A domain adaptation method and system for predicting hypertension

Abstract

The application discloses a kind of domain self-adapting hypertension prediction method and system.The method obtains the basic feature vector of the object to be measured and inputs basic prediction model, and outputs basic hypertension prediction probability vector.Clinical prior dynamic factor of the object to be measured is collected simultaneously and is mapped into probability weight vector.Further, the basic hypertension prediction probability vector and the probability weight vector are linearly weighted and fused to generate a fusion probability vector.Finally, the hypertension prediction diagnosis result is determined according to the fusion probability vector.The basic prediction model uses cascaded long short-term memory network layer and transformer encoder layer, and the grid search mechanism based on private validation set is introduced in the fusion process to dynamically optimize the weight coefficient.The application realizes the organic combination of data-driven model and clinical prior knowledge, significantly improves the accuracy, generalization ability and domain self-adapting optimization performance of hypertension prediction.

Description

Technical Field

[0001] This invention relates to the field of interdisciplinary technology of artificial intelligence and medical health, specifically to a domain-adaptive method and system for predicting hypertension. Background Technology

[0002] Hypertension, a common chronic cardiovascular disease, is a major risk factor for serious complications such as cardiovascular and cerebrovascular events and kidney failure. Early and accurate risk prediction and diagnosis of hypertension are of significant clinical importance for developing personalized intervention strategies and reducing the incidence of cardiovascular disease. With the rapid development of medical informatics and artificial intelligence technologies, the intelligent prediction of hypertension risk using machine learning and deep learning algorithms has become a research hotspot in the current medical and health field.

[0003] In the evolution of existing hypertension prediction technologies, early research mainly relied on traditional machine learning models (such as decision trees, random forests, support vector machines, and logistic regression). These methods typically perform classification and prediction based on collected basic physical examination or demographic characteristics. However, traditional machine learning models are highly dependent on manual feature engineering and struggle to automatically capture the implicit relationships between high-dimensional features in complex medical data, thus significantly limiting their fitting ability and generalization performance.

[0004] To overcome the limitations of traditional machine learning, deep learning-based hypertension prediction schemes have been widely applied in recent years. Recurrent Neural Networks (RNNs) and their variant, Long Short-Term Memory (LSTM), have been introduced to model medical feature data, effectively capturing short-range dependencies and local evolutionary patterns in feature sequences. On the other hand, the Transformer architecture, with its unique multi-head self-attention mechanism, demonstrates significant advantages in capturing global cross-feature interactions and long-range correlations between data. However, existing deep learning prediction models generally construct LSTM or Transformer as a single, independent architecture, resulting in inherent deficiencies in feature representation capabilities: a single LSTM architecture lacks effective modeling of global cross-feature interactions, while a single Transformer architecture is inefficient in extracting short-range correlations of local features. Neither can simultaneously consider the local temporal evolution and global structural correlations of features.

[0005] More importantly, existing hypertension prediction models face severe technical bottlenecks in practical clinical applications and cross-scenario deployments, specifically in the following aspects: First, existing technologies fix and close the input feature space after model training. During inference or clinical application, the model cannot flexibly incorporate or integrate new clinical variables, such as family genetic risk, dynamic lifestyle habits, and real-time sleep quality. This fixed input space makes it difficult for the model to adapt to differentiated diagnostic needs and clinical contexts when deployed to different medical institutions or specific clinical scenarios, resulting in a severe deficiency in the model's domain adaptability.

[0006] Secondly, existing deep learning models typically employ an end-to-end black-box prediction mechanism. The model directly outputs classification results or risk probabilities but fails to provide clear decision-making criteria, making it difficult for doctors to intuitively understand the specific contribution of each input feature or clinical factor to the final prediction result. This lack of interpretability in the decision-making process significantly limits the credibility and acceptance of deep learning-assisted diagnostic systems in serious clinical scenarios.

[0007] Furthermore, when faced with real-world scenarios where there are differences in data distribution between private and public clinical data, i.e., domain shifts, existing technologies lack effective mechanisms to smoothly integrate clinical prior factors for specific scenarios without sacrificing the generalization accuracy of the original basic model. This results in poor feature scalability and predictive stability of the prediction model on private clinical datasets.

[0008] These factors mean that existing hypertension prediction programs, when faced with complex individual differences in clinical practice, often suffer from inaccurate predictions and high false-positive rates due to the lack of effective integration of key genetic and lifestyle factors. This high false-positive rate not only increases unnecessary clinical follow-up costs but may also impose unnecessary psychological burdens on examinees, failing to meet the actual needs of high-precision clinical screening and diagnosis.

[0009] In summary, existing hypertension prediction technologies still have significant shortcomings in terms of the comprehensiveness of feature extraction capabilities, dynamic fusion and domain adaptation capabilities of clinical prior factors, and interpretability of prediction results. There is an urgent need for a hypertension prediction method and system that can take into account efficient feature representation, dynamic clinical factor fusion capabilities, transparent and interpretable prediction decisions, and greater accuracy and stability. Summary of the Invention

[0010] The purpose of this invention is to overcome the technical defects of existing hypertension prediction technologies, such as incomplete feature extraction capabilities, fixed input space leading to the inability to dynamically integrate clinical prior factors, black-box model lack of interpretability, high false positive rate, and insufficient prediction accuracy and domain adaptability, and to provide a domain-adaptive hypertension prediction method and system.

