Development and validation of a predictive model for upper gastrointestinal bleeding in peritoneal dialysis patients: a retrospective, multicenter cohort study

Abstract

Background

The incidence of upper gastrointestinal bleeding (UGIB) in peritoneal dialysis (PD) patients was higher than that in the general population. Although there have been studies on the risk factors of UGIB in PD patients, few studies had established the model of UGIB in PD patients. The purpose of this study was to establish and verify a predictive model of UGIB in PD patients.

Methods

This retrospective study was conducted between February 1, 2007, and November 15, 2021. The patients were divided into training and validation sets in a 1:1 ratio. The least absolute shrinkage and selection operator (LASSO) regression was used to screen potential risk factors for UGIB. These risk factors were used to generate a prognostic nomogram. The nomogram model predictive performance was evaluated by receiver operating characteristic (ROC)curves, calibration curves and decision curve analysis (DCA).

Results

A total of 2107 eligible patients were included in this study. 114 patients suffered from UGIB. After LASSO regression analysis, five representative variables were included in the nomogram: the use of calcium, the use of proton pump inhibitors (PPIs), the history of UGIB, hemoglobin, and uric acid. The C-index of the training set was 0.859 (0.810–0.909), and the C-index of the validation set was 0.874 (0.829–0.919).

Conclusion

Using Calcium and PPIs, the history of UGIB, hemoglobin, and uric acid were important predictors of UGIB in PD patients. The nomogram developed and validated in this study has the potential to become an effective tool for clinicians to predict UGIB in PD patients.

Background

Chronic kidney disease (CKD) is a significant public health concern due to its increasing prevalence [1]. A recent national study estimated that the prevalence rate of CKD was 11%, and this prevalence rate was gradually increasing [2]. Peritoneal dialysis (PD) is one of the commonly used renal replacement therapies (RRT). Due to its little effect on hemodynamics, effective protection of residual renal function (RRF) and reduction of hematogenous infection, the use of PD has become increasingly widespread [3, 4]. Upper gastrointestinal bleeding (UGIB) represents a critical complication in dialysis patients, with a risk fivefold higher compared to individuals without CKD [5, 6]. UGIB not only increases all-cause mortality but also increases the risk of dialysis [7].

At present, some studies have found the risk factors of UGIB in dialysis patients, including the increase of uremic toxins, the application of anticoagulants, secondary hyperparathyroidism, malnutrition, inadequate dialysis, anemia, cardiovascular disease (CVD) and hypoalbuminemia [8–10]. Although there are studies investigating risk factors for UGIB in PD patients. To our knowledge, there is no individual prediction model to evaluate the risk of UGIB in PD patients. A nomogram is a tool that can predict the occurrence of a clinical disease objectively, which is a conventional method to predict the occurrence and development of the disease, which has been used in various fields. Therefore, it is necessary to develop a simple and applicable UGIB risk prediction model to help clinicians detect the risk of patients early and provide appropriate treatment.

The purpose of this paper was to establish and validate the predictive nomogram model to guide clinical decision-making, improve the quality of PD treatment and improve the prognosis of patients.

Methods

Patients and study design

A multicenter retrospective study was conducted in PD patients from four PD centers from February 1, 2007 to November 15, 2021. All patients received PD treatment in each center for the first time. Patients were randomly divided into the training set and validation set at a 1:1 ratio to ensure that outcome events are randomly distributed between the two cohorts. The study was performed in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of the Sixth Affiliated Hospital of Sun Yet-Sen University (No.2021SLYEC-177). All patients signed informed consent upon admission.

The inclusion criteria were as follows: (1) End-stage renal disease caused by various CKD requires replacement therapy with PD; (2) PD was maintained for more than 3 months. The exclusion criteria were as follows: (1) age < 18 years old; (2) patients with incomplete vital medical records; (3) liver cirrhosis, esophagogastric varices; (4) congenital coagulation dysfunction; (5) malignant tumor.

Data collection

All clinical data were retrieved from the patients’ medical records. Most data were collected within the first month after PD initiation. In order to ensure the integrity of the data, the collection of some patients had been expanded to three months. The demographic characteristics comprised age, sex, smoking history, alcohol consumption, comorbidities (including CVD, hyperlipidemia, diabetes, and hypertension), history of UGIB, and medication use (proton pump inhibitors [PPIs], Aspirin, calcium, and angiotensin receptor blockers [ARBs]). Laboratory indicators include white blood cells (WBC), hemoglobin, albumin, creatinine, blood urea nitrogen, uric acid, serum calcium, serum potassium, and serum phosphorus.

