehensive prediction confidence, and the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities are weighted and fused to obtain the target cerebrovascular disease risk probability.

[0011] Send the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level to the prediction request end.

[0012] This application also provides a cerebrovascular disease risk prediction device based on a distributed biobank, the device comprising:

[0013] The request receiving unit is used to receive a first collaborative prediction request sent by the prediction request end. The first collaborative prediction request carries a clinical feature vector, a first initial cerebrovascular disease risk probability and its corresponding first prediction confidence. The clinical feature vector is extracted by the prediction request end based on the clinical feature data of the object to be diagnosed. The first initial cerebrovascular disease risk probability and its corresponding first prediction confidence are generated by the prediction request end based on the clinical feature vector and using its locally deployed cerebrovascular disease risk prediction model.

[0014] The matching unit is used to determine at least one target prediction end that matches the clinical feature vector from multiple candidate prediction ends based on the first collaborative prediction request and the local sample data features of multiple candidate prediction ends. The local sample data features are used to characterize the data attributes of the local biobank used by the corresponding candidate prediction end when training its locally deployed cerebrovascular disease risk prediction model.

[0015] The request sending unit is used to send a second collaborative prediction request to each target prediction end. The second collaborative prediction request carries a clinical feature vector so that each target prediction end can generate a corresponding second initial cerebrovascular disease risk probability and its corresponding second prediction confidence based on the clinical feature vector and its locally deployed cerebrovascular disease risk prediction model.

[0016] The data receiving unit is used to receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal.

[0017] The weighted fusion unit is used to fuse the first prediction confidence and each of the second prediction confidence to obtain the comprehensive prediction confidence, and to perform weighted fusion of the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities to obtain the target cerebrovascular disease risk probability.

[0018] The data sending unit is used to send the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence to the prediction request end.

[0019] This application also provides an electronic device, including a processor and a memory, wherein the memory stores multiple instructions; the processor loads instructions from the memory to execute the steps in any of the cerebrovascular disease risk prediction methods based on a distributed biobank provided in this application.

[0020] This application also provides a computer-readable storage medium storing multiple instructions adapted for loading by a processor to execute steps in any of the cerebrovascular disease risk prediction methods based on a distributed biobank provided in this application.

[0021] This application also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the steps in any of the cerebrovascular disease risk prediction methods based on a distributed biobank provided in this application.

[0022] In this application, upon receiving a first collaborative prediction request from a prediction requesting end, based on the clinical feature vector carried in the first collaborative prediction request and combined with the local sample data features of multiple candidate prediction ends, at least one target prediction end that best matches the clinical feature vector is selected. Since the local sample data features are used to characterize the data attributes of the biobank used by the corresponding candidate prediction end when training its locally deployed cerebrovascular disease risk prediction model, this matching mechanism ensures that the training data distribution of the selected target prediction end has a higher similarity to the clinical characteristics of the subject to be diagnosed, thereby ensuring that the locally deployed cerebrovascular disease risk prediction model of the target prediction end has stronger applicability and reliability in the current prediction task.

[0023] Furthermore, after sending a second collaborative prediction request to each target prediction endpoint, each target prediction endpoint can generate a corresponding second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level based on the clinical feature vector and its locally deployed cerebrovascular disease risk prediction model. Since the target prediction endpoints are selected through the aforementioned matching mechanism, their model outputs are more closely aligned with the clinical background of the subject to be diagnosed, thus making the prediction results more targeted and reliable.

[0024] Subsequently, after receiving the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level from each target prediction endpoint, a weighted fusion of the first prediction confidence level and each second prediction confidence level is obtained to arrive at a comprehensive prediction confidence level. Similarly, a weighted fusion of the first initial cerebrovascular disease risk probability and each second initial cerebrovascular disease risk probability is obtained to arrive at the target cerebrovascular disease risk probability. This fusion mechanism effectively integrates the prediction opinions from cerebrovascular disease risk prediction models deployed locally on multiple target prediction endpoints. While fully respecting the data sovereignty and model independence of each target prediction endpoint, it achieves high-quality, high-confidence collaborative decision-making. Finally, the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level are returned to the prediction request endpoint, enabling the request endpoint to obtain a target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level that combines high accuracy and high reliability. This supports more scientific and reliable clinical decision-making, significantly improving the effectiveness and safety of early warning of cerebrovascular diseases.

[0025] This overcomes the problem of insufficient generalization ability caused by the limited sample size of single-center models, and improves the accuracy and reliability of cerebrovascular disease risk prediction by safely coordinating external prediction resources. Attached Figure Description

[0026] To more clearly illustrate the technical solutions in the embodiments of this application, 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 this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating the cerebrovascular disease risk prediction method based on a distributed biobank provided in this application embodiment;

[0028] Figure 2 This is a schematic diagram of the cerebrovascular disease risk prediction device based on a distributed biobank provided in this application embodiment. Detailed Implementation

[0029] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0030] This application provides a method and apparatus for predicting the risk of cerebrovascular diseases based on a distributed biobank.

[0031] Specifically, the cerebrovascular disease risk prediction device based on a distributed biobank can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet, smart Bluetooth device, laptop, or personal computer (PC); the server can be a single server or a server cluster consisting of multiple servers.

[0032] In some embodiments, the cerebrovascular disease risk prediction device based on a distributed biobank can also be integrated into multiple electronic devices. For example, the cerebrovascular disease risk prediction device based on a distributed biobank can be integrated into multiple servers, and the cerebrovascular disease risk prediction method based on a distributed biobank of this application can be implemented by multiple servers.

[0033] In some embodiments, the server may also be implemented as a terminal.

[0034] It is understood that in the specific implementation of this application, data related to the clinical characteristics of the subject to be diagnosed and local biobanks are involved. When the embodiments of this application are applied to specific products or technologies, user permission or consent is required, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0035] The following sections provide detailed descriptions of each example. It should be noted that the sequence numbers of the following embodiments are not intended to limit the preferred order of the embodiments.

[0036] This embodiment provides a method for predicting the risk of cerebrovascular diseases based on a distributed biobank. This method is applicable to a central node with coordination capabilities, which can be a server, a cloud server, or a coordination device deployed in a trusted computing environment, etc. Figure 1 The specific process of this method for predicting the risk of cerebrovascular diseases based on a distributed biobank can be shown as follows:

[0037] 101. Receive the first collaborative prediction request sent by the prediction request end. The first collaborative prediction request carries a clinical feature vector, a first initial cerebrovascular disease risk probability and its corresponding first prediction confidence. The clinical feature vector is extracted by the prediction request end based on the clinical feature data of the object to be diagnosed. The first initial cerebrovascular disease risk probability and its corresponding first prediction confidence are generated by the prediction request end based on the clinical feature vector and using its locally deployed cerebrovascular disease risk prediction model.

