Prediction method for leakage risk of contrast agent and delayed absorption risk after leakage

Abstract

The invention discloses a method for predicting a leakage risk of a contrast agent and a delayed absorption risk after leakage, and the method comprises the steps: obtaining clinical feature data of a patient, including patient level features, contrast agent application features, nursing operation features and features after leakage; preprocessing the clinical feature data, including performing one-hot coding on multi-classification variables, and standardizing all the feature data; inputting the preprocessed feature data into a trained machine learning model, and outputting a contrast agent leakage risk probability and / or a delayed absorption risk probability after leakage; generating an interpretable graph based on the risk probability output by the model; and displaying the interpretable graph and the risk probability in real time. According to the technical scheme, the two key clinical problems of'beforehand 'risk early warning and'afterward' absorption condition judgment of contrast agent leakage are solved at the same time, and more comprehensive decision support is provided.

Description

Technical Field

[0001] This invention relates to the field of medical information technology and machine learning applications, and to a method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage. Background Technology

[0002] Contrast media extravasation (CMEX) is the leakage of contrast agents into surrounding soft tissues during intravenous injection. It is a relatively common adverse event in imaging examinations and nursing care, with the vast majority of cases involving iodine-based contrast agents, reported in the literature as approximately 0.23%-0.94%. Because most CMEX events are administered via high-pressure injection and the contrast agent concentration is relatively high, mild cases present with swelling at the leakage site, which may be accompanied by induration or significant pain. More severe cases may involve skin and subcutaneous ulcers, and in serious cases, vascular and nerve damage may occur, even leading to compartment syndrome and extensive tissue necrosis. Regarding the absorption time of the extravasated contrast agent, it is generally believed that it is largely absorbed within 3-4 days. If absorption exceeds 4 days or even a week, the possibility of complications such as phlebitis, thrombosis, or even more serious complications should be considered. In 2002, the European Society of Urogenital Radiology summarized high-risk factors for contrast agent extravasation (CMEX) from three aspects: patient-related factors, contrast agent type and selection, and nursing skills. This provided a reference standard for the prevention and management of CMEX. While clinical practice has evolved, the guidelines for the management and prevention of contrast agent extravasation have not been significantly updated. In 2022, Roditi et al. published an article in the European Journal of Radiology entitled "Systematic Review of Intravenous Contrast Extravasation and Updates to the ESUR Contrast Safety Committee Guidelines," providing new supplements and references for the prevention and management of CMEX. In clinical practice, despite strict adherence to relevant CMEX prevention and management guidelines, the incidence of CMEX has decreased compared to previous studies, but it cannot be completely avoided. The focus of clinical attention has shifted from simply predicting leakage risk to reducing the incidence of CMEX and minimizing the occurrence of serious complications such as compartment syndrome.

[0003] Currently, in clinical practice, the risk assessment and determination of post-leakage absorption time for CMEX mainly rely on the personal experience and subjective judgment of nursing staff, lacking objective and quantitative individualized predictive tools. Although organizations such as the European Society of Urogenital Radiology have published relevant prevention and management guidelines that summarize high-risk factors, these guidelines cannot achieve precise risk quantification for specific patients.

[0004] Patent searches revealed that existing related technologies are mostly focused on hardware improvements, such as injection devices to prevent leakage, limb fixation devices, or drug applications to promote absorption, such as patent CN217365894U, which describes a detection alarm device for extravasation of contrast agent during CT enhancement injection. Searches on CNKI, PubMed, and other databases revealed that current research reports on contrast agent extravasation are mainly clinical experience reports, reviews, and meta-analyses. The latest clinical research article is "Zhang L, Yan HF. Application value of high-pressure-resistant peripherally inserted centralcatheters in enhanced computer tomography of diabetic patients with malignant tumors. World J Diabetes. 2024 Dec 15;15(12):2293-2301. doi: 10.4239 / wjd.v15.i12.2293. PMID: 39676811; PMCID: PMC11580592". This study primarily explored the differences in contrast agent extravasation and image quality between high-pressure resistant central venous catheterization and conventional central venous catheterization in patients with diabetes and lung cancer. The study suggests that high-pressure resistant central venous catheterization helps reduce the risk of CMEX in critically ill patients; however, it is specifically targeted at patients with diabetes and lung cancer, and high-pressure resistant central venous catheterization is expensive, hindering its widespread application. Currently, no solution has been found that employs machine learning algorithms to construct a dual-task model simultaneously predicting leakage risk and delayed absorption risk, and integrates interpretable analysis with real-time mobile output.

