Cervical cancer patient VTE monitoring lightweight model and construction method and application thereof

Abstract

The invention provides a lightweight model for monitoring VTE of a cervical cancer patient based on fusion of time features and immune indexes. The model is characterized in that eight key features with the highest contribution degree, namely D-dimers, FDP, WBC, HB, PLT, CD4, CD8 and the ratio of CD4 to CD8, are extracted through a pre-trained CliTsRNN model; encoding the time sequence of the key features into a time feature vector of'static value + dynamic change rate '; the method is obtained by adopting logistic regression training and has the characteristics of accuracy, high efficiency, interpretability, light weight and the like, the AUC reaches 0.918, the accuracy rate reaches 0.866, the F1 score reaches 0.864, the recall rate reaches 0.854, the accuracy rate reaches 0.875, the inference speed is smaller than 0.0001 s / example, the model is smaller than 50 MB, key factor sorting is supported, and the auditing requirements of evidence-based medicine are met.

Description

Technical Field

[0001] This invention relates to the field of medical testing technology, specifically to a lightweight model for monitoring VTE in cervical cancer patients, its construction method, and its application. Background Technology

[0002] Venous thromboembolism (VTE) is one of the most common life-threatening complications in gynecological cancer patients, directly affecting treatment tolerance and long-term prognosis. In major gynecological malignancies such as cervical cancer, endometrial cancer, and ovarian cancer, the incidence of VTE reaches 5.6%-14.6%, 6.8%, and 7.9%, respectively. Among these, the one-year mortality rate for symptomatic VTE patients is as high as 47.8%, and regardless of whether symptoms are present, VTE significantly reduces patient survival. A seven-year longitudinal study confirmed that the five-year survival rate for cervical cancer patients with VTE is less than 40%, far lower than the 80% for patients without thrombotic complications. This data highlights the urgent clinical need for early and accurate assessment of VTE risk; however, current technologies still have multi-dimensional shortcomings and are insufficient to meet the requirements of clinical practice.

[0003] To achieve early assessment of cervical thrombosis (VTE), researchers have explored the application of biomarkers. D-dimer, a classic screening indicator, has an optimal cutoff value of 3.1 μg / mL for VTE in cervical cancer. While the combined use of D-dimer with other biomarkers such as fibrin degradation products (FDP) and tissue plasminogen activator inhibitor complex (t-PAIC) can improve diagnostic sensitivity and specificity, it has two major limitations: first, the lack of standardized testing and interpretation procedures; and second, reliance on static biomarker combinations without considering the dynamic changes of indicators over time, and neglecting the mechanism that "thrombosis is the result of the interaction between the immune and coagulation systems"—currently, very few studies have incorporated immune characteristics (such as immune cell ratios and immune molecule levels) into VTE prediction systems, thus limiting the predictive efficacy of biomarkers.

[0004] The VTE risk assessment tools commonly used in clinical practice also have significant shortcomings. While the Khorana score is applicable to the general cancer population, it lacks specificity in the specific gynecological cancer population. Although the Caprini score and its modified version (which adds the platelet-lymphocyte ratio PLR) have some applications in cervical cancer, they are essentially "single-time-point static scores": they cannot capture the dynamic progression of patients' physiological indicators during VTE formation, nor do they integrate key immune features (such as the CD4 / CD8 ratio and immune function status), resulting in their predictive sensitivity and positive predictive value for VTE in gynecological cancer patients failing to meet clinical needs.

[0005] In recent years, deep learning has provided a new approach to processing complex medical data, especially demonstrating advantages in time series analysis (such as survival analysis). However, the application of existing sequence deep learning models in VTE prediction faces clinical translation bottlenecks: on the one hand, the models rely on high-dimensional computation, resulting in complex structures and high resource consumption, making them difficult to adapt to rapid clinical decision-making scenarios (the need for lightweight design is not met); on the other hand, the decision-making process is a "black box," failing to explain key predictive criteria and reducing clinical trust. More importantly, existing models have not yet systematically integrated immune characteristics with temporal dynamic information, failing to fill the aforementioned technological gaps.