[0038] Among them, the prediction requester refers to the party that initiates the collaborative prediction task, which is usually a medical institution or computing node that conducts initial screening or treatment of the subject to diagnosis, and has deployed a local cerebrovascular disease risk prediction model.

[0039] The first collaborative prediction request is a collaborative prediction trigger message sent from the prediction requesting end to the central node. It includes at least a clinical feature vector, a first initial cerebrovascular disease risk probability and its corresponding first prediction confidence level. It is used to request the central node to coordinate multiple external prediction ends to jointly participate in the cerebrovascular disease risk assessment of the subject to be diagnosed.

[0040] The subject of diagnosis refers to individuals who need to undergo cerebrovascular disease risk assessment, such as patients seeking medical treatment or people undergoing health checkups.

[0041] Clinical characteristic data refers to structured or semi-structured medical data related to the subject of diagnosis that can be used for cerebrovascular disease risk assessment, including but not limited to age, gender, blood pressure, blood glucose, blood lipids, smoking history, past medical history, family history, imaging indicators, and laboratory test results.

[0042] The clinical feature vector is a numerical vector extracted by the prediction requester based on clinical feature data through feature engineering or embedding encoding, and serves as the input to the cerebrovascular disease risk prediction model.

[0043] The cerebrovascular disease risk prediction model is a machine learning or statistical model deployed at various prediction endpoints (including prediction request endpoints and candidate prediction endpoints). It is used to output the probability of a subject developing a cerebrovascular disease (such as ischemic stroke, cerebral hemorrhage, etc.) within a specific future time window, based on the input clinical feature vector. This cerebrovascular disease risk prediction model can take any form, such as logistic regression, random forest, XGBoost, or neural networks, and the model structure at each prediction endpoint can be heterogeneous.

[0044] The initial cerebrovascular disease risk probability is calculated by the prediction requester using its locally deployed cerebrovascular disease risk prediction model based on clinical feature vectors, representing a preliminary risk estimate of the cerebrovascular disease risk of the subject to be diagnosed.

[0045] The first prediction confidence score is a reliability assessment value of the initial cerebrovascular disease risk probability output by the prediction requester, reflecting the credibility of the prediction result under the current input conditions. The confidence score can be generated based on model uncertainty estimation (such as Bayesian methods), calibration curves, historical performance statistics, etc.

[0046] 102. Based on the first collaborative prediction request and the local sample data features of multiple candidate prediction ends, determine at least one target prediction end that matches the clinical feature vector from the multiple candidate prediction ends. The local sample data features are used to characterize the data attributes of the local biobank used by the corresponding candidate prediction end when training its locally deployed cerebrovascular disease risk prediction model.

[0047] In this context, candidate prediction endpoints refer to other participating nodes besides the requesting node that possess local cerebrovascular disease risk prediction capabilities after receiving the first collaborative prediction request. Each of these nodes deploys an independently trained cerebrovascular disease risk prediction model and possesses a corresponding local biosample library. In the collaborative prediction process, the central node excludes the requesting node itself from all available prediction endpoints, treating the remaining prediction endpoints as candidate objects for subsequent matching and screening.

[0048] A local biobank refers to the collection of biomedical data possessed by each prediction endpoint (including prediction request endpoints and candidate prediction endpoints) for training its local cerebrovascular disease risk prediction model. This biobank typically contains historical patients' clinical characteristics, follow-up records, imaging data, laboratory test results, genetic information, etc., and is associated with specific medical institutions or research cohorts, possessing inherent attributes such as regionality, population specificity, and standardized data collection.

[0049] Data attributes are metadata information used to describe the overall characteristics of a local biobank, reflecting key quality and distribution characteristics during the construction and use of the biobank. Typical data attributes include, but are not limited to: population composition (such as age, gender, and ethnicity), geographical coverage, disease spectrum characteristics, completeness of clinical variables, missing data rate, follow-up duration, annotation consistency, and sample size.

[0050] Local sample data features are structured or vectorized representations of data attributes from the local biobank, pre-extracted by each candidate predictor and registered to the central node. These local sample data features characterize the data attributes of the local biobank that the corresponding candidate predictor relies on when training its locally deployed cerebrovascular disease risk prediction model, and serve as the basic input for the central node to perform matching calculations.

[0051] The target prediction endpoints refer to one or more candidate prediction endpoints that best match the current subject to diagnosis, selected by the central node after similarity assessment based on clinical feature vectors and the local sample data features of each candidate prediction endpoint. The local biobanks of these target prediction endpoints are highly similar to the subject to diagnosis in terms of population distribution or clinical feature coverage; therefore, their locally deployed cerebrovascular disease risk prediction models are expected to have higher applicability and predictive reliability in this prediction task.

[0052] In some embodiments, an intelligent screening mechanism based on data distribution similarity ensures that the external models participating in collaborative prediction are highly adapted to the target diagnostic object in terms of training data distribution, effectively avoiding prediction bias caused by domain shift. Compared with the traditional method of randomly selecting or fixing partners, this scheme significantly improves the relevance and credibility of external prediction results, providing high-quality input for subsequent weighted fusion, thereby enhancing the overall accuracy, generalization ability, and clinical applicability of cerebrovascular disease risk assessment. Based on the first collaborative prediction request and the local sample data features of multiple candidate prediction ends, at least one target prediction end matching the clinical feature vector is determined from multiple candidate prediction ends, including:

[0053] Based on the first collaborative prediction request, the prediction ends other than the prediction request end among the multiple prediction ends are identified as candidate prediction ends;

[0054] Determine the similarity between the local sample data features and clinical feature vectors of each candidate prediction endpoint;

[0055] Multiple candidate prediction ends are sorted according to similarity, and the top N candidate prediction ends are selected as the target prediction ends, where N is a positive integer.

[0056] In this context, the prediction endpoint refers to a medical institution or computing node that has deployed a local cerebrovascular disease risk prediction model and connected to the collaborative prediction network, such as the data platform of a top-tier hospital, regional medical center, or research institution. Each prediction endpoint has its own independent local biobank and trains its own cerebrovascular disease risk prediction model based on this biobank.

[0057] Similarity is used to measure the degree of matching between the distribution of clinical features of a subject to be diagnosed and the overall data distribution of a local sample library of a candidate prediction end. In specific implementation, the clinical feature vectors can first be mapped to a feature space (e.g., through a pre-trained medical embedding model or principal component analysis) to obtain a feature embedding representation with uniform dimensions; at the same time, the local sample data features of each candidate prediction end (such as statistical indicators such as median age, prevalence of hypertension, and coverage of blood lipid testing) are converted into sample library embedding representations of the same dimension; then, the cosine similarity, Euclidean distance, or correlation coefficient between the two are calculated as the matching basis.

