A tumor patient PICC catheterization upper limb venous thrombosis risk prediction system

Abstract

This invention discloses a system for predicting the risk of upper limb venous thrombosis in cancer patients after PICC placement, relating to the field of medical risk prediction. The system includes: a data acquisition unit for acquiring the patient's anticoagulant usage status, triglyceride levels, nutritional risk screening score, and surgical history; a processing unit embedded with a risk stratification prediction model trained and validated using the CHAID decision tree algorithm, the model having a defined four-level hierarchical judgment structure; the processing unit is configured to sequentially call each judgment factor according to the hierarchical judgment structure, outputting the corresponding thrombosis risk level layer by layer; and an output unit for presenting the thrombosis risk level. This invention enables rapid stratified and graded identification of upper limb venous thrombosis risk in cancer patients after PICC placement, significantly improving the predictive ability of the model on new clinical samples.

Description

Technical Field

[0001] This invention relates to the field of medical risk prediction, and in particular to a system for predicting the risk of upper limb venous thrombosis after PICC placement in cancer patients. Background Technology

[0002] Peripherally inserted central catheters (PICCs) are crucial vascular access points for long-term intravenous chemotherapy, nutritional support, and anti-infection treatment in cancer patients. However, PICC-related venous thrombosis (PICC-CRT) is its most common serious complication, with an incidence rate as high as 78.1%, and most cases are asymptomatic. PICC-CRT can lead to catheter dysfunction, unplanned extubation, secondary infections, and in severe cases, pulmonary embolism, significantly increasing the risk of death in cancer patients. Therefore, early identification of high-risk groups for thrombosis and accurate prediction of thrombosis risk are of significant clinical importance in reducing the incidence of PICC-CRT. Currently, various PICC-CRT risk prediction models have been developed both domestically and internationally, mainly including nomogram models based on traditional logistic regression and prediction models based on machine learning algorithms such as XGBoost and support vector machines.

[0003] Existing PICC-CRT risk prediction models have the following shortcomings: First, the risk factors screened by traditional logistic regression and nomogram models are greatly affected by sample size, ethnic differences, and heterogeneity of study subjects, resulting in low predictive efficiency. The area under the ROC curve for most models is between 0.786 and 0.794, which is insufficient to meet the needs of clinical application. Second, existing models generally do not screen for specific clinical characteristics of cancer patients, lacking core risk factors with high identifiability in cancer patients, such as anticoagulant use status, nutritional status, and surgical history, leading to insufficient clinical relevance. Third, most models remain at the theoretical construction and validation stage, without forming a standardized, directly applicable stratified prediction process. Medical staff need to perform complex calculations involving multiple indicators, which is cumbersome and makes it difficult to achieve rapid screening of high-risk groups. Fourth, existing models rely excessively on high-cost laboratory test indicators such as D-dimer and fibrinogen, while paying insufficient attention to easily obtainable and low-cost indicators such as triglycerides and nutritional risk screening scores, increasing clinical testing costs and operational complexity. Furthermore, existing prediction methods mostly rely on manual assessment or single-indicator judgment, lacking a hierarchical prediction logic. This makes it impossible to accurately classify thrombosis risk levels based on individual patient characteristics, leading to over-intervention for low-risk patients and under-intervention for high-risk patients. Therefore, based on these challenges, this invention proposes a risk prediction system for upper limb venous thrombosis in cancer patients after PICC placement. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a risk prediction system for upper limb venous thrombosis in cancer patients after PICC placement. This system addresses the problems of low predictive efficacy, lack of specific risk factors for cancer patients, inability to achieve stratified prediction, and high testing costs in existing PICC-related venous thrombosis risk prediction technologies. The goal is to enable rapid and accurate identification of high-risk individuals for thrombosis and provide stratified guidance for clinical intervention.

[0005] To achieve the above objectives, this invention provides a risk prediction system for upper limb venous thrombosis after PICC placement in cancer patients. This invention employs the CHAID decision tree algorithm to construct a risk prediction model. First, based on clinical data of cancer patients with PICC placement, univariate analysis is used to screen candidate variables statistically associated with thrombosis occurrence. Then, by setting constraints such as the maximum tree growth depth, node splitting significance level, and minimum node sample size, the algorithm automatically selects the optimal splitting variable for each layer based on the chi-square test significance of the sample data, training to generate a decision tree model with a four-level hierarchical judgment structure. After training and validation, the judgment factors of this model, in hierarchical order, are anticoagulant use status, triglyceride level, nutritional risk screening score, and surgical history.

[0006] In a first aspect, the present invention provides a system for predicting the risk of upper limb venous thrombosis after PICC placement in cancer patients, comprising: The data acquisition unit is used to acquire patients' anticoagulant use status, triglyceride levels, nutritional risk screening scores, and surgical history information; The processing unit embeds a risk stratification prediction model trained and validated using the CHAID decision tree algorithm. This model is optimized for PICC line placement in cancer patients and has a defined four-level hierarchical judgment structure. The first-level judgment factor is anticoagulant use status, the second-level judgment factor is triglyceride level, the third-level judgment factor is nutritional risk screening score, and the fourth-level judgment factor is surgical history. The processing unit is configured to sequentially call each judgment factor according to the hierarchical judgment structure to perform discretization judgment and output the corresponding thrombosis risk level layer by layer. The risk stratification prediction model includes an anticoagulant protection dynamic attenuation module. This module calculates an anticoagulant protection attenuation coefficient based on the patient's prescribed anticoagulant dosing frequency, actual dosing frequency, and the time interval between the most recent dosing and PICC placement. Based on this coefficient, the anticoagulant usage status is mapped from a binary category to three categories: adequate protection, partial protection, and no protection. The anticoagulant protection attenuation coefficient is positively correlated with the proportion of actual dosing frequency and negatively correlated with the time interval. The mapped new category replaces the original anticoagulant usage status and serves as the first-layer judgment factor input into the decision tree model. The output unit is used to present the thrombosis risk level.

[0007] Furthermore, in the four-level hierarchical judgment structure, each judgment factor is pre-assigned a discrete category value: the anticoagulant usage status is divided into two categories: "used" and "not used"; the triglyceride level is divided into three categories: "above the normal range", "within the normal range" and "below the normal range"; the nutritional risk screening score is divided into two categories: "reached or exceeded the preset risk threshold" and "below the preset risk threshold"; and the surgical history is divided into two categories: "present" and "absent".