[0006] In summary, current VTE assessment technologies suffer from four major pain points: static assessments cannot capture dynamic disease progression, the immune-coagulation interaction mechanism is not reflected in predictions, cumbersome models are difficult to deploy clinically, and decision-making lacks interpretability. These shortcomings result in insufficient accuracy and practicality in early VTE prediction, necessitating an innovative testing solution that can simultaneously achieve "time-dynamic capture, immune feature integration, lightweight deployment, and interpretable decision-making."

[0007] To fill the aforementioned gap, the inventors of this invention constructed a lightweight VTE detection model, TempVTE, that effectively utilizes patient time information. By analyzing the performance and parameters of a pre-trained deep learning model, the core features that contribute most to VTE prediction were extracted, confirming the high correlation between immune features and VTE (consistent with the immune-coagulation interaction mechanism). Based on a dataset of 2971 cervical cancer patients, time-series data was encoded into time feature vectors with additional interpretability, achieving model lightweighting. Experimental results show that this model significantly outperforms single-time-point detection schemes, achieving an area under the curve (AUC) of 0.918 and an accuracy of 0.866 after incorporating immune features. To further improve the model's interpretability, the inventors further analyzed the model parameters. The results show that the model emphasizes some important features, such as the high growth rate of D-dimer, the high decline rate of hemoglobin (HB), and the increase in the CD4 / CD8 ratio, providing a valid basis for the model's decision criteria. Ultimately, the goal of "accurate, efficient, and interpretable" VTE detection was achieved. Summary of the Invention

[0008] In view of this, the purpose of this invention is to provide a lightweight model for monitoring VTE in cervical cancer patients, its construction method, and its application.

[0009] To achieve the above objectives, the present invention provides the following technical solution:

[0010] This invention provides a lightweight model construction method for VTE monitoring in cervical cancer patients. The method extracts the eight key features with the highest contribution through a pre-trained CliTsRNN model: D-dimer, FDP, WBC, HB, PLT, CD4, CD8, and CD4 / CD8 ratio. The time series of the key features are encoded into a time feature vector of "static value + dynamic rate of change". Logistic regression is used for training to obtain a lightweight model for VTE monitoring in cervical cancer patients based on the fusion of time features and immune indicators.

[0011] Furthermore, the model construction method includes the following steps:

[0012] (1) Screening of the study population and data collection: Patients diagnosed with cervical cancer by pathological histology, who received treatment in medical institutions and had complete clinical information were selected, excluding patients with a history of VTE before treatment, secondary cervical cancer, or other primary malignant tumors; clinical information and laboratory indicators of patients were collected; VTE diagnosis results were jointly confirmed by two physicians; the laboratory indicators included coagulation indicators: D-dimer, fibrin degradation products FDP, prothrombin time PT, activated partial thromboplastin time APTT, fibrinogen FIB; routine blood indicators: white blood cell count WBC, hemoglobin HB, platelet count PLT; immune indicators: CD4 cell count, CD8 cell count, CD4 / CD8 ratio; liver and kidney function indicators: blood urea nitrogen, creatinine, bicarbonate;

[0013] (2) Key feature extraction: The pre-trained CliTsRNN deep learning model was used to analyze the weights of the input layer of the model. The features were sorted according to the absolute value of the coefficients and the eight key features with the highest contribution were selected, namely D-dimer, fibrin degradation product FDP, white blood cell count WBC, hemoglobin HB, platelet count PLT, CD4 cell count, CD8 cell count and CD4 / CD8 ratio.

[0014] (3) Construction of time feature vector: The time series data of the 8 key features are converted into time feature vectors containing static and dynamic information; the static information is the initial and final values ​​of the features, and the dynamic information is the total rate of change and the stage rate of change of the features;

[0015] (4) Model training and validation: The time feature vectors are divided into training set and validation set in a 9:1 ratio. The logistic regression model is used to train the model, and the model performance is evaluated by AUC, accuracy, F1 score, recall and precision.

[0016] Furthermore, the clinical information mentioned in step (1) includes age, body mass index (BMI), comorbidities, and treatment history; the comorbidities include hypertension, diabetes, and hypercholesterolemia; the treatment history includes radiotherapy, chemotherapy, targeted therapy, and immunotherapy.

[0017] Furthermore, the diagnostic criteria for VTE in step (1) are as follows: deep vein thrombosis (DVT) is confirmed by vascular ultrasound or computed tomography (CTA), and pulmonary embolism (PTE) is confirmed by chest CT pulmonary angiography or enhanced CT scan.