Follow-up and outcomes

Each center conducted a comprehensive medical assessment of patients every quarter, and all follow-up works were carried out by trained nurses over the phone. The primary endpoint was UGIB. This study focused on clinically significant UGIB events, operationally defined as hemorrhage cases necessitating hospitalization and specialized gastroenterology care. The diagnosis of UGIB should meet the following criteria: (a) there is clinical evidence of hematemesis, melena, or grossly visible blood in the gastric lavage fluid; (b) actively bleeding lesions are visualized endoscopically during the upper gastrointestinal tract examination [11]. The follow-up endpoint was any of the following: death, transfer to hemodialysis, transfer to another centers, transfer to kidney transplant, or follow-up to March 1, 2022.

Statistical analysis

The cohort was divided into the training set and the validation set. Variables exceeding a 10% missing data threshold were systematically excluded. The remaining missing variables were substituted with the mean of the respective variables. The normality of continuous variables was tested, and the variables of normal distribution were represented by mean ± standard deviation (SD), while those of skewed variables were represented by median (quartile). The Student’s T-test was used to determine the significance of the normality variables. The significance of skewed variables was determined using the Mann-Whitney U test. Frequency data were presented as numbers and percentages and compared using the Chi-square test. All the collected baseline variables were included in the Least absolute shrinkage and selection operator (LASSO) regression analysis to determine potential predictors, which effectively eliminated several irrelevant or multicollinearity independent variables to reduce high-dimensional data. All continuous variables were standardized before being included in the LASSO regression. The variables with non-zero coefficients identified by LASSO regression were included in a multivariate logistic regression model for validation. The prediction model nomogram was established using the variables screened by LASSO regression and verified by multiple logistic regression. A nomogram was made as a prediction model of UGIB, and the accuracy of the prediction was verified by receiver operating characteristic (ROC) curve analysis, C-Index, and calibration plot. The calibration degree of the model was judged by the calibration curve, and the decision curve analysis (DCA) was drawn to judge the clinical efficacy of the model.

All statistical analyses were carried out using SPSS (version 25.0; SPSS Company, Chicago, IL, USA) and R software version 4.2.0. All tests were bilateral, and P < 0.05 was considered to be statistically significant.

Results

Baseline characteristics of patients

The demographic and baseline clinical data of the patients were shown in Table 1. A total of 2107 eligible patients with PD were enrolled in the statistical analysis. Using R random sampling method, the dataset was partitioned into a training set (n = 1053) and a validation set (n = 1054). During the follow-up period, 62 patients in the training set experienced UGIB, and 52 patients in the test set experienced UGIB (Fig. 1). There was no statistically significant difference between the baseline data of the training set and the validation set, except for the history of hyperlipidemia. The incidence of hyperlipidemia in the training set was higher than that in the validation set (P = 0.035). The median age in both groups was 54 years and there were 598 males in both groups. There were no significant differences between the two groups in medication history, smoking history, drinking history, CVD, diabetes, hypertension, and UGIB history (Table 1)

Table 1.

Characteristics of the participants in different groups

Variables	training set (1053)	validation set (1054)	P-value
Demographics
Age(years)	54(41–65)	54(42–65)	0.955
Male (%)	598(56.8%)	598(56.7%)	0.960
Smoking (%)	72(6.8%)	59(5.6%)	0.239
Drinking (%)	25(2.4%)	30(2.8%)	0.497
Comorbid
Diabetes (%)	251(23.8%)	262(24.9%)	0.585
Hypertension (%)	541(51.4%)	572(54.3%)	0.184
CVD (%)	216(20.5%)	228(21.6%)	0.529
UGIB (%)	35(3.3%)	34(3.2%)	0.899
Hyperlipidemia (%)	272(25.8%)	231(21.9%)	0.035
Medication history
ARB/ACEI (%)	312(29.7%)	316(30.0%)	0.894
Calcium (%)	179(17%)	189(17.9%)	0.573
Aspirin (%)	79(7.5%)	81(7.7%)	0.874
PPIs (%)	136(12.9%)	140(13.3%)	0.803
Laboratory indicators
WBC (109 /L)	6.69(5.47–8.13)	6.46(5.29–7.93)	0.085
Hemoglobin (g/L)	105(88–120)	107(89–122)	0.730
Albumin (g/L)	34(30.2–37.3)	33.66(29.7–37.11)	0.904
Creatinine (umol/L)	747(569–992)	764(585.5–999)	0.395
Uric acid (umol/L)	388(331–454)	378(323.5–440)	0.055
Urea nitrogen (mmol/L)	17.9(14–22.6)	17.63(13.68–22.6)	0.541
Serum calcium (mmol/L)	2.17(2.03–2.3)	2.15(2.01–2.29)	0.430
Serum potassium (mmol/L)	4.02(3.53–4.53)	4.06(3.60–4.61)	0.174
Serum phosphorus (mmol/L)	1.4(1.1–1.72)	1.38(1.12–1.77)	0.904