[0008] Furthermore, the processing unit determines the thrombosis risk level according to the following logic: When the anticoagulant is in use, a low risk level is output. When the anticoagulant is in the "not used" status and the triglyceride level is "above normal range", an extremely high risk level is output. When the anticoagulant usage status is "not used" and the triglyceride level is "below the normal range", if the nutritional risk screening score is "reached or exceeded the preset risk threshold", an extremely high risk level will be output; if the nutritional risk screening score is "below the preset risk threshold", a high risk level will be output. When the anticoagulant usage status is "not used", the triglyceride level is "within the normal range", and the nutritional risk screening score is "reached or exceeded the preset risk threshold", an extremely high risk level is output. When the anticoagulant usage status is "not used", the triglyceride level is "within the normal range", the nutritional risk screening score is "below the preset risk threshold", and the surgical history is "none", a high risk level is output; When the anticoagulant usage status is "not used", the triglyceride level is "within the normal range", the nutritional risk screening score is "below the preset risk threshold", and the surgical history is "yes", a medium risk level is output.

[0009] This logic does not require continuous numerical calculations; it can classify risk levels through only a finite number of discrete judgments, significantly reducing computational complexity and improving the interpretability of prediction results.

[0010] Furthermore, the node splitting criterion of the CHAID decision tree algorithm is based on significance testing, and the preset significance level of node splitting and merging categories is the same; the maximum growth depth of the tree during algorithm training is limited to four layers, and the model obtained after training corresponds to a four-level hierarchical judgment structure; the minimum sample size required by the parent node in the model is greater than the minimum sample size required by the child node.

[0011] Furthermore, the CHAID decision tree algorithm used to train the model can be replaced with other decision tree algorithms that use information gain ratio or chi-square test as node splitting criteria, including the C4.5 decision tree algorithm or the C5.0 decision tree algorithm; after the replacement algorithm is trained on the same tumor PICC catheterization population sample, a four-level hierarchical prediction model with similar performance can be obtained.

[0012] Furthermore, the risk stratification prediction model has seven terminal nodes, each terminal node uniquely corresponding to a thrombosis risk level. The thrombosis risk level includes four different levels: low risk, medium risk, high risk, and extremely high risk. The extremely high risk level corresponds to three different terminal nodes, and the high risk level corresponds to two different terminal nodes. The same level identifier is output for the same risk level.

[0013] Furthermore, the preset risk threshold of the nutritional risk screening score can be adjusted within a clinically acceptable range, and the normal range boundary value of the triglyceride level can be adaptively set according to the laboratory reference values ​​of different medical institutions. The adjusted model maintains the same hierarchical judgment logic and risk level classification standard, thereby enabling the system to have cross-institutional and cross-population clinical adaptability and reducing prediction bias caused by differences in detection standards.

[0014] Furthermore, the data acquisition unit interfaces with the hospital nursing information system or electronic medical record system, and can automatically read the patient's anticoagulant usage status, triglyceride level, nutritional risk screening score, and surgical history information; the output unit is also used to present clinical intervention recommendations that match the thrombosis risk level.

[0015] The clinical intervention recommendations are based on PICC nursing industry standards and correspond one-to-one with the four thrombosis risk levels output by this system. For low-risk patients, maintain the original anticoagulation regimen. If anticoagulation is discontinued due to bleeding risk, the risk must be reassessed immediately. A routine upper limb venous ultrasound examination is performed on day 7 after catheter insertion. Patients with normal results should be followed up every 4 weeks until catheter removal. The standard PICC maintenance procedure is followed, instructing patients to perform daily fist-clenching exercises and monitoring coagulation function weekly. For intermediate-risk patients, low molecular weight heparin prophylaxis is initiated within 24 hours of catheter insertion, with the anticoagulation course covering the indwelling period and continuing until 7 days after catheter removal. Ultrasound examinations are performed on days 3 and 7 after insertion, with patients with normal results followed up every 2 weeks. The catheter insertion site is assessed daily to avoid infusion of highly irritating drugs. Patients are instructed to limit weight-bearing on the affected limb to no more than 5 kg and to monitor coagulation function every 3 days. For high-risk patients, intensive prophylactic low-molecular-weight heparin therapy should be initiated immediately after catheter placement, combined with intermittent pneumatic compression therapy if there are no contraindications. Ultrasound examinations should be performed on days 1, 3, and 7 after catheter placement, and weekly follow-up examinations should be conducted until 14 days after catheter removal if no abnormalities are found. Catheter maintenance should be performed by experienced nurses, and forced flushing of the catheter is prohibited. Patients should be instructed to elevate the affected limb while lying in bed, and coagulation function should be monitored once daily. For very high-risk patients, low-molecular-weight heparin prophylaxis should be initiated 12 hours before catheter placement, and then switched to therapeutic dose anticoagulation after placement, combined with graded compression stockings and intermittent pneumatic compression therapy. The first ultrasound examination should be performed within 6 hours after catheter placement, followed by daily follow-up examinations for the first 3 days, and then every 3 days until 21 days after catheter removal. A special nursing record should be established, and blood products and fat emulsions should be avoided. If thrombosis is diagnosed, vascular surgery consultation should be requested immediately. Patients should be instructed to strictly adhere to bed rest and immobilize the affected limb, and coagulation function should be monitored every 6 hours. Anticoagulant antagonists should be readily available. This system allows medical institutions to adjust intervention plans according to their own standards and individual patient conditions. It can be built into templates for hospitals of different levels and automatically record intervention execution data.

[0016] Furthermore, the prediction model embedded in the processing unit is a static model formed by performing univariate analysis, CHAID decision tree training and validation on clinical data of patients with PICC catheterization for tumors, and then solidifying it. The four-level hierarchical judgment structure, factor classification thresholds, sample grouping ratios, node splitting significance levels, and maximum growth depth parameters obtained during model training remain unchanged after the model is solidified, thereby ensuring the predictive stability and repeatability of the model after clinical deployment and avoiding risk level drift caused by parameter changes.

