A classification system for chronic venous insufficiency
By constructing a multi-task classification and grading system, combined with Transformer and BERT encoders, the problems of subjectivity and insufficient data integration in the assessment of chronic venous insufficiency were solved. This enabled end-to-end assistance from classification and grading to treatment recommendations, improving diagnostic and treatment efficiency and accuracy, and adapting to dynamic guideline updates.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- BEIJING LUHE HOSPITAL AFFILIATED TO CAPITAL MEDICAL UNIV
- Filing Date
- 2025-09-28
- Publication Date
- 2026-06-23
AI Technical Summary
Existing technologies for assessing chronic venous insufficiency (CVI) suffer from problems such as high subjectivity, insufficient integration of multidimensional data, disconnect between treatment recommendations and guideline updates, and low efficiency in multitasking, resulting in insufficient diagnostic and treatment efficiency and accuracy.
A multi-task classification and grading system is constructed by employing a data acquisition module, a shared feature extraction module, an attention perception module, and a multi-task classification and grading model, combined with a Transformer encoder and a BERT encoder. This system provides end-to-end assistance from data acquisition to treatment recommendation. Gradient masking technology and a triplet loss function are used to ensure the accuracy and interpretability of the system.
It enables end-to-end assistance for CVI from classification and grading to treatment recommendations, improving diagnostic efficiency and accuracy, adapting to dynamic guideline updates, and enhancing the system's robustness and real-time performance.
Smart Images
Figure CN121148732B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical system technology, and more specifically to a classification and grading system for chronic venous insufficiency. Background Technology
[0002] The pathophysiological mechanism of chronic venous insufficiency (CVI) is multifactorial, with its core pathological feature being chronic venous hypertension and the subsequent irreversible damage it causes. This hemodynamic disturbance can ultimately lead to characteristic skin lesions. Major risk factors include demographic characteristics (advanced age, female sex), behavioral factors (history of smoking, prolonged standing or sitting), pathophysiological factors (obesity, gastrocnemius pump dysfunction, limited ankle joint mobility), and medical history (deep vein thrombosis, phlebitis). Epidemiological data show that the global prevalence of CVI varies significantly geographically, ranging from 1% to 40%, and the incidence is higher in women across all age groups than in men.
[0003] The accurate assessment of CVI severity in current clinical practice remains a significant challenge because the functional state of the venous system is determined by complex hemodynamic mechanisms and cannot be assessed using a single indicator. From a pathophysiological perspective, the effectiveness of venous return depends not only on the integrity of venous valves but also on the synergistic effects of a dynamic equilibrium system involving vascular wall compliance, muscle pump efficiency, venous volume load, and microcirculation regulation. Currently, the widely used CEAP (Clinical-Etiology-Anatomic-Pathophysiologic Classification) system primarily relies on static assessment based on clinical manifestations (such as varicose veins, edema, and skin pigmentation), and has the following significant limitations:
[0004] The CEAP classification system is highly subjective and lacks dynamic assessment: It relies on the physician's subjective judgment of clinical manifestations and only reflects the state of the disease at a certain point in time. It cannot capture the dynamic changes of parameters (such as venous reflux pressure URP, maximum venous pressure MWP, etc.) before and after treatment. For example, a patient with C3 edema may be downgraded due to improvement in venous pressure after treatment, but the current system cannot quantify this dynamic change.
[0005] Insufficient integration of multi-dimensional data: Traditional classification and grading rely solely on single clinical indicators, failing to fully integrate hemodynamic parameters (such as FWP, RT90), anatomical information, and individual characteristics (such as age and comorbidities). For example, while venous Doppler imaging is the gold standard, it cannot directly correlate anatomical classification with CVI grading.
[0006] Treatment recommendations are out of sync with guideline updates: Current treatment protocols are largely based on static guideline rules, lacking fusion analysis of individual patient characteristics and real-time data. For example, after clinical guidelines are updated, traditional systems require manual rule adjustments and cannot automatically adapt to the latest diagnostic and treatment standards.