[0005] Therefore, there is an urgent need in the field for a personalized prediction scheme that can provide high accuracy, interpretability and ease of clinical use based on readily available clinical characteristics. Summary of the Invention

[0006] The purpose of this invention is to provide a method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage. Its core lies in constructing an end-to-end prediction and decision support process based on machine learning. Furthermore, it simultaneously addresses two key clinical issues in a single technical solution: "pre-leakage" risk warning and "post-leakage" absorption assessment, providing more comprehensive decision support. Through rigorous feature engineering, a feature set with strong predictive power is selected, and a Logistic Regression model suitable for small sample sizes and low-dimensional data is chosen, resulting in excellent predictive performance and stability.

[0007] This invention is achieved through the following technical solution:

[0008] A method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage includes the following steps:

[0009] Acquire the patient's clinical characteristic data, which includes patient-level characteristics, contrast agent application characteristics, nursing operation characteristics, and post-leakage characteristics;

[0010] The clinical feature data is preprocessed, including one-hot encoding of multi-category variables and standardization of all feature data;

[0011] The preprocessed feature data is input into a trained machine learning model, which outputs the probability of contrast agent leakage risk and / or the probability of delayed absorption risk after leakage.

[0012] Generate interpretable graphs based on the risk probabilities output by the model;

[0013] The interpretable graphs and risk probabilities are displayed in real time.

[0014] To better implement this solution, the sources of the clinical characteristic data further include:

[0015] A CMEX cohort consisting of patients who had experienced contrast agent extravasation events;

[0016] A non-CMEX cohort consisting of a random sample of patients who have never experienced contrast agent extravasation;

[0017] The data in the CMEX queue is also labeled for predicting delayed absorption risk based on whether the absorption time after leakage is greater than 4 days.

[0018] To better implement this plan, the patient-level characteristics further include 10 categories: age, gender, height, weight, chronic underlying diseases such as diabetes, other comorbidities, history of long-term intravenous chemotherapy, PS score of physical condition, patient cooperation level, and nurse's assessment and grading of vascular conditions at the puncture site.

[0019] The contrast agent application characteristics include four categories: the type of CT enhanced scan examination method, the contrast agent concentration, the contrast agent dosage, and the contrast agent injection rate.

[0020] The nursing operation characteristics include three categories: indwelling needle puncture site, number of punctures, and whether a pre-injection test is positive.

[0021] The post-leakage characteristics include seven categories: leakage type, leakage volume, long axis of limb swelling, baseline assessment after leakage, post-leakage treatment, assessment status at 24 and 48 hours after leakage, and duration of subcutaneous leakage.

[0022] To better implement this solution, a feature selection process is further performed before inputting the feature data into the machine learning model. This process includes:

[0023] Initial screening is performed by calculating the correlation coefficients between each feature and the task label. Specifically, the Pearson / Spearman correlation coefficient ρ between each feature and the task label needs to be calculated, and variables with |ρ|>0.9 are retained. For continuous variables and binary variables, the Pearson correlation coefficient is used; for multi-category (including ordinal) variables, the Spearman correlation coefficient is used. That is, all pairs containing "ordinal / ranked" variables are matched using the Spearman correlation coefficient, and the remaining variables are matched using the Pearson correlation coefficient.

[0024] The formula for calculating the Pearson correlation coefficient is:

[0025] The calculation process of Spearman correlation coefficient: First, take the rank of each feature (such as X, Y) to obtain R. X R Y Then, the inter-rank correlation coefficient is calculated according to the Pearson formula, i.e.

[0026]

[0027] For example, Pearson correlation coefficient is used for age (continuous-binary); Spearman correlation coefficient is used for physical condition PS score (ordination-multiclass).

