Thromb Haemost 2022; 122(09): 1558-1566
DOI: 10.1055/a-1754-7551
Stroke, Systemic or Venous Thromboembolism

Bleeding Risk Assessment in End-Stage Kidney Disease: Validation of Existing Risk Scores and Evaluation of a Machine Learning-Based Approach

Stephan Nopp
1   Clinical Division of Haematology and Hemostaseology, Department of Medicine I, Medical University of Vienna, Vienna, Austria
,
Clemens P. Spielvogel
2   Division of Nuclear Medicine, Department of Biomedical Imaging and Image-guided Therapy, Medical University of Vienna, Vienna, Austria
3   Christian Doppler Laboratory for Applied Metabolomics, Vienna, Austria
,
Sabine Schmaldienst
4   Department of Medicine I, Clinic Favoriten, Vienna, Austria
,
Renate Klauser-Braun
5   Department of Medicine III, Clinic Donaustadt, Vienna, Austria
,
Matthias Lorenz
6   Vienna Dialysis Centre, Vienna, Austria
,
Benedikt N. Bauer
1   Clinical Division of Haematology and Hemostaseology, Department of Medicine I, Medical University of Vienna, Vienna, Austria
,
Ingrid Pabinger
1   Clinical Division of Haematology and Hemostaseology, Department of Medicine I, Medical University of Vienna, Vienna, Austria
,
Marcus Säemann
7   Department of Medicine VI, Clinic Ottakring, Vienna, Austria
,
Oliver Königsbrügge
1   Clinical Division of Haematology and Hemostaseology, Department of Medicine I, Medical University of Vienna, Vienna, Austria
,
Cihan Ay
1   Clinical Division of Haematology and Hemostaseology, Department of Medicine I, Medical University of Vienna, Vienna, Austria
› Author Affiliations
Funding The VIVALDI study was supported by an unrestricted grant from the Austrian National Bank (Jubiläumsfond Österreichische Nationalbank, Project Number 16433), Joseph-Skoda-Prize of the Austrian Association of Internal Medicine (ÖGIM), and the Austrian Science Fund (FWF), Special Research Program (SFB) 54.
 

Abstract

Background Patients with end-stage kidney disease (ESKD) on hemodialysis (HD) are at increased risk for bleeding. However, despite relevant clinical implications regarding dialysis modalities or anticoagulation, no bleeding risk assessment strategy has been established in this challenging population.

Methods Analyses on bleeding risk assessment models were performed in the population-based Vienna InVestigation of Atrial fibrillation and thromboemboLism in patients on hemoDialysIs (VIVALDI) study including 625 patients. In this cohort study, patients were prospectively followed for a median observation period of 3.5 years for the occurrence of major bleeding. First, performances of existing bleeding risk scores (i.e., HAS-BLED, HEMORR2HAGES, ATRIA, and four others) were evaluated in terms of discrimination and calibration. Second, four machine learning-based prediction models that included clinical, dialysis-specific, and laboratory parameters were developed and tested using Monte Carlo cross-validation.

Results Of 625 patients (median age: 66 years, 37% women), 89 (14.2%) developed major bleeding, with a 1-year, 2-year, and 3-year cumulative incidence of 6.1% (95% confidence interval [CI]: 4.2–8.0), 10.3% (95% CI: 8.0–12.8), and 13.5% (95% CI: 10.8–16.2), respectively. C-statistics of the seven contemporary bleeding risk scores ranged between 0.54 and 0.59 indicating poor discriminatory performance. The HAS-BLED score showed the highest C-statistic of 0.59 (95% CI: 0.53–0.66). Similarly, all four machine learning-based predictions models performed poorly in internal validation (C-statistics ranging from 0.49 to 0.55).

Conclusion Existing bleeding risk scores and a machine learning approach including common clinical parameters fail to assist in bleeding risk prediction of patients on HD. Therefore, new approaches, including novel biomarkers, to improve bleeding risk prediction in patients on HD are needed.


