Critical risk attributes associated with important decision rules

The Caprini score is a widely verified VTE risk assessment model with high reliability that is widely used in various surgical specialties and is not limited to urology. There were 37 variables in the Caprini scale; however, some were not associated with urological diseases. Machine learning was used to screen variables, reduce the number of variables in the Caprini scale, and focus on variables highly correlated with urological diseases, thus guiding urologists to predict VTE risk more efficiently and accurately.

The Caprini scale is divided into three risk levels, each of which has different important decision rules. At the low-risk level, among the five important decision rules, there were nine risk attributes: “(C18) (five times),” “(C16) (four times),” “(C21) (five times),” “(C19) (five times),” “(C9) (five times),” “(C1) (five times),” “(C3) (four times),” “(C10) (three times),” and “(C20) (three times)”. At the mid-risk level, among the three important decision rules, there were 11 risk attributes: “(C31) (three times),” “(C11) (three times),” “(C10) (three times),” “(C16) (three times),” “(C21) (three times),” “(C2) (three times),” “(C18) (three times),” “(C1) (three times),” “(C3) (three times),” “(C9) (two times),” and “(C34) (two times)”. For the high-risk level, among the four important decision rules, there were four risk attributes: “(C21) (four times),” “(C1) (three times),” “(C18) (two times),” and “(C19) (seven times).” Based on the above results, “(C1) (three times),” “(C18) (two times),” and “(C21) (seven times)” were critical risk attributes for VTE risk stratification when the Caprini score was used for VTE risk assessment of urological inpatients. In addition to these three critical risk attributes, obesity, history of VTE, and lung disease were also important factors. Previous studies have confirmed that these factors have an important impact on VTE risk classification. Therefore, we discuss this according to the following five factors:

Age

Age is a key characteristic variable for VTE risk. In a study on risk factors for VTE in different majors and directions, it was proven that age is an important factor influencing VTE risk classification. For people aged > 40 years, VTE risk gradually increases with age and doubles every 10 years [32]. One study showed that among VTE events in the community, the proportion of people aged > 65 years was as high as 60% [33].

In addition, with increasing age, other factors affecting VTE risk will also change, especially in older patients aged > 65 years; with organ aging, the pathophysiology of the coagulation system changes, and the incidence of chronic diseases affecting VTE risk, such as cardiovascular, cerebrovascular, and lung diseases, will also increase significantly. Therefore, age is an important factor influencing VTE risk, with VTE risk increasing significantly with increasing age. The influence of age on VTE risk can be mainly explained by two factors: traditional and unconventional risk factors.

Traditional risk factors include the following: (1) Immobility or decreased activity: With a gradual increase in age, the amount of activity tends to decrease, and the long-term immobility of older patients leads to an increase in blood viscosity, which is also an important factor influencing VTE risk. In addition, older patients have an increased risk of stroke and fracture, which can increase the probability of bed rest. Notably, VTE risk is highest in the first 4 weeks of bed rest. (2) Increases in complications: With increasing age, the incidence of diseases related to VTE, such as cancer, heart failure, stroke, and diabetes, increases significantly.

Unconventional risk factors include reduced muscle strength and venous insufficiency. Muscle strength begins to decline from the age of 50–55 years [16], with degenerative changes in the lower limb joints, leading to the loss of leg muscles and a decrease in nerve regulation function, further leading to a decrease in muscle strength and seriously affecting the blood pump function of calf muscles. A review showed that thrombosis in older patients is associated with blood stasis and reflux caused by venous dysfunction, which may be caused by the decreased blood pumping function of the calf muscles [34].