[0011] To achieve the aforementioned objectives, the technical solution of this invention revolves around a two-layer architecture of basic model feature extraction and dynamic fusion of clinical prior factors. First, a cascaded architecture of a Long Short-Term Memory (LSTM) network and a Transformer encoder is constructed. This architecture utilizes the LSTM layer to model short-range dependencies and local evolutionary patterns in the input feature sequence, and leverages the multi-head self-attention mechanism in the Transformer encoder layer to capture global cross-feature interactions and long-range structural associations. The two complement each other to jointly represent basic physical examination and demographic characteristics. Second, this invention also designs a domain-adaptive percentage calibration and quantified weighted fusion mechanism based on clinical prior factors. A structured fusion interface for external prior knowledge is reserved in the model architecture. During the inference phase, two clinical prior factors from the target clinical scenario are dynamically introduced: family genetic risk value and sleep quality score over the past three months. A quantifiable weight assignment function is designed to linearly weight and fuse the predicted probability of the basic model with the contribution of the two clinical prior factors, generating the final domain-adaptive predicted probability. Third, this invention also proposes a grid search optimization mechanism adapted to different clinical scenarios. This invention automatically obtains the optimal weight configuration for a specific clinical scenario by traversing and constraining the weight coefficients of the basic model, family genetics, and sleep quality within a preset candidate search space, and using the F1 score or positive accuracy on the target clinical private validation set as an evaluation index.

[0012] Specifically, the first aspect of this invention provides a domain-adaptive hypertension prediction method. This method first obtains the basic feature vector of the subject and inputs it into a trained basic prediction model to output a basic hypertension prediction probability vector. Simultaneously, it collects the clinical prior dynamic factors of the subject and quantizes and maps these factors into probability weight vectors. Then, it linearly weights and fuses the basic hypertension prediction probability vector with the probability weight vector to calculate a fused probability vector. Finally, it determines the hypertension prediction diagnosis result of the subject based on the fused probability vector. This approach overcomes the limitations of traditional single prediction models, which over-rely on data distribution and struggle to incorporate expert experience, by linearly fusing the data-driven model output with clinical prior knowledge along a probabilistic dimension. This significantly improves the model's decision-making rationality and generalization ability in clinical applications.

[0013] Furthermore, in the aforementioned domain-adaptive hypertension prediction method, the basic prediction model employs a cascaded Long Short-Term Memory (LSTM) network layer and a Transformer encoder layer. Specifically, the basic feature vector is reshaped into a feature sequence and input into the LTM network layer. Its temporal modeling mechanism extracts local dependencies between features and outputs a hidden state sequence. This hidden state sequence is then input into the Transformer encoder layer, which utilizes its multi-head self-attention mechanism to capture global interactions between features and outputs a global feature tensor. By cascading the LTM network and the Transformer, both local short-range evolution patterns and long-range global cross-feature correlations are extracted, effectively avoiding the loss of high-dimensional physiological feature information and significantly enhancing the model's ability to represent complex health data.

[0014] Furthermore, in the aforementioned cascaded prediction model, the basic prediction model also includes a fully connected classification layer connected to the back end of the Transformer encoder layer. During feature mapping, feature vectors from a preset time step in the global feature tensor are extracted and fed into the fully connected classification layer. After classification mapping and activation function normalization, the basic hypertension prediction probability vector is finally output. By extracting feature vectors from representative time steps of the sequence and performing smooth probability mapping, a precise conversion from a high-dimensional global feature tensor to a two-dimensional class probability distribution is achieved, ensuring the rigor and smoothness of the basic model's prediction probability distribution.

[0015] Preferably, in processing clinical prior factors, the clinical prior dynamic factors specifically include familial genetic risk values ​​and sleep quality scores. When mapping the clinical prior dynamic factors to probability weight vectors, a corresponding genetic contribution weight vector is generated based on the collected familial genetic risk values, and a corresponding sleep contribution weight vector is generated based on the sleep quality scores. This transforms qualitative or discrete key clinical prior indicators, such as genetics and sleep, into two-dimensional probability vectors with the same dimension as the model output, laying a unified data representation foundation for seamless integration of subsequent algebraic dimensions.

[0016] Furthermore, regarding the multi-source information fusion mechanism, the linear weighted fusion of the baseline hypertension prediction probability vector and the probability weight vector is specifically manifested as follows: a linear weighted sum is performed on the baseline hypertension prediction probability vector, the genetic contribution weight vector, and the sleep contribution weight vector to calculate the fused probability vector. In this process, the sum of the weight coefficients corresponding to each of the baseline hypertension prediction probability vector, the genetic contribution weight vector, and the sleep contribution weight vector strictly satisfies the normalization constraint. Thus, by introducing a multi-source linear weighted summation mechanism with normalization constraints, the contribution ratio of the model confidence and each prior factor is reasonably balanced, effectively eliminating the negative interference of dimensional differences on multi-source decision fusion.

[0017] Furthermore, to enable the model to flexibly adapt to different specific clinical scenarios, a private dataset of the target clinical scenario is obtained before linearly weighting and fusing the baseline hypertension prediction probability vector with the probability weight vector, and a validation set is constructed from this private dataset. Next, multiple candidate weight combinations are traversed within a predefined candidate search space, and the target classification evaluation index corresponding to each candidate weight combination is evaluated on the validation set. Finally, the candidate weight combination that optimizes the target classification evaluation index is determined as the optimal weight coefficient for the target clinical scenario. By introducing a grid search optimization mechanism based on a private validation set, the system is given the ability to automatically find the optimal scenario weights, greatly enhancing the domain adaptive tuning capability and scenario migration stability of the prediction model when deployed across hospitals and departments.

[0018] Preferably, in the basic data acquisition and preprocessing stage, before obtaining the basic feature vector of the object to be tested, basic health data containing multiple discrete and numerical features is first collected. Then, numerical encoding mapping is performed on the discrete features, and a standardization transformation to eliminate dimensional differences is performed on the numerical features. Finally, the basic feature vector is constructed by combining these elements. By performing standardized categorical encoding mapping and standardized dimensional elimination processing on heterogeneous health data, the consistency of the input feature format and the stability of the numerical distribution are ensured, thereby accelerating the gradient convergence of the subsequent neural network and improving feature extraction efficiency.

[0019] In the final decision output stage, determining the hypertension prediction diagnosis result of the test subject based on the fused probability vector specifically includes: firstly, re-normalizing the fused probability vector to ensure that the sum of its components satisfies rigorous probabilistic semantics. Then, comparing the re-normalized hypertension positive probability value with a preset judgment threshold, and accurately determining and outputting the diagnosis result of hypertension positive or non-hypertension positive based on the comparison result. By reshaping the probabilistic semantics and performing dual-threshold verification on the algebraically weighted vector, the physical semantic legitimacy of the output probability value can be ensured, providing a highly reliable and intuitive binary classification decision basis for clinical diagnosis.