#

Introduction

Chronic kidney disease, and especially end-stage kidney disease (ESKD), is associated with an increased risk for bleeding. The bleeding tendency in patients with ESKD is multifactorial and influenced by dysfunctional platelets, impaired platelet–vessel wall interactions, renal anemia, and treatment effects including anticoagulation therapy.[1] The bleeding rate in patients on hemodialysis (HD) has been reported to be 9 per 100 patient-years and 15% of the bleeds result in death.[2] [3] [4] [5] [6] [7] [8] Stratification of bleeding risk could be useful to individualize patient management; however, no valid approach to assess the risk of bleeding in this challenging patient population has been established and recommended to date.

Existing bleeding risk assessment tools, such as the HAS-BLED (Hypertension, Abnormal renal/liver function, Stroke, Bleeding history or predisposition, Labile international normalized ratio, Elderly, Drugs/alcohol concomitantly) score, are widely used in nondialysis patients and have been incorporated in several guidelines to assess bleeding risk in patients on antithrombotic therapy, particularly in patients with atrial fibrillation (AF) on oral anticoagulation.[9] [10] [11] [12] Optimal bleeding risk assessment in patients on dialysis is desirable for several reasons during patient management: (1) bleeding risk could be considered to enable informed decision making such as the ideal dialysis treatment modality as several studies highlighted significantly lower bleeding rates in patients on peritoneal dialysis compared with HD.[13] [14] [15] (2) High-risk patients might benefit from close monitoring, control of risk factors, and decision of anticoagulation regimens for HD. (3) Further, AF is common in patients on HD and oral anticoagulation for stroke prevention is indicated in the vast majority of ESKD patients with AF. However, studies on the benefit of anticoagulation in dialysis patients are inconclusive and excess risk of bleeding during vitamin K antagonist treatment is observed.[16] [17] [18] [19] Therefore, bleeding risk assessment has the potential to identify patients eligible for anticoagulation in the AF population and guide patient management in the general dialysis population.

Thus, we aimed at improving bleeding risk prediction in HD patients using a two-step approach. First, we evaluated known bleeding risk assessment tools in a prospective and observational study of patients with ESKD on HD; second to the standard statistical models, we developed four machine learning-based prediction models to identify novel approaches and predictors for bleeding risk assessment.


#

Methods

Study Design, Population, and Outcome

The Vienna InVestigation of AtriaL Fibrillation and thromboembolism in hemoDIalysis (VIVALDI) study is a prospective multicenter population-based cohort study, in which 625 patients were recruited in a cross-sectional fashion between April 2014 and July 2015 at seven dialysis centers in Vienna, Austria. All adult patients requiring chronic HD except patients who were pregnant or lactating were eligible for inclusion. A detailed description of the study design and procedures has been published elsewhere.[16] [20] The VIVALDI study was approved by the local ethics committees (No. 1146/2014) and all patients provided written informed consent.

Patient demographics (age, gender, body mass index, smoking status) and clinical characteristics (primary kidney disease, AF, diabetes, history of bleeding or stroke, and other comorbidities) were assessed in a structured interview verified against medical records. Dialysis-specific parameters (e.g., remaining diuresis, ultrafiltration rate) and laboratory measurements were extracted from the electronic medical documentation system. After inclusion, patients were prospectively followed for a maximum of 1,350 days for the occurrence of major bleeding as defined by the International Society on Thrombosis and Hemostasis (ISTH)[21] and all-cause death. All outcomes were adjudicated by independent experts upon chart review and based on imaging evidence or autopsy findings.

In this prediction model validation and development study, we adhered to the Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis (TRIPOD) statement.[22]


#

Bleeding Risk Assessment Tools

Currently existing bleeding risk scores were selected from recent reviews on bleedings risk assessment tools in various diseases.[23] [24] [25] Seven bleeding risk prediction scores potentially applicable to the HD setting were identified and evaluated. The HAS-BLED,[26] ATRIA,[27] ORBIT,[28] OBRI,[29] and mORBI[30] scores were calculated based on the variables collected at study inclusion for each patient as defined in their original development studies. For the HEMORR2HAGES risk score, two variables (i.e., “excessive fall risk” and “genetic factors”) could not be considered due to the lack of baseline information with respect to fall risk and CYP2C9 single nucleotide polymorphisms. Further, history of bleeding was used as a proxy for “recent bleeding” for the score developed by Shireman et al.[31] A detailed explanation on the calculation of each risk score is available in the [Supplementary Methods] (available in the online version).