[0007] Multitasking is inefficient: Traditional methods typically handle classification and grading tasks separately, leading to repetitive model training and low feature utilization. For example, the correlation between anatomical classification and CVI grading is not fully explored, which may affect the final diagnostic accuracy. Summary of the Invention
[0008] To address the shortcomings of the existing technology, this invention aims to provide a classification and grading system for chronic venous insufficiency (CVI) to assist in the entire process from classification and grading to treatment recommendations.
[0009] To solve the above problems, the present invention adopts the following technical solution:
[0010] This invention provides a classification and grading system for chronic venous insufficiency, including a data acquisition module, a shared feature extraction module, an attention perception module, and a multi-task classification and grading model output module;
[0011] The data acquisition module is configured to collect patient data for chronic venous insufficiency, including pre- and post-treatment parameters URP, MWP, MSP, FWP, and RT90, as well as individual data on the classification, grading, and treatment methods of chronic venous insufficiency.
[0012] The shared feature extraction module is configured to model a time series based on the patient data using a Transformer encoder, capture the dependency relationship between the time series data and the classification and grading of chronic venous insufficiency, and connect to a fully connected layer to map the features and output a shared feature vector.
[0013] The attention perception module is configured to generate a type- and graded feature vector by linearly transforming each shared feature vector through the task attention matrix and normalizing it.
[0014] The multi-task classification and grading model output module is configured to construct a multi-task classification and grading model based on Softmax, and to train the multi-task classification and grading model using the weighted feature vectors of historical patient data. The input parameters are the pre-treatment parameters URP, MWP, MSP, FWP, and RT90 corresponding to the weighted feature vectors of the patient to be evaluated, as well as individual data. The first task head outputs the anatomical classification of chronic venous insufficiency, and the second task head outputs the CVI classification of chronic venous insufficiency.
[0015] Furthermore, it also includes a treatment recommendation module; the treatment recommendation module is configured to construct a treatment recommendation model using XGBoost, and takes the anatomical classification and CVI classification output by the multi-task classification and grading model, as well as the pre-treatment parameters URP, MWP, MSP, FWP and RT90, and the rule feature vectors of individual data and clinical guideline treatment methods as input, and outputs the recommended treatment method.
[0016] Furthermore, it also includes a clinical rules module; the clinical rules module is configured to receive structured clinical guideline text at the input layer, extract semantic features through a BERT encoder, and then generate rule feature vectors of clinical guideline treatment methods through a Transformer decoder. When the guideline is updated, a contrastive learning mechanism is triggered to fine-tune the encoder parameters through a triplet loss function.
[0017] Furthermore, the treatment recommendation model includes an auxiliary prediction head and a treatment recommendation task prediction head; the auxiliary prediction head is used to predict potential conflicts between the recommended treatment plan and the rules.
[0018] Furthermore, the triplet loss function includes:
[0019]
[0020] in, Let be the rule feature vector of the i-th unupdated guide. Let be the rule feature vector of the i-th updated guide, N be the total number of rule feature vectors to be compared, and m be the marginal parameter. Let be the rule feature vector of the j-th unupdated guide.
[0021] Furthermore, a dual-channel loss function is constructed based on the auxiliary prediction head and the treatment recommendation task prediction head:
[0022]
[0023]
[0024] in, For the total loss function, To determine the loss function for the recommended task prediction head, For conflict loss weights, To assist in predicting the head loss function, α is a hyperparameter. Let h be the rule constraint matrix, and h be the rule eigenvector;
[0025] By using gradient masking techniques, a rule-sensitive gradient threshold θ is set during backpropagation, which affects the gradient of the prediction head for the treatment recommendation task. When the gradient of the current task optimization is strong, the gradient of the rule-related parameters is masked by gradient shading technology. Only parameters that are not related to the rule are updated, so as to avoid the rule-related parameters being over-adjusted by the task gradient, thereby maintaining the effectiveness of the rule constraints.
[0026] when At this time, the task gradient is relatively mild, and both rule-related and non-rule-related parameters are updated simultaneously, taking into account both task optimization and rule constraints.
[0027] Furthermore, the classification of chronic venous insufficiency includes no, superficial venous reflux, deep venous reflux, or deep venous reflux combined with superficial venous reflux.