[0058] For example, if the patient to be diagnosed is a 65-year-old male with a history of diabetes residing in a northern city, his clinical feature vector, after mapping, reflects characteristics such as "advanced age, metabolic abnormalities, and cold climate." Meanwhile, the local biobank of candidate predictor A mainly consists of young people from the south, with its embedding indicating a bias towards "younger age and no chronic diseases." Candidate predictor B, on the other hand, comes from a large stroke registry in the north, containing a large number of elderly diabetic patients. In this case, the calculated similarity between the embedding of candidate predictor B and the clinical feature vector of the patient to be diagnosed is significantly higher than that of candidate predictor A. Therefore, candidate predictor B is prioritized as the target predictor.

[0059] By employing an intelligent screening mechanism based on data distribution similarity, this approach ensures that the external models participating in collaborative prediction are highly adapted to the target diagnostic subjects in terms of training data distribution, effectively avoiding prediction bias caused by domain shift. Compared to traditional methods of randomly selecting or fixing partners, this solution significantly improves the relevance and reliability of external prediction results, providing high-quality input for subsequent weighted fusion, thereby enhancing the overall accuracy, generalization ability, and clinical applicability of cerebrovascular disease risk assessment.

[0060] In some embodiments, an embedding alignment mechanism enables precise matching across granularity (individual vs. group) and modality (clinical records vs. metadata), significantly improving the scientific rigor and rationality of target prediction selection. This not only enhances the targeting of collaborative prediction but also lays a high-quality input foundation for subsequent high-confidence fusion, ultimately effectively improving the overall accuracy and robustness of cerebrovascular disease risk assessment. It determines the similarity between the local sample data features and clinical feature vectors of each candidate prediction endpoint, including:

[0061] The clinical feature vectors are mapped to the feature space to obtain the feature embedding representation;

[0062] Obtain the local sample data features corresponding to each candidate prediction end, and convert the local sample data features into a sample library embedding representation of the same dimension as the feature embedding representation;

[0063] Calculate the similarity between the feature embedding representation and the embedding representation of each sample library.

[0064] Feature embedding refers to projecting the clinical feature vector of the subject to be diagnosed into a dense vector in a low-dimensional continuous vector space through a pre-defined mapping function (such as a fully connected neural network, principal component analysis, or medical knowledge graph embedding model), which is used to more effectively characterize the comprehensive clinical status and potential risk patterns of the subject to be diagnosed.

[0065] Sample library embedding representation refers to converting the local sample data features of a candidate prediction end (such as the average age of patients in the institution's biobank, the prevalence of hypertension, the completeness rate of blood lipid testing, the entropy value of regional distribution, and other statistical metadata) into a vector form with the same or compatible dimensions in the same vector space as the above feature embedding representation through the same or compatible mapping mechanism, so that the two can be compared in a unified semantic space.

[0066] Understandably, by mapping the clinical feature vector of the subject to be diagnosed to the local sample data features corresponding to the candidate prediction model into the same embedding space, it is possible to quantify, in a numerical and computable way, whether "the patient belongs to the typical population of a certain institution's sample database". Based on this, the similarity between the two can be calculated by using methods such as cosine similarity, Euclidean distance, or inner product, which can objectively reflect the applicability potential of the target prediction model for the patient.

[0067] 103. Send a second collaborative prediction request to each target prediction terminal. The second collaborative prediction request carries a clinical feature vector so that each target prediction terminal can generate a corresponding second initial cerebrovascular disease risk probability and its corresponding second prediction confidence based on the clinical feature vector and its locally deployed cerebrovascular disease risk prediction model.

[0068] The second collaborative prediction request is a collaborative reasoning trigger message sent by the central node to the selected target prediction end. It is used to request the target prediction end to perform local risk prediction based on the clinical feature vector of the current object to be diagnosed. The request includes at least the clinical feature vector and optionally also auxiliary information such as task identifier, encryption key or timestamp.

[0069] In some embodiments, the clinical feature vector is encrypted before transmission to protect patient privacy.

[0070] The second initial cerebrovascular disease risk probability refers to the preliminary risk estimate of the subject's potential cerebrovascular disease (such as ischemic stroke, cerebral hemorrhage, etc.) calculated by any target prediction terminal after receiving the second collaborative prediction request, using its locally deployed cerebrovascular disease risk prediction model and based on the clinical feature vectors it carries. This probability value is generated independently and reflects the risk judgment of the target prediction terminal under its own training data distribution.

[0071] The second prediction confidence score refers to the reliability assessment value of the second initial cerebrovascular disease risk probability output by the corresponding target prediction end, used to characterize the credibility of the prediction result under the current input conditions. The confidence score can be generated based on various methods such as internal model uncertainty (such as the variance of a Bayesian neural network), calibrated probability confidence intervals, accuracy-confidence consistency on the historical validation set, or sample library quality weighting. The higher the value, the more reliable the prediction.

[0072] In some embodiments, the core challenge of "data availability but invisibility" in multi-center medical AI collaboration is effectively addressed. This design enhances system compliance and user trust by sending a second collaborative prediction request to each target prediction endpoint, including:

[0073] The clinical feature vector is encrypted to obtain the encrypted clinical feature vector;

[0074] A second collaborative prediction request is generated based on the encrypted clinical feature vector;

[0075] Send a second collaborative prediction request to each target prediction endpoint.

[0076] Encrypted clinical feature vectors refer to ciphertext data generated after protecting clinical feature vectors using cryptographic methods (such as homomorphic encryption, symmetric encryption, asymmetric encryption, or secure multi-party computation protocols).

[0077] Understandably, although clinical feature vectors are structured representations extracted from the clinical feature data of the subject to be diagnosed by the prediction requester, they still contain a large amount of identifiable or sensitive personal health information (such as age, medical history, and laboratory indicators). If transmitted in plaintext over the network, there is a risk of interception, misuse, or reconstruction, potentially leading to privacy breaches of the subject to be diagnosed. Therefore, this application encrypts the clinical feature vectors before sending them to the target prediction end, generating encrypted clinical feature vectors. This encryption mechanism ensures that the target prediction end can only use the data under authorized conditions—for example, by performing inference directly on the ciphertext using a model that supports homomorphic encryption, or by restoring the plaintext in a trusted execution environment (such as a TEE) with secure decryption permissions before making predictions. Furthermore, the central node only accesses and forwards the encrypted clinical feature vectors throughout the entire collaborative prediction process, without accessing, decrypting, or storing any original plaintext patient data. Thus, the exposure path of sensitive information is effectively isolated, significantly improving the privacy and security of the data throughout its entire lifecycle while ensuring the normal operation of the multi-center collaborative prediction function. This mechanism effectively meets the core security requirement of "data usable but not visible" in cross-institutional collaboration in the medical and health field, and provides reliable technical support for the compliant deployment and large-scale application of cerebrovascular disease risk prediction systems.

