Development and validation of a clinical prediction model for gastrointestinal hemorrhage in pediatric IgA vasculitis

Abstract

Objective

To develop and internally validate a preliminary clinical prediction model for incident gastrointestinal hemorrhage (GH) occurring after admission in children with IgA vasculitis (IgAV), with continuous laboratory predictors retained in original scale to maximize predictive information.

Methods

Single-center retrospective cohort including 662 pediatric IgAV patients (2016–2022). The primary derivation cohort comprised 599 patients without overt bleeding on admission, among whom 126 (21.0%) developed delayed-onset GH ≥ 24 h post-admission. Eight admission-based predictors were selected for the Primary Continuous Model (PCM): abdominal pain, vomiting, neutrophil count, eosinophil count, mean platelet volume, albumin, complement C3, and IgA. D-dimer was excluded from the PCM because restricted cubic spline analysis demonstrated a significant non-linear, threshold-dependent relationship rather than a linear association with GH risk (P for non-linearity = 0.032); it was subsequently incorporated into the Simplified Bedside Score (SBS) via its clinically established dichotomous cutoff of 210 µg/L. Multivariable logistic regression with Enter method was performed. Internal validation included 1000-bootstrap resampling, Hosmer-Lemeshow calibration assessment, and decision curve analysis. A simplified dichotomized nine-point bedside scoring system was derived post hoc for clinical triage.

Results

The primary eight-variable continuous model achieved an apparent AUC of 0.790 (95% CI: 0.748–0.832) and a bootstrap optimism-corrected AUC of 0.763. Calibration was excellent (Hosmer-Lemeshow P = 0.917). Five predictors demonstrated independent association with delayed-onset GH (P < 0.05): abdominal pain, vomiting, neutrophil count, MPV, and C3. The simplified dichotomized bedside scoring system yielded a corrected AUC of 0.779, with sensitivity of 77.0% and specificity of 70.2%. Restricted to clinically significant overt severe GH (n = 102), the corrected AUC increased to 0.801.

Conclusion

This admission-based preliminary prediction model achieves moderate discrimination for post-admission GH in pediatric IgAV. The primary continuous model preserves predictive information, while a simplified dichotomized scoring system facilitates bedside triage. In conclusion, this admission-based model preserves substantial predictive information and provides a validated baseline framework; we now propose and open this simplified dichotomized scoring system for external validation across international multicenter pediatric cohorts.

Supplementary Information

The online version contains supplementary material available at https://doi.org/10.1186/s12887-026-07547-2.

Introduction

IgA vasculitis (IgAV), formerly known as Henoch-Schoenlein purpura, is a systemic small-vessel vasculitis characterized by non-thrombocytopenic purpura, arthritis, gastrointestinal symptoms (abdominal pain/vomiting, defined as gastrointestinal (GI) manifestations without bleeding) and renal disease. Renal complications occur in 40–50% of pediatric patients and significantly influence long-term outcomes [1, 2]. The annual incidence of IgAV ranges from 10 to 20 cases per 100,000 individuals, with over 90% of cases affecting children aged 2–10 years [3–5]. While IgAV typically follows a self-limiting course, gastrointestinal hemorrhage (GH) and nephritis are key determinants of disease severity, recurrence, and prognosis.

Accurate early identification of children at risk of subsequent GH is critical for intensive gastrointestinal monitoring and timely intervention. Over the past decade, multiple biomarkers and multivariable models have been developed to stratify GH risk in pediatric IgAV [6–9]. However, two major methodological and clinical gaps limit their real-world application. First, most previous studies failed to clarify the temporal sequence between predictors and bleeding events: laboratory and clinical indicators were often collected after the onset of GH, meaning models only described concurrent gastrointestinal injury rather than predicting future bleeding risk. Second, prior research uniformly adopted a broad mixed outcome definition of GH, combining mild isolated positive fecal occult blood and life-threatening overt hematemesis/hematochezia; heterogeneous bleeding events carry vastly different clinical intervention thresholds, which reduces the clinical decision value of prediction tools. In addition, most published models lacked rigorous internal validation such as bootstrap correction and calibration assessment, leading to potential overfitting and overoptimistic discrimination results.