[0020] Based on the same inventive concept, another aspect of this invention provides a domain-adaptive hypertension prediction system. This system mainly includes a basic feature acquisition module, a cascaded prediction module, a prior vectorization module, a weighted fusion module, and a decision output module. The basic feature acquisition module acquires the basic feature vector of the test subject. The cascaded prediction module is connected to the basic feature acquisition module and internally includes a long short-term memory network layer, a Transformer encoder layer, and a fully connected classification layer, used to extract features from the basic feature vector and output a basic hypertension prediction probability vector. The prior vectorization module collects the clinical prior dynamic factors of the test subject and maps them to probability weight vectors. The weighted fusion module is connected to both the cascaded prediction module and the prior vectorization module, used to perform weighted fusion of the basic hypertension prediction probability vector and the probability weight vector and output a fused probability vector. The decision output module is connected to the weighted fusion module and used to determine and output the hypertension prediction diagnosis result based on the fused probability vector. This system, through explicit modular division and streaming data interoperability, achieves efficient hardware-level collaborative operation throughout the entire process from data acquisition, high-order feature abstraction, prior knowledge injection to final decision output.

[0021] Furthermore, the domain-adaptive hypertension prediction system integrates a grid search optimization module, which is connected to the weighted fusion module. This module is specifically designed to acquire a private validation set for the target clinical scenario and traverse multiple candidate weight combinations within a preset candidate search space. Based on the classification evaluation indicators on the private validation set, it automatically determines and dynamically updates the weight coefficients in the weighted fusion module. The addition of this module enables the prediction system to possess hardware-level domain-adaptive parameter self-learning capabilities, allowing for rapid and automated calibration of fusion weights for data characteristics of heterogeneous terminals or specific deployment scenarios. This significantly reduces the system's on-site maintenance costs and deployment threshold.

[0022] One or more technical solutions provided in the embodiments of the present invention have at least the following technical effects or advantages: 1. This invention achieves joint representation of local temporal dependencies and global cross-feature interactions in feature sequences by constructing a cascaded architecture of a Long Short-Term Memory (LSTM) network and a Transformer encoder. The LSM layer accurately captures short-range evolutionary patterns between features, while the Transformer encoder layer models long-range global associations through a multi-head self-attention mechanism, resulting in complementary enhancements. On the same publicly available test set, the prediction accuracy of the basic model of this invention reached 0.96, significantly higher than the approximately 0.85 accuracy of existing single LSM networks or single Transformer models, verifying the substantial improvement in feature extraction and global feature representation capabilities of the cascaded architecture.

[0023] 2. This invention possesses superior dynamic fusion and domain adaptation capabilities for clinical prior factors. Existing technologies fix the input feature space after model training, preventing the inference phase from incorporating any new clinical variables. This makes it difficult for deployed models to adapt to the diverse prediction needs of different medical scenarios. This invention, by reserving a structured fusion interface for external prior knowledge within the model architecture and designing a percentage calibration module, enables the model to dynamically incorporate two clinical prior factors—family genetic risk value and sleep quality scores over the past three months—during the inference phase. Adjustable weighting coefficients are used to achieve quantitative integration of multi-source information. This mechanism fundamentally breaks the architectural constraints of a fixed input space in traditional models, significantly improving the generalization and feature scalability of the prediction system across different clinical scenarios.

[0024] 3. This invention significantly improves the clinical interpretability and transparency of prediction results. Existing deep learning models are mostly typical black-box systems, making it impossible for clinicians to understand the decision-making basis of the prediction results. This invention uses a linear weighted fusion formula to combine the three sources—basic model probability, family genetic contribution, and sleep quality contribution—with clearly defined weight coefficients, ensuring that the contribution of each factor to the final prediction result is fully quantifiable. Doctors can intuitively understand the overall decision-making logic of the model through the weight proportions of each factor, overcoming the unknowability of traditional model systems and greatly enhancing the credibility and clinical acceptability of AI-assisted diagnostic systems in serious clinical scenarios.

[0025] 4. This invention effectively solves the false positive problem that easily occurs in traditional hypertension prediction, reducing clinical misdiagnosis and unnecessary medical follow-up costs. Experimental data on a private clinical test set shows that the percentage calibration module and grid search weight optimization mechanism proposed in this invention can maintain a stable positive prediction accuracy at a high level of 0.875, while the F1 score and recall rate both meet clinical requirements. This fully demonstrates the excellent performance of this scheme in smoothly integrating additional clinical prior feature parameters without sacrificing the prediction accuracy of the original model, and possesses extremely high feasibility and stability for practical clinical deployment. Attached Figure Description

[0026] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a structural diagram of the basic LSTM-Transformer model. Detailed Implementation

[0027] To better understand the technical solutions of the present invention, exemplary embodiments of the present invention will be described in more detail below with reference to the accompanying drawings. Although exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the present invention to those skilled in the art.

[0028] This invention provides a domain-adaptive method and system for predicting hypertension, the overall method flow of which is as follows: Figure 1 As shown, it mainly includes two stages: basic model construction and training based on public datasets and domain adaptive calibration and decision-making based on private clinical datasets. It aims to solve the problems of incomplete feature extraction capabilities, fixed input space leading to the inability to dynamically integrate clinical prior factors, lack of interpretability of black-box decision-making, and high false positive rate in traditional hypertension prediction models.

[0029] In the first stage, the offline training stage, an open-source health dataset is first collected, such as a subset of the SSC dataset containing over 500 records. Through preprocessing and numerical mapping of the data, a basic feature vector containing 13 fundamental features is constructed. These features specifically include gender, age, weight, height, body mass index (BMI), alcohol consumption level, exercise level, stress level, salt intake, smoking status, marital status, and systolic blood pressure. Subsequently, the preprocessed feature vector is sequentially input as follows: Figure 2 In the constructed cascaded deep learning model shown, a Long Short-Term Memory (LSTM) network captures short-range dependencies and local evolution patterns in the feature sequence. The Transformer encoder layer utilizes multi-head self-attention to model global cross-feature interactions and long-range associations, achieving joint representation of local and global features. Finally, a fully connected classification layer calculates and outputs the hypertension prediction probability vector P of the base model. base Finally, based on the binary cross-entropy loss function, 150 rounds of iterative training were performed on the training set to optimize the model parameters, enabling the basic model to achieve a high classification accuracy of 0.96 on the public test set.

