Abstract
Background
Bleeding is a clinically important complication during percutaneous left ventricular assist devices (pLVAD) support, yet its determinants remain unclear.
Objectives
The objective of the study was to identify predictors of bleeding after pLVAD implantation using a nationwide Japanese registry, J-PVAD (Japanese registry for Percutaneous Ventricular Assist Device).
Methods
A total of 5,608 patients with complete follow-up were analyzed and stratified into acute coronary syndrome (ACS) and non-ACS strata. The primary outcome measure was any bleeding within 30 days. Multivariable Cox proportional hazards model were used to identify predictors within each stratum, with competing-risk analyses using Fine-Gray models performed for sensitivity. Machine learning (ML) with Shapley Additive exPlanations analysis was conducted as a complementary approach.
Results
The cohort included 2,667 ACS and 2,941 non-ACS patients. The 30-day incidence of bleeding was similar between the ACS and Non-ACS strata (25.7% vs 26.2%; P = 0.951), with most events occurring within several days after implantation. Additional mechanical circulatory support (MCS) was the strongest factor associated with bleeding in both strata (ACS: adjusted HR: 1.60; non-ACS: adjusted HR 1.75; both P < 0.001). Competing-risk analyses accounting for early mortality yielded consistent results (ACS: subdistribution HR: 1.52; non-ACS: subdistribution HR: 1.74). ML analyses consistently identified additional MCS use as the dominant contributor to bleeding. Shapley Additive exPlanations–based clustering suggested 2 high-risk phenotypes characterized by procedural complexity including additional MCS use and inflammatory-hemodynamic instability.
Conclusions
Bleeding after pLVAD support is common and occurs early, with comparable incidence across ACS and non-ACS strata. Additional MCS use is the dominant determinant of bleeding risk. ML-based phenotyping provides complementary insights into heterogeneous bleeding mechanisms.
Key words: artificial intelligence, bleeding complication, Impella, machine learning, mechanical circulatory support, percutaneous left ventricular assist device
Central Illustration
Cardiogenic shock remains a highly lethal condition despite advances in pharmacologic and mechanical therapies, with contemporary in-hospital mortality rates exceeding 40%.1,2 Mechanical circulatory support (MCS) is frequently required to stabilize hemodynamics and preserve end-organ perfusion. Percutaneous left ventricular assist devices (pLVADs), such as Impella, have become an increasingly utilized option for patients with drug-refractory acute heart failure and cardiogenic shock, providing active left ventricular unloading and forward circulatory support.
Despite their hemodynamic benefits, pLVADs are associated with clinically significant complications, among which bleeding is one of the most frequent and consequential adverse events. Bleeding during pLVAD support may result from large-bore arterial access, systemic anticoagulation, heparinized purge solutions, and device-related alterations in hemostasis.3, 4, 5 Bleeding complications have been associated with adverse clinical outcomes, including increased mortality.6 However, reported bleeding incidence varies substantially across studies, reflecting heterogeneity in patient populations, clinical practice patterns, and event definitions.2, 3, 4, 5, 6, 7, 8, 9 Thus, contemporary real-world data are needed to better characterize the incidence, timing, and factors associated with bleeding after pLVAD implantation.
Several clinical and procedural factors may modify bleeding risk in patients undergoing pLVAD support. Patients with acute coronary syndrome (ACS) frequently receive dual antiplatelet therapy and undergo emergent revascularization, whereas non-ACS presentations encompass more heterogeneous etiologies with differing inflammatory and hemodynamic profiles.5 In addition, venoarterial extracorporeal membrane oxygenation (VA-ECMO) is often used in patients with severe cardiogenic shock, and combined support with VA-ECMO and pLVAD may improve hemodynamics and confer a potential survival benefit in selected patients with acute myocardial infarction and severe cardiogenic shock, including those requiring cardiopulmonary resuscitation.10 However, escalation to multiple MCS devices may increase procedural complexity and bleeding risk. Whether bleeding risk differs between ACS and non-ACS presentations, and how additional MCS use is associated with bleeding outcomes, remain incompletely characterized in large contemporary pLVAD cohorts.
Using data from the nationwide J-PVAD (Japanese registry for Percutaneous Ventricular Assist Device), we sought to provide a comprehensive real-world assessment of bleeding events following pLVAD implantation. Specifically, we aimed to: 1) characterize the incidence and early timing of bleeding in a national cohort of pLVAD implantation; 2) compare bleeding risk between ACS and non-ACS presentations; and 3) identify clinical and procedural predictors of bleeding using multivariable models, with complementary exploratory machine learning (ML) analysis.
Methods
Study design
The J-PVAD is a multicenter, prospective, nationwide observational registry that captures nearly all pLVAD implantations performed in Japan. In Japan, pLVAD, Impella (Abiomed), for patients with cardiogenic shock was approved in 2017, and all patients requiring pLVAD have been consequently registered.3,4 The study was registered with the University hospital Medical Information Network Clinical Trials Registry in Japan (UMIN000033603). The registry complied with the Declaration of Helsinki, and was approved by the Central Institutional Review Board of Osaka University (Approval No. 17232) and institutional review boards at each participating centers. Written informed consent was waived due to the retrospective nature and opt-out of this study. Individual patient data including baseline characteristics (documented coronary risk factors, heart failure, pulmonary hypertension, and other comorbidities), laboratory data, concomitant procedures, and clinical outcomes were entered by investigators at each center into a centralized electronic database. Laboratory variables were derived from the most recent laboratory assessment before pLVAD implantation.
For the present analysis, we included all patients who underwent successful pLVAD implantation between January 2020 and December 2022. Patients without follow-up information or without bleeding data were excluded. Follow-up data were collected at hospital discharge and at 30 days after pLVAD implantation by participating sites. The primary indication for pLVAD use was classified by treating physician as ACS (ST-segment elevation myocardial infarction [STEMI], non-STEMI, or unstable angina pectoris), or non-ACS (including decompensated heart failure, valvular heart disease, arrhythmia, myocarditis, cardiac arrest, chronic coronary syndrome, and other presentations). These categories were used to define the ACS and non-ACS strata. Additional MCS was defined as the use of VA-ECMO or other temporary circulatory assist devices during the index episode of pLVAD support. Devices recorded as being used only after pLVAD removal were not counted as additional MCS exposure. The registry captured categorical timing relative to pLVAD support (before pLVAD use, during pLVAD support, or after pLVAD removal), but exact initiation and discontinuation times were not systematically recorded. Therefore, additional MCS use was analyzed as a time-fixed, episode-level variable. Additional MCS was treated as an independent variable because our previous observational studies have demonstrated higher bleeding rates in patients supported with combined MCS modalities.11
The primary outcome was any bleeding within 30 days after pLVAD implantation. The secondary outcome was major bleeding, defined as a composite of cerebral bleeding, bleeding requiring blood transfusion, or protocol-defined severe bleeding (defined as life-threatening, requiring hospitalization or prolongation of treatment, or disability-causing events).