[0017] Furthermore, the risk stratification prediction model also includes a metabolic risk correction module. This module calculates a metabolic risk index based on the patient's body mass index, age, and the deviation of the measured triglyceride value from the normal reference range. For the category of triglycerides "above the normal range," it is further divided into a mildly elevated subtype and a significantly elevated subtype based on the metabolic risk index. The metabolic risk index is positively correlated with the patient's body mass index, negatively correlated with age, and positively correlated with the upward deviation of triglycerides. For the category of triglycerides "below the normal range," it is further divided into a mildly decreased subtype and a significantly decreased subtype based on the metabolic risk index. The metabolic risk index is negatively correlated with the patient's body mass index, positively correlated with age, and positively correlated with the downward deviation of triglycerides. The corrected triglyceride subtype replaces the original triglyceride category and is used as the second-layer judgment factor input into the decision tree model.

[0018] Furthermore, the risk stratification prediction model also includes a nutrition-inflammation joint pathway scoring module. This module is used to jointly determine the patient's nutritional risk screening score with the level of at least one serum inflammatory marker, generating four joint pathway phenotypes: high nutritional risk-high inflammation, high nutritional risk-low inflammation, low nutritional risk-high inflammation, and low nutritional risk-low inflammation. These four phenotypes replace the original binary classification result of the nutritional risk screening score and are used as the third-layer judgment factor input into the decision tree model.

[0019] Secondly, a method for predicting the risk of upper limb venous thrombosis after PICC placement in cancer patients is also provided. This method is based on the system described in the first aspect and includes: A risk stratification prediction model is provided, which is trained and validated by the CHAID decision tree algorithm and then solidified. The model is optimized for the PICC catheterization population of cancer patients and has a defined four-level hierarchical judgment structure. The first-level judgment factor is the anticoagulant use status, the second-level judgment factor is the triglyceride level, the third-level judgment factor is the nutritional risk screening score, and the fourth-level judgment factor is the surgical history. Obtain information on the target patients' anticoagulant use status, triglyceride levels, nutritional risk screening scores, and surgical history. According to the order of the four-level hierarchical judgment structure, the acquired information is input into the model in sequence. Each level determines whether to proceed to the next level of judgment or directly output the corresponding thrombosis risk level based on the assigned value category of the judgment factor at that level. The final output is the thrombosis risk level.

[0020] This invention employs the CHAID decision tree algorithm to mine a four-level hierarchical judgment structure, which, after training and validation, is solidified into a static risk stratification prediction model. Using anticoagulant usage status, triglyceride levels, nutritional risk screening scores, and surgical history as core judgment factors, it achieves rapid stratification and classification of upper limb venous thrombosis risk in cancer patients after PICC placement, significantly improving the predictive model's generalization ability and interpretability on new clinical samples. This approach reduces the model's dependence on original detection accuracy and computational complexity by replacing the comprehensive calculation of continuous indicators with discrete category assignments, enabling risk level classification without complex calculations, making it suitable for real-time screening scenarios in frontline clinical settings. Simultaneously, the system's ability to interface with hospital nursing information systems or electronic medical record systems, along with its clinically adaptable adjustment mechanism for risk thresholds and reference ranges, effectively reduces errors and time costs associated with manual data entry, ensuring the stability of prediction results and possessing cross-institutional and cross-population clinical promotion value. Furthermore, by outputting clinical intervention recommendations matched to risk levels, this approach avoids over-intervention for low-risk patients and under-intervention for high-risk patients, achieving individualized and precise thrombosis prevention.

[0021] Beneficial effects By implementing the above-described PICC placement risk prediction system for upper limb venous thrombosis in cancer patients provided by the present invention, the following technical effects are achieved: (1) The four-level hierarchical judgment structure mined by the CHAID algorithm uses anticoagulant use status, triglyceride level, nutritional risk screening score, and surgical history as the first four judgment layers in sequence. After solidifying this structure into a clinical prediction process, this judgment order can screen out low-risk individuals using anticoagulants at early nodes, avoiding subsequent invalid calculations, while also providing refined triage for patients not using anticoagulants layer by layer. This solidified hierarchical judgment process significantly reduces computational redundancy in clinical applications and ensures the clinical interpretability of the prediction path, enabling medical staff to intuitively understand the clinical basis of each judgment layer.

[0022] (2) The four indicators—anticoagulant use status, triglyceride level, nutritional risk screening score, and surgical history—were discretized into two or three categories of values, and this four-factor combination was used as the entire input variable of the model, replacing the complex comprehensive calculation of more than ten continuous indicators in the traditional model. This four-factor combination focuses on the core pathological pathway of thrombosis after PICC placement in cancer patients. Its discretization process allows the model to complete the prediction without relying on precise detection values, requiring only routine clinical interpretation, thereby significantly reducing the dependence on detection accuracy and data collection costs, while ensuring the consistency of the model across different medical institutions. Three independent stratified pathways were set up for the triglyceride index, covering both the high risk of thrombosis associated with hyperlipidemia and the thrombosis risk associated with low triglycerides accompanied by malnutrition and cachexia, avoiding the omission of risk information in the conventional dichotomy method and improving the accuracy of risk identification for patients in different metabolic states.

[0023] (3) By calculating the anticoagulation protection attenuation coefficient based on medication adherence and time interval, the original binary anticoagulant usage status is expanded into three dynamic categories: adequate protection, partial protection, and no protection. This module addresses the shortcomings of the static "use / not use" label, which fails to reflect actual efficacy attenuation and medication fluctuations. It enables the model to identify high-risk patients who, although using anticoagulants, suffer from insufficient protection due to poor adherence or long intervals between doses. Compared to the static binary classification model, the introduction of the dynamic attenuation module significantly reduces the thrombosis misdiagnosis rate in the anticoagulant user subgroup and simultaneously decreases the actual thrombosis incidence rate in the low-risk group, achieving accurate quantification of the patient's true coagulation protection status.