[0030] In the second stage, the online inference and calibration stage, to ensure the trained base model can seamlessly adapt to specific clinical application scenarios, the two stages are connected through a base model parameter file (.pth). This allows the second stage to directly reuse the feature extraction capabilities gained from the first stage, thereby achieving domain-adaptive percentage calibration and final decision-making. In this stage, a private clinical dataset of the target model is first collected. In addition to the 13 basic features, two clinical prior factors—family genetic risk value and sleep quality score over the past three months—are introduced. A grid search mechanism is used to traverse and optimize three weight coefficients within a pre-defined candidate search space: the base model weight coefficient, the family genetic weight coefficient, and the sleep quality weight coefficient. The optimal weight configuration is determined using the F1 score or positive accuracy on the target scenario validation set. Finally, the probability vector, genetic weight vector, and sleep weight vector output by the base model are linearly weighted and fused to generate a fused probability vector. The probability is then re-normalized, and finally, a classification decision is performed based on the re-normalized positive probability value to output the hypertension diagnosis result.

[0031] By organically combining the above two stages, this invention not only ensures the basic model's high-precision extraction capability of general features, but also endows the model with the domain adaptation capability to dynamically absorb clinical prior knowledge during the inference stage. At the same time, it ensures the interpretability of the decision logic through a clear linear weighting mechanism.

[0032] Example 1 Specifically, the domain-adaptive hypertension prediction method provided by this invention mainly includes the following steps: Step 1: Basic Data Collection and Preprocessing. An open-source health dataset containing over 500 records, such as a subset of the SSC dataset, was collected and divided into training and testing sets in an 8:2 ratio. Thirteen basic features in the data, including gender, age, weight, height, body mass index (BMI), alcohol consumption level, exercise level, stress level, salt intake, smoking status, marital status, and systolic blood pressure, were preprocessed using standardization and one-hot encoding or numerical mapping to construct basic feature vectors and corresponding real hypertension diagnostic labels.

[0033] Step 2: Feature extraction from the LSTM layer. The preprocessed 13-dimensional basic feature vector is input into the Long Short-Term Memory (LSTM) network layer. The temporal and sequence modeling mechanism of LSTM is used to capture the short-range dependencies and local evolution patterns between features, and the intermediate hidden layer feature matrix is ​​output.

[0034] Step 3: Global Interaction of Transformer Encoder. The feature matrix output from the LSTM layer is input into the Transformer encoder layer. The multi-head self-attention mechanism is used to model the global feature association, capture the global cross-feature interaction relationship, realize the joint representation of local features and global features, and output the global feature representation vector.

[0035] Step 4: Fully Connected Classification Layer and Base Probability Calculation. The global feature representation vector is input into the fully connected classification layer. After processing with the Sigmoid activation function, the hypertension prediction probability vector P of the base model is calculated and output. base .

[0036] Step 5: Iterative training of the basic model. The model is trained iteratively for, for example, 150 rounds based on the binary cross-entropy loss function to optimize the model parameters. After training, the model performance is verified on the test set of the open-source dataset. Finally, the classification accuracy of the basic model in this embodiment can reach 0.96.

[0037] Step 6: Three-source weighted fusion, probability renormalization, and classification decision. The probability vector P output by the base model is... base Genetic weight vector and sleep weight vector Perform three-source linear weighted fusion to obtain the fusion probability vector P. final ; For P final The system is then renormalized to maintain probabilistic semantics; finally, a classification decision is made based on the renormalized positive probability value, and the final hypertension diagnosis result is output.

[0038] Step 7: Clinical Private Data Collection and Grid Search Optimization of Weight Coefficients. Collect a private clinical dataset from the target hospital, such as data from 20 outpatient cases. In addition to the 13 basic features mentioned above, introduce two additional clinical prior factors: a family genetic risk value (G) and a sleep quality score (S) over the past three months. Optimize the basic model weight coefficients (λ) by traversing the pre-defined candidate search space using a grid search mechanism. model Family genetic weight coefficient λ G and the sleep quality weighting coefficient λ S The optimal weight configuration is selected by using the F1 score or positive accuracy on the validation set to adapt to specific clinical application scenarios.

[0039] Through the coordinated operation of steps 1 to 7 above, this invention completes the entire process from the efficient extraction of general basic features to the dynamic fusion of prior factors and accurate decision-making in specific clinical scenarios. The implementation details of each of the above specific steps will be elaborated below.

[0040] In step 1, the present invention first collects and preprocesses publicly available health datasets. Specifically, step 1 includes the following steps 1.1 to 1.5: Step 1.1: Collection of Public Health Datasets. Individual health record data is obtained from a public health database. This embodiment uses a subset of the open-source health dataset SSC, which contains over 500 records. Each record should contain at least 13 basic features for hypertension prediction, specifically including gender, age, weight, height, body mass index (BMI), alcohol consumption level, exercise level, stress level, salt intake, smoking status, marital status, and systolic blood pressure, as well as the systolic blood pressure measurement used to generate labels. The data format is structured tabular data, such as a comma-separated value file in CSV format.

[0041] Step 1.2: Feature Selection and Categorical Encoding. Thirteen features were selected from the original data as model input. Numerical encoding mapping was performed on categorical features: gender (M→0, F→1), smoking status (Y→1, N→0), and marital status (Y→1, N→0). Other ordinal features, such as BMI, alcohol consumption, exercise level, stress level, and salt intake, maintained their ordinal numerical encoding. Numerical features, such as age, weight, height, body mass index, and systolic blood pressure, were directly input as their original values.