[0028] To eliminate collinearity, the variance inflation factor (VIF) of the surviving features is calculated. Specifically, the variance inflation factor (VIF) is calculated for the surviving variables, and variables with VIF > 10 are successively eliminated until all variables have VIF < 10. Then, all remaining variables are considered to have no collinearity.

[0029] The Least Absolute Shrinkage and Selection Operator (LASSO) feature compression method is used to preserve features with non-zero regression coefficients. Specifically, all variables remaining after eliminating collinearity are incorporated into the L2-Logistic loss function with an L1 penalty. λ is selected using 10-fold cross-validation based on the "1-standard error" criterion, retaining features with non-zero coefficients. LASSO feature compression is achieved through the formula... Solving for the regression coefficients of the feature variables yields estimates. Here, argmin represents "the parameter vector that minimizes the objective function"; The coefficients to be estimated (including the intercept β0); For log-likelihood; Let β be the probability of the i-th event; λ≥0 is the L1 penalty strength selected according to the "accuracy-error" criterion through 10-fold cross-validation; finally, only β is retained. j Characteristics of ≠ 0.

[0030] To better implement this solution, the feature screening process ultimately selects the following features from the contrast agent leakage risk prediction model: chronic underlying diseases, other comorbidities of the patient, vascular condition assessment, puncture site, number of punctures, and iodine contrast agent concentration.

[0031] The features selected for the delayed absorption risk prediction model include: the volume of contrast agent leaked, the long axis of limb swelling, the type of enhanced examination, the number of indwelling needle punctures, the treatment method taken after leakage, and the assessment status 24 hours after leakage.

[0032] To better implement this solution, the machine learning model is a Logistic regression model, which is trained by maximum likelihood estimation and the model parameters are determined by cross-validation.

[0033] To better implement this solution, the interpretable graph further includes a nodal plot, a SHAP summary plot, or a SHAP waterfall plot;

[0034] The nomogram maps the regression coefficients of the model to a visualized score scale for individual risk visualization.

[0035] The SHAP summary plot shows the average contribution of each feature to the model output; the SHAP waterfall plot provides an individualized interpretation of the prediction results for a single patient.

[0036] To better implement this solution, the real-time display is further achieved through a mobile terminal mini-program. The mini-program front-end calls a local inference script to complete real-time calculations, and the data does not leave the user terminal.

[0037] To better implement this plan, it further includes closed-loop intervention steps: based on the leakage risk level predicted by the model, it suggests proactive nursing measures such as reducing the contrast agent injection flow rate, reducing the contrast agent dosage, selecting a lower concentration of contrast agent, or strengthening rounds during the examination; or based on the delayed absorption risk level predicted by the model, it suggests adjusting the frequency of nursing follow-up after leakage.

[0038] An electronic device includes a processor and a memory, the memory storing a computer program, and the processor executing the computer program to implement the method described in any of the preceding claims.

[0039] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage described in this plan mainly includes the following steps:

[0040] 1. Data Preparation and Preprocessing: Clinical data from CMEX patients and non-CMEX patients were collected to form the basis of modeling. Data cleaning and preprocessing were performed on the collected multidimensional clinical features (including patient factors, contrast agent factors, procedural factors, and post-leakage factors). Preprocessing included one-hot encoding of multi-category variables (such as other comorbidities and PS scores) and standardization of all numerical features to eliminate the influence of dimensions.

[0041] 2. Feature Engineering and Model Building: A rigorous feature selection process was employed to extract the most critical feature subset from a large pool of initially selected features. This process combined correlation coefficient analysis, variance inflation factor collinearity testing, and LASSO regression feature compression. Ultimately, six core modeling features were identified for each of the two tasks: leakage risk prediction and delayed absorption risk prediction. Based on these features, the Logistic Regression algorithm was chosen to construct the prediction model. This model was selected because it is stable on low-dimensional data after high-dimensional feature selection, is less prone to overfitting, and its output has inherent probabilistic interpretability and ease of visualization.

[0042] 3. Model Output and Interpretability: The risk probabilities output by the model are presented to the user through various interpretable graphs. This includes:

[0043] 3.1. Nonograph: The coefficients of the Logistic regression model are linearly mapped to scores, allowing medical staff to intuitively estimate individual risk by simply summing the scores.