[0028] The beneficial effects of this invention are as follows: This invention realizes full-process assistance for CVI from classification and grading to treatment recommendations. Combining dynamic guideline updates and interpretable design, it takes into account both accuracy and clinical applicability, and can effectively assist doctors in improving diagnosis and treatment efficiency. In actual deployment, parameters can be fine-tuned according to the hardware conditions and data scale of medical institutions to ensure the robustness and real-time performance of the system. Attached Figure Description
[0029] Figure 1 This is a schematic diagram of a classification and grading system for chronic venous insufficiency according to the present invention. Detailed Implementation
[0030] The present invention will be further described in detail below with reference to specific embodiments.
[0031] It should be noted that these embodiments are only used to illustrate the present invention and are not intended to limit the present invention. Simple improvements to the method under the premise of the present invention are all within the scope of protection claimed by the present invention.
[0032] This classification and grading system for chronic venous insufficiency comprises, in sequence, a data acquisition module, a shared feature extraction module, an attention perception module, a multi-task classification and grading model output module, a treatment recommendation module, and a clinical rule module connected to the treatment recommendation module. These modules work collaboratively to achieve end-to-end processing from patient data acquisition to classification and grading diagnosis and treatment recommendation. The system architecture is as follows: Figure 1 As shown.
[0033] First, this classification and grading system for chronic venous insufficiency includes a data acquisition module 100, a shared feature extraction module 200, an attention perception module 300, and a multi-task classification and grading model output module 400.
[0034] The data acquisition module 100 is configured to collect patient data on chronic venous insufficiency, including pre- and post-treatment parameters URP, MWP, MSP, FWP, and RT90, as well as individual data on the classification, grading, and treatment of chronic venous insufficiency.
[0035] Venous function parameters before and after treatment included URP (upright resting pressure), MWP (mean walking pressure), MSP (mean systolic blood pressure), FWP (percentage decrease in walking pressure), and RT90 (90% pressure recovery time). These parameters were recorded in time series (e.g., continuous values at 1 day, 1 week, 4 weeks before treatment, and 1 day, 1 week, 4 weeks, and 12 weeks after treatment), while historical data were collected from electronic medical records.
[0036] Individual data includes basic patient information (age, gender, BMI), medical history (comorbidities such as diabetes and hypertension), current clinical classification (such as superficial venous reflux, deep venous reflux, and deep venous reflux combined with superficial venous reflux in the CEAP anatomical classification), current CVI classification (such as symptom-based C0-C6 levels), and treatment methods (such as medication, surgical procedures, compression therapy, etc.).
[0037] Data preprocessing:
[0038] Interpolation methods are used for missing values (e.g., linear interpolation for missing time series data, and mode filling for missing individual data); parameters are normalized (e.g., Z-score standardization); unstructured individual data (e.g., medical history text) are converted into structured features (e.g., "whether or not one has diabetes" is encoded as 0 / 1).
[0039] The shared feature extraction module 200 is configured to model time series data based on the patient data using a Transformer encoder, capture the dependency relationship between the time series data and the classification and grading of chronic venous insufficiency, and connect to a fully connected layer to map the features and output a shared feature vector.
[0040] Input: Preprocessed time series data (dimension T×D), where T is the time step, e.g., 4 measurement time points; D=5 represents 5 parameters including URP and MWP), overlaid with location encoding (using sine and cosine functions: , , =512) to preserve time sequence information.
[0041] Encoder layer: Contains 6 Transformer encoder layers, each consisting of a "multi-head self-attention mechanism" and a "feedforward network".
[0042] Multi-head self-attention: With 8 heads, each head having a feature dimension of 64, self-attention weights are used... Capture the dependencies between parameters at different time points (such as the association between changes in URP before and after treatment and CVI classification).
[0043] Feedforward network: consists of two linear layers (with ReLU activation function in between), with dimensions (512→2048→512).
[0044] Output: The time series feature matrix (dimension T×512) is obtained after processing by the encoder.
[0045] Shared feature mapping:
[0046] The time series feature matrix is subjected to global average pooling (T×512→512), and then mapped to a shared feature vector through two fully connected layers (512→256→256). Each layer employs LeakyReLU activation (negative slope 0.2) and Dropout (probability 0.1) to prevent overfitting and enhance the generalization ability of features.