Statistical analysis
Categorical variables were expressed as numbers and percentages and compared with the chi-square test or Fisher exact test as appropriate. Continuous variables were expressed as median with IQR or mean and SD and were compared using the Wilcoxon rank sum test or Student t-test based on their distribution. Cumulative incidences were estimated using the Kaplan-Meier method, and differences among each group were assessed using a log-rank test. Primary analyses were conducted using Cox proportional hazards models stratified by ACS and non-ACS presentations. The 15 risk-adjusting covariates were selected based on clinical relevance and prior literature; covariates included age ≥75 years, female sex, body mass index <25 kg/m2, hypertension, diabetes, dyslipidemia, chronic kidney disease, smoking status, history of heart failure, history of coronary artery disease, prior stroke or transient ischemic attack, out-of-hospital cardiac arrest, use of vasopressor or inotrope, lactate >5 mmol/L, and use of additional MCS. The proportional hazards assumption was assessed graphically using log-log survival plots and Schoenfeld residuals. When evidence of nonproportionality was observed for additional MCS use, the primary exposure of interest, piecewise Cox models were fitted as sensitivity analyses, allowing the association between additional MCS use and bleeding to differ between 0 to 3 days and >3 to 30 days after pLVAD implantation. Missingness rates for variables included in the primary and sensitivity analyses were summarized overall and according to ACS and non-ACS strata (Supplemental Table 1). Differences in missingness between strata were assessed using the chi-square test or Fisher exact test, as appropriate. Primary Cox analyses were performed as complete-case analyses. Sensitivity analyses using multiple imputation were performed to assess the influence of missing covariate data.
As a sensitivity analysis for missing covariate data, multivariable Cox models were repeated using multiple imputation by chained equations. Continuous variables were imputed using predictive mean matching, and binary variables were imputed using logistic regression. Fifty imputed data sets were generated, and estimates were combined using Rubin rules. The imputation model included variables used in the primary Cox models, bleeding status, follow-up time, and ACS stratum; bleeding status, follow-up time, and ACS stratum were included as predictors but were not imputed.
As a sensitivity analysis addressing potential information loss due to dichotomization of continuous covariates, age, body mass index, and lactate were modeled as continuous variables using natural cubic splines with 3 degrees of freedom in the multivariable Cox models. These analyses were performed separately within the ACS and non-ACS strata using the same complete-case population as the primary models, and the results were compared with those of the primary models using dichotomized covariates.
To evaluate the robustness of the primary analyses to the competing risk of early death, competing-risk analyses were performed. Death occurring within 30 days after pLVAD implantation without prior bleeding was treated as a competing event. Subdistribution HRs (SHRs) were calculated using the Fine-Gray proportional subdistribution hazards model. Results were compared with those from the primary Cox models. Statistical analyses were performed using JMP (version 18.2.2; SAS institute Japan) and R (version 4.3.3; R Foundation for Statistical Computing) software. All tests were 2-sided, and P values <0.05 were considered statistically significant.
Development of prediction models using machine learning
As complementary analyses, we developed a ML model to explore nonlinear associations and complex interactions among predictors. To minimize bias from missingness, only variables with <30% missing data were included. In the ML model, all explanatory variables were imputed for missing values before use in the model development. Missing values were imputed using multivariate feature imputation, which is an imputation method that uses all other variables as well as those with missing values. The original data set was randomly split into training and validation data sets in a 1:1 ratio to maintain consistent rates of bleeding events between the 2 data sets. The hyperparameters of all models were tuned using Bayesian optimization with stratified 5-fold cross-validation, a widely recognized automated technique for hyperparameter tuning in artificial intelligence models.12 Prediction model was developed using the Light Gradient-Boosting Machine algorithm.13 Model performance was assessed in the validation data set using the area under the receiver operating characteristic curve (AUC), sensitivity, specificity, calibration slope, calibration intercept, and the Brier score.
To improve interpretability, Shapley Additive Explanations (SHAP) were used to quantify the contribution of individual variables to the predicted bleeding risk.14 The absolute SHAP values of the relevant variables indicate the magnitude of their influence on the prediction outcome. As complementary analyses to evaluate model robustness and potential overfitting, we performed calibration analyses and SHAP-based backward elimination. Calibration was assessed by comparing observed and predicted bleeding risks across deciles of predicted probability. For SHAP-based backward elimination, variables were sequentially removed according to their mean absolute SHAP values, and model performance was reassessed after each iteration. A parsimonious model was defined as the smallest variable set whose validation AUC remained within 0.01 of that of the full model. Detailed methods are provided in the Supplemental Appendix.
Hierarchical clustering based on the SHAP values was used to identify prognostic phenotypes, and the patients were classified into 3 clusters. Based on the incidence of bleedings from highest to lowest, the clusters were labeled as clusters 1 to 3. A dendrogram was created using the Euclidean method and appropriate cluster divisions were created. The number of clusters was determined based on dendrogram structure and clinical interpretability. For each cluster, the top 20 features based on mean absolute SHAP values were evaluated. Time-to event curves for bleeding were generated to assess the prognoses of each cluster. Detailed ML methods, including preprocessing steps, model parameters, and SHAP procedures, are provided in the Supplemental Appendix.
Results
From January 2020 to December 2022, 5,721 patients across 239 participating sites were registered in the J-PVAD registry. After exclusion of 113 patients without follow-up or bleeding data, 5,608 patients were included in this analysis (Figure 1). Among them, 2,667 patients (47.6%) were assigned to the ACS stratum, and 2,941 patients (52.4%) to the non-ACS stratum. The median follow-up duration was 34 days (IQR: 14-52 days).
Figure 1.
Patient Flow Diagram
From January 2020 to December 2022, 5,721 consecutive Japanese patients who received pLVAD support were registered. 113 patients had no follow-up or bleeding data, thus finally 5,608 patients were included in this analysis. ACS = acute coronary syndrome.
The mean age was 66.2 ± 14.5 years, and 1,290 patients (23.0%) were females (Table 1). Patients in the ACS stratum were older (69.8 ± 11.7 vs 62.9 ± 15.9 years, respectively; P < 0.001) and more frequently had cardiovascular risk factors including smoking, diabetes, hypertension, and dyslipidemia (Table 1). The non-ACS stratum had a higher prevalence of prior heart failure, coronary artery disease, and valvular heart disease, and lower left ventricular ejection fraction before pLVAD implantation (32.0% ± 17.6% vs 27.5% ± 14.9%; P < 0.001). In the ACS stratum, patients were diagnosed as STEMI, non-STEMI, unstable angina pectoris, and unknown in 79.8%, 14.5%, 3.8%, and 1.9%, respectively. Heart failure was the most frequent indication of the pLVAD use in the non-ACS stratum (774 patients, 26.3%). Impella CP was the predominant device in both strata (ACS: 93.7%; non-ACS: 87.0%). Femoral artery access was used in more than 90% of cases (ACS: 96.1%; non-ACS: 90.3%). Additional MCS, primarily VA-ECMO, was more commonly used in the non-ACS stratum (14.3% vs 17.0%; P = 0.005). The recorded timing categories of additional MCS relative to pLVAD support are summarized in Supplemental Table 2. Total support duration was longer in non-ACS stratum (3.8 [1.8-7.3] days vs 4.9 [2.1-8.4] days; P < 0.001).