[0044] 3.2. SHAP Analysis: Provides a global SHAP summary plot to show the importance of features, as well as a SHAP waterfall plot for a single sample, clearly revealing how each feature affects the final prediction result for the patient, thus achieving "white-box" modeling.

[0045] 4. Real-time Application and Closed-Loop Intervention: The trained model is embedded into a WeChat mini-program, enabling bedside, zero-training, real-time prediction. After inputting key features, healthcare professionals can instantly obtain risk prediction results and visual explanations. The system also provides specific intervention recommendations based on the predicted risk level (e.g., low, medium, high risk). For example, for high-risk patients, it suggests reducing the injection flow rate and increasing rounds; for high-risk patients with delayed absorption, it suggests increasing follow-up frequency, thus forming a closed-loop nursing care system from prediction to intervention.

[0046] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0047] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage, as described in this invention, solves two key clinical problems simultaneously in one technical solution: "pre-existing" risk warning and "post-existing" absorption assessment of contrast agent leakage, providing more comprehensive decision support.

[0048] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage described in this invention selects a feature set with strong predictive power through rigorous feature engineering, and selects a Logistic regression model suitable for small sample and low-dimensional data, so that the model exhibits excellent predictive performance and stability.

[0049] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage, as described in this invention, innovatively combines the traditional clinical statistical tool of nomograms with the cutting-edge SHAP interpretation method. This makes the machine learning model no longer a black box, greatly enhancing medical staff's trust and understanding of the model's prediction results, and is conducive to clinical promotion.

[0050] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage described in this invention can be implemented through a WeChat mini program. It does not require the installation of additional software, does not rely on deep integration with the hospital information system, and achieves real-time bedside calculation with "data not leaving the terminal". This greatly improves the ease of use and popularity of the technology, and is flexible in deployment and highly practical.

[0051] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage described in this invention directly links the prediction results with specific clinical interventions, promoting the transformation of nursing work from passive treatment to proactive and predictive intervention, which helps to reduce the incidence of CMEX, alleviate serious complications, and improve patient safety. Attached Figure Description

[0052] Figure 1 This is a flowchart illustrating the overall model development, design, and application process provided in this embodiment of the invention.

[0053] Figure 2 This is a schematic diagram of the feature weights of the CMEX risk prediction model;

[0054] Figure 3 This is the ROC curve of the CMEX risk prediction model;

[0055] Figure 4 This is a nodal chart of the CMEX risk prediction model;

[0056] Figure 5 This is a summary diagram of the SHAP features of the CMEX risk prediction model;

[0057] Figure 6 This is an individualized example of the CMEX risk prediction model SHAP interpreting the waterfall chart;

[0058] Figure 7 This is a schematic diagram of the feature weights of the CMEX delayed absorption risk model;

[0059] Figure 8 This is the ROC curve of the CMEX delayed absorption risk model;

[0060] Figure 9 This is a nodal graph of the CMEX delayed absorption risk model;

[0061] Figure 10 This is a summary diagram of the SHAP features of the CMEX delayed absorption risk model;

[0062] Figure 11 This is an individualized example of the waterfall plot interpretation of the CMEX delayed absorption risk model SHAP. Detailed Implementation

[0063] The following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0064] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, unless otherwise expressly indicated by the invention, the singular form is intended to include the plural form as well. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0065] For ease of description, the words "up," "down," "left," and "right" appearing in this invention only indicate that they are consistent with the up, down, left, and right directions of the accompanying drawings themselves, and do not limit the structure. They are merely for the purpose of facilitating the description of this invention and simplifying the description, and do not indicate or imply that the device or component referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0066] Terminology Explanation: The terms "installation," "connection," "linking," and "fixing" in this invention should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal connection of two components or the interaction relationship between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0067] Example 1:

[0068] Methods for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage, such as... Figure 1 As shown, it includes the following steps:

[0069] Acquire the patient's clinical characteristic data, which includes patient-level characteristics, contrast agent application characteristics, nursing operation characteristics, and post-leakage characteristics;

[0070] The clinical feature data is preprocessed, including one-hot encoding of multi-category variables and standardization of all feature data;

[0071] The preprocessed feature data is input into a trained machine learning model, which outputs the probability of contrast agent leakage risk and / or the probability of delayed absorption risk after leakage.