[0078] 104. Receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal.

[0079] 105. The first prediction confidence and each of the second prediction confidences are weighted and fused to obtain the comprehensive prediction confidence, and the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities are weighted and fused to obtain the target cerebrovascular disease risk probability.

[0080] The overall prediction confidence level refers to a unified credibility index generated by weighted fusion based on the original confidence levels (i.e., the first prediction confidence level and each second prediction confidence level) provided by the prediction requester and each target predictioner, combined with the data quality, matching degree, or reliability weights of their corresponding models. This value is used to quantify the overall credibility of the final risk assessment result.

[0081] The target cerebrovascular disease risk probability refers to the final cerebrovascular disease risk estimate generated by fusing local and multiple external highly fitted predictive endpoints. It reflects the comprehensive probability of a potential diagnosed individual experiencing a cerebrovascular event (such as ischemic stroke or cerebral hemorrhage) within a specific future time window. This probability is not a simple average but a dynamically weighted result based on the prediction confidence of each participant and the quality of their sample database, thus possessing stronger generalization ability and clinical representativeness.

[0082] It is understandable that due to differences in the distribution of training data, sample quality, and applicable scenarios of models at different prediction endpoints, the reliability of their output initial risk probabilities and confidence levels also varies. If equal-weighted averaging or ignoring confidence levels is used for fusion, it is susceptible to interference from low-quality or mismatched nodes, leading to biased final results. This application, however, uses confidence levels as weights for weighted fusion, ensuring that highly reliable and well-fitting predictions dominate the final decision-making process. This effectively suppresses noise influence while preserving multi-source diversity, thereby improving the overall accuracy, robustness, and clinical interpretability of the predictions.

[0083] In some embodiments, the interference of low-quality prediction ends on the final result is effectively avoided, significantly improving the accuracy and robustness of the fusion result. Simultaneously, the comprehensive prediction confidence not only reflects the model's confidence level but also embeds data quality information, enabling clinicians to more comprehensively assess the credible basis of risk judgment, thereby supporting safer and more individualized intervention decisions. The comprehensive prediction confidence is obtained by weighted fusion of the first prediction confidence and each of the second prediction confidences, and the target cerebrovascular disease risk probability is obtained by weighted fusion of the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities, including:

[0084] Obtain the first sample quality score corresponding to the local sample data features of the prediction request end, and the second sample quality score corresponding to the local sample data features of each target prediction end.

[0085] The first sample quality score is multiplied by the first prediction confidence level to obtain the first composite weight, and each second sample quality score is multiplied by its corresponding second prediction confidence level to obtain multiple second composite weights.

[0086] The first composite weight and each of the second composite weights are normalized to obtain the first normalized weight and multiple second normalized weights;

[0087] The first initial cerebrovascular disease risk probability is weighted based on the first normalization weight to obtain the first weighted cerebrovascular disease risk probability, and the corresponding second initial cerebrovascular disease risk probability is weighted based on each second normalization weight to obtain multiple second weighted cerebrovascular disease risk probabilities.

[0088] The first weighted cerebrovascular disease risk probability and multiple second weighted cerebrovascular disease risk probabilities are fused to obtain the target cerebrovascular disease risk probability;

[0089] Based on the first normalized weight and each of the second normalized weights, the first prediction confidence and each of the second prediction confidence are weighted and summed to obtain the comprehensive prediction confidence.

[0090] The first sample quality score is used to quantify the overall data quality level of the local biobank on which the prediction request relies. Its value is calculated based on multiple preset quality evaluation dimensions (such as sample size, population diversity, completeness of clinical variables, follow-up duration, and labeling consistency), reflecting the reliability and representativeness of the model training data on this end.

[0091] The second sample quality score is similar to the first sample quality score, used to quantify the data quality of the local biobanks of the corresponding target prediction end, serving as an important basis for evaluating the reliability of its prediction results.

[0092] The first composite weight is obtained by multiplying the first sample quality score of the prediction request end with its first prediction confidence score, which comprehensively reflects the dual reliability of the end in the current task: "data quality × model confidence".

[0093] The second composite weight is obtained by multiplying the second sample quality score of each target prediction end with its corresponding second prediction confidence level, and is used to characterize the comprehensive credible contribution of the corresponding target prediction end in this collaboration.

[0094] The first normalized weight and the second normalized weight are weight coefficients obtained by normalizing all composite weights (including the first normalized weight and each second normalized weight) (such as softmax or linear normalization), ensuring that the sum of the weighting coefficients of each participant is 1, which facilitates subsequent probability fusion and confidence aggregation.

[0095] The first weighted probability of cerebrovascular disease risk refers to the weighted component obtained by multiplying the first initial probability of cerebrovascular disease risk by its corresponding normalized weight, representing the contribution of the prediction requester to the final result.

[0096] The second weighted cerebrovascular disease risk probability refers to the weighted component obtained by multiplying the second initial cerebrovascular disease risk probability of each target prediction end by its corresponding normalized weight, representing the contribution value of each external prediction end in the final result.

[0097] Understandably, traditional multi-model fusion methods typically rely solely on the confidence scores of model outputs or simple averaging, ignoring the quality differences in the training data behind the models. In medical scenarios, a high-confidence prediction derived from a small sample size, a single population, or a low-quality sample library lacking key variables may have a much lower actual reliability than expected. This application introduces a joint weighting mechanism of sample quality scores and prediction confidence scores, achieving an intelligent fusion strategy of "high-quality data + high-confidence output = high-weighted adoption."

[0098] In some embodiments, the core challenge of "model usable but data incomparable" in multi-center medical AI collaboration is effectively solved. Without sharing original data, intelligent collaboration based on data quality awareness is achieved, significantly improving the accuracy, robustness, and clinical feasibility of cerebrovascular disease risk prediction. Before obtaining the first sample quality score corresponding to the local sample data features of the prediction request end and the second sample quality score corresponding to the local sample data features of each target prediction end, the method further includes:

[0099] Based on multiple preset quality evaluation dimensions, the actual feature measurement value corresponding to each preset quality evaluation dimension is extracted from the local sample data features of each prediction end.

[0100] For each prediction end, the evaluation score of the actual feature metric corresponding to each preset quality evaluation dimension is determined based on the mapping relationship between the feature metric value and the preset score of the preset quality evaluation dimension.

[0101] The sample quality score is obtained by summing the evaluation scores of the actual feature measures corresponding to each preset quality evaluation dimension.

[0102] Obtain the first sample quality score corresponding to the local sample data features of the prediction request end, and the second sample quality score corresponding to the local sample data features of each target prediction end, including:

[0103] Obtain the first sample quality score corresponding to the local sample data features of the prediction request end and the second sample quality score corresponding to the local sample data features of each target prediction end from multiple sample quality scores.