#

Statistical Analysis

Patient characteristics are presented as frequencies (percentage), mean (±standard deviation), or median (interquartile range), as appropriate. Follow-up time was calculated using the reverse Kaplan–Meier method and bleeding rates are presented using a Poisson model as well as cumulative incidences at 1, 2, and 3 years after study inclusion. We accounted for death and renal transplantation as a competing risk.

Performance of the bleedings risk scores was assessed using discrimination and calibration. Discriminatory ability of each risk score was evaluated using C-statistics including 95% confidence intervals (CIs). Calibration is presented visually using calibration plots comparing predicted versus actual bleeding risks. In addition, the Brier score, which is a measure for accuracy of probabilistic predictions, was calculated for each bleeding risk model. Further, net reclassification improvement (NRI) and integrated discrimination improvement (IDI) were computed for each risk score when compared with the HAS-BLED score. Missing values (displayed in [Table 1]) were imputed using single imputation by predictive mean matching. Statistical analysis was conducted using R (version 3.6.2; R Core Team, 2019) and the following add-on packages were used: survival, cmprsk, riskRegression, pec, prodlim, epitools, predictABEL, and mice.

Table 1

Patient characteristics

Total study cohort

( n  = 625)

Subgroup of patients with atrial fibrillation ( n  = 165)

Demographics

 Age, y

66 (54.5–75.0)

72 (64.0–78.0)

 Male sex

394 (63.0%)

116 (70.3%)

 BMI, kg/m2 (2)

25.7 (22.4–29.4)

25.8 (22.8–29.6)

Etiology of end-stage kidney disease

 Diabetic nephropathy

160 (25.6)

43 (26.1)

 Vascular nephropathy

121 (19.4)

39 (23.6)

 Glomerulonephritis

81 (13.0)

16 (9.7)

 Atrophic nephropathy

57 (9.1)

16 (9.7)

 Cystic nonhereditary nephropathy

36 (5.8)

10 (6.1)

 Hereditary nephropathy

31 (5.0)

5 (3.0)

 Nephrectomy

20 (3.2)

10 (6.1)

 Toxic nephropathy

28 (4.5)

9 (5.4)

 Other causes

91 (14.6)

17 (10.3)

Dialysis-specific parameters

 Cumulative time on hemodialysis, y

2.7 (1.0–5. 0)

3.0 (1.1–6.0)

 Ultrafiltration rate, mL (35)

2000 (982.5–3,000)

2100 (997.5–3,000)

 Remaining diuresis, mL/d (24)

500 (0–1,000)

325 (0–1,000)

 History of kidney transplantation

90 (14.4)

26 (15.8)

 History of peritoneal dialysis (15)

46 (7.5)

11 (6.7)

Comorbidities

 Atrial fibrillation

165 (26.4%)

165 (100%)

 Diabetes mellitus (2)

237 (38.0)

69 (42.1)

 Heart failure

183 (29.3)

70 (42.4)

 Coronary artery disease

232 (37.1)

83 (50.3)

 Peripheral artery disease

197 (31.5)

57 (34.6)

 Artificial heart valve

43 (6.9)

21 (12.7)

 Hypertension

574 (91.8)

152 (92.1)

 Active or history of cancer[a]

152 (24.3)

56 (33.9)

 Prior stroke/TIA/systemic embolism

137 (21.9)

48 (29.1)

 Prior clinically relevant bleeding

147 (23.5)

51 (30.9)

 Prior major bleeding

67 (10.7)

21 (12.7)

 Prior myocardial infarction

104 (16.6)

35 (21.1)

 Prior venous thromboembolism

61 (9.8)

25 (15.2)