[0072] Generate interpretable graphs based on the risk probabilities output by the model;

[0073] The interpretable graphs and risk probabilities are displayed in real time.

[0074] This solution addresses two key clinical issues—pre-contrast risk warning and post-contrast absorption assessment—within a single technical approach, providing more comprehensive decision support.

[0075] Example 2:

[0076] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage in this embodiment is an improvement on that in Embodiment 1. The sources of the clinical characteristic data include:

[0077] A CMEX cohort consisting of patients who had experienced contrast agent extravasation events;

[0078] A non-CMEX cohort consisting of a random sample of patients who have never experienced contrast agent extravasation;

[0079] The data in the CMEX queue is also labeled for predicting delayed absorption risk based on whether the absorption time after leakage is greater than 4 days.

[0080] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0081] Example 3:

[0082] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage in this embodiment is an improvement on the method in embodiment 1. The patient-level characteristics include 10 categories: age, gender, height, weight, underlying disease (diabetes), other comorbidities, history of long-term intravenous chemotherapy, PS score of physical condition, patient cooperation level, and nurse's assessment and grading of vascular conditions at the puncture site.

[0083] The contrast agent application characteristics include four categories: the type of CT enhanced scan examination method, the contrast agent concentration, the contrast agent dosage, and the contrast agent injection rate.

[0084] The nursing operation characteristics include three categories: indwelling needle puncture site, number of punctures, and whether a pre-injection test is positive.

[0085] The post-leakage characteristics include seven categories: leakage type, leakage volume, limb swelling extent (long axis), baseline assessment after leakage, post-leakage treatment, assessment status at 24 and 48 hours after leakage, and duration of subcutaneous leakage.

[0086] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0087] Example 4:

[0088] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage in this embodiment is an improvement on Embodiment 1. Before inputting the feature data into the model, a feature screening process is performed, which includes:

[0089] Calculate the correlation coefficients between each feature and the task label for initial screening; calculate the Pearson / Spearman correlation coefficient ρ between each feature and the task label, and retain variables with |ρ|>0.9. For continuous variables and binary variables, use the Pearson correlation coefficient; for multi-category (including ordinal) variables, use the Spearman correlation coefficient. That is, all pairs containing "ordinal / ranked" variables use the Spearman correlation coefficient, and the remaining variables use the Pearson correlation coefficient.

[0090] The formula for calculating the Pearson correlation coefficient is:

[0091] The calculation process of Spearman correlation coefficient: First, take the rank of each feature (such as X, Y) to obtain R. X R Y Then, the inter-rank correlation coefficient is calculated according to the Pearson formula, i.e.

[0092]

[0093] For example, Pearson correlation coefficient is used for age (continuous-binary); Spearman correlation coefficient is used for physical condition PS score (ordination-multiclass).

[0094] Calculate the variance inflation factor of the surviving features to eliminate collinearity; more specifically, calculate the variance inflation factor (VIF) for the surviving variables and successively eliminate variables with VIF>10 until all variables with VIF<10 are retained.

[0095] The Least Absolute Shrinkage and Selection Operator (LASSO) feature compression method is used to preserve features with non-zero regression coefficients. Specifically, all variables remaining after eliminating collinear features are incorporated into the L2-Logistic loss function with an L1 penalty. λ is selected using 10-fold cross-validation based on the "1-standard error" criterion, retaining features with non-zero coefficients. LASSO feature compression is achieved through the formula... Solving for the regression coefficients of the feature variables yields estimates. Here, argmin represents "the parameter vector that minimizes the objective function"; The coefficients to be estimated (including the intercept β0); For log-likelihood; Let β be the probability of the i-th event; λ≥0 is the L1 penalty strength selected according to the "accuracy" rule after 10-fold cross-validation; finally, only β is retained. j Characteristics of ≠ 0.