[0047] The attention perception module 300 is configured to generate a feature vector with specific attention weights for fractal and hierarchical classification by linearly transforming each shared feature vector through the task attention matrix and normalizing it.
[0048] Task attention matrix design:
[0049] Attention matrices are defined for the two tasks: "anatomical classification" and "CVI grading". (t=1 corresponds to fractal, t=2 corresponds to hierarchical), and the matrix parameters are learned through training.
[0050] Perform a linear transformation on the shared feature vector s: Then, the task attention weights are obtained through Softmax normalization. .
[0051] Weighted feature generation:
[0052] Fractal task characteristics: , It is the element-wise product;
[0053] Characteristics of hierarchical tasks: ; in The characteristics that are more important for classification and grading are highlighted respectively (such as the greater impact of pressure parameter changes on grading).
[0054] The multi-task classification and grading model output module 400 is configured to construct a multi-task classification and grading model based on Softmax. The multi-task classification and grading model is trained using the weighted feature vectors of historical patient data. The input parameters are the pre-treatment parameters URP, MWP, MSP, FWP, and RT90 corresponding to the weighted feature vectors of the patient to be evaluated, as well as individual data. The first task head outputs the anatomical classification of chronic venous insufficiency, and the second task head outputs the CVI classification of chronic venous insufficiency.
[0055] Model structure:
[0056] The hierarchical task head (first task head) consists of a fully connected layer (256→128→4) and Softmax, and outputs four types of CVI anatomical classifications (no 1, superficial venous reflux 2, deep venous reflux 3, deep venous reflux combined with superficial venous reflux 4).
[0057] The typing task head consists of a fully connected layer (256→128→7) and a Softmax layer, and outputs 7 anatomical typing categories (corresponding to C0-C6 of the CEAP typing).
[0058] Model training:
[0059] Training data: historical patients' f1 and f2 scores and corresponding subtype and grading labels;
[0060] Loss function: Multi-task cross-entropy loss is used. ,in , For classification labels, (for predicting probabilities) Similarly;
[0061] Optimizer: AdamW (learning rate 1e-4, weight decay 1e-5), 50 training epochs, batch size 32, using an early stopping strategy (stop if validation set accuracy does not improve after 3 epochs).
[0062] Reasoning process:
[0063] Input: Pre-treatment parameters (URP, MWP, etc.) and individual data of the patient to be evaluated (preprocessed and then input into the shared feature extraction module);
[0064] Output: Anatomical classification (e.g., "C3") and CVI grade (e.g., "2"), with the final result being the category with a predicted probability higher than 0.5.
[0065] The clinical rules module 500 is configured to receive structured clinical guideline text as input, extract semantic features through a BERT encoder, and then generate rule feature vectors of clinical guideline treatment methods through a Transformer decoder. When the guideline is updated, a contrastive learning mechanism is triggered to fine-tune the encoder parameters through a triplet loss function.
[0066] Clinical guideline handling:
[0067] Input: Structured guideline text (such as the treatment rule in the 2023 ESC Guidelines for the Diagnosis and Treatment of Venous Diseases, formatted as "If CVI grade ≥3 and venous reflux is present, surgical treatment is recommended").
[0068] BERT Encoder: A pre-trained model, bert-base-uncased (12 layers, 768 dimensions), is trained on clinical guideline text for domain adaptation. The input format is "[CLS] rule text [SEP]". Input IDs are generated through WordPiece word segmentation, encoded by BERT to obtain word embeddings (768 dimensions), and then pooled to obtain initial semantic features. ;
[0069] Transformer decoder: A 2-layer decoder (4-head attention) decodes g into a regular feature vector. The vector dimension matches the patient features, facilitating subsequent concatenation. BERT fine-tuning parameters: learning rate 5e-5, AdamW optimizer (weight decay 0.01), batch size 16, 5 training epochs, gradient accumulation (accumulated 4 times) to improve the stability of small sample learning.