CVD: cardiovascular disease; ARB: angiotensin receptor blocker; ACEI: angiotensin converting enzyme inhibitor; UGIB: Upper gastrointestinal bleeding; PPIs: proton-pump inhibitors; WBC: white blood cell

Fig. 1.

Flowchart of participants enrollment and outcomes

Development of nomogram

The LASSO regression analysis was carried out on the training set to screen the risk factors of UGIB occurrence in PD patients. In LASSO regression, variables with non-zero coefficients are retained (Fig. 2a). Finally, five variables were screened out: the use of calcium, the use of PPIs, a history of UGIB, hemoglobin, and uric acid (Fig. 2b). These five variables were verified by multivariate logistic regression. According to the results of multivariate logistic regression analysis, it is verified that each variable has statistical significance (Table 2). Finally, the above five risk factors were used to establish a prediction model to predict the risk of UGIB in PD patients, which was shown in the form of the nomogram (Fig. 3).

Fig. 2.

(a). LASSO coefficient profiles of the 33 features. (b) Optimal parameter (lambda) selection in the LASSO model used five-fold cross-validation based on minimum criteria

Table 2.

LASSO regression and Multivariate logistic regression

Characteristic	LASSO regression	Multivariate logistic regression
	Coefficients	OR	95% CI	P-value
History of gastrointestinal bleeding	1.290847	5.906	3.220–10.833	<0.001
Calcium	0.115147	2.970	1.952–4.517	<0.001
PPIs	0.107966	2.133	1.353–3.362	<0.001
Hemoglobin (g/L)	−0.005344	0.984	0.974-0-995	0.003
Uric acid (umol/L)	0.000957	1.002	1.000–1.004	0.045

PPIs, Proton pump inhibitors

Fig. 3.

Nomogram for assessing the risk of UGIB

Model validation

In order to verify the reliability of the prediction model, the model is tested using data from training and validation sets. ROC curves were drawn with UGIB as the outcome variable. According to the performance of the model in the ROC curve, it can be seen that the differentiation of the model is high (Fig. 4a and Fig. 4b). The C-index of the training set was 0.859 (95%CI, 0.810–0.909), and the C-index of the validation set was 0.874 (95%CI, 0.829–0.919).

Fig. 4.

(a) ROC of nomogram model in training set. (b) ROC of nomogram model in validation set

The calibration curve is shown in Fig. 5. The Hosmer-Lemeshow test was performed to assess the calibration of the prediction model, demonstrating a good fit between the observed and predicted probabilities (P = 0.755), which indicated excellent calibration performance of the model.

Fig. 5.

Calibration curve for UGIB, Hosmer-Lemeshowtest, P = 0.7547

Clinical utility

Clinical practicability is a key index to evaluate the applicability of the predictive model in the clinical environment and its benefits to patients. To demonstrate the clinical utility of the model, we assessed whether integrating DCA with model-derived line chart-assisted decision-making would improve patient prognosis. The decision curve of the model was in the range of the prediction thresholds of 0–0.25, and the net benefit curve is higher than the two invalid lines (Fig. 6).

Fig. 6.

Decision curve analysis of a nomogram of UGIB

The model can be used for patient education and individualized treatment in PD patients at high risk of UGIB, thereby reducing the overall risk of UGIB. When identifying patients at high risk for UGIB, nephrologists should promptly conduct a medication review for nonsteroidal anti-inflammatory drugs (NSAIDs) and anticoagulants, optimize anemia management and calcium-phosphate balance, and consider early endoscopic evaluation.