[0096] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0097] Example 5:

[0098] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage in this embodiment is an improvement on the method in embodiment 4. The feature screening process ultimately selects the following features from the contrast agent leakage risk prediction model: chronic underlying diseases, other comorbidities of the patient, vascular condition assessment, puncture site, number of punctures, and iodine contrast agent concentration.

[0099] The features selected for the delayed absorption risk prediction model include: the volume of contrast agent leaked, the long axis of limb swelling, the type of enhanced examination, the number of indwelling needle punctures, the treatment method taken after leakage, and the assessment status 24 hours after leakage.

[0100] The other parts of this embodiment are the same as those in embodiment 4, so they will not be described again.

[0101] Example 6:

[0102] The method for predicting the contrast agent leakage risk and the delayed absorption risk after leakage in this embodiment is an improvement on that in Embodiment 5. The machine learning model is a Logistic Regression model, which is trained by maximum likelihood estimation and the model parameters are determined by cross-validation. The features of the two sets of models mentioned in Embodiment 5 are used to establish a risk probability model using Logistic Regression (L2 regularization). Regarding the selection of the model algorithm, considering the characteristics of the research task and the data type, we chose the LR model as the implementation model. This is because the features used to establish the contrast agent leakage risk prediction model and the delayed absorption risk prediction model are both 6-item, belonging to a low-dimensional data scenario. In this case, using tree models (RF, LightGBM, GBDT) and deep networks (MLP) is prone to overfitting, reducing the model's generalization ability. In the process of model optimization, in addition to LR, we also used SVM, MLP, Random Forest, Light GBM, Gradient Boosting, and other models. LR, SVM, and MLP outperformed other models (delong test, P < 0.05). Among these three models, the LR model had the smallest differences in performance indicators such as AUC among the subgroups, and was therefore considered the optimal model. From the perspective of matching technical solutions, Logistic Regression can provide explicit coefficients β, which can be directly mapped to clinical scales (risk probabilities) to meet the needs of subsequent model visualization and individualized interpretation; other models have no analytical coefficients or coefficients that are nonlinear and cannot be used to directly generate nomograms.

[0103] The interpretable graphs include nomograms, SHAP summary graphs, or SHAP waterfall graphs;

[0104] The nomogram maps the regression coefficients of the model to a visual score scale for individual risk visualization. Specifically, it maps the Logistic regression coefficients corresponding to each modeling feature to a 0–100 score scale in a linear proportion, retrieves Points for each feature value from the graph, and accumulates them to Total Points. The corresponding probability P is then read directly from the Total→Risk axis, enabling the visualization and interpretation of individual risk.

[0105] The SHAP summary plot shows the average contribution of each feature to the model output; the SHAP waterfall plot provides an individualized interpretation of the prediction results for a single patient. For any sample i, according to the Shapley formula φ... ij The marginal contribution of the j-th feature to the model output of sample i is calculated, with the sign indicating whether the feature increases or decreases the risk probability; based on the average |φ of the features of each variable. ij| Generate horizontal bar charts; for a single case, use a waterfall chart to stack the bars one by one to show the extent to which the relevant features increase or decrease the predicted probability, thus achieving personalized local interpretation.

[0106] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0107] Example 7:

[0108] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage in this embodiment is an improvement on Embodiment 1. The real-time display is achieved through a mobile terminal applet. The applet's front end calls a local JavaScript inference script to complete millisecond-level real-time calculations, ensuring that the data does not leave the terminal, thus facilitating bedside communication and nursing decisions for both doctors and patients. In addition, the model data is stored locally with anonymized information, and sensitive information such as the patient's name, gender, and age is not collected.

[0109] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0110] Example 8:

[0111] The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage in this embodiment is an improvement on the method in Embodiment 1, and also includes a closed-loop intervention step: based on the leakage risk level predicted by the model, it suggests reducing the contrast agent injection flow rate, reducing the contrast agent dosage, using a lower concentration of contrast agent, or strengthening the rounds during the examination; or based on the delayed absorption risk level predicted by the model, it suggests adjusting the frequency of nursing follow-up after leakage.