 Current or past smoker (11)

305 (49.7)

75 (46.3)

 CHA2DS2-VASc score

3 (2–5)

4 (3–5)

 HAS-BLED score

3 (2–4)

4 (3–5)

Laboratory parameters

 Hemoglobin, g/dL (27)

10.9 (10.0–11.7)

11.1 (10.1–11.9)

 Hematocrit, % (27)

33.3 (30.3–35.9)

33.4 (30.7–36.0)

 Thrombocytes, G/L (28)

210 (169–257)

197 (153–232)

 Leukocytes, G/L (27)

6.4 (5.2–8.0)

6.4 (5.2–7.8)

 Albumin, g/L (117)

36.0 (33.0–39.2)

35.0 (32.0–37.5)

 C-reactive protein, mg/dL (96)

0.66 (0.29–1.9)

0.72 (0.29–2.20)

Antithrombotic therapy

 Anticoagulation therapy

139 (22.2%)

83 (50.3%)

 Platelet inhibitor therapy

345 (55.2)

90 (54.6)

Abbreviation: BMI, body mass index.


Note: Italic number in brackets indicates the number of patients with missing values. Continuous variables are presented as median (25th–75th percentile) and categorical variables as absolute frequencies and percentages.


a Excluding nonmelanoma skin cancer. Patients with multiple malignancies were counted once.



#

Machine Learning Methods

Four machine learning algorithms were employed for the binary classification of whether a patient experienced major bleeding within 24 months after study inclusion. Patients who were lost to follow-up or received renal transplantation were excluded from the subsequent analysis. The development of the supervised models consisted of three steps: data preprocessing, model training, and model validation ([Fig. 1]). In the data-preprocessing step, handling of missing patient characteristics and laboratory parameters was conducted using a k-nearest neighbor imputation approach. Further, potential issues arising from class distribution imbalances were mitigated using a combination of random undersampling and the synthetic minority oversampling technique (SMOTE).[32] Both, imputation and resampling were performed after splitting into training and test set. Resampling was only performed on the training data. The k-nearest neighbor model for imputation was trained based on the training data and applied to training and test set to avoid data leakage. The training was performed using four classification algorithms, including a k-nearest neighbor algorithm, a decision tree algorithm, a random forest algorithm, and a neural network algorithm to derive a total of four classification models. A total of 25 features consisting of patient demographics, clinical characteristics, dialysis-specific parameters, and laboratory parameters were used for the training procedure. For the validation, 100-fold Monte Carlo (MC) cross-validation with 85% of samples in the training set was employed to obtain a highly robust and repeatable estimation of the predictive performance of the model. Preplanned feature selection, feature importance measurements, and hyperparameter tuning were not performed due to the insufficient performance of all models. To identify potential clusters within the data, principal component analysis (PCA) was performed to reduce the dataset to two dimensions. A detailed list of parameters and features used for the analysis can be found in [Supplementary Tables S1] and [S2] (available in the online version).

Zoom Image
Fig. 1 Machine learning strategy.

The analysis was performed using Python 3.9.5 and the external packages Scikit-Learn 0.24.2,[33] Seaborn 0.11.1,[34] Matplotlib 3.3.4,[35] Numpy 1.20.2,[36] Pandas 1.2.4,[37] and Imblearn 0.8.0.[38]


#
#

Results

Clinical characteristics of all 625 patients are presented in [Table 1]. In brief, the median age was 66 (54.5–75.0) years, 231 (37%) patients were female, and 26% had AF at baseline. During a median follow-up time of 3.47 (3.38–3.58) years, 89 (14.2%) patients developed major bleeding comprising of 41 gastrointestinal bleeds, 14 intracranial hemorrhagic events, 13 other critical organ hemorrhage events, and 21 other major bleeding events. This accounts for a major bleeding rate of 6.8 (95% CI: 5.5–8.4) per 100 patient-years or a cumulative incidence of 6.1% (95% CI: 4.2–8.0) at 1 year, 10.3% (95% CI: 8.0–12.8) at 2 years, and 13.5% (95% CI: 10.8–16.2) at 3 years. One patient (0.16%) was lost to follow-up and 113 (18.1%) received kidney transplantation.