Table 1.
Patient Characteristics in Each Stratum
| All (N = 5,608) | ACS Stratum (n = 2,667) | Non-ACS Stratum (n = 2,941) | P Value | |
|---|---|---|---|---|
| Age (y)a | 66.2 ± 14.5 (n = 5,608) | 69.8 ± 11.7 (n = 2,667) | 62.9 ± 15.9 (n = 2,941) | <0.001 |
| Femalea | 1,290 (23.0%) (n = 5,608) | 532 (19.9%) (n = 2,667) | 758 (25.8%) (n = 2,941) | <0.001 |
| Body mass index (kg/m2)a | 23.6 ± 4.4 (n = 5,358) | 23.8 ± 3.9 (n = 2,529) | 23.5 ± 4.7 (n = 2,829) | <0.001 |
| Out-of-hospital cardiac arresta | 1,077/5,574 (19.3%) | 399/2,655 (15.0%) | 678/2,919 (23.2%) | <0.001 |
| In-hospital cardiac arrest | 1,590/5,606 (28.4%) | 713/2,667 (26.7%) | 877/2,939 (29.8%) | 0.010 |
| Vasopressor/inotrope usea | 4,224/5,605 (75.4%) | 1,873/2,667 (70.2%) | 2,351/2,938 (80.0%) | <0.001 |
| Arrival to pLVAD insertion (min) | 221 [99-1,837] (n = 5,540) | 139 [77-451] (n = 2,635) | 599 [140-4,458] (n = 2,905) | <0.001 |
| Shock to pLVAD insertion (min) | 142 [72-340] (n = 3,417) | 114 [54-225] (n = 1,742) | 180 [97-530] (n = 1,675) | <0.001 |
| Smoking (former and current)a | 2,950/4,804 (61.4%) | 1,501/2,316 (64.8%) | 1,449/2,488 (58.2%) | <0.001 |
| Diabetesa | 2,188/5,388 (40.6%) | 1,149/2,573 (44.7%) | 1,039/2,815 (36.9%) | <0.001 |
| Hypertensiona | 3,254/5,337 (61.0%) | 1,738/2,540 (68.4%) | 1,516/2,797 (54.2%) | <0.001 |
| Pulmonary hypertension | 230/5,214 (4.4%) | 62/2,491 (2.5%) | 168/2,723 (6.2%) | <0.001 |
| Dyslipidemia | 2,418/5,331 (45.4%) | 1,320/2,543 (51.9%) | 1,098/2,788 (39.4%) | <0.001 |
| Chronic kidney disease | 1,770/5,342 (33.1%) | 765/2,536 (30.2%) | 1,005/2,806 (35.8%) | <0.001 |
| Dialysis | 340/5,337 (6.4%) | 114/2,535 (4.5%) | 226/2,802 (8.1%) | <0.001 |
| eGFR <30 or dialysis | 1,500/5,316 (28.2%) | 589/2,546 (23.1%) | 911/2,770 (32.9%) | <0.001 |
| Prior stroke/TIAa | 451/5,391 (8.4%) | 235/2,577 (9.1%) | 216/2,814 (7.7%) | 0.056 |
| Previous coronary artery diseasea | 1,678/5,334 (31.5%) | 686/2,543 (27.0%) | 992/2,791 (35.5%) | <0.001 |
| Myocardial infarction | 1,128/5,356 (21.1%) | 480/2,568 (18.7%) | 648/2,788 (23.2%) | <0.001 |
| Angina pectoris | 879/5,278 (16.7%) | 368/2,528 (14.6%) | 511/2,750 (18.6%) | <0.001 |
| History of heart failurea | 1,554/5,354 (29.0%) | 391/2,560 (15.3%) | 1,163/2,794 (41.6%) | <0.001 |
| Valvular heart disease | 546/5,363 (10.2%) | 71/2,564 (2.8%) | 475/2,799 (17.0%) | <0.001 |
| LVEF before pLVAD insertion | 29.4 ± 16.3 (n = 3,187) | 32.0 ± 17.6 (n = 1,335) | 27.5 ± 14.9 (n = 1,852) | <0.001 |
| Systolic blood pressure (mmHg) | 90.8 ± 35.6 (n = 5,599) | 93.7 ± 35.9 (n = 2,666) | 88.2 ± 35.2 (n = 2,933) | <0.001 |
| Diastolic blood pressure (mmHg) | 58.9 ± 25.2 (n = 5,599) | 60.8 ± 25.3 (n = 2,666) | 57.1 ± 24.9 (n = 2,933) | <0.001 |
| Heart rate | 89.3 ± 32.9 (n = 5,601) | 89.0 ± 31.4 (n = 2,665) | 89.6 ± 34.2 (n = 2,936) | 0.498 |
| eGFR (mL/min/1.73 m2) | 45.6 ± 24.0 (n = 5,497) | 46.9 ± 22.2 (n = 2,636) | 44.3 ± 25.4 (n = 2,861) | <0.001 |
| Lactate (mmol/L)a | 4.2 [2.1-8.9] (n = 4,137) | 4.2 [2.3−8.2] (n = 1,964) | 4.2 [1.9–9.7] (n = 2,173) | 0.800 |
| CRP (mg/dL) | 0.9 [0.1-5.4] (n = 5,244) | 0.4 [0.1–3.2] (n = 2,510) | 1.7 [0.2–7.0] (n = 2,734) | <0.001 |
| Type of pLVAD | (n = 5,608) | (n = 2,667) | (n = 2,941) | <0.001 |
| Impella 2.5 | 216 (3.9%) | 85 (3.2%) | 131 (4.5%) | |
| Impella CP | 5,069 (90.4%) | 2,503 (93.9%) | 2,566 (87.3%) | |
| Impella 5.0 | 248 (4.4%) | 57 (2.1%) | 191 (6.5%) | |
| Impella 5.5 | 75 (1.3%) | 22 (0.8%) | 53 (1.8%) | |
| Access site of pLVAD | (n = 5,605) | (n = 2,666) | (n = 2,939) | <0.001 |
| Femoral artery | 5,210 (93.0%) | 2,559 (96.0%) | 2,651 (90.2%) | |
| Subclavian artery | 365 (6.5%) | 91 (3.4%) | 274 (9.3%) | |
| Others | 30 (0.5%) | 16 (0.6%) | 14 (0.5%) | |
| PCI during hospitalization | 3,568/5,606 (63.7%) | 2,401/2,667 (90.0%) | 1,167/2,939 (39.7%) | <0.001 |
| Use of right heart catheter | 3,557/5,582 (63.7%) | 1,626/2,656 (61.2%) | 1,931/2,926 (66.0%) | <0.001 |
| Additional MCS usea | 880/5,608 (15.7%) | 380/2,667 (14.3%) | 500/2,941 (17.0%) | 0.005 |
| Total support duration (days) | 4.4 [1.9-7.9] (n = 2,938) | 3.8 [1.8-7.3] (n = 1,355) | 4.9 [2.1-8.4] (n = 1,583) | <0.001 |
ACS = acute coronary syndrome; CRP = C-reactive protein; eGFR = estimated glomerular filtration rate; LVEF = left ventricular ejection fraction; MCS = mechanical circulatory support; PCI= percutaneous coronary intervention; pLVAD = percutaneous left ventricular assist device; TIA = transient ischemic attack.