[0024] (4) By introducing a metabolic risk index that integrates body mass index, age, and the degree of triglyceride deviation, the "higher than normal" triglyceride category is further subdivided into two subtypes: mildly elevated and significantly elevated. This correction enables the model to distinguish between patient groups within the same abnormal range but with different metabolic burdens, avoiding overclassification of those with mild metabolic disorders as extremely high-risk, while significantly improving the detection sensitivity for patients with truly high metabolic burdens. Compared to the model that only uses the original three-category triglyceride classification, the corrected model shows a clear improvement in both positive predictive value and sensitivity in extremely high-risk populations, while the false positive rate decreases simultaneously.

[0025] (5) By combining nutritional risk screening scores with serum inflammatory marker levels, four combined nutrition-inflammation phenotypes are generated, replacing the original single binary classification of nutritional scores. This module identifies a high-risk subgroup of latent inflammation: "low nutritional risk - high inflammation state." These patients are often missed in conventional models due to normal nutritional scores, but their actual thrombosis rate is comparable to that of high-nutritional-risk patients. Simultaneously, the module directly outputs the "high nutritional risk - high inflammation state" phenotype as extremely high risk, avoiding delays caused by its inclusion in subsequent surgical history assessments. This module significantly improves the model's sensitivity in identifying latent inflammatory risks and enhances its ability to capture the synergistic effects of multiple pathological pathways. Attached Figure Description

[0026] To make the above-described upper limb venous thrombosis risk prediction system for PICC placement in cancer patients more clearly and understandable, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0027] Figure 1 This is a flowchart illustrating the method of the present invention. Figure 2 This diagram illustrates the system structure of the present invention. Figure 3 This represents the integration relationship diagram of the enhancement modules. Detailed Implementation

[0028] Example 1: This embodiment uses patients with PICC line insertion in the oncology department of a tertiary-level Class A hospital as the research subjects to exemplify the construction, validation, and clinical application of a model and system for predicting the risk of upper limb venous thrombosis after PICC line insertion in cancer patients. The method flow is as follows: Figure 1 As shown, the system structure is as follows: Figure 2 As shown.

[0029] First, clinical data were collected. Patients with tumors who had undergone PICC placement and met the inclusion / exclusion criteria were selected as study subjects. Their clinical data were collected, including general demographic characteristics, disease characteristics, laboratory test results, and clinically relevant indicators. Among the laboratory test results, triglyceride levels and NRS2002 nutritional risk screening scores were collected. Clinically relevant indicators included anticoagulant use and surgical history. A total of 178 patients were included. Patients who developed PICC-related venous thrombosis were assigned to the thrombosis group, and those who did not develop thrombosis were matched 1:1 to the non-thrombosis group. This study used a 1:1 case-control matched design. Matching factors included patient age, sex, tumor type, and placement site. The purpose of this design was to effectively control the influence of these confounding factors on the risk of thrombosis, improve the testing power for thrombosis-related risk factors, and avoid insufficient sample size in the thrombosis group due to the low baseline incidence of PICC-related venous thrombosis in the real clinical population, which would prevent the screening of statistically significant core risk factors. Of the 178 patients included in this study, 89 were in the thrombosis group and 89 in the non-thrombosis group. This sample ratio is an artificially balanced ratio under the study design and does not represent the baseline incidence of thrombosis in the real clinical population. Stratified random sampling was used during model validation to ensure that the ratio of thrombosis to non-thrombosis groups remained consistent between the training and validation sets. The area under the receiver operating characteristic (ROC) curve, sensitivity, and specificity indices reflect the model's relative discriminative ability, not the absolute predictive probability for the real population. In clinical application, this system will automatically calibrate the predicted probability based on the baseline incidence of thrombosis in the target medical institution's real population to ensure the accuracy of the final output thrombosis risk classification. This case-control study design is a common standard method in the field of medical risk prediction model construction and has been widely used and recognized in numerous similar studies both domestically and internationally. Subsequently, univariate analysis was performed using chi-square test, t-test, or Wilcoxon rank-sum test to screen variables statistically significant with the occurrence of PICC-CRT.

[0030] Next, model training was performed. Variables selected from univariate analysis were used as predictors and imported into the CHAID package of R4.2.3 software. The chaid() function was used for chi-square automatic interactive detection decision tree analysis. The model parameters were set as follows: the significance level for node splitting and merging categories was 0.05, the tree growth depth was limited to 4 levels, the minimum number of cases per parent node was 20, and the minimum number of cases per child node was 5. Training was conducted using the occurrence of PICC-CRT as the outcome metric. The algorithm automatically performed variable selection and node splitting based on the sample data, ultimately generating a decision tree model with 4 levels, containing 5 growth nodes and 7 terminal nodes. The model automatically identifies four core risk factors from numerous clinical variables and establishes a corresponding hierarchical judgment order: the first layer is anticoagulant use (values: not used = 1, used = 0); the second layer is triglyceride level (values: above normal = 1, normal = 2, below normal = 3); the third layer is NRS2002 nutritional score (values: ≥3 = 1, <3 = 2); and the fourth layer is surgical history (values: none = 0, present = 1). This model can progressively assess a patient's thrombosis risk level based on these four indicators, ultimately outputting the corresponding risk level.

[0031] To verify the predictive performance of the model, the total sample was randomly divided into a training set (125 cases) and a validation set (53 cases) in a 7:3 ratio. The area under the receiver operating characteristic (AUC), sensitivity, and specificity were used for evaluation. The results showed that the AUC of the training set was above 0.909, with a sensitivity of at least 87.32% and a specificity of at least 96.23%; the AUC of the validation set was above 0.837, with a sensitivity of at least 78.57% and a specificity of at least 88.46%. These indicators demonstrate that the model has good predictive accuracy and generalization ability, and it was selected as the final prediction model after passing the validation.

[0032] Based on the decision tree model described above, this embodiment further constructs a thrombosis risk prediction system. The system includes a computer host, an input module, a display module, and a storage module. The input module is a keyboard or touchscreen, used by medical staff to input four core indicators: the patient's anticoagulant usage, triglyceride level, NRS2002 nutritional score, and surgical history. The storage module stores the trained CHAID decision tree prediction model, patient input data, prediction results, and PICC-CRT clinical intervention guidelines for cancer patients. The computer host has a built-in prediction program that, after receiving data from the input module, automatically calculates the patient's thrombosis risk level according to the model's preset hierarchical judgment logic. The display module is a screen that outputs the prediction results in real time, clearly indicating the patient's risk type and corresponding clinical intervention recommendations.

