Machine learning models predict the risk of deep vein thrombosis in hospitalized patients for spontaneous intracephalic hemorrhage

Patients with natural ICH are at a higher risk of developing lower extremity DVT, mainly due to factors such as extended bed rest during hospitalization and limited mobility. Studies have shown that the incidence of DVT in hospitalized ICH patients is approximately 20-40%. [1]. These patients often fail to perform normal limb activity due to neurological disorders, which leads to blood stagnation and increases the risk of DVT. Currently, there are relatively few risk prediction models for the development of lower extremity DVT in patients with spontaneous ICH, and most of the available risk assessment models (e.g., Caprini, Padua, and Autor) have been developed based on a general hospitalized patient population and specific disease groups, such as patients with acute ischemic stroke. Although these models have some validity in predicting VTE risk, they have not been specifically optimized for specific clinical features of patients with intracerebral hemorrhage. The development of a DVT risk prediction model for patients with intracerebral hemorrhage will provide a more accurate assessment of the risk of patients developing blood clots during hospitalization, thus providing individualized clinical prevention. This helps improve the efficacy of thrombosis, reduce unnecessary interventions, reduce the risk of bleeding, and ultimately improve the long-term prognosis of the patient.

To predict the risk of lower extremity DVT in patients with spontaneous intracerebral hemorrhage, data from the second associated hospital at Fujian Medical University were retrospectively analyzed and screened for independent predictors. Based on these factors, we applied five machine learning algorithms (RF, LR, SVM, LGBM, and XGB) to construct a DVT prediction model. The AUCS for the LGBM, XGBoost, RF, SVM, and LR models were 0.998, 0.981, 0.839, 0.754, and 0.729, respectively. The performance of the model was evaluated by the ROC curve, the model was evaluated using the calibration curve and bootstrap method (b = 1000) for internal validation, the LGBM model showed optimal Brier score, DCA analysis evaluated the clinical values ​​of the model, and the LGBM showed good results with accuracy (0.9433) (0.9433). (0.9504), MCC (0.8863) and Auroc (0.9881) metrics. Therefore, LGBM was selected as the best model for predicting DVT after intracerebral hemorrhage in this study. The application of SHAP values ​​further elucidates the extent to which several key contributors affect the risk of DVT.

This study identified six independent predictors that were mechanically and clinically plausible. Intracerebral hemorrhages can trigger a systemic inflammatory response that activates leukocytes and releases cytokines and chemotactic factors, thereby promoting a thrombotic tendency. Inflammatory mediators such as tumor necrosis factor (TNF-α) and interleukins (IL-6, IL-8) activate endothelial cells, promote the secretion of coagulation substances, and increase blood clotting activity¹⁶. In particular, TNF-α and CD40 ligands (TNF family membrane glycoproteins on platelets) induce tissue factor (TF) expression in endothelial cells and monocytes via nuclear factor κB-dependent mechanisms.¹⁷. The protein C pathway is particularly sensitive to downregulation of IL-1β and TNF-α. IL-1β promotes thrombin accumulation by limiting the activation of protein C and inhibiting TM and EPCR transcription¹⁸. Furthermore, IL-1β causes shedding of EPCR and generates SEPCR that inhibits protein C activation¹⁹. TNF-α acts directly on endothelial cells, releasing histological plasminogen activator (TPA) and urokinase plasminogen activators, inhibiting fibrinolysis and increasing PAI-1, which removes fibrin removal.²⁰. Studies have shown that leukocyte activation and aggregation caused by inflammatory responses are one of the key factors in deep venous thrombosis.^{twenty one}. Thrombosis is not only dependent on the traditional Wilkou triangle (deceleration of blood flow, changes in blood composition, vascular damage), but is also closely related to the systemic inflammatory response.^{twenty one}. In particular, leukocyte aggregation and activation increase the risk of venous thrombosis. Especially in areas where venous blood flow is slow, such as the lower limbs,^{twenty two}. The intense systemic inflammatory response induced by heavy bleeding further increases the chance of thrombosis by activating the coagulation system and inhibiting fibrolysis.^{twenty three}. Furthermore, bleeding-induced blood pressure fluctuations and other circulatory stress responses can alter blood flow dynamics, thereby exacerbating the risk of DVT^{twenty four}.

