Cross-center deep vein thrombosis risk prediction method based on semi-supervised optimal transmission
By aligning cross-center deep vein thrombosis risk data using a semi-supervised optimal transfer method and the DeepJDOT algorithm, high-confidence pseudo-labels are generated, solving the problems of prediction bias and label utilization across medical institutions, and achieving higher prediction accuracy and risk score credibility.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- ZHEJIANG UNIV
- Filing Date
- 2026-02-03
- Publication Date
- 2026-05-15
AI Technical Summary
Existing methods for assessing the risk of deep vein thrombosis suffer from poor model transferability and large prediction bias when deployed across medical institutions. They also struggle to effectively utilize heterogeneous data and limited labels, failing to meet the requirements for reliable deployment in cross-center environments.
A semi-supervised optimal transmission-based approach is adopted, combining the DeepJDOT joint distribution alignment algorithm and a high-confidence pseudo-label generation mechanism. By constructing a composite cost matrix and solving iteratively with Sinkhorn-Knopp, pseudo-labels are generated and the feature extraction network and classification head are optimized to achieve cross-center data embedding alignment and risk score calibration.
It improves the prediction accuracy and risk score credibility of heterogeneous data across centers, enhances the model's generalization ability in a few-label environment, and provides stable clinical early warning evidence.
Smart Images
Figure CN122050819A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of clinical risk prediction technology, and in particular to a method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transmission. Background Technology
[0002] Deep vein thrombosis (DVT) is a type of thrombotic disease that occurs in the deep veins of the lower extremities, characterized by its insidious nature and rapid progression. If not identified and intervened in a timely manner, the thrombus can detach and enter the pulmonary artery, causing pulmonary embolism and resulting in serious consequences. Clinically commonly used risk assessment methods are mostly based on fixed scoring rules, such as the Wells score and the Autar score. These methods, based on static clinical indicators and fixed calculation formulas, are difficult to adapt to rapid changes in patients' conditions and fail to fully consider complex factors such as individual differences and the interaction of multiple clinical parameters in practical applications. Therefore, they suffer from insufficient accuracy and poor adaptability.
[0003] With the introduction of artificial intelligence technology, researchers have attempted to use machine learning models to model and analyze large-scale clinical data, improving the ability to perceive deep vein thrombosis (DVT) risk through feature selection and model training. Although some methods have used fully supervised learning models to achieve individualized predictions, when deployed across medical institutions, the model's transferability and generalization performance significantly decrease due to differences in sample distribution, number of labels, and feature systems among centers, making prediction bias more likely. Some methods introduce unsupervised optimal transfer strategies that require data alignment between different centers, but without effective pseudo-label constraints and label structure guidance, it is difficult to guarantee the accuracy of pseudo-labels and the rationality of medical interpretation. Existing methods lack a cross-center prediction mechanism that balances data distribution alignment, limited label utilization, and maintenance of medical risk ranking, making it difficult to meet the reliable deployment requirements of intelligent DVT risk assessment in multi-institutional environments.
[0004] Therefore, how to provide a method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transmission is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention
[0005] One objective of this invention is to propose a cross-center deep vein thrombosis risk prediction method based on semi-supervised optimal transmission. This invention combines the DeepJDOT joint distribution alignment algorithm with a high-confidence pseudo-label generation mechanism, and details the complete process of embedding alignment, label migration, and risk score calibration across heterogeneous medical centers. It possesses advantages such as strong prediction stability, high cross-domain generalization ability, and high accuracy in risk stratification.
[0006] The transcentral deep vein thrombosis risk prediction method based on semi-supervised optimal transport according to embodiments of the present invention includes the following steps: Step 1: Collect source center labeled feature sets, target center labeled feature sets, and target center unlabeled feature sets, and perform preprocessing to obtain the corresponding preprocessed feature sets. Step 2: Input the preprocessed feature sets into the feature extraction network, which outputs source center embedding sets, target center labeled embedding sets, and target center unlabeled embedding sets. Step 3: Based on the source center embedding sets and target center embedding sets, construct a composite cost matrix using the DeepJDOT joint distributed optimal transport algorithm to generate a transport plan. Step 4: Solve the transport plan using the Sinkhorn-Knopp iterative algorithm, ensuring a matching confidence of at least 80% for target center unlabeled embeddings. Step 5: Assign pseudo-labels to samples, calculate the weights of source center samples and unlabeled target center samples; Step 6: Merge source center samples, labeled target center samples, and pseudo-labeled target center samples along with their corresponding weights into an expanded training set, and simultaneously optimize the feature extraction network and classification head using the optimal transmission loss, source center cross-entropy loss, and target center cross-entropy loss, outputting the original predicted probabilities; Step 7: Repeat steps 3 to 5 to iteratively update the feature extraction network and classification head until the negative log-likelihood of the target center validation set converges, obtaining the probabilities before calibration; Step 8: Calibrate the probabilities on the target center validation set using the temperature scaling method, obtain the calibrated probability values, and classify the risk levels according to the threshold.
[0007] Optionally, step one specifically includes: acquiring clinical data including gender, age, height, weight, family history of coagulation dysfunction, history of thrombosis, smoking history, drinking history, surgical history, pregnancy history, platelet count, prothrombin time, thrombin time, activated partial thromboplastin time, fibrinogen concentration, international normalized ratio, D-dimer level, coagulation factor level, antithrombin level, magnetic resonance venous imaging results, and heparin drug use; performing data cleaning, data aggregation, and data annotation on the data to obtain a labeled dataset; performing feature extraction and feature selection on the labeled dataset to generate a processed dataset; randomly dividing the processed dataset into a training set and a test set; training the model on the training set using a random forest model, and calculating the prediction error based on the test set; repeating cross-validation until the prediction error meets the preset convergence condition; outputting the trained random forest model and using the results as an initialization reference for the source center labeled feature set, the target center labeled feature set, and the target center unlabeled feature set.