[0033] The standardized clinical prediction process of the model is as follows: First, medical staff input the patient's four core indicators through the input module. The system first determines the use of anticoagulants. If the patient is using anticoagulants, they are directly classified as low-risk for thrombosis; if not, the system proceeds to the next step. Next, triglyceride levels are assessed: if triglycerides are higher than normal, the patient is classified as extremely high-risk for thrombosis; if triglycerides are lower than normal, the system proceeds to a nutritional score assessment; if triglycerides are within the normal range, the system proceeds to the next step. For patients with triglycerides below normal, the NRS2002 nutritional score is assessed: if the score is ≥3, the patient is classified as extremely high-risk for thrombosis; if the score is <3, the patient is classified as high-risk for thrombosis. For patients with triglycerides within the normal range, the NRS2002 nutritional score is assessed: if the score is ≥3, the patient is classified as extremely high-risk for thrombosis; if the score is <3, the system continues to assess surgical history: if there is no surgical history, the patient is classified as high-risk for thrombosis; if there is a surgical history, the patient is classified as intermediate-risk for thrombosis. Through the above-mentioned hierarchical process, patients are classified into four risk levels: low, medium, high, and very high, with each level corresponding to a different clinical intervention strategy.

[0034] In terms of predictive efficacy, the CHAID decision tree algorithm can capture the nonlinear relationship between clinical indicators and thrombosis occurrence, achieving a training set AUC of 0.909, significantly higher than the traditional Logistic regression model (usually ≤0.794), with greatly improved sensitivity and specificity. Regarding clinical relevance, the study subjects were pure tumor patients with PICC catheters. The four risk factors selected closely matched the typical clinical characteristics of tumor patients, such as lipid metabolism disorders and nutritional status decline caused by radiotherapy and chemotherapy, coagulation dysfunction requiring assessment of anticoagulation therapy, and a history of tumor-related surgery, avoiding insufficient relevance due to the heterogeneity of the study subjects. In terms of ease of operation, only four routine indicators are required, and the stratified judgment logic enables rapid screening, with a single patient prediction time of no more than 10 seconds, far superior to the multi-indicator comprehensive calculation of existing models. Regarding stratified guidance, the four risk levels avoid a "one-size-fits-all" intervention model; low-risk individuals avoid the risk of bleeding due to excessive anticoagulation, while extremely high-risk individuals initiate emergency preventive treatment. In terms of system implementation, the integrated hardware and software design allows for seamless integration with hospital nursing information systems or electronic medical record systems, enabling automatic data retrieval and result writing back, thus reducing manual data entry errors. Regarding testing costs, the testing costs for triglycerides and the NRS2002 nutritional score are significantly lower than those for indicators such as D-dimer and fibrinogen, and no additional testing is required.

[0035] Furthermore, this embodiment also provides several alternative solutions. Regarding algorithm replacement, the CHAID decision tree algorithm used in the model training phase can be replaced with the C4.5 decision tree algorithm (using information gain ratio for node splitting) or the C5.0 decision tree algorithm (adding pruning optimization to C4.5). Both of these algorithms belong to the classic decision tree class. After training and validation on the same clinical dataset of patients with PICC line placement in tumors, both can yield thrombosis risk prediction models with hierarchical and graded judgment capabilities. Finally, the validated model structure can be solidified into the prediction system, achieving prediction performance similar to the original solution. These algorithms can be implemented in R4.2.3 software using corresponding toolkits. Specifically, the C4.5 decision tree can be implemented using the J48() function of the RWeka package, the C5.0 decision tree can be implemented using the C5.0() function of the C50 package, and the conditional inference decision tree can be implemented using the ctree() function of the party package. Regarding indicator fine-tuning and substitution, while maintaining the same hierarchical judgment position, the threshold for "above normal value" triglycerides can be fine-tuned to 1.7 mmol / L, or the risk threshold for the NRS2002 nutritional score can be adjusted from 3 to 4 points to accommodate differences in reference values ​​among different medical institutions. In terms of system hardware substitution, the input module can use a barcode scanner instead of a keyboard to automatically scan the graphic codes on the test reports; the display module can use a touchscreen on a hospital nursing workstation instead of a regular display screen; and the storage module can use a hospital cloud server instead of a local hard drive. All of these substitutions will not affect the system's working logic and predictive functions.

[0036] Example 2: In the aforementioned embodiments, the anticoagulant usage status was binarized into "used" or "not used," without considering the actual duration of anticoagulant efficacy and fluctuations in patient medication adherence. Cancer patients often use anticoagulants intermittently due to thrombocytopenia, bleeding events, or treatment cycle adjustments; a single "used" label is insufficient to reflect the current true level of thrombotic protection. Therefore, this embodiment introduces an anticoagulant protection attenuation coefficient. Based on the time interval between the patient's most recent anticoagulant use and PICC placement, the prescribed dosing frequency, and actual adherence, an effective protection factor between 0 and 1 is calculated. This expands the original binary variable into a continuous variable, and then remaps it to three new categories—"sufficient protection," "partial protection," and "no protection"—through a preset threshold, thereby improving the model's adaptability to dynamic medication patterns.

[0037] During the data acquisition phase, in addition to collecting data on "whether anticoagulants were used", the following parameters were also collected: the prescribed medication frequency; the actual number of medications administered in the past 7 days; and the time interval between the most recent medication administration and PICC placement.

[0038] Calculate the basic compliance rate:

[0039] In the formula, The basic compliance ratio is dimensionless and ranges from 0 to 1. It represents the proportion of the patient's actual medication use relative to the prescribed number of medication uses. If the ratio exceeds 1, it is counted as 1. The number of times the patient actually used anticoagulants in the past 7 days, expressed as times per week; The frequency of anticoagulant use as prescribed by the doctor, in units of times per week.