Obesity (body mass index)

“Obesity (body mass index [BMI] > 25 kg/m2; C3)” is a high-risk factor for VTE [35]. An increase in BMI is directly proportional to VTE risk. In one study, VTE risk in obese people with a BMI ≥ 30 kg/m2 increased twofold compared with that in the general population [36]. Obesity is an important risk factor for myocardial infarction [35]. Considering VTE risk, obesity and myocardial infarction are associated with hypersuperposition. A study showed that the incidence of VTE in patients exposed to two factors increased threefold compared with that in patients exposed to a single factor, and the comprehensive effect of the two exposures exceeded the sum of the individual effects [35, 37, 38]; obesity is a prerequisite for the two factors. Obesity is often accompanied by a hypercoagulable and inflammatory state in the blood, and a hypercoagulable state is an important condition for VTE. A meta-analysis of the correlation between C-reactive protein levels and VTE risk showed that the inflammatory state was positively correlated with VTE risk [39]. Therefore, similar to age, obesity is an important variable associated with VTE risk in urological patients. Based on this factor, medical staff can quickly and conveniently obtain corresponding information, perform a rough analysis of VTE risk in patients, and increase attention to VTE risk.

Surgical factors

In the present study, the characteristic variables associated with surgery included “(C2),” “(C18),” and “(C19).” Surgery is the main treatment method used in urology; however, surgery is recognized as the main risk factor for VTE, and the incidence of VTE increases significantly after major abdominal and pelvic surgeries [40]. From a pathophysiological perspective, surgery-induced vascular injury can easily lead to platelet aggregation and fibrosis repair induced by anticoagulant factors [41].

In addition, surgery is associated with risk factors for VTE, such as braking, hypercoagulability, and an inflammatory state. For new minimally invasive surgical methods, such as laparoscopy and robotic surgery, long-term pneumoperitoneum and positioning compress the main veins in the abdominal cavity [42], aggravating venous blood stasis and overlapping with other risk factors.

The incidence of VTE differs among different urological surgical types. A study on the incidence of DVT and its influencing factors in urology showed that prostatectomy (including traditional laparotomy and transurethral resection) exhibited the highest incidence, followed by cystectomy and urinary calculus surgery. Prostatectomies and cystectomies are complicated pelvic surgeries performed in urology [43].

Malignant tumors

In the present study, “(C21)” was a risk factor for VTE, and many studies have confirmed that malignant tumors are an independent risk factor for VTE [39, 44]. According to relevant research, the incidence of VTE in patients with malignant tumors is 4–5 times greater than that in patients with nonmalignant tumors [45]. Approximately 20% of patients with VTE have malignant tumors, which are a critical cause of death in patients with VTE. At the pathophysiological level, malignant tumors, as exogenous factors, activate coagulation factor X and promote platelet activation and fibrin synthesis [46]. The risk factors for malignant tumors are also associated with tumor stage and treatment methods such as surgery, central venous catheter placement or infusion port placement, and chemotherapy (a risk factor for venous endothelial injury). Each factor is an independent risk factor for VTE that produces a risk superposition effect [45, 46]. In urology, malignant tumors in the bladder, prostate, and kidney are common. Treatment methods for malignant tumors in the urinary system mainly include surgery, chemotherapy, and central venous catheterization. In addition, long-term catheter and central venous catheter indwelling increases the risk of catheter-related infection, which indirectly leads to inflammation-related VTE risk, and the comprehensive risk is far greater than the cumulative sum of the individual risks. In addition, weakness, immobility (including postoperative immobilization), and pain (including cancer and postoperative pain) caused by malignant tumors increase the risk of VTE [44].

Lung disease

In the present study, the important factors associated with lung diseases were “(C10)” and “(C11)”, which mainly included abnormal lung function, chronic obstructive pulmonary disease (COPD), and other lung diseases. According to a previous study on the correlation between COPD and VTE, COPD is an independent risk factor for VTE. Pulmonary embolism (PE) is the primary manifestation of VTE in patients with COPD compared with other patients. The incidence of cerebral vein thrombosis is lower in patients with COPD than in patients with PE [47], and the recurrence, bleeding, and death risks associated with VTE in patients with COPD are greater than those in patients without COPD [47,48,49].