[0018] Furthermore, the pre-training dataset of the CliTsRNN model described in step (2) contains VTE time series data of cervical cancer, and the pre-training performance is AUC=0.941 and accuracy=87%.

[0019] Of course, the models constructed by the lightweight model construction method for VTE monitoring of cervical cancer patients described in any of the above-mentioned methods are also within the scope of protection of this invention.

[0020] This invention also provides the application of the lightweight model construction method for VTE monitoring in cervical cancer patients described in any of the above-mentioned methods in VTE risk assessment products for cervical cancer patients or in the study of the pathogenesis of thrombosis.

[0021] The present invention also provides a computer device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the model constructed by the lightweight model construction method for VTE monitoring in cervical cancer patients as described in any of the preceding claims to predict and assess the VTE risk in cervical cancer patients.

[0022] The present invention also provides a non-transitory computer-readable storage medium storing instructions that, when executed by a processor, perform a model constructed by the lightweight model construction method for VTE monitoring in cervical cancer patients as described above to predict and assess the VTE risk in cervical cancer patients.

[0023] The present invention also provides system software for predicting and assessing the risk of VTE in cervical cancer patients. The system software is deployed on a computer-readable storage medium and performs the following functions during operation:

[0024] (1) Data input and preprocessing: Receive clinical information and laboratory test data of cervical cancer patients; the laboratory test data includes at least D-dimer, fibrin degradation product FDP, white blood cell count (WBC), hemoglobin (HB), platelet count (PLT), CD4 cell count, CD8 cell count and CD4 / CD8 ratio.

[0025] (2) Key feature extraction: The pre-trained CliTsRNN deep learning model was called to extract the 8 key features that contribute the most to VTE detection: D-dimer, fibrin degradation product FDP, white blood cell count (WBC), hemoglobin (HB), platelet count (PLT), CD4 cell count, CD8 cell count, and CD4 / CD8 ratio.

[0026] (3) Construction of time feature vector: The time series data of the eight key features are converted into time feature vectors containing static and dynamic information; the static information is the initial and final detection values ​​of each feature, and the dynamic information is the total rate of change of each feature and the stage rate of change of different detection stages;

[0027] (4) VTE risk inference: Load the model constructed by the lightweight model construction method for VTE monitoring of cervical cancer patients described above, input the time feature vector into the model, calculate and output the VTE risk prediction results of cervical cancer patients;

[0028] (5) Results output: The VTE risk prediction results and key influencing factors of the model decision are presented in a visual form.

[0029] The beneficial effects of this invention are as follows: This invention provides a lightweight model for monitoring VTE in cervical cancer patients based on the fusion of time features and immune indicators. The model extracts the eight key features with the highest contribution through a pre-trained CliTsRNN model: D-dimer, FDP, WBC, HB, PLT, CD4, CD8, and CD4 / CD8 ratio. The time series of the key features are encoded into a time feature vector of "static value + dynamic rate of change". The model is trained using logistic regression and has the characteristics of accuracy, efficiency, interpretability, and lightweightness. The AUC reaches 0.918, the accuracy reaches 0.866, the F1 score reaches 0.864, the recall reaches 0.854, the precision reaches 0.875, the inference speed is <0.0001s / case, the model size is <50MB, it supports key factor ranking, and meets the requirements of evidence-based medicine audit. Attached Figure Description

[0030] Figure 1 This diagram illustrates the deep learning-based feature importance extraction, time series data visualization of VTE patients, and temporal feature transformation. (a) Diagram of the deep learning model and feature importance extraction; (b) Sample time series data of a VTE patient (feature values ​​have been normalized for visualization); (c) Temporal feature transformation process.

[0031] Figure 2 This study analyzes the dynamic evolution of VTE risk characteristics and the weighting of key indicators. Specifically, it examines (a) the development trends of key characteristics in VTE and non-VTE populations; and (b) the importance scores of these characteristics.