[0112] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0113] Example 9:

[0114] An electronic device includes a processor and a memory. The memory stores a computer program, and the processor executes the computer program to implement the method described in any one of Embodiments 1-8. When the processor executes the program, it can perform all functions from data preprocessing, model calling, result calculation to graphics generation and interface display. The device can be a server, workstation, or a mobile terminal such as a smartphone or tablet computer with the WeChat mini-program installed.

[0115] Example 10:

[0116] This embodiment combines Figures 2-6 This is used to illustrate how to predict the risk of contrast agent leakage in clinical practice.

[0117] Data Acquisition: Data was collected from the CMEX queue (n=210) and non-CMEX queues (n=316). Data without leakage was labeled "0" for n=316; data with leakage was labeled "1" for n=210. The training, test, and validation sets were divided in a ratio of 0.6:0.25:0.15 (316, 132, and 78 examples respectively). The data source for the CMEX queue was:

[0118] From January 2023 to October 2024, 210 patients who experienced contrast agent extravasation (CMEX) adverse events constituted the CMEX cohort data. All patients were followed up until the outcome was clear, namely whether the leakage was absorbed early (≤4 days), the final absorption time, and whether there were adverse events such as phlebitis, skin ulceration, vascular nerve injury, or compartment syndrome.

[0119] The non-CMEX cohort data consists of 31,612 patients who did not experience CMEX between January 2023 and October 2024. 1% of these patients were randomly selected to form the non-CMEX cohort data.

[0120] Feature input: Medical staff can input or select key features determined by feature screening through the mini-program interface: chronic underlying diseases, other comorbidities of the patient, vascular condition assessment, puncture site, number of punctures, and iodine contrast agent concentration.

[0121] Real-time prediction and interpretation: The mini-program backend calls a localized Logistic Regression model for real-time calculation, instantly outputting the leakage risk probability. Specifically, we choose Logistic Regression (L2 regularization, or LR for short) to build the risk probability model.

[0122] Meanwhile, the interface can display, for example Figure 4 The nomogram shown is for medical staff to manually check and calculate scores, or to display information such as... Figure 6 The SHAP waterfall chart shown visually illustrates how much each feature (such as "Vascular condition assessment = poor") increases the final risk probability.

[0123] Closed-loop intervention: If the system predicts that the patient is at high risk (e.g., probability > 70%), the interface will prominently display the message "High risk, recommendation: reduce the injection flow rate to X ml / s and closely monitor the patient throughout the process".

[0124] After testing, using AUC, accuracy, precision, and F1 score as the main performance indicators, the CMEX risk prediction model achieved AUCs of 0.946, 0.927, and 0.935 on the training, test, and validation sets, respectively; accuracy of 0.873, 0.871, and 0.863; precision of 0.832, 0.789, and 0.727; and F1 scores of 0.856, 0.841, and 0.814. All performance indicators met the modeling requirements.

[0125] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0126] Example 11:

[0127] This embodiment combines Figure 1 and Figures 7-11 This is used to illustrate how to predict the absorption of contrast agent after a contrast agent leakage event occurs.

[0128] Data Acquisition: After leakage occurred, post-leakage characteristics were collected from CMEX cohort patients (210 CMEX cohort patients as described in Example 10, i.e., n=210). Data meeting the early absorption criteria (≤4 days) after leakage were labeled as "0", n=146; data with an absorption time >4 days after leakage were labeled as "1", n=64; the training and test sets were divided in an 8:2 ratio, with 168 cases in the training set and 42 cases in the test set.

[0129] Feature input: Through the mini-program interface, input key features: volume of leaked contrast agent, long diameter of limb swelling, type of enhanced examination, number of indwelling needle punctures, treatment methods taken after leakage, and assessment status 24 hours after leakage.

[0130] Real-time prediction and interpretation: The model outputs the probability of risk of delayed absorption after calculation. Figure 9 nomogram and Figure 11 The SHAP waterfall diagram provides both global and individual interpretations.

[0131] Closed-loop intervention: If the predicted risk of delayed absorption is high, the system will suggest: "Increase daily follow-up to X times, focusing on changes in the extent of swelling and pain."