[0008] Optionally, step two specifically includes: inputting the source center preprocessed feature set, the target center labeled preprocessed feature set, and the target center unlabeled preprocessed feature set into a feature extraction network, wherein the feature extraction network adopts a structure consisting of an input layer, multiple one-dimensional convolutional layers, a normalization layer, an activation function layer, and a fully connected layer connected sequentially; the input layer is used to receive structured clinical feature tensors, and the input dimension of each sample is consistent with the number of fields contained in the feature set; the one-dimensional convolutional layer is used to extract local feature patterns, and each convolutional layer is connected to a batch normalization layer and a non-linear activation function layer; the multi-layer convolutional feature mapping is compressed into a fixed-dimensional embedding vector through the fully connected layer, thereby generating the source center embedding set, the target center labeled embedding set, and the target center unlabeled embedding set, respectively.
[0009] Optionally, step three specifically includes: A cost matrix is constructed to measure the joint distribution deviation between the source center embedding set and the target center embedding set. The cost matrix includes three components: the Euclidean distance between the source and target sample embeddings serves as the fundamental geometric distance term in the embedding space; for labeled samples in the target center, the true labels of the labeled samples in the target center are compared with the labels of the source samples; if the labels are inconsistent, a penalty factor is introduced into the cost matrix, forming a label consistency sub-item; key clinical indicator fields highly correlated with deep vein thrombosis are extracted, including D-dimer level, prothrombin time, international normalized ratio, activated partial thromboplastin time, and fibrinogen concentration. These indicators are input into the medical ranking gating mapping module, which performs indicator ranking calculations and calculates ranking penalty factors based on the ranking differences. Based on preset clinical risk stratification standards, the ranking differences are segmented and mapped to penalty amplitudes, enabling the medical ranking gating mapping module to dynamically control the ranking relationship of high-risk indicators. A label-driven regularization sub-item is introduced, which adds a label-driven term to the original DeepJDOT entropy regularization loss structure. In the process of calculating the transmission distribution entropy, the label-driven term gives low-entropy encouragement to sample pairs that match the same type of label and imposes high-entropy penalty on sample pairs with inconsistent labels, thereby adjusting the stability of the pseudo-label propagation path in the transmission plan. A sample reliability weighting mechanism is constructed, which uses the training error of the source center sample and the deviation of the embedding distribution of the target center sample as a reliability estimation index. Learnable weights are applied to each sample pair in the cost matrix to control the actual impact of the sample in the transmission plan. The learnable weights are kept updated in the Sinkhorn-Knopp iteration. Based on a composite cost matrix that includes Euclidean distance, label consistency, medical ranking penalty, label-guided regularization sub-item, and sample reliability weights.
[0010] Optionally, the process of constructing the penalty factor includes: For the label consistency term, a sample pair is constructed between the source center embedding set and the target center labeled embedding set. The source center sample label is compared with the target center labeled sample label. If the labels are consistent, the label consistency term at the corresponding cost matrix position is assigned a value of zero. If the labels are inconsistent, a fixed non-zero penalty value is assigned to the corresponding position. The label consistency term constrains the transmission plan to prioritize matching sample pairs with consistent risk levels. For the medical ranking penalty item, key clinical indicator fields are extracted from the source center embedding set and the target center embedding set respectively. The key clinical indicator fields include D-dimer concentration, prothrombin time, activated partial thromboplastin time, fibrinogen concentration and international normalized ratio. Each key indicator is numerically ranked in the source center sample set and the target center sample set respectively. For the constructed sample pairs, the absolute value of the ranking difference between the source center sample and the target center sample is calculated. The absolute value of the ranking difference of each key indicator is multiplied by the corresponding weight coefficient and then summed to obtain the medical ranking penalty value of the sample pair. The weight coefficients corresponding to each key indicator are obtained through feature importance analysis during the source center training process. The label consistency term and the medical ranking penalty term are superimposed on the basic Euclidean distance term to form the joint distributed optimal transmission cost matrix used for optimal transmission solution.
[0011] Optionally, step four specifically includes: Based on the optimal transmission cost matrix, the Sinkhorn-Knopp iterative algorithm with entropy regularization is used to solve the transmission plan. In each iteration, the following operations are performed: Initialize the source center marginal distribution and the target center marginal distribution as uniform distributions; use an exponential function to scale the cost matrix element by element to form an initial transmission kernel; sequentially perform vector normalization operations on the transmission kernel according to the source center marginal distribution and the target center marginal distribution, update the transmission matrix until the preset convergence condition is met, and form a normalized transmission plan; In the normalized transmission plan, a matching confidence guidance mechanism is introduced. Specifically, at the end of each iteration, the confidence of the sample is calculated based on the maximum transmission probability of the column vector corresponding to the unlabeled sample at the target center. If the confidence is not less than 80%, the unlabeled sample at the target center is assigned the label of the source center sample corresponding to the maximum matching probability, forming a pseudo-label set. Based on the normalized transmission plan, a probability normalization redistribution strategy is introduced. The transmission probabilities of each row of the source center and each column of the target center are summed and normalized respectively. The transmission weights of the source center samples and the pseudo-label samples of the target center are calculated. Optionally, step five specifically includes: The source center samples, target center labeled samples, and target center unlabeled samples with pseudo-labels obtained through the transmission plan are merged to form an expanded training set; transmission weights are added to each source center sample, pseudo-label weights are added to each pseudo-labeled target center sample, and target center labeled samples are set to unit weights. Based on expanding the training set and simultaneously optimizing the feature extraction network and the classification head, a joint loss function is constructed. The joint loss function consists of the following three parts: optimal transmission loss, which calculates the distribution difference between the source center and the target center in the joint embedding space and the predicted label space based on the joint distribution optimal transmission cost matrix generated in step three and the transmission plan matrix updated in step four; source center cross-entropy loss, which calculates cross-entropy using the true labels of the source center samples and the output results of the classification head, and introduces the source center sample weights to weight the loss value; and target center cross-entropy loss, which calculates cross-entropy for the labeled target center samples and the pseudo-labeled target center samples respectively, and works together with their respective sample weights to optimize the overall loss. By jointly minimizing the above three losses, the parameters of the feature extraction network and the classification head are updated synchronously through backpropagation until the prediction loss of the expanded training set converges, and the original prediction probability corresponding to each input sample is output.