Application of a practical prevention strategy based on important decision rules

Combining VTE risk assessment results with clinical preventive measures is important in VTE risk assessment. In the present study, the Caprini assessment scale was used to assess the risk of VTE in urological inpatients. There were four decision rules with support sizes > 35 in the high-risk group and three with support sizes > 35 in the medium-risk group. Clinicians, nurses, and medical administrators should pay attention to the decision-making information contained in these rules and implement appropriate preventive measures. The three risk levels are as follows:

High VTE risk decision rules

There were four high-risk decision rules for VTE: IF [C18 is 1] and [C1 is 3], IF [C21 is 1] and [C18 is 1], IF [C21 is 1] and [C1 is 3], and IF [C21 is 1] and [C19 is 1]. Based on these four decision rules, it can be concluded that the characteristics of patients with a high risk of VTE in urology include older patients aged ≥ 75 years who underwent large-scale open or tumor surgery and patients who underwent large-scale tumor or laparoscopic surgery. With the continuous development of surgical techniques, instruments, and materials, large-scale open surgery in urology has gradually been replaced by endoscopy. Minimally invasive surgery and minimally invasive surgery performed through the natural lumen have become the mainstream surgical methods in urology. However, the degree of internal injury in complex endoscopic surgery in urology is still relatively large. The variable “(C19)” is rare; however, it is suggested that “(C18)” be treated equally to “(C19)” in practical work.

Patients with high VTE risk are subdivided into two categories according to bleeding risk: (i) high VTE and low bleeding risks and (ii) high VTE and bleeding risks.

According to relevant research reports, advanced age, malignant tumors, diabetes, and recent surgical history are high-risk factors for bleeding [1, 2, 41, 46] and VTE. Therefore, appropriate preventive measures should be taken according to the results of the bleeding evaluation. Clinicians should be aware of the standardized diagnosis and treatment of VTE, the indications and contraindications for anticoagulation and thrombolytic therapy, and the corresponding drugs (including unfractionated heparin, low-molecular-weight heparin, warfarin, and new oral anticoagulants) and treatment equipment.

For patients with high VTE and low bleeding risk, based on their conditions, medical staff should actively conduct basic preventive measures against VTE (including health education, avoiding dehydration and breaking, getting out of bed early, engaging in functional exercise, avoiding lower limb vascular puncture, and raising the lower limbs in a timely manner) to reduce some risk factors. In addition, based on basic preventive measures, drug use alone or drug use combined with mechanical prevention (including elastic socks, intermittent inflation and compression devices, and plantar venous pumps) should be adopted [7, 48].

Patients with high VTE and bleeding risks were treated with mechanical prevention methods, the course of prevention was generally 7–14 days postoperatively, and major tumor surgery was postponed until 28 days postoperatively [7]. Urologists should prioritize older patients undergoing tumor surgery in their departments. If necessary, they should invite VTE multidisciplinary diagnosis and treatment teams or specialists to participate in preoperative discussions, perioperative treatment plan formulation, informed notification, and other work and incorporate multidisciplinary team opinions into the quality control index system for such patients. Urologists and nursing staff should closely observe the clinical manifestations of VTE, such as cough, hemoptysis, chest pain, and lower limb swelling.

Medium VTE risk decision rules

There were three high-risk decision rules for VTE. By combining these findings, we can conclude that the characteristics of patients at risk of VTE in the urology department was 1 for “(C18)” and 2 for “(C1)”, indicating that surgical patients aged 61–74 years were included. These patients mainly undergo bladder and prostate resection, urinary calculi, and other nonmalignant tumor operations [50,51,52]. In patients at risk of VTE, the risk of bleeding should also be assessed. Drug or mechanical prevention methods have been adopted for patients with VTE who are at low risk of bleeding. However, patients with VTE with a high risk of bleeding should adopt mechanical prevention methods, and the course of prevention is the same as that in patients with a high risk of VTE.

Comparative analysis of the results of different machine learning methods

In this study, regarding the sampling method, each data point had an 80% probability of being sampled to the training set and a 20% probability of being sampled to the test set. Therefore, the amount of data in the training and test sets of each model changed dynamically to increase the amount of data in different modeling and testing situations. Based on this amount of data, the RS, RF, SVM, and BPANN algorithms were used to establish a prediction model and calculate the accuracy. In terms of prediction accuracy, the RF, SVM, and BPANN models reached > 85% accuracy, with the SVM model reaching 92.6% accuracy and the BPANN model reaching 97.2% accuracy, which is excellent. The accuracy of the RS method was close to 80%, indicating poorer accuracy than that of the other three machine learning methods.