Variables included in the multivariable analysis.
The 30-day incidence of any bleeding was 26.0% in the overall cohort, and the incidence of major bleeding was 24.0%. Most bleeding events occurred within a few days after the pLVAD insertion (The median interval of any bleeding was 1 [0-3] day). Bleeding rates did not differ significantly between ACS and non-ACS strata (any bleeding: ACS stratum 655 cases [25.7%] vs non-ACS stratum 719 cases [26.2%]; P = 0.951, and major bleeding: ACS stratum 600 cases [23.5%] vs non-ACS stratum 672 cases [24.5%]; P = 0.731) (Figure 2). The early clustering of bleeding events was consistent across both strata. Across strata, patients who received additional MCS had substantially higher bleeding rates than those supported with pLVAD alone (ACS stratum: 131 [37.5%] vs 524 [23.8%] at 30 days; P < 0.001, and non-ACS stratum: 184 [39.6%] vs 535 [23.4%]; P < 0.001, respectively) (Figure 3). This pattern remained consistent for major bleeding events. To further characterize the non-ACS population, registry-recorded clinical presentations within the non-ACS stratum are summarized in Supplemental Table 3. The non-ACS stratum was heterogeneous and included decompensated heart failure, valvular heart disease, arrhythmia, myocarditis, cardiac arrest, chronic coronary syndrome, and other presentations.
Figure 2.
Cumulative Incidence of Bleeding in ACS and Non-ACS Strata
The cumulative incidence of any and major bleeding did not differ significantly between the ACS and non-ACS stratum. Abbreviation as in Figure 1.
Figure 3.
Cumulative Incidence of Bleeding According to Additional MCS Use
The cumulative incidence of any bleeding was higher in the additional MCS support group than pLVAD-alone group. MCS = mechanical circulatory support; pLVAD = percutaneous left ventricular assist device; other abbreviation as in Figure 1.
Multivariable Cox models were constructed separately within each stratum (Table 2). Across both strata, the strongest factor associated with bleeding was the use of additional MCS: ACS stratum: HR 1.60 (95% CI: 1.24-2.06), P < 0.001, non-ACS stratum: HR 1.75 (95% CI: 1.41-2.18), P < 0.001. Additional MCS use was associated with higher 30-day bleeding risk, corresponding to Kaplan-Meier–estimated absolute risk differences of 13.7 percentage points in the ACS stratum and 16.2 percentage points in the non-ACS stratum, alongside adjusted HRs of 1.60 and 1.75, respectively (Supplemental Table 4). Other covariates including age, diabetes, hypertension, chronic kidney disease, body mass index, and coronary artery disease were not consistently associated with bleeding in adjusted analyses. The inclusion of additional clinically relevant covariates (estimated glomerular filtration rate <30 mL/min/1.73 m2 or dialysis, systolic blood pressure <100 mmHg, right heart catheterization, and percutaneous coronary intervention) did not materially alter the results (Supplemental Table 5). Sensitivity analyses using multiple imputation for missing covariate data yielded results consistent with the complete-case analyses. Additional MCS use remained significantly associated with bleeding in both the ACS stratum (HR: 1.58; 95% CI: 1.30-1.92; P < 0.001) and the non-ACS stratum (HR: 1.70; 95% CI: 1.43-2.02; P < 0.001) (Supplemental Table 6). Sensitivity analyses modeling age, body mass index, and lactate using natural cubic splines yielded results consistent with the primary models. Additional MCS use remained significantly associated with bleeding in both the ACS stratum (HR: 1.63; 95% CI: 1.26-2.09; P < 0.001) and the non-ACS stratum (HR: 1.75; 95% CI: 1.41-2.18; P < 0.001) (Supplemental Table 7). Although the proportional hazards assumption was not fully satisfied for additional MCS use, piecewise Cox models demonstrated that additional MCS remained significantly associated with bleeding during both the early (0-3 days) and later (>3-30 days) periods after pLVAD implantation in both strata (Supplemental Table 8). Competing-risk analyses accounting for 30-day mortality yielded results that were directionally and quantitatively consistent with the primary Cox models. In Fine-Gray models, additional MCS use remained significantly associated with an increased cumulative incidence of bleeding in both strata (ACS: SHR: 1.52; 95% CI: 1.18-1.94; non-ACS: SHR: 1.74; 95% CI: 1.42-2.14]) (Supplemental Table 9). The direction and magnitude of association were similar to those observed in the primary analyses.
Table 2.
Potential Risk Factors for Any Bleeding in Each Stratum
| ACS Stratum |
Non-ACS Stratum |
|||
|---|---|---|---|---|
| HR (95% CI) | P Value | HR (95% CI) | P Value | |
| Additional MCS use | 1.60 (1.24-2.06) | <0.001 | 1.75 (1.41-2.18) | <0.001 |
| Age ≥75 year | 0.92 (0.73-1.14) | 0.44 | 0.98 (0.77-1.25) | 0.81 |
| Female | 1.01 (0.77-1.32) | 0.96 | 0.90 (0.70-1.14) | 0.37 |
| Body mass index <25 kg/m2 | 0.88 (0.72-1.09) | 0.25 | 1.03 (0.83-1.27) | 0.81 |
| Smoking (former and current) | 0.88 (0.70-1.11) | 0.28 | 0.85 (0.69-1.05) | 0.14 |
| Dyslipidemia | 1.00 (0.81-1.24) | 0.97 | 1.19 (0.95-1.50) | 0.12 |
| Hypertension | 1.33 (1.06-1.68) | 0.02 | 1.15 (0.92-1.43) | 0.22 |
| Diabetes | 0.81 (0.66-1.01) | 0.06 | 0.89 (0.71-1.12) | 0.31 |
| Coronary artery disease | 1.03 (0.79-1.34) | 0.84 | 0.90 (0.70-1.15) | 0.40 |
| History of heart failure | 0.87 (0.61-1.23) | 0.43 | 0.81 (0.64-1.01) | 0.06 |
| Chronic kidney disease | 1.14 (0.90-1.43) | 0.28 | 1.17 (0.93-1.48) | 0.18 |
| Prior stroke/transient ischemia attack | 0.85 (0.58-1.23) | 0.38 | 1.05 (0.73-1.51) | 0.79 |
| Out-of-hospital cardiac arrest | 0.99 (0.73-1.35) | 0.96 | 1.26 (0.99-1.61) | 0.07 |
| Vasopressor/inotrope use | 1.01 (0.81-1.27) | 0.91 | 1.18 (0.90-1.55) | 0.22 |
| Lactate >5 mmol/L | 1.25 (1.01-1.55) | 0.04 | 1.09 (0.89-1.35) | 0.40 |
Abbreviations as in Table 1.