[0012] Optionally, step six specifically includes: The expanded training set is input into the feature extraction network and the classification head, and the process of minimizing the joint loss function is performed. Gradient backpropagation and iterative updates are performed simultaneously on the parameters of the feature extraction network and the parameters of the classification head. After each round of training iterations, the negative log-likelihood loss of the current model on the validation set is calculated based on the target center validation set, the validation loss is recorded and compared with the result of the previous round of iterations; When the negative log-likelihood loss of the target center validation set reaches the preset stopping condition during several consecutive rounds of iterative training, the iterative training process is terminated, and the model parameters of the current round are saved as the optimal model parameters. Based on the trained feature extraction network and classification head, forward reasoning is performed on all samples in the expanded training set and the target center validation set, and the original predicted probability corresponding to each sample is output.
[0013] Optionally, step seven specifically includes: The original predicted probabilities are paired with the true labels in the target center validation set to construct a temperature scaling model with negative log-likelihood loss as the objective function. All parameters of the feature extraction network and the classification head are fixed, and only the temperature parameters in the temperature scaling model are trained. The negative log-likelihood loss is iteratively optimized until convergence. The original predicted probability is scaled using the temperature parameters obtained from training. The scaled result is then input into the sigmoid function, which outputs a calibrated probability value. Fixed thresholds of 0.25, 0.5, and 0.75 are set, and the risk level of the sample is classified as low risk, medium risk, high risk, or extremely high risk based on the calibrated probability value. When a sample is classified as high risk or extremely high risk, an early warning trigger operation is executed.
[0014] The beneficial effects of this invention are: (1) Improve the prediction accuracy of heterogeneous data across multiple medical centers. This invention constructs a composite cost matrix that integrates Euclidean distance, label consistency and medical ranking penalty, and uses a joint distribution optimal transmission strategy guided by high-confidence pseudo-labels to align the distribution deviations of source centers and target centers in the embedding space and label space, thereby improving the stability and prediction accuracy of the model in the migration of heterogeneous data across multiple medical centers; (2) Enhance the credibility of risk scores. Use temperature scaling calibration strategy to post-process the predicted probability and combine iterative optimization of temperature parameters with target validation set to make the risk probability output by the model closer to the real label distribution, improve the numerical reliability and interpretability of risk level classification, and provide a more stable quantitative basis for clinical early warning. (3) Make full use of a small number of labeled samples. In the case of scarce labels in the target center, this invention integrates a semi-supervised learning mechanism, combines transmission plan optimization and weight adjustment methods, constructs a high-confidence pseudo-label expanded training set, improves the model generalization ability in a low-label environment, and reduces the dependence on manually labeled data. Attached Figure Description
[0015] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is an overall flowchart of the transcenter deep vein thrombosis risk prediction method based on semi-supervised optimal transmission proposed in this invention; Figure 2 The DeepJDOT cost matrix construction and optimal transmission structure diagram are shown for the transcenter deep vein thrombosis risk prediction method based on semi-supervised optimal transmission proposed in this invention. Figure 3 This is a diagram showing the temperature scaling calibration and risk level classification structure of the transcenter deep vein thrombosis risk prediction method based on semi-supervised optimal transmission proposed in this invention. Figure 4 This is a schematic diagram of the system user interface for the transcenter deep vein thrombosis risk prediction method based on semi-supervised optimal transmission proposed in this invention. Detailed Implementation
[0016] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.
[0017] refer to Figure 1-4 A method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transport includes the following steps: Step 1: Collect source center labeled feature sets, target center labeled feature sets, and target center unlabeled feature sets, and perform preprocessing to obtain the corresponding preprocessed feature sets. Step 2: Input the preprocessed feature sets into the feature extraction network, which outputs source center embedding sets, target center labeled embedding sets, and target center unlabeled embedding sets. Step 3: Based on the source center embedding sets and target center embedding sets, construct a composite cost matrix using the DeepJDOT joint distributed optimal transport algorithm to generate a transport plan. Step 4: Solve the transport plan using the Sinkhorn-Knopp iterative algorithm, ensuring a matching confidence of at least 80% for target center unlabeled embeddings. Step 5: Assign pseudo-labels to samples, calculate the weights of source center samples and unlabeled target center samples; Step 6: Merge source center samples, labeled target center samples, and pseudo-labeled target center samples along with their corresponding weights into an expanded training set, and simultaneously optimize the feature extraction network and classification head using the optimal transmission loss, source center cross-entropy loss, and target center cross-entropy loss, outputting the original predicted probabilities; Step 7: Repeat steps 3 to 5 to iteratively update the feature extraction network and classification head until the negative log-likelihood of the target center validation set converges, obtaining the probabilities before calibration; Step 8: Calibrate the probabilities on the target center validation set using the temperature scaling method, obtain the calibrated probability values, and classify the risk levels according to the threshold.
[0018] In this embodiment, step one specifically includes: acquiring historical data containing gender, age, height, weight, thrombotic tendency and genetic history, past thrombotic history, smoking history, surgical type, pregnancy status, platelet count, prothrombin time, thrombin time, activated partial thromboplastin time, fibrinogen concentration, international normalized ratio, D-dimer level, coagulation factor level, antithrombin level, magnetic resonance venography examination results, and heparin drug use records; performing data cleaning, data aggregation, and data annotation on the historical data to obtain a labeled dataset; performing feature extraction and feature selection on the labeled dataset to generate a processed dataset; randomly dividing the processed dataset into a training set and a test set; training a random forest model based on the training set and calculating the prediction error based on the test set; repeating cross-validation until the prediction error meets the preset convergence condition; outputting the trained random forest model and using the results as an initialization reference for the source center labeled feature set, the target center labeled feature set, and the target center unlabeled feature set.