Performance of Bleeding Risk Prediction Tools

Discriminatory performances of seven currently existing bleeding risk prediction scores are shown in [Table 2]. In summary, none of the risk assessment models had a C-statistic above 0.60 in the total study population and in the AF population, CIs of all scores overlapped 0.50. Of all scores, the HAS-BLED score[26] showed the highest discriminatory performance in the total population with a C-statistic of 0.59 (95% CI: 0.53–0.66). C-statistics of the ATRIA (Anticoagulation and Risk Factors in Atrial Fibrillation),[27] HEMORR2HAGES (Hepatic or renal disease, Ethanol abuse, Malignancy, Older age, Reduced platelet count or function, Rebleeding risk, Hypertension, Anemia, Genetic factors, Excessive fall risk, and Stroke),[39] ORBIT (Outcomes Registry for Better Informed Treatment),[28] OBRI (Outpatient Bleeding Risk Index),[29] mOBRI (modified Outpatient Bleeding Risk Index),[30] and Shireman et al[31] scores ranged between 0.54 and 0.59. Calibration plots are presented in [Fig. S1]. Based upon visual inspection and the Brier score, calibration performance for bleeding risk was moderate across all prediction scores. Detailed comparisons between risk scores reporting on NRI and IDI including decision curve analysis are available in [Supplementary Table S3] and [Fig. S2] (available in the online version).

Table 2

Discriminatory performance of bleeding risk scores in patients on hemodialysis

Bleeding risk score

Total study cohort ( n  = 625)

C-statistics (95% CI)

Subgroup of patients with atrial fibrillation ( n  = 165)

C-statistic (95% CI)

HAS-BLED

0.59

(0.53–0.66)

0.54

(0.42–0.66)

ATRIA

0.55

(0.48–0.62)

0.58

(0.46–0.70)

HEMORR2HAGES

0.58

(0.51–0.65)

0.56

(0.44–0.69)

ORBIT

0.59

(0.52–0.66)

0.61

(0.47–0.74)

OBRI

0.54

(0.47–0.61)

0.54

(0.44–0.64)

mOBRI

0.54

(0.47–0.60)

0.55

(0.44–0.65)

Shireman et al

0.59

(0.52–0.67)

0.62

(0.50–0.74)

Abbreviation: CI, confidence interval.


Note: In the total cohort, C-statistics of existing bleeding risk scores ranged between 0.54 and 0.59, indicating poor discriminatory performance. In the subgroup of patients with atrial fibrillation, C-statistics ranged between 0.54 and 0.62, while confidence intervals overlapped 0.5, which indicates that no model performed better than chance.



#

Performance of Machine Learning-Based Bleeding Risk Models

The MC cross-validation procedure yielded area under the receiver operating characteristic curve (AUROC) values close to random predictions for all four machine learning algorithms. AUROC performances were 0.55, 0.51, 0.50, and 0.49 for k-nearest neighbor, decision tree, neural network, and random forest algorithm, respectively ([Fig. 2]). Accuracies ranged from 0.86 for the random forest to 0.53 for the neural network, indicating imbalanced predictions of the two classes. Sensitivities for decision tree and random forest were low with 0.17 and 0.02, respectively. The other algorithms performed better at handling the imbalanced classes with sensitivities of 0.43 and 0.46 as well as specificities of 0.68 and 0.53 for the k-nearest neighbor algorithm and neural network, respectively. The cross-validation performance metrics, including accuracy, AUROC, sensitivity, specificity, positive predictive value, and negative predictive value for the four algorithms are shown in [Supplementary Table S4] (available in the online version). Overall, the performance metrics indicated that none of the machine learning algorithms was able to pick up any patterns associated with the bleeding risk of the patients. These results were further supported by the PCA-based dimensionality reduction, which did not show any cluster formation after reduction to two dimensions based on visual assessment as shown in [Fig. 3].