[0032] Figure 3 The performance, convergence, inference time, and coefficients of the models are compared. (a) The receiver operating characteristic (ROC) curve of the model; (b) The change in the loss function during training, which indicates the model's convergence; (c) The relationship between the model's average inference time and the length of the input time series; and (d) The model coefficients. Detailed Implementation

[0033] The technical solution of the present invention will be further described in detail below with reference to specific embodiments. It should be understood that the following embodiments are merely illustrative and explanatory of the present invention and should not be construed as limiting the scope of protection of the present invention. All technologies implemented based on the above content of the present invention are covered within the scope of protection intended by the present invention. It should be noted that the experimental materials whose source is not specified in the embodiments of the present invention are all commercially available, and the experimental methods whose specific conditions are not specified in the embodiments of the present invention are generally performed according to conventional experimental methods or according to the methods recommended by the experimental material manufacturers.

[0034] Example 1: Construction of a lightweight model for VTE monitoring in cervical cancer patients: TempVTE

[0035] This invention constructs a lightweight and highly interpretable lightweight model, TempVTE, for monitoring cervical ectasia (VTE) in cervical cancer patients through a process of "data screening - feature extraction - time encoding - model training - validation and optimization". The specific steps are as follows:

[0036] (1) Screening of the research population and data collection

[0037] Study population: 4307 cervical cancer patients who underwent surgery at Chongqing University Cancer Hospital between January 2017 and January 2024 were selected.

[0038] Inclusion criteria: ① Cervical cancer confirmed by pathological histology; ② Cervical cancer treatment completed at this hospital; ③ Complete clinical and laboratory test information.

[0039] Exclusion criteria: ① History of VTE (non-cervical cancer related) before treatment; ② Secondary cervical cancer or coexisting primary malignant tumors;

[0040] Data type:

[0041] Clinical information: age, body mass index (BMI), comorbidities (hypertension, diabetes, hypercholesterolemia), treatment history (radiotherapy, chemotherapy, targeted therapy, immunotherapy);

[0042] Laboratory indicators: coagulation-related indicators (D-dimer, FDP, prothrombin time PT, activated partial thromboplastin time APTT, fibrinogen FIB), routine blood indicators (white blood cell count WBC, hemoglobin HB, platelet count PLT), immune indicators (CD4 cell count, CD8 cell count, CD4 / CD8 ratio), liver and kidney function indicators (blood urea nitrogen, creatinine, bicarbonate);

[0043] VTE diagnostic criteria: Deep vein thrombosis (DVT) is confirmed by vascular ultrasound or computed tomography angiography (CTA), and pulmonary embolism (PE) is confirmed by chest CT pulmonary angiography or enhanced CT. The diagnosis results must be confirmed by two experienced clinicians. All patient data have been anonymized and comply with medical ethics requirements.

[0044] (2) Extracting feature importance from pre-trained deep learning models

[0045] Pre-trained model selection: The CliTsRNN model, which is based on recurrent neural networks (RNN) combined with autoencoders, was used. This model was pre-trained on the VTE time series dataset containing cervical cancer patients and showed stable performance (AUC = 0.941, accuracy = 87%).

[0046] Feature importance assessment: The weights of the input layer of the CliTsRNN model were analyzed, and all indicators were ranked according to the absolute value of the coefficients corresponding to the features. The larger the absolute value of the coefficient, the higher the contribution of the feature to VTE detection. Finally, the 8 key features with the highest contribution were selected: D-dimer, FDP, WBC, HB, PLT, CD4 cell count, CD8 cell count, and CD4 / CD8 ratio. Figure 1 a, Figure 2 ab).

[0047] Because some patients had too much missing data for these indicators, making it impossible to calculate their dynamic changes, they were removed from the dataset, resulting in a reduction in the sample size from 4307 to 2971 (Table 1). Among them, there were 2791 cases in the non-venous thromboembolism (non-VTE) group and 180 cases in the venous thromboembolism (VTE) group. Compared with the non-VTE group, the VTE group had significantly higher levels of D-dimer, FDP, WBC, PLT, and CD4 / CD8 ratio, while the hemoglobin (HB) level was lower; the differences in CD4 and CD8 cell counts between the two groups were relatively small.

[0048] Table 1 Valid Sample Data

[0049]

[0050] (3) Construction of time feature vectors

[0051] The time series data of eight key features are transformed into time feature vectors containing both dynamic and static information. Figure 1 bc):

[0052] Preserve the original values ​​of each feature (such as initial and final values);

[0053] By incorporating dynamic trends over time, including the growth rate of features (such as the increase in D-dimer) and the decrease rate (such as the decrease in HB), dual information encoding of "static values ​​+ dynamic trends" is achieved, capturing the progression pattern of VTE.