[0019] In this embodiment, step two specifically includes: inputting the source center preprocessed feature set, the target center labeled preprocessed feature set, and the target center unlabeled preprocessed feature set into the feature extraction network, respectively. The feature extraction network adopts a structure consisting of an input layer, multiple one-dimensional convolutional layers, a normalization layer, an activation function layer, and a fully connected layer connected sequentially. The input layer is used to receive structured clinical feature tensors, and the input dimension of each sample is consistent with the number of fields contained in the feature set. The one-dimensional convolutional layer is used to extract local feature patterns. Each convolutional layer is followed by a batch normalization layer to stabilize the feature distribution and a nonlinear activation function layer to enhance the nonlinear expression capability of the network. In this embodiment, the core principle of the batch normalization layer is to standardize the output features of the intermediate layer according to the mean and standard deviation of the batch in each training batch, so that the distribution tends to be balanced, thereby alleviating the problem of gradient vanishing or exploding. The operation is performed after the convolutional layer. Common activation functions in nonlinear activation function layers include ReLU, Sigmoid, and Tanh. Activation functions map the linear output of convolutional layers to a nonlinear space, enabling neural networks to fit complex nonlinear relationships and improve the model's expressive power and prediction performance. Fully connected layers compress multi-layer convolutional feature maps into fixed-dimensional embedding vectors, generating source center embedding sets, target center labeled embedding sets, and target center unlabeled embedding sets, respectively.
[0020] In this embodiment, step three specifically includes: A cost matrix is constructed to measure the joint distribution deviation between the source center embedding set and the target center embedding set. The cost matrix includes three components: the Euclidean distance between the source sample embedding and the target sample embedding is used as the basic geometric distance term in the embedding space; for the labeled samples in the target center, the true labels of the labeled samples in the target center are compared with the labels of the source samples. If the labels are inconsistent, a penalty factor is introduced into the cost matrix to form a label consistency sub-item. The consistency sub-item is used to explicitly guide the priority matching of samples with the same risk level in the transmission plan; key clinical indicator fields highly correlated with deep vein thrombosis are extracted, including D-dimer, prothrombin time, international normalized ratio, activated partial thromboplastin time, and fibrinogen concentration. This set of key indicators is input into the medical ranking gating mapping module to perform indicator ranking calculation. The ranking penalty factor is calculated based on the ranking difference. The medical ranking gating mapping module maps the ranking difference into segments as penalty amplitude based on the preset clinical risk stratification standard to achieve dynamic control of the ranking relationship of high-risk indicators. A label-driven regularization sub-item is introduced, which adds a label-driven term to the original DeepJDOT entropy regularization loss structure. In the process of calculating the transmission distribution entropy, the label-driven term gives low-entropy encouragement to sample pairs that match the same type of label and imposes high-entropy penalty on sample pairs with inconsistent labels, thereby adjusting the stability of the pseudo-label propagation path in the transmission plan. This invention introduces a label-guided regularization sub-term based on the DeepJDOT entropy regularization term to improve the stability of false label propagation. Specifically, during the calculation of the optimal transmission plan, the source center sample label is used as a reference, and the matching relationship of each pair of source center and target center samples is divided into two categories: label consistent and label inconsistent. Then, in the information entropy regularization term, sample pairs with consistent labels are assigned a lower entropy weight to encourage them to occupy a higher matching probability in the transmission plan; sample pairs with inconsistent labels are assigned a higher entropy weight, thereby actively compressing their transmission quality during iterative updates and reducing the impact of false matches. A sample reliability weighting mechanism is constructed, which uses the training error of the source center sample and the deviation of the embedding distribution of the target center sample as a reliability estimation index. Learnable weights are applied to each sample pair in the cost matrix to control the actual impact of the sample in the transmission plan. The learnable weights are kept updated in the Sinkhorn-Knopp iteration. Based on a composite cost matrix that includes Euclidean distance, label consistency, medical ranking penalty, label-guided regularization sub-item, and sample reliability weights.
[0021] In this embodiment, in the DeepJDOT joint distributed optimal transmission structure of the present invention, a composite cost matrix is constructed to measure the joint matching cost between the source center sample and the target center sample. The definition is as follows: ; in, Indicates the source center sample index. Indicates the target center sample index; Indicates the first Embedding vectors of each source center sample Indicates the first Embedding vectors of each target center sample; For the Euclidean distance term in the embedded space; This is a label consistency indicator function; it takes the value when labels are inconsistent. Otherwise This constitutes a label consistency item; The number of key medical indicators; and These are the original feature vectors of the source center and target center samples, respectively; Indicates the first The ranking values of key indicators in the sample set; For the first The importance weights of each key indicator can be obtained through feature importance analysis or SHAP value evaluation during the training phase. , which is the weight adjustment coefficient for each term in the cost matrix, is a positive real number, and is used to balance the overall weights of the Euclidean distance term, label consistency term, and medical ranking penalty term in the cost matrix; The composite cost matrix, as the core part of the optimal transmission loss calculation, is input into the Sinkhorn-Knopp algorithm with entropy regularization to complete the joint alignment of cross-center source target sample pairs and the optimization of pseudo-label propagation paths.
[0022] In this embodiment, the process of constructing the penalty factor includes: For the label consistency term, a sample pair is constructed between the source center embedding set and the target center labeled embedding set. The source center sample label is compared with the target center labeled sample label. If the labels are consistent, the label consistency term at the corresponding cost matrix position is assigned a value of zero. If the labels are inconsistent, a fixed non-zero penalty value is assigned to the corresponding position. The label consistency term constrains the transmission plan to prioritize matching sample pairs with consistent risk levels. For the medical ranking penalty item, key clinical indicator fields are extracted from the source center embedding set and the target center embedding set respectively. The key clinical indicator fields include D-dimer level, prothrombin time, activated partial thromboplastin time, fibrinogen concentration and international normalized ratio. Each key indicator is numerically ranked in the source center sample set and the target center sample set respectively. For the constructed sample pairs, the absolute value of the ranking difference between the source center sample and the target center sample is calculated. The absolute value of the ranking difference of each key indicator is multiplied by the corresponding weight coefficient and then summed to obtain the medical ranking penalty value of the sample pair. The weight coefficients corresponding to each key indicator are obtained through feature importance analysis during the source center training process. The feature importance analysis method adopts the feature importance score based on the random forest model or the global feature contribution distribution calculated based on SHAP value. The higher the value of the key indicator, the greater the ranking penalty weight. The label consistency term and the medical ranking penalty term are superimposed on the basic Euclidean distance term to form the joint distributed optimal transmission cost matrix used for optimal transmission solution.