[0042] Step 1.3: Label generation. Using a systolic blood pressure ≥140 mmHg as the threshold for positivity of hypertension (according to the "Chinese Guidelines for the Prevention and Treatment of Hypertension"), a binary label is generated: if the systolic blood pressure is ≥140 mmHg, the generated binary label is 1 (positive for hypertension); otherwise, it is 0 (positive for non-hypertension).

[0043] Step 1.4: Dataset Splitting. A stratified random sampling strategy is used to divide all samples into training and test sets in an 8:2 ratio, and a random seed is set to ensure the reproducibility of the splitting results.

[0044] Step 1.5: Data Standardization. Perform Z-score standardization on all numerical features in the training set. This involves subtracting the mean from each numerical feature and then dividing by the standard deviation to eliminate dimensional differences between features. The standardization formula is as follows: X norm = (X - μ) / σ Where X is the original numerical feature value, μ is the mean of the feature in the training set, and σ is the standard deviation of the feature in the training set. norm These are the standardized feature values. The test set is standardized using the mean μ and standard deviation σ of the training set, thus constructing the basic feature vectors for subsequent model input.

[0045] like Figure 2 As shown, the basic prediction model adopts a deep neural network structure of cascaded LSTM and Transformer encoder, which includes the following layers from input to output: input layer, LSTM layer, Transformer encoder layer and fully connected classification layer.

[0046] In step 2, the present invention constructs the LSTM layer in the basic prediction model, focusing on the construction and feature extraction of the LSTM layer to extract the short-range and temporal dependencies between the preprocessed 13-dimensional basic feature vectors.

[0047] The 13-dimensional basic feature vectors, preprocessed and standardized in step 1, are fed into the input layer, and the 13 features are treated as a sequence of length 13, i.e., each feature is a time step. The LSTM layer extracts the temporal dependencies and local evolution patterns among the 13 input features through its sequence modeling mechanism.

[0048] In this embodiment, the specific network structure parameters of the LSTM layer are configured as follows: the number of hidden units is set to 128, the number of stacked layers is set to 3, and the parameter batch_first=True is set. When the three LSTM layers are stacked, the next LSTM layer receives the hidden state output of the previous LSTM layer as its own input sequence, thereby realizing multi-layer feature abstraction.

[0049] After feature extraction and information transfer through three LSTM layers, the LSTM layer finally outputs a hidden state sequence tensor:

[0050] In the formula, batch_size represents the number of samples in the batch, 13 represents the time step length (i.e., 13 basic features), and 128 represents the dimension of the hidden units. This hidden state sequence tensor H fully preserves the short-range correlation information between features and serves as the input to the subsequent Transformer encoder layer.

[0051] In step 3, the present invention constructs a Transformer encoder layer in the basic prediction model to receive the feature tensors output by the LSTM layer and capture the global cross-feature interaction relationships among 13 features.

[0052] The Transformer encoder layer directly receives the hidden state sequence tensor output by the LSTM layer in step 2 as its input. Through a multi-head self-attention mechanism, the Transformer encoder layer breaks the distance limitation in sequence processing, efficiently capturing the global interaction relationships and long-range structural associations among 13 features, thus complementing and enhancing the local short-range dependency extraction of the LSTM layer.

[0053] In this embodiment, the specific network parameters of the Transformer encoder are configured as follows: the model dimension is set to 128, consistent with the number of hidden units in the LSTM; the number of multi-head attention heads is set to 8; and the number of encoder layers is set to 2. Each encoder layer contains a multi-head self-attention sublayer and a feedforward network sublayer, and residual connections and layer normalization are used after each sublayer to ensure smooth gradient propagation and the stability of feature representation. The core calculation formula for scaled dot product attention is:

[0054] In the formula, Q, K, and V are the query matrix, key matrix, and value matrix obtained by mapping from the input sequence, respectively, and d k The dimension is the key vector. After multi-head self-attention fusion and feedforward network transformation through two Transformer encoder layers, the final output is a global feature tensor containing global cross-feature interaction representations:

[0055] The global feature tensor T is input into the subsequent fully connected classification layer to calculate the hypertension prediction probability of the base model.

[0056] In step 4, this invention constructs a fully connected classification layer in the basic prediction model, maps the high-dimensional global feature representation output by the Transformer encoder layer into a two-dimensional class probability distribution, and calculates and generates the hypertension prediction probability vector P of the basic model. base Furthermore, the specific implementation process of step 4 is as follows: First, the global feature tensor T output by the Transformer encoder layer in step 3 is received. In order to extract representative representations of the sequence dimension, the feature vector T[:,-1,:] of the last time step in the output sequence is extracted by tensor slicing.

[0057] Secondly, the extracted feature vector T[:,-1,:] is fed into a fully connected classification layer for linear transformation and dimension mapping, outputting a two-dimensional logits vector.

[0058] Finally, the two-dimensional logits vector is input into the Softmax function for normalization, transforming it into a two-dimensional probability vector P that conforms to the semantics of a probability distribution. base = [p0, p1] In the formula, p0 represents the probability that the sample predicted by the base model is not hypertension-positive, and p1 represents the probability that the sample predicted by the base model is hypertension-positive (i.e., the probability of hypertension positivity). The probability vector P output by the base model is... baseThis will be used for subsequent weighted fusion with clinical prior factors.

[0059] In step 5, based on the preprocessed training data, the present invention performs training configuration, training process iteration, and model parameter saving on the constructed cascaded basic prediction model. Further, step 5 specifically includes the following implementation process: During the training and configuration phase, this invention selects cross-entropy loss as the objective loss function, uses the Adam Optimizer as the optimizer, sets the initial learning rate to 0.0005, and sets the number of iterations to 150. The batch size is determined using either full training or mini-batch gradient descent; if mini-batch gradient descent is used, the preferred batch size is 32-64.