[0104] The pre-defined quality assessment dimensions refer to a set of standardized indicators used to systematically evaluate the data quality of a local biobank. These dimensions are pre-defined by those skilled in the art according to medical research standards (such as STROBE, TRIPOD, etc.). Typical dimensions include, but are not limited to: sample size, population diversity (such as age / gender / geographical distribution entropy), completeness of clinical variable coverage, missing rate of key risk factors, follow-up duration, accuracy of event labeling, and consistency of data collection.

[0105] The actual feature measure refers to the specific quantitative value calculated for the characteristics of local sample data at a certain prediction end under a specific preset quality assessment dimension. For example, under the dimension of "completeness of clinical variable coverage", the actual feature measure can be "9 out of 10 core stroke risk factors are fully recorded, with a completeness rate of 90%".

[0106] The mapping relationship between the feature metrics of a preset quality evaluation dimension and the preset scores refers to the pre-established scoring rules or functions used to convert the actual feature metrics of a certain dimension into the corresponding evaluation scores. This mapping relationship can be a piecewise linear function, a lookup table method, a threshold grading system (e.g., 5 points for completeness ≥ 90%, 3 points for 80%~90%, and 1 point for < 80%), or a machine learning calibration model. Its purpose is to transform objective data quality into comparable numerical scores.

[0107] The sample quality score is a comprehensive quantitative assessment of the overall data quality of a local biobank for a prediction endpoint. It is obtained by summing (or weighted fusion) the scores of each preset quality evaluation dimension. The higher the score, the more reliable and representative the training data of the prediction endpoint is, and the more trustworthy the prediction results output by its local model are.

[0108] It is understandable that biobanks at different prediction endpoints (medical institutions) vary significantly in terms of construction standards, data governance levels, and clinical record standards. Treating all prediction endpoints equally in collaborative prediction without considering their underlying data quality could lead to models trained on low-quality biobanks having a misleading impact on the final results. This application introduces a sample quality score mechanism based on multidimensional quality assessment, enabling an objective, quantitative, and interpretable reliability evaluation of the prediction capabilities of each participant.

[0109] In some embodiments, by introducing a dynamic participation mechanism based on timeout detection, only high-quality prediction results that are actually usable and returned in a timely manner are integrated, while ensuring the timeliness of prediction. This effectively avoids system blockage or result distortion caused by node failure or communication anomalies, significantly improving the robustness, real-time performance, and fault tolerance of the collaborative prediction system. It ensures that highly reliable cerebrovascular disease risk assessment results can still be stably output in heterogeneous and unreliable multi-center environments, meeting the actual clinical needs for "fast, reliable, and uninterrupted" intelligent auxiliary decision-making. The system receives the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal, including:

[0110] The response status of each target prediction end is detected;

[0111] If at least one target prediction end fails to return the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence within the preset response time, the unresponsive target prediction end will be removed from the participation list of the current collaborative prediction task.

[0112] Receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by the remaining target prediction terminals that have responded.

[0113] The response status refers to the status indicator of whether the target prediction end has successfully completed local inference and returned a valid prediction result after receiving the second collaborative prediction request, including "responded" (successfully returned) or "not responded" (timeout, failure, or no return).

[0114] The preset response time refers to the maximum waiting time (e.g., 5 seconds, 10 seconds, etc.) set by the central node for this collaborative prediction task, used to determine whether the target prediction end has completed the calculation and feedback within a reasonable time. This time can be dynamically configured based on network conditions, model complexity, or historical performance.

[0115] Unresponsive target prediction ends refer to candidate target prediction ends that fail to return a valid second initial cerebrovascular disease risk probability and a second prediction confidence level to the central node within the preset response time. Their failure to respond may be due to reasons such as network interruption, insufficient computing resources, model abnormality or service unavailability.

[0116] The current collaborative prediction task participation list refers to the set of all target prediction ends selected in this prediction process. It is used to track the task execution status of each participant and serves as the input source for subsequent fusion calculations.

[0117] The target prediction end that has responded refers to the target prediction end that successfully returns a valid prediction result within the preset response time. Its output will be included in the weighted fusion stage and participate in the generation of the final risk assessment.

[0118] Understandably, in distributed medical collaboration scenarios spanning multiple institutions and regions, there are significant differences in the computing power, network stability, and system load of each target prediction endpoint, and some nodes may experience delays or failures. Forcing all nodes to respond will result in excessively high overall prediction latency, affecting clinical timeliness; while blindly using partially missing results for fusion may introduce bias or reduce confidence.

[0119] 106. Send the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level to the prediction request end.

[0120] In some embodiments, the clinical acceptability and decision support value of the cerebrovascular disease risk prediction system are significantly enhanced. Simultaneously, by highlighting key risk factors and their contribution, it can also assist doctors in providing personalized health education to patients or developing precise prevention strategies, promoting a shift from "passive treatment" to "active intervention," and improving overall public health prevention and control effectiveness. This involves sending the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level to the prediction request end, including:

[0121] Generate risk interpretation information, which includes several clinical features that have the greatest impact on the probability of target cerebrovascular disease risk and their contribution.

[0122] The risk interpretation information, along with the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level, are sent to the prediction request end.

[0123] Among them, risk interpretation information refers to interpretable auxiliary information used to explain the causes of the final cerebrovascular disease risk assessment results. It aims to reveal which clinical factors play a key role in the current high (or low) risk judgment, thereby enhancing the transparency and credibility of the model output.

[0124] Clinical characteristics refer to the original medical attributes of the subject to diagnosis, such as age, systolic blood pressure, history of diabetes, smoking status, total cholesterol level, and history of stroke. These characteristics constitute the input basis for risk prediction.

[0125] Contribution refers to the quantitative impact of a clinical feature on the probability of target cerebrovascular disease risk, which can be calculated using model interpretation techniques (such as SHAP value, LIME, gradient significance, feature ablation analysis, etc.). For example, a contribution of +0.15 for the "age" feature means that this feature increases the probability of risk by 15 percentage points.

[0126] Understandably, while deep learning or ensemble models exhibit high accuracy in predicting cerebrovascular disease risk, their "black box" nature often makes it difficult for clinicians to trust or accept the prediction results. This is especially true in high-risk warning scenarios, where the inability to explain "why a high risk was determined" severely limits the clinical applicability of AI systems. This application addresses this by automatically generating clinically interpretable feedback along with the final prediction result, enabling the requesting end (such as a doctor's workstation or health management platform) to not only know "how high the risk is" but also "where the risk comes from." This not only enhances the model's transparency and traceability but also allows doctors to cross-validate or adjust the prediction results based on their professional knowledge.