Models used for medical evaluation and decision-making should be transparent and easy to use. Medical staff should be able to compare the results of the decision rules and key characteristic variables based on their knowledge. A high degree of transparency and interpretability may increase the trust of medical staff in machine learning for building models. Additionally, the RS method can reveal the laws and potential causal relationships underlying the data. Moreover, RF, SVM and BPANN are “black-box models,” and the internal calculation model and decision rules cannot be explained, which is not easily accepted by clinical medical staff and managers. Such models cannot reveal the laws and potential causal relationships underlying the data. Therefore, as an interpretable machine learning method, the accuracy of the RS model is acceptable in the present study. The model may help urological medical staff explore the characteristics of patients at high risk of VTE, establish clear decision rules, quickly identify patients at high risk of VTE, conduct follow-up preventive measures accurately and in a timely manner, improve VTE evaluation accuracy, standardize prevention rates, reduce VTE incidence, and achieve standardized prevention and treatment.

Explainable machine learning models fall into two categories. The first category includes intrinsically interpretable machine learning models, such as logistic regression, decision trees, Bayesian models, and machine learning models based on decision rules. These models directly provide a certain degree of information through method features and are easily understood by decision makers. The second category is the postevent interpretation method of prediction models, which provides other supplementary explanatory information for most prediction models, such as the SHapley Additive exPlanation (SHAP) model.

Compared with non-rule-based classifiers, decision-rule-based methods all show a certain degree of performance degradation [53]. Clinical credibility and the application of prediction models largely rely on how well doctors understand and interpret models. Evaluation indicators include the accuracy of predictions and the complexity of interpreting the results. Sometimes, achieving high accuracy conflicts with the difficulty of explanation, necessitating a balance between the two.

In this study, RS is an interpretable machine learning method falling into the first category. RS resolves data ambiguities using set theory, providing clear decision rules.

The RS-based predictive model provides physicians with decision rules that improve model traceability and information content. The RS model has further advantages in that dominance-based decision rules condense a range of attribute values into each rule, thereby maximizing information density. Despite potential slight performance differences, the RS model is valued for its accessibility, simplicity, and ease of interpretation. Rule-based approaches have further benefits in that they clearly indicate the patient characteristics most relevant to VTE risk. The rules are simple and easily understood, particularly for high-risk patients, and can be enforced by medical personnel, improving the transparency and interpretability of the classification process and strengthening the accessibility of the model, thus enhancing its credibility.

Limitations

This study has several limitations. (i) The data were collected from inpatients in the Urology Department of a general hospital in Zhejiang Province from December 2019 to July 2022; this was a single-center retrospective study. Since single-center studies cannot represent the Chinese urology population, a multicenter study is needed to verify the generalizability of our findings.

Additionally, machine learning has several limitations. (ii) The accuracy of machine learning predictions depends on the data quality of the Caprini assessment. Laboratory examination data are relatively objective and have few influencing factors. However, observation and subjective evaluations by doctors, nurses or patient’s self-reported medical history and family history may lead to deviations between clinical data and actual conditions. Variability in observation information among different personnel can affect prediction models, which rely on historical data [54]. The development of a multicenter risk prediction model using Caprini assessment data from various institutions may mitigate some limitations and enhance the advantages of machine learning in big data mining [55].

(iii) Artificial intelligence (AI) can uncover subtle patterns or relationships hidden in traditional VTE risk modeling, providing more effective decision-making information. Since large amounts of information are being condensed, performance degradation may still occur during model development. Even so, machine learning remains important for understanding and assessing real-world VTE risk. Improved data quality can further enhance predictive performance [56].

(iv) Medical ethics, laws and regulations limit the widespread adoption of AI decision-making models in clinical practice. Even so, machine learning-based predictive models may improve the efficiency of medical staff. The acceptance of AI by medical professionals is also critical for the successful application of predictive models in clinical practice [57].

Finally, (v) the decision rules of different VTE risk levels identified in this study require further development for better interpretation. Further research is needed to establish accurate preventive measures based on these decision rules.