[0060] During the training phase, the complete training process of the model includes four stages: forward propagation, loss calculation, backpropagation, and parameter update, iterating repeatedly until a preset number of 150 iterations is reached. In each training iteration, preprocessed training data is fed into the model according to a set batch size to perform forward propagation, calculating the cross-entropy loss between the model's predicted output and the true diagnostic label. Subsequently, the gradients of the parameters of each network layer in the model are calculated using the backpropagation algorithm, and the network weights are updated using the Adam optimizer. During training, the current loss value is output every 10 training iterations to monitor the model's training convergence status in real time.

[0061] During the model saving phase, once the training reaches the preset 150 rounds and the model converges, all trainable parameters of the model are saved as .pth format files so that they can be loaded and used at any time during the subsequent clinical private data inference and calibration phases.

[0062] Validation on the public test set shows that the basic model trained and saved using the above process achieves a hypertension binary classification prediction accuracy of 0.96, verifying the basic model's effective extraction of 13 basic features and high-precision classification performance.

[0063] In step 6, this invention collects clinical prior dynamic factors and vectorizes them as percentages, thereby achieving three-source linear weighted fusion of the basic model's predicted probability and the prior dynamic factors, probability re-normalization, and the final hypertension classification decision. Further, the specific implementation process of step 6 is as follows: In the clinical prior dynamic factor collection and vectorization phase, two additional dynamic factors were collected for each subject using a standardized questionnaire or electronic medical record system: a family genetic risk value (G) and a sleep quality score (S) over the past three months. The family genetic risk value (G) was assigned as follows: G=0 represents no family history of the disease, and G=1 represents a family history of hypertension. The sleep quality score (S) over the past three months was assigned as follows: S=0 represents normal or good sleep quality, and S=1 represents poor sleep quality or significant impairment, specifically assessed according to the Pittsburgh Sleep Quality Index or the Clinical Short Form Scale. Subsequently, the percentage calibration module quantified these two dynamic factors into two-dimensional probability weight vectors, specifically defined as: a genetic contribution weight vector. When G=0, it maps to [1.0, 0.0], and when G=1, it maps to [0.0, 1.0]; Sleep contribution weight vector When S=0, it is mapped to [1.0, 0.0], and when S=1, it is mapped to [0.0, 1.0].

[0064] In the three-source linear weighted fusion stage, the probability vector P output by the basic model in step 4 is... base The genetic contribution weight vector constructed above and sleep contribution weight vector Perform three-source linear weighted fusion to calculate the fusion probability vector P. final Its core linear weighted fusion calculation formula is expressed as: P final = λ model ·P base + λ G · + λ S · In the formula, λ mode The confidence retention coefficient for the basic model is set to 0.6 by default; λ G λ represents the family genetic weighting coefficient, with a default value set to 0.15. S The sleep quality weighting coefficients are set to a default value of 0.25. Furthermore, the three weighting coefficients strictly satisfy the normalization constraint λ. model + λ G + λ S = 1.

[0065] In the probability renormalization and classification decision stage, to ensure that the weighted fused vector satisfies probabilistic semantics, the P obtained from the weighted fusion is calculated. finalA re-normalization process is performed to ensure that the sum of all components is strictly equal to 1. After probability re-normalization, the category corresponding to the maximum value in the re-normalized probability vector is taken as the final hypertension prediction output. In actual judgment, this is equivalent to judging the hypertension positive probability component P in the fused probability vector. (1) final Does P satisfy? (1) final If the value is ≥0.5, the model will predict a positive result for hypertension; otherwise, it will predict a positive result for non-hypertension.

[0066] In step 7, this invention collects private clinical datasets from target hospitals and dynamically optimizes the three weight coefficients using a grid search mechanism, enabling the hypertension prediction model to automatically adapt to different specific clinical application scenarios and achieve accurate domain-adaptive decision-making. Further, the specific implementation process of step 7 is as follows: In the clinical private data acquisition and preprocessing stage, the specific clinical private dataset of the target hospital is first collected. In this private dataset, each patient health record includes, in addition to the 13 basic characteristics mentioned in step 1 (gender, age, weight, height, body mass index, BMI, alcohol consumption level, exercise level, stress level, salt intake, smoking status, marital status, and systolic blood pressure), the patient's family genetic risk value G, sleep quality score S over the past three months, and clinical hypertension diagnostic label are also collected. The data preprocessing process for these 13 basic characteristics in the private dataset, including feature encoding, mapping, and standardization transformation based on training set parameters, is exactly the same as in step 1.

[0067] In the grid search optimization stage for weight coefficients, to enable the model to flexibly adapt to the differences in data distribution among different target hospitals or clinical departments, this invention provides a grid search mechanism based on validation set evaluation to automatically find the optimal weight coefficient configuration. Specifically, it iterates through all possible candidate weight combinations in a preset candidate search space: where the basic model weight coefficients λ model The candidate set is set as {0.5, 0.6, 0.7, 0.8, 0.9}; the family genetic weight coefficient λ G The candidate set is set as {0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35}; the sleep quality weighting coefficient λ S Then, based on the normalization constraints, it can be directly calculated from the formula, i.e., λ. S = 1 - λ model - λ G Furthermore, the search process strictly limits the value of λ. S The range is [0, 0.3].

[0068] When iterating through each group of candidate weight combinations, the validation set data of the target clinical scenario is used to perform weighted fusion and decision evaluation on each candidate combination, and the F1 score or positive precision on the validation set is calculated respectively. Finally, the candidate combination that maximizes the target evaluation index, such as the F1 score or positive precision, is selected as the optimal weight configuration parameter for that specific clinical scenario, and is used for subsequent actual prediction and diagnosis of hypertension in that scenario.

[0069] The technical solution provided by this invention has undergone thorough experimental construction and data verification. The following provides a detailed description of the experimental and deployment environment, the test results of the public dataset, and the verification results of the real clinical private dataset in this embodiment.

[0070] In terms of experimental environment configuration, this embodiment uses the Ubuntu 18.04 operating system environment, builds and trains the model based on the PyTorch deep learning framework, and uses a single NVIDIA P100 graphics processing unit (GPU) for accelerated computing to ensure efficient operation of model training and inference evaluation.