Zoom Image
Fig. 2 Performance of four machine learning-based prediction models to predict major bleeding in patients on hemodialysis.
Zoom Image
Fig. 3 Principal component analysis in patients on hemodialysis.

#
#

Discussion

In this prospective study of patients with ESKD on HD, we evaluated currently existing bleeding risk scores, developed in a nondialysis population, for their suitability in this special patient cohort. None of the seven bleeding risk assessment tools showed a C-statistic above 0.60, indicating poor performance to identify patients at low or high risk for bleeding. The machine learning-based prediction models derived from and internally validated in our dataset showed similar poor predictive abilities. The four commonly used machine learning models, which were based on 25 features comprising patient demographics, clinical characteristics, and laboratory measurements (shown in [Supplementary Table S2], available in the online version) did not predict major bleeding within 24 months better than chance.

Bleeding risk assessment poses a special challenge to the treating physician when deciding on anticoagulation modalities during HD or anticoagulation for treatment and prevention of thromboembolic events. Therefore, different models have been proposed to assess bleeding risk in clinical practice in the general patient population.[40] [41] However, potentially due to the heterogeneity in bleeding type and causality of bleeding events, risk prediction performs only poorly to moderately across different diseases and clinical settings.[24] [42] Individual bleeding risk prediction is especially challenging in patient populations at high bleeding risk.[43] [44] In our prevalent cohort (i.e., patients who were on HD at enrollment into the study), risk of developing major bleeding was 10% after 2 years. This risk estimate highlights the burden of bleeding in HD patients and is in line with previous findings of similar settings and bleeding definitions.[2] [4] [8]

Of note, none of the existing bleeding risk assessment strategies withstood external validation in our HD cohort. This is exemplified by the HAS-BLED score, which performed “best” in our cohort. The HAS-BLED score showed a C-statistic of 0.59 in our total cohort of patients with ESKD on HD and 0.54 in the subgroup of patients with AF, indicating poor discrimination of patients who developed major bleeding compared with those who did not. Further, expected event rates differed from observed event rates as shown in [Supplementary Fig. S1] (available in the online version). The HAS-BLED score is recommended by several guidelines as the best model to predict bleeding with C-statistics ranging up to 0.80 in some studies.[9] [10] [11] [12] [45] [46] Other bleeding risk scores do not offer advantages over the HAS-BLED score.[47] [48] [49] In a randomized controlled trial of AF patients, regular bleeding risk evaluation was associated with a reduction in bleeding events and increased anticoagulation usage.[50] Thus, appropriate and responsible use of bleeding risk scores is suggested to identify high-risk patients and address modifiable bleeding risk factors to improve patient outcome. However, previous studies already highlighted the difficulty of bleeding risk assessment in patients with chronic kidney disease[44] and a prior evaluation of bleeding risk scores (i.e., HAS-BLED, ATRIA, HEMORR2HAGES, ORBIT) in HD patients showed similarly poor predictive results.[51] In our study, we could confirm that these bleeding risk scores perform poorly in ESKD patients on HD. In addition, we have analyzed other bleeding risk scores, such as the OBRI, mOBRI, and the score of Shireman et al, which did not show an improved predictive value. Taken together, all bleeding risk scores performed worse in cohort studies in dialysis patients when compared with their development and validation studies, with most scores not predicting bleeding better than chance. This is true for the overall population of patients with ESKD on HD as well as the subgroup of AF patients and suggests that standard statistical models might not be appropriate for predicting bleeding in this specific patient population.