Abnormal systolic blood pressure, particularly in the acute phase, is strongly associated with the risk of thrombosis^{twenty five}. High systolic blood pressure increases intravascular blood pressure, leading to hemodynamic changes. During acute intracerebral hemorrhage, patients often have hypertensive fluctuations. High systolic blood pressure increases turbulence in blood flow and increases contact with the vessel wall, making it more likely to cause thrombus formation.²⁶. High acute systolic blood pressure causes mechanical damage to the vascular endothelium, and endothelial damage is the basis of thrombosis. After endothelial cell damage, exposed collagen and other substances can activate platelets and contribute to thrombus formation. This injury is particularly pronounced in low flow areas of the lower extremity vein, increasing the risk of DVT. High systolic blood pressure is often slower, especially in the lower limb veins, and when blood flow is slow, coagulation factors and platelets in the blood tend to aggregate and form a blood clot²⁷. High systolic blood pressure also tends to accumulate in slow blood flow veins, which can ultimately lead to the formation of microtrombi, leading to DVT.²⁸. High systolic blood pressure can increase systemic vascular resistance, which can further affect venous revenues in the lower extremities, causing a blood pool and increasing the chance of venous thrombosis.²⁹. After intracerebral hemorrhage, the patient's activity is limited and abnormal systolic blood pressure slows further, making it very easy to form a clot. Blood clotting ability tends to increase with age, and the activity of coagulation factors in the blood of older adults may increase, but the activity of the fibrolytic system may decrease.³⁰. This change increases the chances of blood clotting and increases the risk of thrombosis. Among older adults, venous vessels can exhibit some degree of sclerosis and reduced elasticity, reducing the elasticity and contractility of the vessel wall, slowing blood flow, increasing the stagnation time of blood in the vein, increasing the chances of thrombus formation.³¹. Furthermore, patients with significant midline shifts usually require surgical intervention, and these patients often require long-term bedridden after surgery, with limited limb movement, especially in lower extremities where blood flow is poor. Similarly, this condition significantly increases the risk of lower limb thrombosis in postoperative patients due to occlusion of venous reversal.

Machine learning shows great potential for disease diagnosis and prognostic evaluation by recognizing the relationships between complex clinical data³². Common clinical conditions, past medical history, imaging characteristics, and biochemical properties can be incorporated, resulting in an “information gain” compared to models that use only these individual data categories.³³. The study can quantify the contribution of each variable to predicting outcomes via SHAP values, with positive SHAP values ​​showing negative values ​​representing risk-promoting and protective effects. This approach improves the interpretability of the model. This method not only increases model transparency, but also supports individual risk assessments. In recent years, relevant models have gradually achieved web-side deployment, and are expected to achieve real-time identification of high-risk groups in the future, promote early intervention and thus improve patient clinical prognosis.^34,35.

Our research has several strengths. We included not only clinical laboratory and laboratory data, but also imaging information to validate and cross-support patient diagnosis from multiple perspectives. Combining imaging data with clinical and laboratory data provides more accurate analysis and results. On the other hand, collecting data from the perspective of multiple variables provides a basis for subsequent big data analysis and mining of potential health patterns or disease prediction models. This not only provides accurate medical solutions to current patients, but also informs of treatment for similar patients in the future. By considering multiple potential confounding factors simultaneously and applying statistical methods to control and adjust them, bias can be reduced and accuracy of study findings can be ensured. For predictive analytics (for example, using machine learning or regression modeling), these multivariate data can help you build more robust predictive models. Furthermore, by combining machine learning with SHAP descriptive modeling, the percentage of contributions of each factor within the model can be more clear.

Furthermore, this study still has some limitations. For example, data were collected at admission and was unable to analyze changes in outcomes or prognosis for patients undergoing treatment. Longitudinal data (e.g., follow-up data) provide more comprehensive information and helps to better understand disease progression, treatment outcomes, and long-term patient health trends. Furthermore, this study did not include patients with intracerebral hemorrhage with small amounts of bleeding that are usually seen by the Department of Neurology. Therefore, there are several biases in the selection of study populations, which can affect the generalizability of the results, which limits the applicability of the conclusions to other groups. Although every effort was made in the present study to include various variables related to DVT, due to the large number of missing values, some variables were ultimately excluded from the analysis (e.g., patients with a high NIHSS score are more likely to experience severe neurological defects, including motor functions with impaired limbs, which significantly increases the risk of deep thrombosis of decline). The SHAP approach has the right intuition to explain high risk factors for DVT in ICH patients and helps clinicians understand important predictors, but improving model performance and generalizability is required to better respond to a diverse range of clinical settings. Future research should prioritize optimization of model utilities and developing user-friendly interfaces that allow clinicians to access risk prediction in real time.