[0040] The anticoagulation protection attenuation coefficient is constructed using the following formula:

[0041] In the formula, This is the anticoagulation protection attenuation coefficient, dimensionless, ranging from 0 to 1, used to quantify the overall level of protection provided by anticoagulants at the time of PICC placement. A higher value indicates more adequate protection. This is an empirical decay constant, reflecting the rate at which the protective effect of anticoagulants decays over time. Its specific value depends on the type of anticoagulant: for low molecular weight heparin, Regarding warfarin, If the patient is using multiple anticoagulants simultaneously, take the highest value of k. This refers to the time interval between the patient's most recent anticoagulant use and the PICC insertion date, expressed in days. If medication was used on the day of insertion, then... .

[0042] Will The value is mapped to three new categories: like If so, it is judged as "fully protected"; like If so, it is determined to be "partially protected"; like If so, it is judged as "unprotected".

[0043] The original binary anticoagulant usage status is replaced by the three new categories mentioned above, serving as the first-level decision factors in the CHAID decision tree model. The model's root node then splits into three branches, each corresponding to a different subsequent decision path. For example, those with "sufficient protection" are directly classified as low-risk; those with "partial protection" proceed to triglyceride assessment; and those with "no protection" have their risk weights increased for subsequent factors.

[0044] In clinical practice, the above parameters can be automatically retrieved from medication records in the hospital information system, or easily entered by nurses through the system interface. The system automatically calculates these parameters in the background. It completes category mapping without any manual intervention.

[0045] A CHAID decision tree model, which does not introduce an anticoagulant protection attenuation coefficient and only uses a binary classification of whether or not anticoagulant is used, was used as the conventional comparison method. The comparison process used the same dataset and model parameters, only replacing the first-level decision factor with a binary variable. The mapped three-category classification. To ensure fairness in the comparison, both models used only publicly available medication records and did not introduce additional detection indicators. The comparison criteria were set as follows: among the 82 patients with a history of anticoagulant use included in the analysis, 34.1% had actual medication adherence of less than 50% or an interval of more than 3 days between recent doses. The comparison results showed that in the conventional method, all patients who "used" anticoagulants were judged as low-risk, resulting in an actual thrombosis rate of 19.5% in this group; introducing... Subsequently, the incidence of thrombosis was 31.2% in patients classified as "partially protected" and 52.9% in patients classified as "unprotected." Based on this, the system redirected these two types of patients to subsequent hierarchical assessments instead of directly outputting low risk. Ultimately, the overall model reduced the missed diagnosis rate in the anticoagulant user subgroup from 19.5% to 6.1%, while the actual incidence of thrombosis in the low-risk group decreased from 19.5% to 4.3%, achieving a precise response to dynamic medication patterns.

[0046] Example 3: Cancer patients exhibit significant individual differences in their triglyceride metabolism due to interventions such as chemotherapy and hormone therapy. In the aforementioned embodiments, triglyceride levels were categorized into three types: "above normal," "normal," and "below normal." However, even within the "above normal" category, different patients still showed significant differences in metabolic load. Therefore, this embodiment introduces a metabolic risk index. This index integrates the nonlinear mapping relationship between the patient's body mass index, age, and the deviation of triglycerides from normal levels to dynamically weight and correct the original triglyceride category. This allows the model to distinguish between "mildly elevated" and "significantly elevated" subtypes, thereby improving the sensitivity of identifying extremely high-risk individuals.

[0047] During the data acquisition phase, in addition to collecting measured triglyceride values, patients' body mass index and age were also collected.

[0048] Define the normal reference upper limit for triglycerides. (The clinically routine value is 1.7 mmol / L), calculate the relative deviation of triglycerides:

[0049] In the formula, This represents the relative deviation of triglycerides, dimensionless, indicating the relative extent to which the measured value exceeds the normal upper reference limit; Serum triglyceride concentration measured in patients, in mmol / L; This is the upper limit of the normal reference value for triglycerides used in laboratories or clinical settings, in mmol / L, usually taken as 1.7 mmol / L.

[0050] when When, it indicates that the normal upper limit has not been exceeded; when When this occurs, it indicates that the normal upper limit has been exceeded.

[0051] The metabolic risk index is further constructed using the following formula:

[0052] In the formula, This is a dimensionless metabolic risk index used to further subdivide the "higher than normal" triglyceride category into subtypes, under normal metabolic conditions. A value close to 1 indicates a higher metabolic risk, with larger values ​​suggesting greater metabolic risk. The patient's body mass index is expressed in kg / m² and is calculated by dividing weight by the square of height. The patient's age.

[0053] In this formula The reference center value of 22 kg / m² is the ideal weight standard for Asian populations recommended by the World Health Organization. The age reference center value of 60 years is the median age of onset of cancer patients included in this study. Selecting these reference centers ensures that the formula achieves the best correction effect in the target population. The triglyceride term uses... The nonlinear design aims to avoid the imbalance of correction weights caused by excessively high triglyceride levels, while ensuring its positive contribution to metabolic risk. The study uses a relatively mild correction margin because the risk of thrombosis in cancer patients is primarily driven by chemotherapy-induced lipid metabolism disorders rather than simple obesity, and excessive amplification would be detrimental. The influence of this can lead to an overestimation of the risk in obese patients by the model. For For extremely obese patients with a weight of ≥40 kg / m² and no obvious malnutrition, the results calculated by this formula can be multiplied by a correction factor of 1.1 in clinical application to further improve the accuracy of risk identification.

[0054] The metabolic risk index was fused and corrected with the original triglyceride category: If the original category is "below normal" or "normal", keep the original category unchanged; If the original category is "above normal", then further judgment is made: when If the patient is in the "significantly elevated triglyceride subtype", then the patient is in the "mildly elevated triglyceride subtype".

[0055] The corrected subtype information is used as a new input feature, replacing the original triglyceride category, and fed into the second-level decision node of the CHAID decision tree model. During training, the model automatically splits into new child nodes based on this subtype, achieving refined stratification of the extremely high-risk population.