[0023] In this embodiment, step four specifically includes: Based on the optimal transmission cost matrix, the Sinkhorn-Knopp iterative algorithm with entropy regularization is used to solve the transmission plan. In each iteration, the following operations are performed: Initialize the source center marginal distribution and the target center marginal distribution as uniform distributions; use an exponential function to scale the cost matrix element by element to form an initial transmission kernel; sequentially perform vector normalization operations on the transmission kernel according to the source center marginal distribution and the target center marginal distribution, update the transmission matrix until the preset convergence condition is met, and form a normalized transmission plan; In the normalized transmission plan, a matching confidence guidance mechanism is introduced. Specifically, at the end of each iteration, the confidence of the sample is calculated based on the maximum transmission probability of the column vector corresponding to the unlabeled sample at the target center. If the confidence is not less than 80%, the unlabeled sample at the target center is assigned the label of the source center sample corresponding to the maximum matching probability, forming a pseudo-label set. Based on the normalized transmission plan, a probability normalization redistribution strategy is introduced. The transmission probabilities of each row of the source center and each column of the target center are summed and normalized respectively. The transmission weights of the source center samples and the pseudo-label samples of the target center are calculated. The probability normalization redistribution strategy refers to the process of performing weighted summation and normalization on the transmission matrix after completing the entropy-regularized Sinkhorn-Knopp iteration and obtaining the initial transmission plan matrix. To ensure the interpretability and numerical stability of the weight allocation during training, the transmission matrix is processed by weighted summation and normalization on both rows and columns: For each source center sample, the cumulative transmission quality of the source sample is obtained by summing all transmission probability values of the corresponding row vector in the transmission matrix, and this is normalized to serve as the sample weight of the source sample; for each unlabeled target center sample, the acceptance probability quality of the unlabeled sample is obtained by summing all transmission probability values of the corresponding column vector in the transmission matrix, and this is normalized to serve as the training weight of the pseudo-labeled sample. Through this explicit probability accumulation and normalization process, the probability normalization redistribution strategy ensures that the matching degree of each sample pair in the transmission plan is transformed into a continuous weight signal during the training phase, thereby enhancing the stability and effectiveness of sample weighted training. The matching confidence guidance mechanism and probability normalization redistribution strategy, as improvements to the standard Sinkhorn-Knopp iterative process, enhance the stability of high-confidence paths in the transmission plan and suppress the misleading effect of low-confidence matching, thereby improving the accuracy of pseudo-label generation and the robustness of cross-center data alignment.
[0024] In this embodiment, step five specifically includes: The source center samples, target center labeled samples, and target center unlabeled samples with pseudo-labels obtained through the transmission plan are merged to form an expanded training set; transmission weights are added to each source center sample, pseudo-label weights are added to each pseudo-labeled target center sample, and target center labeled samples are set to unit weights. Based on expanding the training set and simultaneously optimizing the feature extraction network and the classification head, a joint loss function is constructed. The joint loss function consists of the following three parts: optimal transmission loss, which calculates the distribution difference between the source center and the target center in the joint embedding space and the predicted label space based on the joint distribution optimal transmission cost matrix generated in step three and the transmission plan matrix updated in step four; source center cross-entropy loss, which calculates cross-entropy using the true labels of the source center samples and the output results of the classification head, and introduces the source center sample weights to weight the loss value; and target center cross-entropy loss, which calculates cross-entropy for the labeled target center samples and the pseudo-labeled target center samples respectively, and works together with their respective sample weights to optimize the overall loss. By jointly minimizing the above three losses, the parameters of the feature extraction network and the classification head are updated synchronously through backpropagation until the prediction loss of the expanded training set converges, and the original prediction probability corresponding to each input sample is output.
[0025] In this embodiment, the transcentral deep vein thrombosis risk prediction method based on semi-supervised optimal transmission described in this invention constructs the following joint loss function during training: ; The optimal transmission loss, used to measure the difference in the joint distribution of source-center samples and target-center samples in the embedding space and the predicted label space, is defined as: ;in, Indicates the number of samples at the source center. Indicates the number of samples at the target center. Indicates source sample With target sample Transmission plan matrix elements between This represents the cost term for the corresponding sample pair in the composite cost matrix; The source-center cross-entropy loss is defined as: ;in, For source samples Transmission weight, Its embedding vector, To give it its true label, This represents the output function of the classification head. Represents the cross-entropy loss function; The target center cross-entropy loss, comprising two parts: labeled target samples and pseudo-labeled samples, is defined as follows: ;in, This indicates the number of labeled samples at the target center. and These are the embedding vector and the true label of the sample, respectively. Indicates the number of pseudo-labeled samples. Represents its embedding vector, Indicates a pseudo tag. For the transmission weights of pseudo-labeled samples, and These are the loss weight coefficients corresponding to the source center and target center monitoring signals, respectively, both of which are positive real numbers; The joint loss function is used during training to simultaneously optimize the parameters of the feature extraction network and the classification head, ultimately outputting the raw predicted probability for each input sample.