[0070] Guideline dynamic update mechanism:
[0071] When the guide is updated (e.g., a new version is released), comparative learning is triggered:
[0072] Triple sampling: Positive sample pairs are new and old versions of the regular feature vectors from the same guideline. Negative sample pairs are old features of the regular feature vectors of other randomly selected guidelines. ;
[0073] Triple loss function: ;
[0074] N is the total number of regular feature vectors to be compared, and m is the marginal parameter, which can be 0.5.
[0075] Training strategy: 128 triples are generated in each batch. Online hard sample mining (OHEM) is used to improve efficiency. The BERT encoder parameters are fine-tuned by loss to make the distance between features of the same rule before and after the update closer and the distance between features of other rules farther, thus preserving the stability of the core rules.
[0076] The treatment recommendation module 600 is configured to use XGBoost to build a treatment recommendation model. The model takes the anatomical classification and CVI classification output by the multi-task classification and grading model, as well as the pre-treatment parameters URP, MWP, MSP, FWP and RT90, and the rule feature vectors of individual data and clinical guideline treatment methods as input, and outputs the recommended treatment methods.
[0077] Feature splicing:
[0078] The input features include: typing results (unique heat encoding, 7 dimensions), grading results (unique heat encoding, 4 dimensions), pre-treatment parameters (5 dimensions), individual data (10 dimensions, such as age, comorbidities, etc.), and regular feature vector r (128 dimensions). The total dimensions after concatenation are 7+4+5+10+128=154.
[0079] XGBoost Treatment Recommendation Model:
[0080] Model structure: 100 decision trees (dynamically adjusted through early stopping strategy), maximum depth 8 (determined by grid search 3-10), learning rate 0.1 (decaying by 0.9 every 10 rounds), regularization parameters reg_alpha=0.1, reg_lambda=1, objective function is multi-class classification (output 3 treatment methods: drug treatment, pressure intervention, surgery).
[0081] Auxiliary prediction head: Fully connected layer (154→64→1), outputs collision probability. (Activated via Sigmoid) to predict conflicts between recommended solutions and guideline rules (e.g., when the rule requires "surgery for C6 level" but the model recommends medication, the probability of conflict increases).
[0082] Loss function and training:
[0083] Total loss: ;
[0084] in: The loss function for the treatment recommendation task prediction head; Let h be the rule constraint matrix, which is randomly initialized and then subjected to adversarial training to enhance rule sensitivity. Let h be the concatenated features and α be the hyperparameter.
[0085] Gradient occlusion: Set a threshold θ=0.1, when When this occurs, it indicates that the gradient for the current task optimization is strong, and the process is frozen. The gradient of the rule-related parameters is masked by gradient masking technology, and only parameters that are not related to the rule are updated. This avoids the rule-related parameters being over-adjusted by the task gradient, thereby maintaining the effectiveness of the rule constraints.
[0086] when At this time, the task gradient is relatively mild, and both rule-related and non-rule-related parameters are updated simultaneously, taking into account both task optimization and rule constraints.
[0087] This system outputs: CVI anatomical classification, CVI grade, recommended treatment plan, and probability of conflict (if any). If so, it will be marked "requires clinical review", etc.
[0088] Explainability visualization:
[0089] Rule constraint matrix heatmap: Rendered using D3.js, with a gradient from red (high conflict) to blue (low conflict) generated via d3.scaleSequential. Threshold classification: Conflict intensity > 0.7 is high risk, 0.3-0.7 is medium risk, and < 0.3 is low risk. Supports hover hints (displaying feature name, conflict value, and corresponding guideline rule) and feature dimension filtering (such as focusing on parameters like "RT90" and "MWP").
[0090] Importance of key features: through Calculate (8 is the number of attention heads) Let the weight of the j-th head be... The SHAP value for this header (calculated using shap.TreeExplainer) highlights the contribution of parameters such as URP variation and MWP to the decision-making process.
[0091] Conflict probability annotation: Annotation of key parameters such as ΔFWP (difference in FWP before and after treatment) It helps doctors determine the source of conflict.
[0092] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described with reference to preferred embodiments, those skilled in the art should understand that various changes in form and detail can be made without departing from the spirit and scope of the invention as defined in the appended claims.