In complementary analyses, a ML model using Light Gradient-Boosting Machine was developed in the training cohort to predict 30-day any bleeding. The model demonstrated moderate discrimination, with an AUC of 0.72 in the training cohort and 0.60 in the validation cohort (Supplemental Figure 1). Calibration analysis demonstrated adequate agreement between predicted and observed risk in the validation cohort (calibration slope 0.96, intercept −0.04, Brier score 0.181) (Supplemental Figure 2). SHAP-based backward elimination showed that model discrimination remained largely unchanged despite substantial variable reduction, and additional MCS use consistently remained among the highest-ranked predictors (Supplemental Figure 3, Supplemental Table 10). SHAP analysis identified variables with the greatest contribution to bleeding prediction. The most influential variables for bleeding included additional MCS use, in-hospital cardiopulmonary arrest, C-reactive protein (CRP) level, right heart catheterization use, and heart rate before pLVAD insertion (Figure 4). Hierarchical clustering based on SHAP value profiles identified 3 distinct clusters with differential bleeding risk (Figure 5A and 5B). The Cluster 1 was characterized by in-hospital cardiopulmonary arrest and elevated CRP levels and showed a high bleeding incidence. Cluster 2 was characterized by frequent additional MCS use and procedural complexity and also demonstrated a high cumulative 30-day incidence of bleeding. In contrast, cluster 3, characterized primarily by elevated preprocedural heart rate but without strong mechanical or inflammatory contributors exhibited a comparatively lower incidence of bleeding than the former 2 clusters (Figure 5C).
Figure 4.
SHAP Summary Plot for Prediction of 30-Day Bleeding
The SHAP summary plots show the 20 most important variables for predicting bleeding complications. A higher positive SHAP value associated with higher incidence of bleeding, whereas a higher negative SHAP value indicates lower incidence of bleeding. In the SHAP summary plots, the length of each bar represents the mean absolute SHAP value of the 20 most important variables in the model. The feature ranking (y-axis) indicates the variables that are important for the prediction outcome and the SHAP value (x-axis) is a unifying indicator of the influence of each variable. Each variable is indicated by differently colored dots for the attribution of all patients to the predicted outcome. For the numerical variables, low actual values are represented by blue dots, whereas high actual values are represented by red dots. For categorical variables processed with one-hot encoding, blue and red dots represent 0 and 1, respectively. ALT = alanine aminotransferase; AST = aspartate aminotransferase; BP = blood pressure; CK = creatine kinase; CPA = cardiopulmonary arrest; CRP = C-reactive protein; eGFR = estimated glomerular filtration rate; HR = heart rate; LDH = lactate dehydrogenase; MI, = myocardial infarction; SHAP = Shapley Additive exPlanations; T-Bil = total bilirubin; other abbreviations as in Figure 2.
Figure 5.
SHAP-Based Clustering and Characteristics of Each Cluster
(a) Cluster map of SHAP values from the model predicting any bleeding in 30 days. (b) Radar charts depict the standardized or proportion of observed values for each cluster derived from SHAP-based clustering. (c) Cumulative distribution functions of clusters based on standardized observed values for bleeding complications. Clusters 1 to 3 were defined using hierarchical clustering based on standardized values. Alb = albumin; other abbreviations as in Figures 2 and 4.
In descriptive landmark analyses at days 1, 3, and 7, patients who experienced bleeding before each landmark had higher subsequent mortality through day 30 than those without bleeding. In overall adjusted landmark Cox models, bleeding before the landmark was associated with higher mortality through day 30 at day 1 (HR: 1.34; 95% CI: 1.13-1.60), day 3 (HR: 1.41; 95% CI: 1.18-1.69), and day 7 (HR: 1.63; 95% CI: 1.34-1.99) (Supplemental Table 11).
Discussion
Bleeding is a frequent and clinically important complication during pLVAD support. In this nationwide registry analysis of more than 5,600 patients, we observed 3 key findings. First, about one-quarter of patients with pLVAD experienced bleeding within 30 days, with events occurring predominantly early after device implantation. Second, bleeding incidence did not differ between ACS and non-ACS presentations. Third, the use of additional MCS with pLVAD, most commonly VA-ECMO, was the strongest and most consistent determinant of bleeding across strata. These findings were consistent in competing-risk analyses accounting for early mortality and were supported by complementary ML analyses (Central Illustration).
Central Illustration.
Real-World Bleeding Burden After pLVAD Support
In a nationwide J-PVAD (Japanese registry for Percutaneous Ventricular Assist Device), including 5,608 patients undergoing percutaneous left ventricular assist device (pLVAD), we evaluated predictors of bleeding. Bleeding occurred in approximately one-quarter of patients and did not differ between acute coronary syndrome (ACS) and non-ACS presentations (25.7% vs 26.2%; P = 0.951). Additional mechanical circulatory support (MCS) was the strongest determinant of bleeding risk, and its association was robust to competing-risk adjustment. Complementary machine learning with Shapley Additive exPlanations analyses suggested 2 high-risk phenotypes characterized by procedural complexity including additional MCS use and inflammatory–hemodynamic instability. CPA = cardiopulmonary arrest; CRP = C-reactive protein; SHAP = Shapley Additive exPlanations.
MCS is essential to maintain the hemodynamics in selected patients with cardiogenic shock; however, MCS use is associated with some adverse events including thromboembolism, limb ischemia, hemolysis, kidney failure, and bleeding events.3, 4, 5, 6, 7, 8, 9, 10 Bleeding complications during temporary MCS have consistently been associated with worse clinical outcomes and increased mortality.15 Intra-aortic balloon pumping (IABP), pLVAD, and ECMO are commonly used as MCS. pLVAD provides superior hemodynamic support than IABP and greater unloading of left ventricle, so its use has been increasing over time. Although data on bleeding complications related to pLVAD use are limited, it has been reported that the most common cause of morbidity and mortality during pLVAD support are bleeding and thrombotic complications.16 A meta-analysis comparing pLVAD with IABP reported that the bleeding risk of pLVAD use was 2.5-fold higher than IABP.15 Prior J-PVAD reports have provided important disease-specific insights into patients with acute myocardial infarction–related cardiogenic shock, including the prognostic implications of bleeding complications.6 The present study addresses a distinct but complementary question by evaluating the incidence, timing, and factors associated with bleeding across a broader real-world pLVAD population, including heterogeneous non-ACS presentations. The non-ACS stratum included heterogeneous registry-recorded presentations, with decompensated heart failure and myocarditis representing clinically important subgroups, but also with substantial proportions of cardiac arrest and other or unclassified presentations. Because these categories were not centrally adjudicated etiologies of cardiogenic shock and some contained limited numbers of patients and events, disease-specific risk modeling was not considered reliable. Accordingly, the primary ACS vs non-ACS framework was retained, whereas non-ACS subgroup findings were presented descriptively.