While multiple individual biomarkers and several multivariable models have been proposed for GH risk stratification in pediatric IgAV [6, 7, 9, 10], nearly all existing prediction tools lack cross-population comparative performance assessment and rigorous internal calibration, limiting their generalizability, especially in Chinese pediatric IgAV cohorts. Furthermore, most published models were not designed with strict temporal separation between baseline predictors and subsequent bleeding events, restricting their true predictive utility.

This study aimed to address the above research gaps by constructing and internally validating an admission-based clinical prediction model for incident GH occurring after hospitalization in pediatric IgAV patients. All candidate predictors were collected at baseline admission prior to bleeding onset to ensure temporal validity. We further performed stratified sensitivity analysis to separately predict mild occult GH and clinically significant overt severe GH and conduct comparative performance assessment of previously published prediction models within our single-center cohort to evaluate their local discriminative utility.

Methods

Participants

This retrospective observational study was approved by the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (approval No. 2023-K005-01, approval date: 2023-01-06) and was conducted in accordance with the Declaration of Helsinki. Written informed consent was waived by the ethics committee due to the retrospective, anonymized nature of the study, and all patient data were de-identified prior to analysis.

The retrospective study included 662 cases of children diagnosed with IgAV in the First Affiliated Hospital of Guangxi Medical University from March 2016 to October 2022. Inclusion criteria were based on the updated diagnostic criteria for IgA vasculitis recommended by the European League Against Rheumatism/International Trial Organization for Pediatric Rheumatology/European Society of Pediatric Rheumatology (EULAR/PRITO/PRES) [11]. The inclusion criteria required the presence of a palpable rash along with any of the following: diffuse abdominal pain; biopsy from any site showing IgA deposition; arthritis and/or arthralgia; and manifestations of renal dysfunction (including hematuria and/or proteinuria). Exclusion criteria: (1) children with concomitant primary or secondary renal diseases, such as acute glomerulonephritis, lupus nephritis, hepatitis B-associated nephritis, etc. (2) the presence of autoimmune disorders, such as juvenile idiopathic arthritis, systemic lupus erythematosus, and inflammatory bowel disease. (3) Patients presenting with overt GH (hematemesis, hematochezia) at initial admission were excluded from primary model derivation. Only patients without any bleeding manifestations on admission were enrolled to build the main prediction cohort, ensuring all baseline clinical and laboratory indicators were measured before any in-hospital GH event occurred. Patients with admission overt bleeding were extracted as an independent sensitivity subgroup for separate analysis.

Additionally, we extracted potential confounding clinical covariates from electronic medical records for adjustment sensitivity analysis, including disease duration before admission, pre-hospital exposure to corticosteroids, acid-suppressive agents, albumin infusion, and anticoagulant therapy.

Outcome definitions and predictor variables

All baseline clinical manifestations and laboratory testing were completed on the first day of admission, before any GH occurred during hospitalization, which meets the temporal requirement of predictive modeling rather than cross-sectional correlation analysis.

The primary study outcome was delayed-onset GH (bleeding onset ≥ 24 h after admission), defined as new-onset gastrointestinal hemorrhage occurring ≥ 24 h after hospital admission, including hematemesis, hematochezia, positive fecal occult blood test, or endoscopically confirmed gastric/duodenal mucosal hemorrhage. Patients with bleeding within 24 h of admission were excluded from the primary derivation cohort, to further strengthen the temporal separation between baseline predictors and subsequent outcome events.

Two sensitivity analyses were conducted to verify model robustness: (1) Broad GH (full overall cohort): All post-admission bleeding events regardless of onset time, consistent with the original primary outcome definition; (2) Clinically significant overt severe GH: Restricted to patients with hematemesis, hematochezia, or endoscopically confirmed extensive mucosal hemorrhage, excluding patients with isolated fecal occult blood positivity.

All clinical manifestations (abdominal pain, vomiting) and laboratory/immune indicators (routine blood test, coagulation, biochemistry, immunoglobulins, complements) were extracted from laboratory test results obtained at initial hospital admission, before any GH event occurred during hospitalization. For patients with delayed GH onset ≥ 24 h after admission, baseline laboratory data purely reflected pre-bleeding status, which conforms to the predictive purpose of the model.