[0054] (4) TempVTE model training and validation

[0055] Data partitioning: The 2971 valid samples were randomly divided into a training set and an internal validation set at a ratio of 9:1.

[0056] Model infrastructure: The logistic regression model is used as the infrastructure for model training. The input is the time feature vector constructed in step (3), and the output is the VTE risk prediction result.

[0057] Performance evaluation: Model performance is assessed using metrics such as Area Under the Curve (AUC), accuracy, F1 score, recall, precision, inference speed, and model size. Simultaneously, changes in the loss function during training are monitored to verify model convergence. Figure 3 The results showed that the TempVTE model had an AUC of 0.918, an accuracy of 0.866, an F1 score of 0.864, a recall of 0.854, a precision of 0.875, an inference speed of <0.0001s / example, and a model size of <50MB.

[0058] Interpretability analysis: Analysis of the coefficients of the TempVTE model shows that the experimental results are consistent with prior knowledge and are interpretable. Figure 3 As shown in Figure d, the model tends to assign higher VTE prediction scores to patients with the following characteristics: high initial D-dimer levels, high final D-dimer levels with an upward trend. Similarly, slightly lower initial hemoglobin (HB) levels, decreasing final HB levels, and a decreasing trend in HB are also important features for obtaining high VTE prediction scores. Furthermore, the change in the CD4 / CD8 ratio—decreasing but with a higher final ratio—is also a key factor in the model's decision-making process. These results not only demonstrate robust performance but also reveal the complex relationship between VTE and key features.

[0059] Example 2 Comparative Example

[0060] Multiple machine learning algorithms were used as baseline models. These baseline models were trained using data from the patient's last observation time point. Model performance was evaluated using metrics such as area under the curve (AUC), precision, F1 score, recall, and accuracy. Results are shown in Table 2 and... Figure 3 As shown in Figure a, decision trees and logistic regression performed relatively poorly. The decision tree's AUC was 0.667, with relatively similar and moderate values ​​across all metrics, indicating a balanced overall performance but lacking a standout feature. Logistic regression achieved an AUC of 0.811, an improvement over decision trees, demonstrating enhanced discrimination ability. The consistent metric values ​​indicate stable and balanced model performance. Support Vector Machines (SVM) achieved an AUC of 0.831, further improving discrimination. All metrics performed well, indicating that the model outperformed decision trees and logistic regression overall. Random Forest achieved an AUC of 0.840, but its accuracy was relatively low, suggesting a weakness in identifying false positives. The convergence of the model was verified by varying the loss function. As shown in Table 2, the baseline model without time information performed only moderately in predicting venous thromboembolism (VTE). While the model incorporating time information but not immune features showed improvement (AUC of 0.878, accuracy of 83.6%), it still did not achieve optimal results. The TempVTE model, which integrated both immune features and time information, performed best, with an AUC as high as 0.918 and other indicators also at high levels. This indicates that the model incorporating both immune features and time information has a significant performance advantage, further highlighting the importance of immune features and time information in VTE prediction.

[0061] Table 2 Comparison of Model Performance Indicators

[0062]

Claims

1. A method for constructing a lightweight model for VTE monitoring in cervical cancer patients, characterized in that, The method extracts the eight key features with the highest contribution through a pre-trained CliTsRNN model: D-dimer, FDP, WBC, HB, PLT, CD4, CD8, and CD4 / CD8 ratio. The time series of the key features are encoded into a time feature vector of "static value + dynamic rate of change". Logistic regression is used to train a lightweight model for monitoring VTE in cervical cancer patients based on the fusion of time features and immune indicators.