[0127] In some embodiments, while ensuring the continuity of the prediction service, the risk of misjudgment due to missing data or model incompatibility is effectively prevented. Data supplementation prompts promote the completeness of clinical information collection, forming a virtuous cycle of "assessment-feedback-optimization." Ultimately, this enhances the practicality and credibility of cerebrovascular disease risk prediction in real medical environments. The target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level are sent to the prediction requester, including:

[0128] When the overall prediction confidence is less than the preset confidence threshold, an unreliable prompt is generated for the target cerebrovascular disease risk probability, as well as a data supplement prompt for clinical feature data.

[0129] Send unreliability alerts and data supplementation alerts to the prediction requester;

[0130] When the overall prediction confidence level is not less than the preset confidence threshold, the target cerebrovascular disease risk probability and its corresponding overall prediction confidence level are sent to the prediction request end.

[0131] The preset reliability threshold refers to a confidence threshold (e.g., 0.7, 0.8, etc.) set by the system or configured by the user to determine whether the current collaborative prediction results meet the clinically usable reliability standard. This threshold can be dynamically adjusted according to the application scenario (e.g., screening vs. diagnosis), disease severity, or institutional strategy.

[0132] An unreliable warning is an automatic alert generated by the system when the overall prediction confidence level falls below a preset confidence threshold. It informs the requesting client that "the current risk assessment results lack sufficient confidence and are not recommended for direct use in clinical decision-making." For example, it might say, "The risk assessment results have low confidence; please use them with caution."

[0133] Data supplementation suggestions refer to further analyzing missing or low-quality clinical feature dimensions when the confidence level is insufficient, and recommending which key information to supplement to improve the reliability of the prediction. For example: "It is recommended to supplement carotid ultrasound examination results" or "The lack of blood lipid test data may affect the accuracy of the assessment."

[0134] Understandably, in real-world clinical scenarios, not all individuals seeking diagnosis can provide complete and high-quality clinical characteristic data. Forcing high confidence risk values ​​onto cases with incomplete information or significant distributional biases may lead to misdiagnosis, causing unnecessary anxiety or missed diagnoses. Traditional systems often provide immediate answers to every question, lacking a self-assessment and feedback mechanism for their predictive reliability. This application introduces a dynamic response strategy based on comprehensive predictive confidence, achieving an intelligent interactive mode that outputs results for high confidence and guides completion for low confidence. This not only avoids the indiscriminate use of low-quality predictive results but also proactively provides doctors or patients with clear guidance on improving data input.

[0135] In some embodiments, while sending unreliable alerts and supplementary data alerts to the prediction requester, a preset strategy may be used to determine whether to send the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level together.

[0136] The preset strategy allows for sending only unreliable alerts and supplementary data to avoid low-confidence results interfering with clinical judgment. Alternatively, it can simultaneously send the target cerebrovascular disease risk probability and the overall predictive confidence level for professionals to use in conjunction with the context.

[0137] In summary, the embodiments of this application overcome the problem of insufficient generalization ability caused by the limited sample size of single-center models, and improve the accuracy and reliability of cerebrovascular disease risk prediction by securely coordinating external prediction resources.

[0138] To better implement the above methods, this application also provides a cerebrovascular disease risk prediction device based on a distributed biobank. This device can be integrated into an electronic device, such as a terminal or server. The terminal can be a mobile phone, tablet computer, smart Bluetooth device, laptop computer, or personal computer; the server can be a single server or a server cluster consisting of multiple servers.

[0139] For example, in this embodiment, a cerebrovascular disease risk prediction device based on a distributed biobank will be specifically integrated into an electronic device to illustrate the method of this application embodiment in detail.

[0140] For example, such as Figure 2 As shown, the cerebrovascular disease risk prediction device based on a distributed biobank may include a request receiving unit 201, a matching unit 202, a request sending unit 203, a data receiving unit 204, a weighted fusion unit 205, and a data sending unit 206, as follows:

[0141] (a) Request receiving unit 201.

[0142] The request receiving unit 201 is used to receive a first collaborative prediction request sent by the prediction request end. The first collaborative prediction request carries a clinical feature vector, a first initial cerebrovascular disease risk probability and its corresponding first prediction confidence. The clinical feature vector is extracted by the prediction request end based on the clinical feature data of the object to be diagnosed. The first initial cerebrovascular disease risk probability and its corresponding first prediction confidence are generated by the prediction request end based on the clinical feature vector and using its locally deployed cerebrovascular disease risk prediction model.

[0143] (ii) Matching unit 202.

[0144] The matching unit 202 is used to determine at least one target prediction end that matches the clinical feature vector from multiple candidate prediction ends based on the first collaborative prediction request and the local sample data features of multiple candidate prediction ends. The local sample data features are used to characterize the data attributes of the local biobank used by the corresponding candidate prediction end when training its locally deployed cerebrovascular disease risk prediction model.

[0145] In some embodiments, based on a first collaborative prediction request and local sample data features of multiple candidate prediction endpoints, determining at least one target prediction endpoint that matches a clinical feature vector from multiple candidate prediction endpoints includes:

[0146] Based on the first collaborative prediction request, the prediction ends other than the prediction request end among the multiple prediction ends are identified as candidate prediction ends;

[0147] Determine the similarity between the local sample data features and clinical feature vectors of each candidate prediction endpoint;

[0148] Multiple candidate prediction ends are sorted according to similarity, and the top N candidate prediction ends are selected as the target prediction ends, where N is a positive integer.

[0149] In some embodiments, determining the similarity between local sample data features and clinical feature vectors for each candidate prediction endpoint includes:

[0150] The clinical feature vectors are mapped to the feature space to obtain the feature embedding representation;

[0151] Obtain the local sample data features corresponding to each candidate prediction end, and convert the local sample data features into a sample library embedding representation of the same dimension as the feature embedding representation;

[0152] Calculate the similarity between the feature embedding representation and the embedding representation of each sample library.

[0153] (III) Request Sending Unit 203.

[0154] The request sending unit 203 is used to send a second collaborative prediction request to each target prediction end. The second collaborative prediction request carries a clinical feature vector, so that each target prediction end can generate a corresponding second initial cerebrovascular disease risk probability and its corresponding second prediction confidence based on the clinical feature vector and its locally deployed cerebrovascular disease risk prediction model.

[0155] In some embodiments, sending a second collaborative prediction request to each target prediction endpoint includes:

[0156] The clinical feature vector is encrypted to obtain the encrypted clinical feature vector;

[0157] A second collaborative prediction request is generated based on the encrypted clinical feature vector;

[0158] Send a second collaborative prediction request to each target prediction endpoint.

[0159] (iv) Data receiving unit 204.

[0160] The data receiving unit 204 is used to receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction end.

[0161] In some embodiments, receiving the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal includes:

[0162] The response status of each target prediction end is detected;

[0163] If at least one target prediction end fails to return the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence within the preset response time, the unresponsive target prediction end will be removed from the participation list of the current collaborative prediction task.