Discussion

This multicenter retrospective cohort study established and validated a clinical prediction model for UGIB in PD patients. Five variables were included in this predictive nomogram model: the use of Calcium, the use of PPIs, a history of UGIB, hemoglobin, and uric acid. The prediction nomogram was verified to have sufficient accuracy and well discrimination ability.

Patients receiving PD exhibit elevated UGIB risk compared to the general population. CKD patients demonstrated GIB incidence of 161/1000 person-years in a retrospective cohort analysis [12]. During follow-up, 114 patients developed UGIB, with an incidence of 5.41%. UGIB contributes to 3%-7% of mortality in maintenance dialysis populations, with associated all-cause in-hospital mortality risk tripling compared to non-UGIB cases [13]. The occurrence of UGIB significantly reduces quality of life in PD patients while increasing hospitalization rates and mortality, thus establishing a UGIB prediction model specifically for this population holds crucial clinical significance.

In this study, the history of UGIB is one of the important risk factors for recurrent UGIB in PD patients. Studies have shown that a history of UGIB is independently associated with a higher risk of UGIB in patients with CKD, and the HR of UGIB history after multivariate analysis was 1.94 (95% CI, 1.36–2.83) [10]. While prior UGIB represents an established risk determinant, current evidence demonstrates incomplete integration of this parameter in predictive modeling frameworks [14, 15], so clinicians should pay more attention to the risk of rebleeding in these patients.

Hyperuricemia is one of the common metabolic disorders in patients with CKD due to the deterioration of renal function [16]. It has been confirmed that the increase of uric acid can significantly increase the risk of CVD and promote the occurrence of CVD [17]. Patients with elevated uric acid had 1.93 times higher risk of hemorrhagic stroke than those with normal uric acid levels [18]. It may indicate that the increase of serum uric acid may be related to hemorrhagic disease. The correlation between uric acid and GIB has been reported in some studies [19]. Our previously published propensity score matching analysis demonstrated that baseline hyperuricemia is an independent risk factor for GIB in PD patients, and revealed a linear association between uric acid and new-onset GIB events [20]. High levels of uric acid are associated with endothelial dysfunction and overactivation of the renin-angiotensin-aldosterone system (RAAS) [21, 22]. Hyperuricemia can also promote oxidative stress and enhance inflammatory response [23, 24]. High levels of uric acid will cause vascular endothelial dysfunction, resulting in a decrease in gastric mucosal blood flow, resulting in a decline in the defense ability of gastric mucosa, which is easy to be attacked by pepsin and cause ulcer bleeding. Although the specific mechanisms underlying the association between uric acid and GIB remain unclear, uric acid levels should still be considered by clinicians when assessing the risk of UGIB in PD patients.

In our study, hemoglobin emerged as an independent risk predictor for UGIB in PD patients. Due to erythropoietin deficiency, iron metabolism abnormalities and malnutrition, anemia is highly prevalent in PD patients and is closely associated with their prognosis [25]. The decrease of hemoglobin is also a risk factor for UGIB which has been reported [26]. Decreased hemoglobin levels indicate impaired oxygen transport capacity, potentially causing localized tissue hypoxia that contributes to gastrointestinal mucosal injury and elevates GIB risk [27, 28]. Hemoglobin levels serve as one of the key indicators for assessing nutritional status. In PD patients, hemoglobin levels are not only associated with disease status but also reflect their nutritional condition [29]. Long-term malnutrition can lead to a decline in gastric mucosal protective and repair functions. Anemia reduces platelet function by reducing platelet-vessel wall interaction, reducing ADP release/inactivation of PGI2 and reducing nitric oxide clearance [30]. Platelet dysfunction is considered to be a key factor in UGIB in patients with uremia [31].