[0026] In this embodiment, step six specifically includes: The expanded training set is input into the feature extraction network and the classification head, and the process of minimizing the joint loss function is performed. Gradient backpropagation and iterative updates are performed simultaneously on the parameters of the feature extraction network and the parameters of the classification head. Specifically, after inputting the expanded training set into the feature extraction network and the classification head, the optimal transmission loss, source center cross-entropy loss, and target center cross-entropy loss are calculated sequentially according to the constructed joint loss function. The three losses are then weighted and summed according to preset weights to form the overall optimization objective. The gradient of the joint loss function with respect to the network parameters is differentiated through an automatic differentiation mechanism. The parameters of the feature extraction network and the classification head are updated respectively using the backpropagation algorithm. This allows the model to gradually adjust the feature representation and classification boundary in each iteration to minimize the overall error of the training samples in the cross-center embedding alignment and label prediction process, thereby improving the model's adaptability and generalization ability. After each round of training iterations, the negative log-likelihood loss of the current model on the validation set is calculated based on the target center validation set, the validation loss is recorded and compared with the result of the previous round of iterations; Specifically, the validation set is input into the feature extraction network and classification head during training to obtain the predicted probability distribution of each validation sample; combined with the real label, the log-likelihood value under the current predicted distribution is calculated, and then the negative value is taken as the loss; the smaller the negative log-likelihood loss value, the closer the model's predicted probability of the real label is to one, and the more reliable the prediction result is. The metric serves as the criterion for stopping training and saving the model, and can dynamically evaluate the stability of pseudo-label transfer and the convergence of the model without introducing test label information. When the negative log-likelihood loss of the target center validation set reaches the preset stopping condition during several consecutive rounds of iterative training, the iterative training process is terminated, and the model parameters of the current round are saved as the optimal model parameters. Based on the trained feature extraction network and classification head, forward reasoning is performed on all samples in the expanded training set and the target center validation set, and the original predicted probability corresponding to each sample is output.
[0027] In this embodiment, step seven specifically includes: The original predicted probabilities are paired with the true labels in the target center validation set to construct a temperature scaling model with negative log-likelihood loss as the objective function. All parameters of the feature extraction network and the classification head are fixed, and only the temperature parameters in the temperature scaling model are trained. The negative log-likelihood loss is iteratively optimized until convergence. In this embodiment, during the probability calibration of the original predicted probability, the original predicted probability is first matched one-to-one with the real label of the corresponding sample in the target center verification set to form a calibration training pair, which is used as the input of the temperature scaling model. The temperature scaling model uses a single learnable parameter to perform division scaling on the original predicted probability in order to adjust the confidence distribution structure of the prediction distribution. When training the temperature scaling model, all parameters in the feature extraction network and the classification head are kept constant, and only the temperature parameter is used as an optimizable variable in backpropagation. The negative log-likelihood loss on the validation set is used as the optimization objective, and the temperature parameter is updated by gradient. Through iterative optimization, the predicted probability is made to be more numerically close to the probability distribution of the true label, thereby achieving the calibration effect. The iterative process continues until the temperature parameter converges or the early termination condition is met. The temperature parameter at this point is the optimal temperature, which is used to uniformly scale all the original predicted probabilities. The calibration process does not change the structure and prediction order of the original model, but only performs a monotonic transformation on the numerical scale of the probability output to ensure the logical consistency of the risk level classification. The original predicted probability is scaled using the temperature parameters obtained from training. The scaled result is then input into the sigmoid function, which outputs a calibrated probability value. Fixed thresholds of 0.25, 0.5, and 0.75 are set, and the risk level of the sample is classified as low risk, medium risk, high risk, or extremely high risk based on the calibrated probability value. When a sample is classified as high risk or extremely high risk, an early warning trigger operation is executed.
[0028] Implementation Case 1: To verify the feasibility of this invention in practice, it was applied to the deep vein thrombosis (DVT) risk prediction task in a medical data integration platform. The platform integrates structured inpatient data from multiple medical centers, which exhibits typical cross-center scenario characteristics such as inconsistent sample label distribution, differences in feature definitions, and a high proportion of unlabeled samples in some centers.
[0029] In the real-world scenario, the source center is a relatively complete historical sample database containing 10,000 labeled DVT-related hospitalization records, while the target center is a newly accessed center containing only 200 labeled records and 1,200 unlabeled records. Each sample includes 23 standardized features such as gender, age, height, weight, family history of thrombosis, D-dimer, coagulation factor levels, anticoagulant use, and MRI venography. Traditional methods, based on single-center logistic regression scoring and using fixed-weight formulas for risk assessment, are unstable on the target center validation set, exhibiting high false positive rates and failing to effectively utilize unlabeled data.
[0030] The above data is processed according to the method of this invention. First, multi-center data cleaning and normalization are completed. Valid fields are selected using predefined rules, and samples with excessively high missing rates are removed. In the feature extraction stage, a lightweight convolutional structure is used to construct embedding representations, and the embedding sets of source centers and target centers are output separately. In the alignment stage, an improved DeepJDOT with label-guided regularization sub-items is used to construct a composite cost matrix, which integrates clinical risk indicator ranking information. The transfer plan is optimized through Sinkhorn-Knopp iteration. After setting the pseudo-label confidence threshold to 0.8, pseudo-labels are automatically assigned to unlabeled samples of the target center. Sample weights are generated according to confidence levels. During training, a joint loss function is used to guide the feature extraction network and the classification head to update synchronously. Finally, the original output probability is calibrated through a temperature scaling module to obtain the probability output used for the final risk level classification.
[0031] The model was evaluated on a target center validation set to compare its accuracy and risk identification capabilities. The results showed that, compared with traditional DVT scoring rules and fully supervised random forest models, the proposed solution exhibits higher sensitivity and stronger risk stratification capabilities in environments with few labeled data, especially in the identification of extremely high-risk populations, where the accuracy is improved by nearly 20%. The calibrated predicted probabilities are more evenly distributed across the four risk levels, effectively supporting clinical early warning procedures.
[0032] Table 1: Comparison of the effectiveness of different methods in predicting DVT risk at target centers
[0033] Table 1 shows the performance comparison of the method of this invention with the traditional scoring method and the fully supervised random forest model on the target center validation set. From the AUC index, the traditional rule scoring method is only 0.654, indicating that the overall discrimination ability is limited and it is difficult to effectively capture the nonlinear relationship between features due to fixed weights and a single rule. The fully supervised random forest model has a slight improvement under the condition of completely relying on labeled samples, with an AUC of 0.731. However, it still has the problem of insufficient prediction generalization ability in target centers with scarce labels.