2. The model construction method according to claim 1, characterized in that, The model construction method includes the following steps: (1) Screening of the study population and data collection: Patients diagnosed with cervical cancer by pathological histology, who received treatment in medical institutions and had complete clinical information were selected, excluding patients with a history of VTE before treatment, secondary cervical cancer, or other primary malignant tumors; clinical information and laboratory indicators of patients were collected; VTE diagnosis results were jointly confirmed by two physicians; the laboratory indicators included coagulation indicators: D-dimer, fibrin degradation products FDP, prothrombin time PT, activated partial thromboplastin time APTT, fibrinogen FIB; routine blood indicators: white blood cell count WBC, hemoglobin HB, platelet count PLT; immune indicators: CD4 cell count, CD8 cell count, CD4 / CD8 ratio; liver and kidney function indicators: blood urea nitrogen, creatinine, bicarbonate; (2) Key feature extraction: The pre-trained CliTsRNN deep learning model was used to analyze the weights of the input layer of the model. The features were sorted according to the absolute value of the corresponding coefficients and the eight key features with the highest contribution were selected, namely D-dimer, fibrin degradation product FDP, white blood cell count (WBC), hemoglobin (HB), platelet count (PLT), CD4 cell count, CD8 cell count and CD4 / CD8 ratio. (3) Construction of time feature vector: The time series data of the 8 key features are converted into time feature vectors containing static and dynamic information; the static information is the initial and final values ​​of the features, and the dynamic information is the total rate of change and the stage rate of change of the features; (4) Model training and validation: The time feature vectors are divided into training and validation sets in a 9:1 ratio. The logistic regression model is used to train the model, and the model performance is evaluated by AUC, accuracy, F1 score, recall and precision.

3. The model construction method according to claim 2, characterized in that, The clinical information mentioned in step (1) includes age, body mass index (BMI), comorbidities and treatment history; the comorbidities include hypertension, diabetes, and hypercholesterolemia; the treatment history includes radiotherapy, chemotherapy, targeted therapy, and immunotherapy.

4. The model construction method according to claim 2, characterized in that, In step (1), the diagnostic criteria for VTE are: deep vein thrombosis (DVT) is confirmed by vascular ultrasound or computed tomography (CTA), and pulmonary embolism (PTE) is confirmed by chest CT pulmonary angiography or enhanced CT scan.

5. The model construction method according to claim 2, characterized in that, The pre-training dataset of the CliTsRNN model mentioned in step (2) contains VTE time series data of cervical cancer, and the pre-training performance is AUC=0.941 and accuracy=87%.

6. The application of the lightweight model construction method for VTE monitoring in cervical cancer patients as described in any one of claims 1 to 5 in VTE risk assessment products for cervical cancer patients or in the study of the pathogenesis of thrombosis.

7. A computer device, comprising: The memory, the processor, and the computer program stored in the memory and executable on the processor are characterized in that the processor executes the computer program to implement the model constructed by the lightweight model construction method for VTE monitoring in cervical cancer patients according to any one of claims 1 to 5 to predict and assess the VTE risk in cervical cancer patients.

8. A non-transitory computer-readable storage medium storing instructions, characterized in that, When the instruction is executed by the processor, the model constructed by the lightweight model construction method for VTE monitoring in cervical cancer patients according to any one of claims 1 to 5 is used to predict and assess the VTE risk in cervical cancer patients.

9. A system software for predicting and assessing the risk of VTE in cervical cancer patients, characterized in that, The system software is deployed on a computer-readable storage medium and can perform the following functions when running: (1) Data input and preprocessing: Receive clinical information and laboratory test data of cervical cancer patients; the laboratory test data includes at least D-dimer, fibrin degradation product FDP, white blood cell count (WBC), hemoglobin (HB), platelet count (PLT), CD4 cell count, CD8 cell count and CD4 / CD8 ratio. (2) Key feature extraction: The pre-trained CliTsRNN deep learning model was called to extract the 8 key features that contribute the most to VTE detection: D-dimer, fibrin degradation product FDP, white blood cell count (WBC), hemoglobin (HB), platelet count (PLT), CD4 cell count, CD8 cell count, and CD4 / CD8 ratio. (3) Construction of time feature vector: The time series data of the eight key features are converted into time feature vectors containing static and dynamic information; the static information is the initial and final detection values ​​of each feature, and the dynamic information is the total rate of change of each feature and the stage rate of change of different detection stages; (4) VTE risk inference: Load the model constructed by the lightweight model construction method for VTE monitoring of cervical cancer patients according to any one of claims 1 to 5, input the time feature vector into the model, calculate and output the VTE risk prediction result of cervical cancer patients; (5) Results output: The VTE risk prediction results and key influencing factors of the model decision are presented in a visual form.

Images