[0056] The CHAID decision tree model, using only the original three-category classification of triglycerides without metabolic risk index correction, was used as the standard comparative method. During the comparison, both models used the same training and validation sets (data from 178 patients with PICC line placement for tumors, randomly split in a 7:3 ratio), the same decision tree parameter settings (4-layer growth depth, minimum number of cases per parent node of 20, minimum number of cases per child node of 5, significance level of 0.05), and the same hierarchical order of the core decision factors. The only difference was whether the second-layer decision factor was replaced with the corrected "triglyceride subtype". The results showed that the standard method had a positive predictive value of 71.2% and a sensitivity of 78.6% for the very high-risk population in the validation set. After introducing metabolic risk index correction, the positive predictive value for the very high-risk population increased to 86.4%, and the sensitivity increased to 89.3%, while the false positive rate decreased from 12.5% ​​to 7.8%. These results indicate that metabolic risk index correction can effectively distinguish metabolic load differences under the "above normal" triglyceride category, reclassifying some mildly elevated patients who were originally misclassified as extremely high risk into the intermediate-high risk category, thereby significantly improving the clinical discrimination accuracy of the prediction model.

[0057] Example 4: In the aforementioned embodiments, only the NRS2002 nutritional score was used as the third-level judgment factor. However, the thrombosis risk in cancer patients is simultaneously influenced by both nutritional status and systemic inflammatory state. Inflammatory factors can activate the coagulation cascade, while malnutrition exacerbates the inflammatory response, creating a vicious cycle. This embodiment integrates NRS2002 with common inflammatory biomarkers to construct a nutrition-inflammation joint risk pathway score. This score does not rely on a specific threshold but instead combines the risk levels of the two dimensions into a matrix, outputting four pathway phenotypes: "low nutrition-low inflammation," "low nutrition-high inflammation," "high nutrition-low inflammation," and "high nutrition-high inflammation." Each phenotype is assigned a different thrombosis risk weight and serves as a new composite judgment factor, replacing the original single nutritional score, thereby more comprehensively reflecting the pathophysiological state of cancer patients.

[0058] In addition to the standard NRS2002 score, an easily obtainable inflammatory marker should be added: serum C-reactive protein (CRP) (mg / L) or the neutrophil / lymphocyte ratio (NLR) calculated from a complete blood count. CRP is preferred as it is a routine preoperative test in most hospitals.

[0059] Inflammatory markers are classified into two categories: according to clinical practice, CRP > 10 mg / L or NLR > 3.0 is considered a "hyperinflammatory state", otherwise it is considered a "hypoinflammatory state".

[0060] A nutrition-inflammation joint matrix was constructed, as shown in Table 1.

[0061] Table 1 Nutrition-Inflammation Combined Matrix ≥3 points (high nutritional risk) High inflammation High nutrition-high inflammation type Extremely high ≥3 points (high nutritional risk) Low inflammation High nutrition - low inflammation Medium and high <3 points (low nutritional risk) High inflammation Low nutrition-high inflammation type high <3 points (low nutritional risk) Low inflammation Low nutrition-low inflammation type medium to low The aforementioned joint pathway phenotype is used as a new discretized variable to replace the NRS2002 binary variable in the original scheme, and placed in the third-level decision position of the decision tree model.

[0062] During decision tree training, the model automatically splits into corresponding child nodes based on the joint phenotype. For example, if a patient entering this layer belongs to the "hypernutrition-hyperinflammation" type, they are directly judged as extremely high risk and do not need to enter the fourth layer for surgical history assessment; if they belong to the "hyponutrition-hypoinflammation" type, they enter the surgical history assessment node.

[0063] During clinical implementation, the system can automatically retrieve CRP or NLR results from the laboratory report system and automatically integrate them with the NRS2002 score to generate a combined pathway phenotype. The entire process is transparent to the operator and does not increase the workload.

[0064] The CHAID decision tree model, which uses only the NRS2002 nutritional score as the third-level decision factor, was used as the standard comparative method. The same dataset and model structure were used in the comparison process, except that the fourth phenotype generated by the nutrition-inflammation combined pathway score was replaced at the third-level decision factor. The comparison conditions were set as follows: serum CRP data were collected simultaneously from patients, with CRP > 10 mg / L defined as a high-inflammatory state; NRS2002 ≥ 3 points indicated high nutritional risk. A total of 162 patients with tumor-associated PICC catheters and CRP test results were included. The results showed that in the standard method, 28 patients had NRS2002 ≥ 3 points and CRP > 10 mg / L, of whom 21 developed thrombosis; however, the standard model only classified 16 of these patients as extremely high risk based on the NRS2002 score, while the remaining 12 patients, due to a nutritional score ≥ 3 without inflammation correction, were still classified as extremely high risk, resulting in no actual stratification difference. After introducing the nutrition-inflammation combined risk pathway score, the model directly outputs extremely high risk for the high-nutrition-high-inflammation phenotype, with a positive predictive value of 84.6%; and for the low-nutrition-high-inflammation phenotype, it is upgraded from "continuing to the fourth level of judgment" in the conventional model to "directly outputting high risk". Overall, the nutrition-inflammation combined risk pathway score improves the model's sensitivity in identifying patients with latent inflammatory risk from 23.1% in the conventional method to 88.5%, significantly reducing the risk of missed diagnosis of this high-risk subgroup.

[0065] Figure 3The integration relationship between the aforementioned anticoagulation protection dynamic attenuation module, metabolic risk correction module, and nutrition-inflammation combined pathway scoring module and the CHAID decision tree model is illustrated. Figure 3 As shown, the three enhancement modules map or correct the original input factors respectively, and the generated new categories replace the original factors as the judgment inputs of the first, second and third layers of the decision tree, and finally output the thrombosis risk level.