[0034] In contrast, the method of this invention, after introducing a semi-supervised optimal transfer mechanism, achieves an AUC of 0.814 even without temperature calibration, which is superior to traditional methods and fully supervised models. This indicates that it has a strong effect on cross-center embedding alignment and pseudo-label generation. In terms of sensitivity index, the traditional rule method and random forest are 0.498 and 0.586 respectively, indicating that there are large omissions in the identification of high-risk samples. The method of this invention achieves 0.713, showing a stronger ability to identify individuals at high risk of DVT, which is helpful for early warning.
[0035] In terms of accuracy for identifying extremely high risks, traditional methods achieve only 0.441, random forests improve this to 0.602, while the method of this invention reaches 0.758. This verifies the advantage of the improved DeepJDOT algorithm in maintaining consistency in medical risk ranking and enabling more accurate differentiation of high-risk groups. Regarding the average calibration error, this invention already outperforms other methods in the uncalibrated state, and further reduces it to 0.024 after temperature scaling, improving the interpretability and reliability of the model's output probability.
[0036] In summary, the above analysis shows that the present invention outperforms existing methods in terms of accuracy, sensitivity, and risk level identification capabilities, and exhibits stronger generalization ability and practical value, especially in cross-center application environments.
[0037] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transport, characterized in that, The process includes the following steps: Step 1: Collect source center labeled feature sets, target center labeled feature sets, and target center unlabeled feature sets, and perform preprocessing to obtain the corresponding preprocessed feature sets; Step 2: Input the preprocessed feature sets into the feature extraction network, and output source center embedding sets, target center labeled embedding sets, and target center unlabeled embedding sets; Step 3: Based on the source center embedding sets and target center embedding sets, construct a composite cost matrix using the DeepJDOT joint distributed optimal transport algorithm to generate a transport plan; Step 4: Solve the transport plan through Sinkhorn-Knopp iterative calculation, and for target centers with a matching confidence of not less than 80%,... Step 5: Assign pseudo-labels to unlabeled samples, calculate the weights of source center samples and unlabeled target center samples; Step 6: Merge source center samples, labeled target center samples, and pseudo-labeled target center samples along with their corresponding weights into an expanded training set, and simultaneously optimize the feature extraction network and classification head using the optimal transmission loss, source center cross-entropy loss, and target center cross-entropy loss, outputting the original predicted probabilities; Step 7: Repeat steps 3 to 5 to iteratively update the feature extraction network and classification head until the negative log-likelihood of the target center validation set converges, obtaining the probabilities before calibration; Step 8: Calibrate the probabilities on the target center validation set using the temperature scaling method, obtain the calibrated probability values, and classify the risk levels according to the threshold.
2. The method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transmission according to claim 1, characterized in that, Step one specifically includes: acquiring clinical data including gender, age, height, weight, family history of coagulation dysfunction, history of thrombosis, smoking history, drinking history, surgical history, pregnancy history, platelet count, prothrombin time, thrombin time, activated partial thromboplastin time, fibrinogen concentration, international normalized ratio, D-dimer level, coagulation factor level, antithrombin level, magnetic resonance venous imaging results, and heparin drug use; performing data cleaning, data aggregation, and data annotation on the data to obtain a labeled dataset; performing feature extraction and feature selection on the labeled dataset to generate a processed dataset; randomly dividing the processed dataset into a training set and a test set; training the model on the training set using a random forest model, and calculating the prediction error based on the test set; repeating cross-validation until the prediction error meets the preset convergence condition; outputting the trained random forest model and using the results as the initialization reference for the source center labeled feature set, the target center labeled feature set, and the target center unlabeled feature set.
3. The method for predicting the risk of transcentral deep vein thrombosis based on semi-supervised optimal transmission according to claim 2, characterized in that, Step two specifically includes: inputting the source center preprocessed feature set, the target center labeled preprocessed feature set, and the target center unlabeled preprocessed feature set into a feature extraction network. The feature extraction network adopts a structure consisting of an input layer, several one-dimensional convolutional layers, a normalization layer, an activation function layer, and a fully connected layer connected sequentially. The input layer is used to receive structured clinical feature tensors, and the input dimension of each sample is consistent with the number of fields contained in the feature set. The one-dimensional convolutional layer is used to extract local feature patterns. Each convolutional layer is followed by a batch normalization layer and a non-linear activation function layer. The fully connected layer compresses the multi-layer convolutional feature mapping into a fixed-dimensional embedding vector, generating the source center embedding set, the target center labeled embedding set, and the target center unlabeled embedding set, respectively.
4. The method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transmission according to claim 3, characterized in that, Step three specifically includes: A cost matrix is constructed to measure the joint distribution deviation between the source center embedding set and the target center embedding set. The cost matrix includes three components: the Euclidean distance between the source sample embedding and the target sample embedding is used as the basic geometric distance term of the embedding space; for the labeled samples in the target center, the true labels of the labeled samples in the target center are compared with the labels of the source samples. If the labels are inconsistent, a penalty factor is introduced into the cost matrix to form a label consistency sub-item; key clinical indicator fields with high correlation to deep vein thrombosis are extracted, including D-dimer level, prothrombin time, international normalized ratio, activated partial thromboplastin time, and fibrinogen concentration. The relevant indicators are input into the medical ranking gating mapping module to perform indicator ranking calculation. The ranking penalty factor is calculated based on the ranking difference. Based on the preset clinical risk stratification standard, the ranking difference is mapped to the penalty amplitude in segments to realize the dynamic control of the ranking relationship of high-risk indicators by the medical ranking gating mapping module. A label-driven regularization sub-item is introduced, which adds a label-driven term to the original DeepJDOT entropy regularization loss structure. In the process of calculating the transmission distribution entropy, the label-driven term gives low-entropy encouragement to sample pairs that match the same type of label and imposes high-entropy penalty on sample pairs with inconsistent labels, thereby adjusting the stability of the pseudo-label propagation path in the transmission plan. A sample reliability weighting mechanism is constructed, which uses the training error of the source center sample and the deviation of the embedding distribution of the target center sample as a reliability estimation index. Learnable weights are applied to each sample pair in the cost matrix to control the actual impact of the sample in the transmission plan. The learnable weights are kept updated in the Sinkhorn-Knopp iteration. Based on a composite cost matrix that includes Euclidean distance, label consistency, medical ranking penalty, label-guided regularization sub-item, and sample reliability weights.