[0071] For validation on public datasets, this embodiment uses an open-source health dataset (a subset of the SSC dataset) containing over 500 records, divided into training and test sets in an 8:2 ratio. The cascaded base model, consisting of LSTM layers and Transformer encoder layers, achieved a 96% binary classification prediction accuracy on the test set after 150 iterations of training on the training set, fully validating the base model's efficient feature extraction and accurate classification capabilities on publicly available general data distributions.

[0072] For validation on a private clinical dataset, to verify the predictive performance and domain adaptation effect of the invention in a real clinical scenario, data from 20 outpatients collected from a hospital were used as the private clinical test set. Each record in this private test set included the aforementioned 13 basic features: gender, age, weight, height, body mass index, BMI, alcohol consumption level, exercise level, stress level, salt intake, smoking status, marital status, and systolic blood pressure; two dynamic factors: family genetic risk level and sleep quality score over the past three months; and a clinically recognized hypertension diagnosis label. Experimental evaluation and comparison showed that when only the 13 basic features were used as input, the basic model's prediction accuracy for patients with hypertension was 0.875. However, after introducing the two dynamic factors of family genetic risk level and sleep quality score over the past three months, the invention, through a percentage calibration module and a grid search weight optimization mechanism, performed weighted fusion prediction, and the positive prediction accuracy remained stable at 0.875. Simultaneously, the F1 score and recall both reached a clinically acceptable high level.

[0073] The experimental results above demonstrate that the percentage calibration module proposed in this invention can effectively and dynamically integrate additional clinical prior feature parameters without sacrificing the original model's prediction accuracy. This confirms that the method has excellent feature scalability and prediction stability on private clinical data, and possesses extremely high feasibility for practical clinical deployment and promotion.

[0074] Example 2 Based on the aforementioned overall methodology and the specific implementation of each step, this invention also provides an embodiment of a domain-adaptive hypertension prediction system. In actual hardware deployment, the system runs on a computer device or dedicated medical terminal device that includes a graphics processor (such as a single NVIDIA P100 GPU), a central processing unit, and memory. The memory stores executable program instructions and trained network parameter weight files (such as .pth format files). The processor, by calling the instructions and data in memory, controls the various software and hardware logic modules within the system to work closely together and coordinate their operation, thereby completing the entire process from general feature extraction to domain-adaptive and accurate decision-making in specific clinical scenarios at the hardware level.

[0075] The system first includes a basic data acquisition and preprocessing module. This module receives publicly available health datasets, such as a subset of the SSC dataset containing over 500 records, and divides it into training and testing sets in an 8:2 ratio. This module performs class encoding mapping and Z-score standardization on 13 basic features collected, including gender, age, weight, height, body mass index (BMI), alcohol consumption level, exercise level, stress level, salt intake, smoking status, marital status, and systolic blood pressure. Simultaneously, it generates binary true diagnostic labels based on the criterion of systolic blood pressure ≥140 mmHg, thus outputting a standardized 13-dimensional basic feature vector.

[0076] The basic feature vectors output by the basic data acquisition and preprocessing module are directly fed into the basic model training and storage module, as well as the cascaded inference network consisting of the LSTM feature extraction module, the Transformer global interaction module, and the fully connected classification and basic probability calculation module. During model training, the basic model training and storage module uses the cross-entropy loss function and the Adam optimizer, setting an initial learning rate of 0.0005 and 150 iterations, and performs backpropagation and gradient updates on the weight parameters of each element in the cascaded inference network. During training, this module monitors the loss convergence status every 10 iterations and, upon completion, solidifies the converged model parameters and saves them as a .pth format file in the system's memory for direct loading and retrieval during the inference phase.

[0077] In the inference and data feature extraction stages, the LSTM feature extraction module and the Transformer global interaction module form a cascaded feature abstraction pipeline. The LSTM feature extraction module, loaded with fixed parameters, receives the preprocessed 13-dimensional basic feature vector, treating it as a time sequence of length 13. Through a neural network structure configured with 128 hidden units and 3 stacked layers, it efficiently extracts the short-range dependencies and local evolution patterns among the 13 basic features and outputs a hidden state sequence tensor H. Next, the Transformer global interaction module takes this hidden state sequence tensor H and, utilizing its internally configured model dimension of 128, 8 attention heads, and 2 encoder layers, captures the global interactions and long-range structural associations across features through a multi-head self-attention mechanism. After residual connections and layer normalization, the global feature tensor T, containing joint local and global representations, is passed to the next module.

[0078] The fully connected classification and basic probability calculation module is connected to the Transformer global interaction module. It extracts the feature vector T[:,-1,:] of the last time step in the global feature tensor T through tensor slicing operations, and feeds it into the fully connected classification layer for binary classification mapping and Softmax function normalization. It then calculates and outputs the hypertension prediction probability vector P of the basic model. base = [p0, p1].

[0079] To achieve domain-adaptive calibration of the model in specific clinical scenarios, the clinical private data acquisition center and grid search optimization module are first involved in the inference preparation phase. This module acquires the target hospital's clinical private dataset (containing 13 basic features, family genetic risk value G, sleep quality scores S over the past three months, and ground truth labels), and traverses candidate weight combinations within a pre-defined candidate search space. Specifically, λ model The candidate set is set as {0.5, 0.6, 0.7, 0.8, 0.9}; the family genetic weight coefficient λ G The candidate set is set as {0.05, 0.10, 0.15, 0.20, 0.25, 0.30, 0.35}; the sleep quality weighting coefficient λ S Then, based on the normalization constraints, it can be directly calculated from the formula, i.e., λ. S = 1 - λ model -λ G Furthermore, the search process strictly limits the value of λ. S The range is [0, 0.3]. By evaluating F1 scores or positive precision on a clinical private validation set, this module automatically finds and locks the weight configuration parameters that optimize the target evaluation indicator, and passes this optimal parameter to the subsequent fusion calculation module.