[0132] Experiments showed that the AUC of the CMEX delayed absorption prediction model in the training and test sets were 0.924 and 0.916, respectively; the accuracy was 0.845 and 0.860, respectively; the precision was 0.70 and 0.75, respectively; and the F1 score was 0.78 and 0.80, respectively. All performance indicators met the modeling requirements.

[0133] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0134] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage, characterized in that, Includes the following steps: Acquire the patient's clinical characteristic data, which includes patient-level characteristics, contrast agent application characteristics, nursing operation characteristics, and post-leakage characteristics; The clinical feature data is preprocessed, including one-hot encoding of multi-category variables and standardization of all feature data; The preprocessed feature data is input into a trained machine learning model, which outputs the probability of contrast agent leakage risk and / or the probability of delayed absorption risk after leakage. Generate interpretable graphs based on the risk probabilities output by the model; The interpretable graphs and risk probabilities are displayed in real time.

2. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 1, characterized in that, The sources of the clinical characteristic data include: A CMEX cohort consisting of patients who had experienced contrast agent extravasation events; A non-CMEX cohort consisting of a random sample of patients who have never experienced contrast agent extravasation; The data in the CMEX queue is also labeled for predicting delayed absorption risk based on whether the absorption time after leakage is greater than 4 days.

3. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 1, characterized in that, The patient-level characteristics include 10 categories: age, gender, height, weight, chronic underlying diseases, other comorbidities, history of long-term intravenous chemotherapy, PS score of physical condition, patient cooperation level, and nurse's assessment and grading of vascular conditions at the puncture site. The contrast agent application characteristics include four categories: the type of CT enhanced scan examination method, the contrast agent concentration, the contrast agent dosage, and the contrast agent injection rate. The nursing operation characteristics include three categories: indwelling needle puncture site, number of punctures, and whether a pre-injection test is positive. The post-leakage characteristics include seven categories: leakage type, leakage volume, limb swelling extent (long axis), baseline assessment after leakage, post-leakage treatment, assessment status at 24 and 48 hours after leakage, and duration of subcutaneous leakage.

4. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 1, characterized in that, Before inputting feature data into the machine learning model, a feature selection process is performed, which includes: Calculate the correlation coefficients between each feature and the task label for initial screening; Calculate the variance inflation factor of the surviving features to eliminate collinearity; We use minimum absolute shrinkage and selection operator features for compression, while retaining the feature that the regression coefficients are non-zero.

5. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 4, characterized in that, The feature screening process ultimately selects features from the contrast agent leakage risk prediction model, including: chronic underlying diseases, other comorbidities of the patient, vascular condition assessment, puncture site, number of punctures, and iodine contrast agent concentration. The features selected for the delayed absorption risk prediction model include: the volume of contrast agent leaked, the long axis of limb swelling, the type of enhanced examination, the number of indwelling needle punctures, the treatment method taken after leakage, and the assessment status 24 hours after leakage.

6. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 1, characterized in that, The machine learning model is a Logistic regression model, which is trained by maximum likelihood estimation and the model parameters are determined by cross-validation.

7. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 1, characterized in that, The interpretable graphs include nomograms, SHAP summary graphs, or SHAP waterfall graphs; The nomogram maps the regression coefficients of the model to a visualized score scale for individual risk visualization. The SHAP summary plot shows the average contribution of each feature to the model output; the SHAP waterfall plot provides an individualized interpretation of the prediction results for a single patient.

8. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 1, characterized in that, The real-time display is achieved through a mobile terminal mini-program. The mini-program front end calls a local inference script to complete real-time calculations, and the data does not leave the user terminal.

9. The method for predicting the risk of contrast agent leakage and the risk of delayed absorption after leakage according to claim 1, characterized in that, It also includes closed-loop intervention steps: based on the leakage risk level predicted by the model, it suggests reducing the contrast agent injection flow rate, reducing the contrast agent dosage, selecting a lower concentration of contrast agent, or strengthening rounds during the examination; or based on the delayed absorption risk level predicted by the model, it suggests adjusting the frequency of nursing follow-up after leakage.

10. An electronic device comprising a processor and a memory, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 9.

Images