5. The method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transmission according to claim 4, characterized in that, The process of constructing the penalty factor includes: For the label consistency term, a sample pair is constructed between the source center embedding set and the target center labeled embedding set. The source center sample label is compared with the target center labeled sample label. If the labels are consistent, the label consistency term at the corresponding cost matrix position is assigned a value of zero. If the labels are inconsistent, a fixed non-zero penalty value is assigned to the corresponding position. The label consistency term constrains the transmission plan to prioritize matching sample pairs with consistent risk levels. For the medical ranking penalty item, key clinical indicator fields are extracted from the source center embedding set and the target center embedding set respectively. The key clinical indicator fields include D-dimer concentration, prothrombin time, activated partial thromboplastin time, fibrinogen concentration and international normalized ratio. Each key indicator is numerically ranked in the source center sample set and the target center sample set respectively. For the constructed sample pairs, the absolute value of the ranking difference between the source center sample and the target center sample is calculated. The absolute value of the ranking difference of each key indicator is multiplied by the corresponding weight coefficient and then summed to obtain the medical ranking penalty value of the sample pair. The weight coefficients corresponding to each key indicator are obtained through feature importance analysis during the source center training process. The label consistency term and the medical ranking penalty term are superimposed on the basic Euclidean distance term to form the joint distributed optimal transmission cost matrix used for optimal transmission solution.
6. The method for predicting the risk of transcentral deep vein thrombosis based on semi-supervised optimal transmission according to claim 5, characterized in that, Step four specifically includes: Based on the optimal transmission cost matrix, the Sinkhorn-Knopp iterative algorithm with entropy regularization is used to solve the transmission plan. In each iteration, the following operations are performed: Initialize the source center marginal distribution and the target center marginal distribution as uniform distributions; use an exponential function to scale the cost matrix element by element to form an initial transmission kernel; sequentially perform vector normalization operations on the transmission kernel according to the source center marginal distribution and the target center marginal distribution, update the transmission matrix until the preset convergence condition is met, and form a normalized transmission plan; In the normalized transmission plan, a matching confidence guidance mechanism is introduced. Specifically, at the end of each iteration, the confidence of the sample is calculated based on the maximum transmission probability of the column vector corresponding to the unlabeled sample at the target center. If the confidence is not less than 80%, the unlabeled sample at the target center is assigned the label of the source center sample corresponding to the maximum matching probability, forming a pseudo-label set. Based on the normalized transmission plan, a probability normalization redistribution strategy is introduced. The transmission probabilities of each row of the source center and each column of the target center are summed and normalized respectively, and the transmission weights of the source center samples and the pseudo-label samples of the target center are calculated.
7. The method for predicting the risk of transcentral deep vein thrombosis based on semi-supervised optimal transmission according to claim 6, characterized in that, Step five specifically includes: The source center samples, target center labeled samples, and target center unlabeled samples with pseudo-labels obtained through the transmission plan are merged to form an expanded training set; transmission weights are added to each source center sample, pseudo-label weights are added to each pseudo-labeled target center sample, and target center labeled samples are set to unit weights. Based on expanding the training set and simultaneously optimizing the feature extraction network and the classification head, a joint loss function is constructed. The joint loss function consists of the following three parts: optimal transmission loss, which calculates the distribution difference between the source center and the target center in the joint embedding space and the predicted label space based on the joint distribution optimal transmission cost matrix generated in step three and the transmission plan matrix updated in step four; source center cross-entropy loss, which calculates cross-entropy using the true labels of the source center samples and the output results of the classification head, and introduces the source center sample weights to weight the loss value; and target center cross-entropy loss, which calculates cross-entropy for the labeled target center samples and the pseudo-labeled target center samples respectively, and works together with their respective sample weights to optimize the overall loss. By jointly minimizing the above three losses, the parameters of the feature extraction network and the classification head are updated synchronously through backpropagation until the prediction loss of the expanded training set converges, and the original prediction probability corresponding to each input sample is output.
8. The method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transmission according to claim 7, characterized in that, Step six specifically includes: The expanded training set is input into the feature extraction network and the classification head, and the process of minimizing the joint loss function is performed. Gradient backpropagation and iterative updates are performed simultaneously on the parameters of the feature extraction network and the parameters of the classification head. After each round of training iterations, the negative log-likelihood loss of the current model on the validation set is calculated based on the target center validation set, the validation loss is recorded and compared with the result of the previous round of iterations; When the negative log-likelihood loss of the target center validation set reaches the preset stopping condition during several consecutive rounds of iterative training, the iterative training process is terminated, and the model parameters of the current round are saved as the optimal model parameters. Based on the trained feature extraction network and classification head, forward reasoning is performed on all samples in the expanded training set and the target center validation set, and the original predicted probability corresponding to each sample is output.
9. The method for predicting the risk of transcenter deep vein thrombosis based on semi-supervised optimal transmission according to claim 8, characterized in that, Step seven specifically includes: The original predicted probabilities are paired with the true labels in the target center validation set to construct a temperature scaling model with negative log-likelihood loss as the objective function. All parameters of the feature extraction network and the classification head are fixed, and only the temperature parameters in the temperature scaling model are trained. The negative log-likelihood loss is iteratively optimized until convergence. The original predicted probability is scaled using the temperature parameters obtained from training. The scaled result is then input into the sigmoid function, which outputs a calibrated probability value. Fixed thresholds of 0.25, 0.5, and 0.75 are set, and the risk level of the sample is classified as low risk, medium risk, high risk, or extremely high risk based on the calibrated probability value. When a sample is classified as high risk or extremely high risk, an early warning trigger operation is executed.