In need of better prognostic models, we applied machine learning to our dataset. Machine learning represents a major branch of artificial intelligence and data science, and holds promising results in predictive modeling with the ability to reveal key features from complex datasets. Its strength lies in handling enormous numbers of predictors and combining them in nonlinear and highly interactive ways.[52] [53] Strikingly, none of four different machine learning-based prediction models, which included 25 features consisting of patient demographics, clinical characteristics, dialysis-specific parameters, and common laboratory parameters, performed better than chance during internal validation in our dataset. Further, the lack of clustering or pattern recognition in PCA confirms the absence of helpful predictors for major bleeding. Earlier studies without machine learning methods concur with our findings while some identified single predictor variables such as recent gastrointestinal bleeding.[54] In summary, no strategy to assess bleeding risk in patients on HD can be recommended to date, as neither standard statistical nor machine-learning methods based on general clinical parameters have proven useful in predicting bleeding events. Thus, novel approaches to characterize bleeding risk in patients on HD are urgently needed. Notably, biomarker-based prediction scores such as the ABC (age, biomarkers, clinical history) bleeding risk score showed promising results in patients with AF and might guide the development of specific bleeding risk scores in patients with ESKD on HD.[55]

Some strengths and limitations must be considered when interpreting our findings. First, the bleeding risk scores evaluated in our study were mainly derived from AF cohorts, and in our population only one-quarter had AF at baseline. Second, the power of machine learning increases by sample size and in contrast to other larger studies that rely on insurance claims databases, our cohort is considered of only moderate size. On the contrary, face-to-face follow-ups and independent adjudication of bleeding events supplied optimal data quality for our analysis. Third, the ISTH definition for major bleeding was used in our study. Therefore, administration of two or more units of packed erythrocytes was counted as major bleeding. However, patients on HD frequently suffer from renal anemia resulting in a low threshold for blood transfusion and would therefore more likely fulfill the criteria for major bleeding. Nevertheless, fatal bleeding and bleeding in a critical organ including intracranial bleeding contributed substantially to the high incidence of major bleeding in our cohort. Furthermore, we could not consider time-to-event data in our machine learning approach, thus patients who were lost to follow-up had to be excluded from model development. Finally, frequent changes in antithrombotic strategies throughout the study period were noted, which might have affected the risk of bleeding. We tried to address this issue by including a baseline variable reflecting the intention-to-treat approach for antithrombotic medications as a feature in the machine learning models, at the risk of oversimplifying the variables for anticoagulation and antiplatelet therapy.


#

Conclusion

Bleeding risk assessment in patients with ESKD on HD is suboptimal and remains challenging, as existing prediction tools do not provide useful information for patient management. Similarly, our machine learning approach failed to identify useful models derived from a set of clinically available markers in current practice. Thus, we conclude that common clinical parameters are insufficient to stratify HD patients by bleeding risk and call for the development of novel approaches including biomarker-based prediction scores to characterize the bleeding risk in patients with ESKD on HD.

What is known about this topic?

  • Patients with end-stage kidney disease on hemodialysis are at very high risk of bleeding.

  • Stratification by bleeding risk could guide clinical decision-making for antithrombotic therapies or dialysis treatment modalities. However, no valid approach toward bleeding risk assessment has been recommended so far.

What does this paper add?

  • Our study highlights that existing bleeding risk assessment models—mainly developed in a nondialysis population—are not useful for predicting bleeding in patients with end-stage kidney disease on hemodialysis.

  • Also machine learning-based prediction models did not help to identify patients at high risk for bleeding.


#
#

Conflict of Interest

None declared.

Acknowledgments

We acknowledge the members of the adjudication committee: Julia Riedl, Johannes Thaler, Christoph Kopp, Thomas Gremmel, and Fritz Leutmezer.

Supplementary Material


Address for correspondence

Cihan Ay, MD
Clinical Division of Haematology and Haemostaseology, Department of Medicine I, Medical University of Vienna
Waehringer Guertel 18-20, A-1090 Vienna
Austria   

Publication History

Received: 08 November 2021

Accepted: 27 January 2022

Accepted Manuscript online:
28 January 2022

Article published online:
19 June 2022

© 2022. Thieme. All rights reserved.

Georg Thieme Verlag KG
Rüdigerstraße 14, 70469 Stuttgart, Germany


Zoom Image
Fig. 1 Machine learning strategy.
Zoom Image
Fig. 2 Performance of four machine learning-based prediction models to predict major bleeding in patients on hemodialysis.
Zoom Image
Fig. 3 Principal component analysis in patients on hemodialysis.