The overall 30-day bleeding incidence of approximately 26% in this study is higher than many rates reported in single-center cohorts and device trials, where patient selection and bleeding definitions may differ.6 Previous reports have described bleeding rates ranging from <10% in selected high-risk percutaneous coronary intervention populations to >40% in cardiogenic shock populations.17 Our study, leveraging a nationwide, unselected cohort, provides one of the most comprehensive assessments of bleeding risk associated with contemporary pLVAD use in real-world practice. Japan’s experience, characterized by an aging population, a high burden of critical illness, and frequent escalation of circulatory support contributes to the substantial bleeding burden observed. These findings highlight the substantial bleeding burden associated with pLVAD use outside of controlled trial settings. Descriptive landmark analyses further supported the clinical relevance of bleeding events by showing that bleeding before prespecified landmarks was associated with higher subsequent mortality through day 30, although these associations should not be interpreted causally.
Bleeding during temporary MCS is likely multifactorial. pLVAD needs large-bore arterial access (12- to 21-F), and procedural complexity, particularly multiple-vessel cannulation, is required when additional MCS is applied. Systemic anticoagulation is also required both for device thrombosis prevention and for the heparinized purge solution. In addition to that, shear stress from high-flow pumps contributes to acquired von Willebrand factor deficiency, further increases the risk of bleeding.18 Critical illness frequently induces a proinflammatory and procoagulant milieu characterized by endothelial injury, hepatic dysfunction, impaired synthesis of coagulation factors, and, in some cases, hyperfibrinolysis. A procoagulant state is often seen in patients with critical illness, because of multiorgan failure and a systemic inflammatory response.19,20 Development of liver failure is associated with depletion of coagulant factors II, V, VII, IX, and X. Studies on out-of-hospital cardiac arrest indicated hyperfibrinolytic state in up to one-half of the patients.21 These overlapping mechanisms likely explain the early clustering of bleeding events observed in our cohort.
Despite differences in antithrombotic therapy and clinical characteristics or situations, bleeding incidence did not differ between ACS and non-ACS presentations. This finding suggests that once pLVAD support is initiated, factors related to hemodynamic instability, systemic inflammation, and device or critical illness related coagulopathy may outweigh presentation-specific influences. These observations underscore the importance of focusing on procedural and physiological factors rather than clinical presentation alone when assessing bleeding risk.
The need for additional MCS emerged as the dominant determinant of bleeding risk. This likely reflects not only procedural complexity, including multiple cannulation sites and intensified anticoagulation requirements, but also the severity of underlying circulatory failure necessitating escalation of support. Previous studies in patients with ACS-related cardiogenic shock have shown that additional MCS use, particularly combined support with VA-ECMO and pLVAD, is associated with higher risk of bleeding complications.7,22,23 These findings collectively highlight the hemostatic vulnerability of patients requiring device escalation and support our observation that multidevice MCS identifies a distinctly high-risk bleeding phenotype. Importantly, the association between additional MCS use and early bleeding remained consistent after accounting for the competing risk of early mortality. This finding suggests that the observed relationship is unlikely to be solely explained by differential early death between groups and supports the robustness of our primary analyses. Therefore, the elevated bleeding risk associated with multidevice MCS support appears to reflect intrinsic procedural and clinical complexity rather than survival-related selection effects. Because detailed bleeding source and procedural attribution were unavailable, these findings should be interpreted as associations with overall early bleeding risk rather than evidence of specific bleeding mechanisms. The consistency of the cause-specific Cox and Fine–Gray analyses suggests that additional MCS use was associated not only with the instantaneous hazard of bleeding among patients who remained alive and free of bleeding, but also with the cumulative occurrence of bleeding in a competing-risk framework that accounted for early mortality.
As a complementary analyses, ML with SHAP-based interpretation provided additional insights into the heterogeneity of bleeding risk. Although predictive performance was modest, these analyses consistently identified additional MCS as a key contributor and suggested 2 distinct phenotypic profiles associated with bleeding risk: a procedural complexity phenotype characterized by additional MCS use and procedural burden, and an inflammatory-hemodynamic instability phenotype characterized by elevated CRP levels and in-hospital cardiopulmonary arrest. In contrast, patients characterized primarily by elevated preprocedural heart rate but without strong mechanical or inflammatory contributors exhibited comparatively lower bleeding incidence. These findings underscore the heterogeneity of bleeding mechanisms beyond device exposure alone and suggest that bleeding risk may reflect converging pathways of procedural complexity, systemic inflammation, and hemodynamic instability. Although prior studies have developed artificial intelligence-based predictive models for outcomes of pLVAD therapy, to our knowledge, no previous study has applied ML approaches specifically to evaluate bleeding risk associated with PVAD use.24 Importantly, the primary objective of the present study was to characterize the real-world incidence, timing, and clinical predictors of bleeding after pLVAD implantation using a nationwide registry. The ML analyses were performed as complementary exploratory analyses to visualize potential heterogeneity in bleeding risk and to generate hypotheses regarding patient phenotypes, rather than to develop a clinically deployable prediction model. These findings support the potential value of ML for exploratory risk stratification, while highlighting that ML-derived phenotypes reflect clinical and procedural complexity that must be interpreted within the broader clinical context. Moreover, the attenuation of discrimination in the validation cohort indicates potential overfitting and limited generalizability, likely reflecting clinical heterogeneity and inherent constraints of registry-based data. Given the modest discrimination in the validation cohort, these ML analyses should be interpreted primarily as exploratory and hypothesis-generating for risk phenotyping rather than as a standalone tool for clinical decision-making.
Several practical implications emerge from these findings. Early onset of bleeding emphasizes the importance of meticulous vascular access techniques, careful device management, and individualized anticoagulation strategies. Recognition that additional MCS substantially increases bleeding risk should inform decisions regarding device escalation and postprocedural monitoring. Although escalation to additional MCS support may be unavoidable in critically ill patients with profound cardiogenic shock, careful prevention and early management of bleeding complications are essential. Procedural strategies including appropriate selection of vascular access, use of ultrasound and fluoroscopic guidance, and adoption of micropuncture techniques with a low stick angle may help reduce access-site complications.25 The U.S. Food and Drug Administration has approved a bicarbonate-based purge solution for cases in which unfractionated heparin is contraindicated or bleeding develops, and its use has been reported to be associated with reduced bleeding in observational analyses.26 Recent data also suggest that anticoagulation protocols guided by parallel monitoring of anti-Xa activity and activated partial thromboplastin time can mitigate bleeding risk.18 Despite these advances, no established strategy exists to prevent bleeding during pLVAD use, accentuating the need for further research to optimize anticoagulation and access management in this high-risk population. An improved understanding of bleeding phenotypes using ML with SHAP may support more tailored management approaches in critically ill patients requiring temporary MCS.