Claims
1. A classification and grading system for chronic venous insufficiency, characterized in that, It includes a data acquisition module, a shared feature extraction module, an attention perception module, and a multi-task classification and hierarchical model output module; The data acquisition module is configured to collect patient data for chronic venous insufficiency, including pre- and post-treatment parameters URP, MWP, MSP, FWP, and RT90, as well as individual data on the classification, grading, and treatment methods of chronic venous insufficiency. The shared feature extraction module is configured to model a time series based on the patient data using a Transformer encoder, capture the dependency relationship between the time series data and the classification and grading of chronic venous insufficiency, and connect to a fully connected layer to map the features and output a shared feature vector. The attention perception module is configured to generate a type- and graded feature vector by linearly transforming each shared feature vector through the task attention matrix and normalizing it. Task attention matrix design: Attention matrices are defined for the two tasks: anatomical classification and CVI grading. t=1 corresponds to fractal, t=2 corresponds to hierarchical, and the matrix parameters are learned through training; Perform a linear transformation on the shared feature vector s: , Then, the task attention weights are obtained through Softmax normalization. ; Weighted feature generation: Fractal task characteristics: , It is the element-wise product; Tiered tasks feature: ; in Each highlights the features that are more important for classification and grading, respectively. The multi-task classification and grading model output module is configured to construct a multi-task classification and grading model based on Softmax, and to train the multi-task classification and grading model using the weighted feature vectors of historical patient data. The input parameters are the pre-treatment parameters URP, MWP, MSP, FWP and RT90 corresponding to the weighted feature vectors of the patient to be evaluated, as well as individual data. The first task head outputs the anatomical classification of chronic venous insufficiency, and the second task head outputs the CVI classification of chronic venous insufficiency. It also includes a treatment recommendation module; the treatment recommendation module is configured to use XGBoost to build a treatment recommendation model, and the anatomical classification and CVI classification output by the multi-task classification and grading model, as well as the pre-treatment parameters URP, MWP, MSP, FWP and RT90, and the rule feature vectors of individual data and clinical guideline treatment methods are concatenated as input, and the treatment recommendation model outputs the recommended treatment method. The treatment recommendation model includes an auxiliary prediction head and a treatment recommendation task prediction head; the auxiliary prediction head is used to predict potential conflicts between the recommended treatment plan and the rules. A dual-channel loss function is constructed based on the aforementioned auxiliary prediction head and treatment recommendation task prediction head: in, For the total loss function, To determine the loss function for the recommended task prediction head, For conflict loss weights, To assist in predicting the head loss function, α is a hyperparameter. Let h be the rule constraint matrix, and h be the rule eigenvector; By using gradient masking techniques, a rule-sensitive gradient threshold θ is set during backpropagation, which affects the gradient of the prediction head for the treatment recommendation task. When the gradient of the current task optimization is strong, the gradient of the rule-related parameters is masked by gradient shading technology. Only parameters that are not related to the rule are updated, so as to avoid the rule-related parameters being over-adjusted by the task gradient, thereby maintaining the effectiveness of the rule constraints. when At this time, the task gradient is relatively mild, and both rule-related and non-rule-related parameters are updated simultaneously, taking into account both task optimization and rule constraints.
2. The classification and grading system for chronic venous insufficiency according to claim 1, characterized in that, It also includes a clinical rules module; the clinical rules module is configured to receive structured clinical guideline text at the input layer, extract semantic features through a BERT encoder, and then generate rule feature vectors of clinical guideline treatment methods through a Transformer decoder. When the guideline is updated, a contrastive learning mechanism is triggered to fine-tune the encoder parameters through a triplet loss function.
3. The classification and grading system for chronic venous insufficiency according to claim 2, characterized in that, The triplet loss function includes: in, Let be the rule feature vector of the i-th unupdated guide. Let be the rule feature vector of the i-th updated guide, N be the total number of rule feature vectors to be compared, and m be the marginal parameter. Let be the rule feature vector of the j-th unupdated guide.
4. The classification and grading system for chronic venous insufficiency according to claim 1, characterized in that, The classification of chronic venous insufficiency includes no, superficial venous reflux, deep venous reflux, or deep venous reflux combined with superficial venous reflux.