[0080] During the actual reasoning and decision-making phase, the clinical prior vectorization and three-source weighted fusion module simultaneously receives the base probability vector P output by the fully connected classification and base probability calculation modules. base A two-dimensional genetic contribution weight vector, collected from electronic medical records or standardized questionnaires and quantified. With sleep contribution weight vector The optimal weight coefficients are passed from the grid search optimization module. This module uses the linear weighted fusion formula P... final = λ model ·P base + λ G · +λ S · Efficiently fuse three sources of data to generate a fusion probability vector P. final .

[0081] Finally, the normalization and classification decision module receives the fusion probability vector P output by the clinical prior vectorization and three-source weighted fusion module. final The sample is then re-normalized to maintain rigorous probabilistic semantics. Subsequently, the module determines whether the re-normalized positive probability component for hypertension satisfies P... (1) final ≥0.5; if satisfied, the hardware output device outputs a positive diagnosis result for hypertension; otherwise, it outputs a non-positive diagnosis result for hypertension, thus completing the closed-loop implementation of the entire process from data input, feature abstraction, prior integration to hardware decision output among the modules.

[0082] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied 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.

[0083] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as 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, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxesFigure 1 A device that provides the functions specified in one or more boxes.

[0084] 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.

[0085] 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.

[0086] It should be noted that any reference signs placed between parentheses in the claims should not be construed as limiting the claims. The word "comprising" does not exclude the presence of components or steps not listed in the claims. The word "a" or "an" preceding a component does not exclude the presence of a plurality of such components. The invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer. In a unit claim enumerating several means, several of these means may be embodied by the same item of hardware. The use of the words first, second, third, etc., does not indicate any order. These words can be interpreted as names.

[0087] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0088] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A domain-adaptive method for predicting hypertension, characterized in that, Includes the following steps: Obtain the basic feature vector of the object under test; The basic feature vector is input into the trained basic prediction model, and the basic hypertension prediction probability vector is output. Collect the clinical prior dynamic factors of the test subject and map the clinical prior dynamic factors into a probability weight vector; The baseline hypertension prediction probability vector and the probability weight vector are linearly weighted and fused to generate a fused probability vector. Based on the fusion probability vector, the hypertension prediction and diagnosis result of the subject to be tested is determined.

2. The hypertension prediction method according to claim 1, characterized in that, The basic prediction model includes a cascaded long short-term memory network layer and a Transformer encoder layer; The basic feature vector is reshaped into a feature sequence and input into the long short-term memory network layer. The long short-term memory network layer is used to extract the local dependencies between features and output the hidden state sequence. The hidden state sequence is input into the Transformer encoder layer, and the global interaction relationship between features is extracted using a multi-head self-attention mechanism, outputting a global feature tensor.

3. The hypertension prediction method according to claim 2, characterized in that, The basic prediction model also includes a fully connected classification layer connected to the back end of the Transformer encoder layer; The feature vectors of the global feature tensor at a preset time step are extracted and input into the fully connected classification layer. After classification mapping and activation function normalization, the basic hypertension prediction probability vector is output.

4. The hypertension prediction method according to claim 1, characterized in that, The clinical prior dynamic factors include family genetic risk value and sleep quality score; Mapping the clinical prior dynamic factors to probability weight vectors includes: generating a genetic contribution weight vector based on the family genetic risk value, and generating a sleep contribution weight vector based on the sleep quality score.

5. The hypertension prediction method according to claim 4, characterized in that, The baseline hypertension prediction probability vector and the probability weight vector are linearly weighted and fused, including: The fusion probability vector is calculated by performing a linear weighted summation on the baseline hypertension prediction probability vector, the genetic contribution weight vector, and the sleep contribution weight vector. The sum of the weight coefficients corresponding to the basic hypertension prediction probability vector, the genetic contribution weight vector, and the sleep contribution weight vector satisfies the normalization constraint condition.

6. The hypertension prediction method according to claim 5, characterized in that, Before linearly weighting and fusing the baseline hypertension prediction probability vector with the probability weight vector, the method further includes: Obtain a private dataset of the target clinical scenario, and construct a validation set from the private dataset; Multiple candidate weight combinations are traversed within a preset candidate search space, and the target classification evaluation index corresponding to each candidate weight combination is evaluated on the validation set. The candidate weight combination that optimizes the target classification evaluation index is determined as the weight coefficient under the target clinical scenario.

7. The hypertension prediction method according to claim 1, characterized in that, Before obtaining the basic feature vector of the object under test, the following steps are also included: Collect basic health data that includes multiple discrete and numerical features; The discrete features are subjected to numerical encoding mapping, and the numerical features are subjected to standardization transformation to eliminate dimensional differences, and the basic feature vector is constructed by combining them.

8. The hypertension prediction method according to claim 1, characterized in that, Determining the hypertension prediction and diagnosis result of the subject based on the fusion probability vector includes: The fused probability vector is then renormalized so that the sum of its components satisfies probabilistic semantics. The normalized positive probability value for hypertension is compared with the preset judgment threshold. Based on the comparison result, the diagnosis result of hypertension positive or non-hypertension positive is determined and output.

9. A domain-adaptive hypertension prediction system, characterized in that, include: The basic feature acquisition module is used to acquire the basic feature vector of the object under test; The cascaded prediction module, connected to the basic feature acquisition module, includes a cascaded long short-term memory network layer, a Transformer encoder layer, and a fully connected classification layer, used to extract features from the basic feature vector and output a basic hypertension prediction probability vector. The prior vectorization module is used to collect the clinical prior dynamic factors of the test subject and map them into probability weight vectors; The weighted fusion module is connected to the cascaded prediction module and the prior vectorization module respectively, and is used to perform weighted fusion of the basic hypertension prediction probability vector and the probability weight vector to output the fused probability vector. The decision output module, connected to the weighted fusion module, is used to determine and output the hypertension prediction and diagnosis results based on the fusion probability vector.

10. The hypertension prediction system according to claim 9, characterized in that, Also includes: The grid search optimization module is connected to the weighted fusion module; The grid search optimization module is used to obtain a private validation set of the target clinical scenario, traverse multiple sets of candidate weight combinations in a preset candidate search space, and determine and update the weight coefficients in the weighted fusion module based on the classification evaluation index on the private validation set.

Images