Although the J-PVAD registry lacks the granularity required to adjudicate bleeding mechanisms, its nationwide scope provides an important opportunity to characterize bleeding complications in contemporary real-world pLVAD practice. The value of this study lies in defining the national burden, early timing, and clinical contexts of bleeding risk across a broad all-comer population, rather than in establishing causal mechanisms. Accordingly, the exploratory subgroup and ML-based analyses should be interpreted as hypothesis-generating approaches that complement conventional regression and help identify clinical settings in which heightened vigilance may be warranted.
Study limitations
Several limitations of the present study should be acknowledged. First, as a registry-based observational study, treatment strategies including device selection, anticoagulation management, and decisions regarding escalation to additional MCS were not standardized across centers, introducing the potential for unmeasured confounding. Data on length of hospital stay, health care costs, resource utilization, renal replacement therapy during pLVAD support, and baseline peripheral artery disease were not systematically available. Although additional MCS use was the strongest factor associated with bleeding, exact initiation and discontinuation times of additional MCS were not systematically captured; therefore, additional MCS could not be modeled as a true time-dependent exposure and should be interpreted as a time-fixed, episode-level marker of procedural complexity and illness severity rather than as a causal exposure. Second, bleeding events were not adjudicated using standardized definitions such as the Bleeding Academic Research Consortium or Global Use of Strategies to Open Occluded Arteries criteria. Although the registry captured any bleeding and registry-defined major bleeding, detailed information regarding bleeding subtype, anatomical source, access-site involvement, procedural attribution, timing relative to specific procedural steps, and bleeding-related hemodynamic consequences was unavailable. Therefore, differentiation between access-site and non–access-site bleeding or procedure-related and non–procedure-related bleeding could not be reliably performed, limiting mechanistic interpretation and comparability with prior studies. However, our definition of major bleeding encompasses Bleeding Academic Research Consortium bleeding criteria of type 3 or 5 bleeding and Global Use of Strategies to Open Occluded Arteries bleeding criteria of moderate or higher bleeding. Thus, we have evaluated bleeding that requires critical medical intervention during pLVAD treatment. Third, detailed laboratory information relevant to bleeding risk, such as hemoglobin level, platelet count, coagulation parameters including prothrombin time-international normalized ratio and activated partial thromboplastin time, liver function, and antiplatelet/anticoagulant therapy use, were not available.5 Consequently, the multivariable models could not account for all clinically relevant confounders of bleeding risk, and residual confounding cannot be excluded. As a result, although ML-derived phenotypes suggested mechanistic heterogeneity, we could not determine whether these phenotypes corresponded to distinct bleeding subtypes. This limitation constrains pathophysiological interpretation and underscores the need for more granular bleeding adjudication in future investigations. Non-ACS presentations were summarized descriptively, the registry did not include centrally adjudicated etiologies of cardiogenic shock or granular disease-specific data. In addition, several non-ACS categories included limited numbers of patients and events. Therefore, subgroup findings within the non-ACS stratum should be interpreted as descriptive and exploratory. Fourth, because the Impella CP was predominantly used as pLVAD in this study population, analysis on other types of pLVAD may have been underpowered. Finally, because Impella CP was predominantly used and external validation of the ML analyses was not performed, generalizability to other pLVAD devices and independent cohorts remains uncertain. ML-derived phenotypes should therefore be interpreted as exploratory and hypothesis-generating rather than as mechanistic classifications or clinically actionable prediction tools.
Conclusions
Bleeding complications post-pLVAD support are common, occur early after device implantation, and pose a significant clinical burden. In this nationwide cohort, bleeding incidence did not differ between ACS and non-ACS presentations, whereas the use of additional MCS emerged as the strongest predictor of bleeding across strata. These findings highlight the need for careful anticoagulation management, meticulous access-site technique, and heightened vigilance in patients requiring device escalation.
Perspectives.
COMPETENCY IN PATIENT CARE AND PROCEDURAL SKILLS: Bleeding during pLVAD support is frequent, occurs early, and represents a major contributor to morbidity. Recognition that additional MCS markedly increases bleeding risk is essential for clinical decision-making. Careful vascular access selection, meticulous procedural technique, and individualized anticoagulation strategies are critical to reducing early bleeding complications in patients requiring temporary mechanical support.
TRANSLATIONAL OUTLOOK: Future studies should focus on developing standardized anticoagulation protocols and optimizing device management strategies to minimize bleeding risk, particularly in patients requiring multidevice support. Prospective validation of risk prediction models, including those informed by ML, may enable more personalized bleeding risk assessment and guide preventive measures in critically ill populations.
Funding support and author disclosures
The authors have reported that they have no relationships relevant to the contents of this paper to disclose.
Footnotes
The authors attest they are in compliance with human studies committees and animal welfare regulations of the authors’ institutions and Food and Drug Administration guidelines, including patient consent where appropriate. For more information, visit the Author Center.
Appendix
For an expanded Methods section, supplemental tables and figures, please see the online version of this paper.
Contributor Information
Toshiaki Toyota, Email: totoyota@kuhp.kyoto-u.ac.jp.
Hirohiko Kohjitani, Email: kohjitani.hirohiko.7z@kyoto-u.ac.jp.