[0164] Receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by the remaining target prediction terminals that have responded.

[0165] (v) Weighted fusion unit 205.

[0166] The weighted fusion unit 205 is used to fuse the first prediction confidence and each of the second prediction confidence to obtain a comprehensive prediction confidence, and to perform weighted fusion of the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities to obtain the target cerebrovascular disease risk probability.

[0167] In some embodiments, a weighted fusion of the first prediction confidence and each of the second prediction confidences is performed to obtain a comprehensive prediction confidence, and a weighted fusion of the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities is performed to obtain a target cerebrovascular disease risk probability, including:

[0168] Obtain the first sample quality score corresponding to the local sample data features of the prediction request end, and the second sample quality score corresponding to the local sample data features of each target prediction end.

[0169] The first sample quality score is multiplied by the first prediction confidence level to obtain the first composite weight, and each second sample quality score is multiplied by its corresponding second prediction confidence level to obtain multiple second composite weights.

[0170] The first composite weight and each of the second composite weights are normalized to obtain the first normalized weight and multiple second normalized weights;

[0171] The first initial cerebrovascular disease risk probability is weighted based on the first normalization weight to obtain the first weighted cerebrovascular disease risk probability, and the corresponding second initial cerebrovascular disease risk probability is weighted based on each second normalization weight to obtain multiple second weighted cerebrovascular disease risk probabilities.

[0172] The first weighted cerebrovascular disease risk probability and multiple second weighted cerebrovascular disease risk probabilities are fused to obtain the target cerebrovascular disease risk probability;

[0173] Based on the first normalized weight and each of the second normalized weights, the first prediction confidence and each of the second prediction confidence are weighted and summed to obtain the comprehensive prediction confidence.

[0174] In some embodiments, before obtaining the first sample quality score corresponding to the local sample data features of the prediction request end and the second sample quality score corresponding to the local sample data features of each target prediction end, the method further includes:

[0175] Based on multiple preset quality evaluation dimensions, the actual feature measurement value corresponding to each preset quality evaluation dimension is extracted from the local sample data features of each prediction end.

[0176] For each prediction end, the evaluation score of the actual feature metric corresponding to each preset quality evaluation dimension is determined based on the mapping relationship between the feature metric value and the preset score of the preset quality evaluation dimension.

[0177] The sample quality score is obtained by summing the evaluation scores of the actual feature measures corresponding to each preset quality evaluation dimension.

[0178] Obtain the first sample quality score corresponding to the local sample data features of the prediction request end, and the second sample quality score corresponding to the local sample data features of each target prediction end, including:

[0179] Obtain the first sample quality score corresponding to the local sample data features of the prediction request end and the second sample quality score corresponding to the local sample data features of each target prediction end from multiple sample quality scores.

[0180] (vi) Data transmission unit 206.

[0181] The data sending unit 206 is used to send the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence to the prediction request end.

[0182] In some embodiments, sending the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level to the prediction request terminal includes:

[0183] Generate risk interpretation information, which includes several clinical features that have the greatest impact on the probability of target cerebrovascular disease risk and their contribution.

[0184] The risk interpretation information, along with the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level, are sent to the prediction request end.

[0185] In some embodiments, sending the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level to the prediction request terminal includes:

[0186] When the overall prediction confidence is less than the preset confidence threshold, an unreliable prompt is generated for the target cerebrovascular disease risk probability, as well as a data supplement prompt for clinical feature data.

[0187] Send unreliability alerts and data supplementation alerts to the prediction requester;

[0188] When the overall prediction confidence level is not less than the preset confidence threshold, the target cerebrovascular disease risk probability and its corresponding overall prediction confidence level are sent to the prediction request end.

[0189] In practice, each of the above units can be implemented as an independent entity or can be arbitrarily combined to be implemented as the same or several entities. For the specific implementation of each of the above units, please refer to the previous method embodiments, which will not be repeated here.

[0190] Therefore, the embodiments of this application overcome the problem of insufficient generalization ability caused by the limited sample size of single-center models, and improve the accuracy and reliability of cerebrovascular disease risk prediction by securely coordinating external prediction resources.

[0191] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be performed by instructions, or by instructions controlling related hardware. These instructions can be stored in a computer-readable storage medium and loaded and executed by a processor.

[0192] Therefore, embodiments of this application provide a computer-readable storage medium storing a plurality of instructions that can be loaded by a processor to execute steps in any of the cerebrovascular disease risk prediction methods based on a distributed biobank provided in embodiments of this application.

[0193] The storage medium may include: read-only memory (ROM), random access memory (RAM), disk or optical disk, etc.

[0194] Since the instructions stored in the storage medium can execute the steps in any of the trash rack cleaning path planning methods provided in the embodiments of this application, the beneficial effects that any of the cerebrovascular disease risk prediction methods based on distributed biobanks provided in the embodiments of this application can achieve can be realized. For details, please refer to the previous embodiments, which will not be repeated here.

[0195] According to one aspect of this application, a computer program product or computer program is provided, comprising a computer program / instructions stored in a computer-readable storage medium. A processor of an electronic device reads the computer program / instructions from the computer-readable storage medium and executes the computer program / instructions, causing the electronic device to perform the method provided in the above embodiments regarding cerebrovascular disease risk prediction based on a distributed biobank.

[0196] The foregoing has provided a detailed description of a method and apparatus for predicting the risk of cerebrovascular disease based on a distributed biobank, as provided in the embodiments of this application. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the method and its core ideas. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this application. Therefore, the content of this specification should not be construed as a limitation of this application.

Claims

1. A method for predicting the risk of cerebrovascular disease based on a distributed biobank, characterized in that, The method includes: The system receives a first collaborative prediction request sent by the prediction request terminal. The first collaborative prediction request carries a clinical feature vector, a first initial cerebrovascular disease risk probability and its corresponding first prediction confidence. The clinical feature vector is extracted by the prediction request terminal based on the clinical feature data of the object to be diagnosed. The first initial cerebrovascular disease risk probability and its corresponding first prediction confidence are generated by the prediction request terminal based on the clinical feature vector using its locally deployed cerebrovascular disease risk prediction model. Based on the first collaborative prediction request and the local sample data features of multiple candidate prediction ends, at least one target prediction end that matches the clinical feature vector is determined from the multiple candidate prediction ends. The local sample data features are used to characterize the data attributes of the local biobank used by the corresponding candidate prediction end when training its locally deployed cerebrovascular disease risk prediction model. A second collaborative prediction request is sent to each target prediction terminal. The second collaborative prediction request carries the clinical feature vector, so that each target prediction terminal can generate a corresponding second initial cerebrovascular disease risk probability and its corresponding second prediction confidence based on the clinical feature vector and its locally deployed cerebrovascular disease risk prediction model. Receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal; The first prediction confidence and each of the second prediction confidences are weighted and fused to obtain the comprehensive prediction confidence, and the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities are weighted and fused to obtain the target cerebrovascular disease risk probability. The target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level are sent to the prediction request terminal.