In PD patients, hyperphosphatemia is a common complication, and Calcium supplements remain one of the standard therapeutic options. Long-term use of Calcium supplements may cause calcium overload, lead to hypercalcemia, increase the risk of vascular calcification, and then increase the incidence of CVD and gastrointestinal side effects [32–34]. The results of this study indicate that the use of Calcium is a risk factor for UGIB in PD patients. Although we did not analyze the dosage of Calcium, it still suggests that clinicians should fully assess the gastrointestinal conditions of PD patients when administering drug therapy. PPIs can reduce gastric acid production and it is used in several upper gastrointestinal diseases, including non-hemorrhagic peptic ulcer, hemorrhagic peptic ulcer, gastroesophageal reflux, Barrett’s esophagus, eosinophilic esophagitis, indigestion and Helicobacter pylori infection. They are also used to prevent upper gastrointestinal ulcers and bleeding from antiplatelet or non-steroidal anti-inflammatory drugs [35]. The majority of patients using PPIs have a history of gastrointestinal diseases, and replacing different types of gastrointestinal diseases with PPIs drugs may simplify the research process.

The effect of antiplatelet drugs such as Aspirin on UGIB is still controversial. This study also reviewed the use of Aspirin, but did not find that it was associated with a high risk of UGIB in patients undergoing PD, which is consistent with previous studies [9, 10].

Figures 5 and 6 show that the model tends to overestimate UGIB risk when predicted probabilities are ≥ 30%. This overestimation may stem from two factors: first, the training set includes a low proportion of high-risk patients-although UGIB is a critical complication affecting outcomes in PD patients, its incidence remains low. Additionally, clinicians may have intensified management of high-risk patients in clinical practice, reducing actual event rates through proactive interventions. Therefore, when managing patients at extremely high UGIB risk, clinicians should incorporate real-time clinical data to avoid overestimation by static models, especially for stable patients.

This study has several strengths. First, we developed a prediction model for UGIB in a multicenter cohort of PD patients. Second, this nomogram model performs well in prediction, correction and clinical application. Finally, the factors included in the model are inexpensive, easily accessible, applicable to hospitals at different levels, and demonstrate excellent clinical promotion value.

However, there are several limitations in this research. First, although UGIB is an important complication in PD patients, its incidence remains relatively low. When a certain number of candidate variables are included, there may be a risk of overfitting. Subsequently, it is necessary to validate the model by incorporating multicenter data to expand the sample size and using an external independent cohort. Second, the prediction model has not undergone external validation. Additionally, although this study incorporated clinical prior knowledge and basic indicators for LASSO variable selection, the final determination of predictors remained dependent on data-driven effect size screening, which may have led to the exclusion of variables with weak effect sizes (such as albumin, diabetes, or other variables). Model validation relied only on a single randomly partitioned 1:1 train-test split. Finally, since the present study was a retrospective study, data on taking steroid and anticoagulant, and Helicobacter pylori infection status were not available, so there may be some bias in the later results. In the future, more center data need to be included for verification, and a large-scale prospective cohort study is also necessary.

Conclusion

In this study, we found that the use of calcium, the use of PPIs, a history of UGIB, hemoglobin, and uric acid were significantly correlated with the risk of UGIB in PD patients. we constructed a simple nomogram model to predict the risk of UGIB in PD patients. Clinicians can use this simple model to evaluate the risk of UGIB on admission in order to make better clinical decisions. It has certain clinical guiding significance for the prevention of UGIB in PD patients.

Abbreviations

ARBs

Angiotensin receptor blockers

CKD

Chronic kidney disease

CVD

Cardiovascular disease

DCA

Decision curve analysis

LASSO

Least absolute shrinkage and selection operator

NSAIDs

Nonsteroidal anti-inflammatory drugs

PD

Peritoneal dialysis

PPIs

Proton pump inhibitors

ROC

Receiver operating characteristic

RAAS

Renin-angiotensin-aldosterone system

UGIB

Upper gastrointestinal bleeding

WBC

White blood cells

Author contributions

Ning Su and Hongrui Shi conceived the design of this manuscript. Sijia Shang and Xing Zhang drafted the original manuscript. Xiaojiang Zhan, Na Tian, Fenfen Peng, Yueqiang Wen, Xiaoyang Wang, Xiaoran Feng Xianfeng Wu, Qingdong Xu provided the data. Qian Zhou checked the appraisals of risk of bias and applicability concerns. Ning Su accessed and was responsible for the raw data associated with the study. All authors commented on drafts of the paper.

Funding

No funding.

Data availability

The data underlying this article will be shared on reasonable request to the corresponding author.

Declarations

Ethics approval and consent to participate

This study was approved by the Scientific Ethics Committee of the Sixth Affiliated Hospital of Sun Yat-sen University (approval number: 2021SLYEC-177) and was conducted in accordance with the principles of the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.