Supplementary material
References
- 1.Naidu S.S., Baran D.A., Jentzer J.C., et al. SCAI SHOCK stage classification expert consensus update: a review and incorporation of validation studies: this statement was endorsed by the American College of Cardiology (ACC), American College of Emergency Physicians (ACEP), American Heart Association (AHA), European Society of Cardiology (ESC) Association for Acute Cardiovascular Care (ACVC), International Society for Heart and Lung Transplantation (ISHLT), Society of Critical Care Medicine (SCCM), and Society of Thoracic Surgeons (STS) in December 2021. J Am Coll Cardiol. 2022;79(9):933–946. doi: 10.1016/j.jacc.2022.01.018. [DOI] [PubMed] [Google Scholar]
- 2.Ardito V., Sarucanian L., Rognoni C., Pieri M., Scandroglio A.M., Tarricone R. Impella versus VA-ECMO for patients with cardiogenic shock: comprehensive systematic literature review and meta-analyses. J Cardiovasc Dev Dis. 2023;10(4):158. doi: 10.3390/jcdd10040158. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Ikeda Y., Ako J., Toda K., et al. Short-Term outcomes of impella support in Japanese patients with cardiogenic shock due to acute myocardial infarction - japanese registry for percutaneous ventricular assist device (J-PVAD) Circ J. 2023;87(5):588–597. doi: 10.1253/circj.CJ-22-0476. [DOI] [PubMed] [Google Scholar]
- 4.Toda K., Ako J., Hirayama A., et al. Three-year experience of catheter-based micro-axial left ventricular assist device, impella, in Japanese patients: the first interim analysis of Japan registry for Percutaneous Ventricular Assist Device (J-PVAD) J Artif Organs. 2023;26(1):17–23. doi: 10.1007/s10047-022-01328-1. [DOI] [PubMed] [Google Scholar]
- 5.Toyota T., Watanabe H., Kim K., Furukawa Y., Kimura T. Antithrombotic therapy for acute coronary syndrome. J Neuroendovasc Ther. 2025;19 doi: 10.5797/jnet.ra.2024-0102. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Suzuki T., Asano T., Shinozaki T., et al. Prognostic implications of bleeding and ischemic complications in acute myocardial infarction-related cardiogenic shock managed with microaxial flow pump. Circ Cardiovasc Interv. 2026;19(4) doi: 10.1161/CIRCINTERVENTIONS.125.015676. [DOI] [PubMed] [Google Scholar]
- 7.Ali J.M., Abu-Omar Y. Complications associated with mechanical circulatory support. Ann Transl Med. 2020;8(13):835. doi: 10.21037/atm.2020.03.152. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Amin A.P., Spertus J.A., Curtis J.P., et al. The evolving landscape of impella use in the United States among patients undergoing percutaneous coronary intervention with mechanical circulatory support. Circulation. 2020;141(4):273–284. doi: 10.1161/CIRCULATIONAHA.119.044007. [DOI] [PubMed] [Google Scholar]
- 9.Miller P.E., Bromfield S.G., Ma Q., et al. Clinical outcomes and cost associated with an intravascular microaxial left ventricular assist device vs intra-aortic balloon pump in patients presenting with acute myocardial infarction complicated by cardiogenic shock. JAMA Intern Med. 2022;182(9):926–933. doi: 10.1001/jamainternmed.2022.2735. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Shibasaki I., Masawa T., Abe S., et al. Benefit of veno-arterial extracorporeal membrane oxygenation combined with impella (ECpella) therapy in acute coronary syndrome with cardiogenic shock. J Cardiol. 2022;80(2):116–124. doi: 10.1016/j.jjcc.2022.02.013. [DOI] [PubMed] [Google Scholar]
- 11.Sano M., Toyota T., Matsuoka Y., et al. Predictors of bleeding complications after extracorporeal cardiopulmonary resuscitation. JACC Asia. 2026;6:314–325. doi: 10.1016/j.jacasi.2025.09.027. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Frazier P.I. A tutorial on Bayesian optimization [Internet]. arXiv [stat.ML] 2018. http://arxiv.org/abs/1807.02811
- 13.Ke G., Meng Q., Finley T., et al. LightGBM: a highly efficient gradient boosting decision tree [Internet] 2017. https://proceedings.neurips.cc/paper_files/paper/2017/file/6449f44a102fde848669bdd9eb6b76fa-Paper.pdf
- 14.Lundberg S., Lee S.I. A unified approach to interpreting model predictions [Internet]. arXiv:1705.07874 [cs.AI] 2017. https://arxiv.org/abs/1705.07874
- 15.Takahashi K., Kubo S., Ikuta A., et al. Incidence, predictors, and clinical outcomes of mechanical circulatory support-related complications in patients with cardiogenic shock. J Cardiol. 2022;79(2):163–169. doi: 10.1016/j.jjcc.2021.08.011. [DOI] [PubMed] [Google Scholar]
- 16.Van Edom C.J., Gramegna M., Baldetti L., et al. Management of bleeding and hemolysis during percutaneous microaxial flow pump support: a practical approach. JACC Cardiovasc Interv. 2023;16(14):1707–1720. doi: 10.1016/j.jcin.2023.05.043. [DOI] [PubMed] [Google Scholar]
- 17.Subramaniam A.V., Barsness G.W., Vallabhajosyula S., Vallabhajosyula S. Complications of temporary percutaneous mechanical circulatory support for cardiogenic shock: an appraisal of contemporary literature. Cardiol Ther. 2019;8(2):211–228. doi: 10.1007/s40119-019-00152-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Vandenbriele C., Arachchillage D.J., Frederiks P., et al. Anticoagulation for percutaneous ventricular assist device-supported cardiogenic shock: JACC review topic of the week. J Am Coll Cardiol. 2022;79(19):1949–1962. doi: 10.1016/j.jacc.2022.02.052. [DOI] [PubMed] [Google Scholar]
- 19.Vandenbriele C., Vanassche T., Price S. Why we need safer anticoagulant strategies for patients on short-term percutaneous mechanical circulatory support. Intensive Care Med. 2020;46(4):771–774. doi: 10.1007/s00134-019-05897-3. [DOI] [PubMed] [Google Scholar]
- 20.Cuinet J., Garbagnati A., Rusca M., et al. Cardiogenic shock elicits acute inflammation, delayed eosinophilia, and depletion of immune cells in most severe cases. Sci Rep. 2020;10(1):7639. doi: 10.1038/s41598-020-64702-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 21.Schochl H., Cadamuro J., Seidl S., et al. Hyperfibrinolysis is common in out-of-hospital cardiac arrest: results from a prospective observational thromboelastometry study. Resuscitation. 2013;84(4):454–459. doi: 10.1016/j.resuscitation.2012.08.318. [DOI] [PubMed] [Google Scholar]
- 22.Freund A., Jobs A., Lurz P., et al. Frequency and impact of bleeding on outcome in patients with cardiogenic shock. JACC Cardiovasc Interv. 2020;13(10):1182–1193. doi: 10.1016/j.jcin.2020.02.042. [DOI] [PubMed] [Google Scholar]
- 23.Bogerd M., Ten Berg S., Peters E.J., et al. Impella and venoarterial extracorporeal membrane oxygenation in cardiogenic shock complicating acute myocardial infarction. Eur J Heart Fail. 2023;25(11):2021–2031. doi: 10.1002/ejhf.3025. [DOI] [PubMed] [Google Scholar]
- 24.Consolo F., D'Andria Ursoleo J., Pieri M., et al. The intelligent Impella: future perspectives of artificial intelligence in the setting of Impella support. ESC Heart Fail. 2024;11(5):2933–2940. doi: 10.1002/ehf2.14865. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.Vetrovec G.W., Kaki A., Wollmuth J., Dahle T.G. Strategies for reducing vascular and bleeding risk for percutaneous left ventricular assist device-supported high-risk percutaneous coronary intervention. Heart Int. 2022;16(2):105–111. doi: 10.17925/HI.2022.16.2.105. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Bergen K., Sridhara S., Cavarocchi N., Silvestry S., Ventura D. Analysis of bicarbonate-based purge solution in patients with cardiogenic shock supported via Impella ventricular assist device. Ann Pharmacother. 2023;57(6):646–652. doi: 10.1177/10600280221124156. [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.