2. The method as described in claim 1, characterized in that, The step of determining at least one target prediction endpoint that matches the clinical feature vector from multiple candidate prediction endpoints based on the local sample data features of the first collaborative prediction request and multiple candidate prediction endpoints includes: Based on the first collaborative prediction request, prediction ends other than the prediction request end among multiple prediction ends are determined as candidate prediction ends; Determine the similarity between the local sample data features of each candidate prediction end and the clinical feature vector; The multiple candidate prediction ends are sorted according to the similarity, and the top N candidate prediction ends are selected as the target prediction ends, where N is a positive integer.

3. The method as described in claim 2, characterized in that, Determining the similarity between the local sample data features of each candidate prediction endpoint and the clinical feature vector includes: The clinical feature vectors are mapped to a feature space to obtain a feature embedding representation; Obtain the local sample data features corresponding to each candidate prediction end, and convert the local sample data features into a sample library embedding representation of the same dimension as the feature embedding representation; Calculate the similarity between the feature embedding representation and each of the sample library embedding representations.

4. The method as described in claim 1, characterized in that, The weighted fusion of the first prediction confidence and each of the second prediction confidences to obtain a comprehensive prediction confidence, and the weighted fusion of the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities to obtain a target cerebrovascular disease risk probability, include: Obtain the first sample quality score corresponding to the local sample data features of the prediction request terminal, and the second sample quality score corresponding to the local sample data features of each target prediction terminal. The first sample quality score is multiplied by the first prediction confidence to obtain the first composite weight, and each second sample quality score is multiplied by its corresponding second prediction confidence to obtain multiple second composite weights. The first composite weight and each of the second composite weights are normalized to obtain the first normalized weight and multiple second normalized weights; The first initial cerebrovascular disease risk probability is weighted based on the first normalized weight to obtain the first weighted cerebrovascular disease risk probability, and the corresponding second initial cerebrovascular disease risk probability is weighted based on each second normalized weight to obtain multiple second weighted cerebrovascular disease risk probabilities. The first weighted cerebrovascular disease risk probability and the plurality of second weighted cerebrovascular disease risk probabilities are fused to obtain the target cerebrovascular disease risk probability; Based on the first normalized weight and each of the second normalized weights, the first prediction confidence and each of the second prediction confidence are weighted and summed to obtain the comprehensive prediction confidence.

5. The method as described in claim 4, characterized in that, Before obtaining the first sample quality score corresponding to the local sample data features of the prediction request terminal and the second sample quality score corresponding to the local sample data features of each target prediction terminal, the method further includes: Based on multiple preset quality evaluation dimensions, the actual feature measurement value corresponding to each preset quality evaluation dimension is extracted from the local sample data features of each prediction end. For each prediction end, the evaluation score of the actual feature metric corresponding to each preset quality evaluation dimension is determined based on the mapping relationship between the feature metric value and the preset score of the preset quality evaluation dimension. The sample quality score is obtained by summing the evaluation scores of the actual feature measurement values ​​corresponding to each preset quality evaluation dimension. The step of obtaining the first sample quality score corresponding to the local sample data features of the prediction request terminal and the second sample quality score corresponding to the local sample data features of each target prediction terminal includes: The first sample quality score corresponding to the local sample data features of the prediction request terminal and the second sample quality score corresponding to the local sample data features of each target prediction terminal are obtained from multiple sample quality scores.

6. The method as described in claim 1, characterized in that, Sending the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level to the prediction request terminal includes: Generate risk interpretation information, which includes several clinical features that have the greatest impact on the probability of the target cerebrovascular disease risk and their contribution; The risk interpretation information, along with the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level, are sent to the prediction request terminal.

7. The method as described in claim 1, characterized in that, The process of receiving the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal includes: The response status of each target prediction end is detected; If at least one target prediction end fails to return the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence within the preset response time, the unresponsive target prediction end will be removed from the participation list of the current collaborative prediction task. Receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by the remaining target prediction terminals that have responded.

8. The method as described in claim 1, characterized in that, Sending a second collaborative prediction request to each target prediction endpoint includes: The clinical feature vector is encrypted to obtain an encrypted clinical feature vector; A second collaborative prediction request is generated based on the encrypted clinical feature vector; Send a second collaborative prediction request to each target prediction endpoint.

9. The method as described in claim 1, characterized in that, Sending the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence level to the prediction request terminal includes: When the overall prediction confidence is less than the preset confidence threshold, an unreliable prompt is generated for the target cerebrovascular disease risk probability, and a data supplement prompt is generated for the clinical feature data. Send the unreliability alert and the data supplementation alert to the prediction request terminal; When the overall prediction confidence level is not less than the preset confidence threshold, the target cerebrovascular disease risk probability and its corresponding overall prediction confidence level are sent to the prediction request terminal.

10. A cerebrovascular disease risk prediction device based on a distributed biobank, characterized in that, The device includes: The request receiving unit is used to receive a first collaborative prediction request sent by the prediction requesting end. The first collaborative prediction request carries a clinical feature vector, a first initial cerebrovascular disease risk probability and its corresponding first prediction confidence. The clinical feature vector is extracted by the prediction requesting end based on the clinical feature data of the object to be diagnosed. The first initial cerebrovascular disease risk probability and its corresponding first prediction confidence are generated by the prediction requesting end based on the clinical feature vector and using its locally deployed cerebrovascular disease risk prediction model. A matching unit is configured to determine at least one target prediction end that matches the clinical feature vector from multiple candidate prediction ends based on the first collaborative prediction request and the local sample data features of multiple candidate prediction ends. The local sample data features are used to characterize the data attributes of the local biobank used by the corresponding candidate prediction end when training its locally deployed cerebrovascular disease risk prediction model. The request sending unit is used to send a second collaborative prediction request to each target prediction end. The second collaborative prediction request carries the clinical feature vector, so that each target prediction end can generate a corresponding second initial cerebrovascular disease risk probability and its corresponding second prediction confidence based on the clinical feature vector and its locally deployed cerebrovascular disease risk prediction model. The data receiving unit is used to receive the second initial cerebrovascular disease risk probability and its corresponding second prediction confidence level returned by each target prediction terminal; The weighted fusion unit is used to fuse the first prediction confidence and each of the second prediction confidence to obtain a comprehensive prediction confidence, and to perform weighted fusion of the first initial cerebrovascular disease risk probability and each of the second initial cerebrovascular disease risk probabilities to obtain a target cerebrovascular disease risk probability. The data sending unit is used to send the target cerebrovascular disease risk probability and its corresponding comprehensive prediction confidence to the prediction request terminal.

Images