Claims

1. A system for predicting the risk of upper limb venous thrombosis after PICC placement in cancer patients, characterized in that, include: The data acquisition unit is used to acquire patients' anticoagulant use status, triglyceride levels, nutritional risk screening scores, and surgical history information; The processing unit embeds a risk stratification prediction model trained and validated using the CHAID decision tree algorithm. This model has a defined four-level hierarchical judgment structure, where the first-level judgment factor is anticoagulant use status, the second-level factor is triglyceride level, the third-level factor is nutritional risk screening score, and the fourth-level factor is surgical history. The processing unit is configured to sequentially call each judgment factor according to the hierarchical judgment structure, outputting the corresponding thrombosis risk level layer by layer. The model training employs a 1:1 case-control matching design, with matching factors including patient age, gender, tumor type, and catheter placement site. Model validation uses a stratified random sampling method to ensure that the ratio of thrombosis to non-thrombosis groups remains consistent between the training and validation sets. In clinical application, the processing unit automatically calibrates the predicted probability based on the baseline incidence of thrombosis in the target medical institution's real population. The risk stratification prediction model includes an anticoagulant protection dynamic attenuation module. This module calculates an anticoagulant protection attenuation coefficient based on the patient's prescribed anticoagulant dosing frequency, actual dosing frequency, and the time interval between the most recent dosing and PICC placement. Based on this coefficient, the anticoagulant usage status is mapped from a binary category to three categories: adequate protection, partial protection, and no protection. The anticoagulant protection attenuation coefficient is positively correlated with the proportion of actual dosing frequency and negatively correlated with the time interval. The mapped new category replaces the original anticoagulant usage status and serves as the first-layer judgment factor input into the decision tree model. The risk stratification prediction model also includes a nutrition-inflammation joint pathway scoring module, which is used to jointly determine the patient's nutritional risk screening score with the level of at least one serum inflammatory marker to generate four joint pathway phenotypes, namely high nutritional risk-high inflammation state, high nutritional risk-low inflammation state, low nutritional risk-high inflammation state, and low nutritional risk-low inflammation state. These four phenotypes replace the original binary classification results of the nutritional risk screening score and are used as the third-layer judgment factors input into the decision tree model; The risk stratification prediction model also includes a metabolic risk correction module. This module calculates a metabolic risk index based on the patient's body mass index (BMI), age, and the deviation of the measured triglyceride value from the normal reference range. For the category of "triglycerides above the normal range," it further classifies them into a mildly elevated subtype and a significantly elevated subtype based on the metabolic risk index. The metabolic risk index is positively correlated with the patient's BMI, negatively correlated with age, and positively correlated with the upward deviation of triglycerides. For the category of "triglycerides below the normal range," it further classifies them into a mildly decreased subtype and a significantly decreased subtype based on the metabolic risk index. The metabolic risk index is negatively correlated with the patient's BMI, positively correlated with age, and positively correlated with the downward deviation of triglycerides. The corrected triglyceride subtype replaces the original triglyceride category and is used as the second-layer judgment factor input into the decision tree model. The output unit is used to present the thrombosis risk level.

2. The system according to claim 1, characterized in that: In the four-level hierarchical judgment structure, each judgment factor is pre-assigned a discrete category value: the anticoagulant usage status is divided into two categories: "used" and "not used"; the triglyceride level is divided into three categories: "above the normal range", "within the normal range" and "below the normal range"; the nutritional risk screening score is divided into two categories: "reached or exceeded the preset risk threshold" and "below the preset risk threshold"; and the surgical history is divided into two categories: "present" and "absent".

3. The system according to claim 1, characterized in that, The processing unit determines the thrombosis risk level according to the following logic: When the anticoagulant is in use, a low risk level is output. When the anticoagulant is in the "not used" status and the triglyceride level is "above normal range", an extremely high risk level is output. When the anticoagulant usage status is "not used" and the triglyceride level is "below the normal range", if the nutritional risk screening score is "reached or exceeded the preset risk threshold", an extremely high risk level will be output; if the nutritional risk screening score is "below the preset risk threshold", a high risk level will be output. When the anticoagulant usage status is "not used", the triglyceride level is "within the normal range", and the nutritional risk screening score is "reached or exceeded the preset risk threshold", an extremely high risk level is output. When the anticoagulant usage status is "not used", the triglyceride level is "within the normal range", the nutritional risk screening score is "below the preset risk threshold", and the surgical history is "none", a high risk level is output; When the anticoagulant usage status is "not used", the triglyceride level is "within the normal range", the nutritional risk screening score is "below the preset risk threshold", and the surgical history is "yes", a medium risk level is output.

4. The system according to claim 1, characterized in that: The node splitting criterion of the CHAID decision tree algorithm is based on significance testing, and the preset significance level of node splitting and merging categories is the same; the maximum growth depth of the tree is limited to four layers during algorithm training, and the model obtained after training corresponds to a four-level hierarchical judgment structure; the minimum sample size required for the parent node in the model is greater than the minimum sample size required for the child node.

5. The system according to claim 4, characterized in that: The CHAID decision tree algorithm used to train the model can be replaced by other decision tree algorithms that use information gain ratio or chi-square test as node splitting criteria, including the C4.5 decision tree algorithm or the C5.0 decision tree algorithm; after the replacement algorithm is trained on the same tumor PICC catheterization population sample, a four-level hierarchical prediction model with similar performance can be obtained.

6. The system according to claim 1, characterized in that: The risk stratification prediction model has seven terminal nodes, each terminal node uniquely corresponding to a thrombosis risk level. The thrombosis risk level includes four different levels: low risk, medium risk, high risk, and extremely high risk. The extremely high risk level corresponds to three different terminal nodes, and the high risk level corresponds to two different terminal nodes. The same level identifier is output for the same risk level.

7. The system according to claim 1, characterized in that: The prediction model embedded in the processing unit is a static model formed by performing univariate analysis, CHAID decision tree training and validation on clinical data of patients with PICC catheterization for tumors, and then solidifying it. The four-level hierarchical judgment structure, factor classification threshold, sample grouping ratio, node splitting significance level, and maximum growth depth parameter obtained during model training remain unchanged after the model is solidified.

8. A method for predicting the risk of upper limb venous thrombosis after PICC placement in cancer patients, characterized in that: The method is implemented based on the system described in any one of claims 1-7: The method includes: A risk stratification prediction model is provided, which is trained and validated by the CHAID decision tree algorithm and then solidified. The model is optimized for the PICC catheterization population of cancer patients and has a defined four-level hierarchical judgment structure. The first-level judgment factor is the anticoagulant use status, the second-level judgment factor is the triglyceride level, the third-level judgment factor is the nutritional risk screening score, and the fourth-level judgment factor is the surgical history. Obtain information on the target patients' anticoagulant use status, triglyceride levels, nutritional risk screening scores, and surgical history. According to the order of the four-level hierarchical judgment structure, the acquired information is input into the model in sequence. Each level determines whether to proceed to the next level of judgment or directly output the corresponding thrombosis risk level based on the assigned value category of the judgment factor at that level. The final output is the thrombosis risk level.

Images