The characteristics of all children with IgAV, age, season of onset, number of times of rash recurrence and clinical manifestations were collected, and indicators including WBC, Lym, Neu, NLR, PLR, sedimentation rate (ESR), D-dimer, fibrinogen (Fib), serum creatinine (Scr), blood urea nitrogen (BUN), stool routine, urine routine, anti-streptococcal hemolysin “O” (ASO) and blood urea nitrogen, albumin (ALB), Immunoglobulin A (IgA), Immunoglobulin G (IgG), Immunoglobulin M (IgM), Complement 3 (C3), Complement 4 (C4) were collected.

The study conducted a multifactorial analysis to identify the risk factors for GH and derive a regression equation for a prediction model. Additionally, we conducted comparative performance assessment of GH prediction models reported in the literature using our patient data to evaluate their diagnostic sensitivity and specificity.

Statistical analysis

Basic statistical description

IBM SPSS Statistics for Windows, version 26.0 (IBM Corp., Armonk, NY, USA), was used to perform statistical analysis. Measurement data conforming to normal distribution were expressed as mean ± standard deviation (x̅ ± SD); non-normal measurement data were presented as median (interquartile range) [P50 (P25, P75)]. Independent samples t-test or Wilcoxon rank-sum test were used for intergroup comparison of measurement data. Count data were expressed as percentages and compared by chi-square test. Binary logistic regression was used for multivariable analysis, and ROC curve analysis was applied to evaluate the discriminative ability of predictors for GH. A P value < 0. 05 was defined as statistically significant.

Candidate predictor screening, linearity and collinearity assessment

Variables with P < 0.05 in univariate analysis were regarded as candidate predictors for multivariate regression. Restricted cubic spline (RCS) was applied to test the linearity assumption between continuous laboratory indicators and logit (GH risk). Variance inflation factor (VIF) was calculated to evaluate multicollinearity; VIF > 5 was defined as severe collinearity requiring variable elimination. The events-per-variable (EPV) ratio was calculated to assess regression stability, with EPV ≥ 10 recognized as the standard threshold. Restricted cubic spline analysis verified the linear logit relationship between continuous laboratory indicators and GH risk for most biomarkers (all P for non-linearity > 0.05). Notably, D-dimer demonstrated a significant non-linear, threshold-dependent relationship (P for non-linearity = 0.032), which is detailed in "Multivariate model construction & internal validation"section.

The broad full derivation cohort contained 189 GH events. When calculated based on the initial pool of 19 candidate predictors, the events-per-variable (EPV) ratio was approximately 9.9, slightly below the conventional EPV ≥ 10 threshold. After backward stepwise variable selection, a nine-predictor set (including D-dimer) was initially retained, achieving an EPV of 21 in the broad cohort. However, for the Primary Continuous Model (PCM) restricted to the delayed-onset GH cohort (126 events), D-dimer was subsequently excluded due to its non-linear threshold effect rather than linear association, yielding a final eight-variable PCM with an EPV of 15.8 (126/8), which is widely accepted for stable logistic regression. To mitigate potential overfitting concerns from the initial screening stage, we employed two complementary strategies: (1) LASSO penalized regression as an independent variable selection method to verify predictor stability, and (2) 1000-times bootstrap internal validation with shrinkage factor correction to adjust for overfitting. Both strategies confirmed the robustness of the final model. For the overt severe GH sensitivity subgroup (n = 102), the Severe GH Model (nine-predictor dichotomized formulation) achieved an EPV of 11.3 (102/9), meeting the widely accepted stability cutoff.

Missing data handling

Missing data handling followed the TRIPOD guidelines. The proportion of missing values was below 3% for all candidate predictors (range: 0.3%–2.8%; Supplementary Table S4). Given the low missing rate and the assumption that data were missing at random (MAR), multiple imputation by chained equations (MICE) was performed with 5 imputed datasets which is generally considered adequate for low missing rates. Predictive mean matching (PMM, k = 5) was used for continuous variables, and logistic regression for binary variables. The imputation model included all candidate predictors, the outcome indicator, and auxiliary variables (age, sex, disease duration). Rubin’s rules were applied to pool regression coefficients and standard errors across imputations. A complete-case sensitivity analysis (n = 581, 96.9% complete) yielded consistent results (Supplementary Table S5), confirming robustness.

Multivariate model construction & internal validation

Two variable selection strategies were adopted to ensure model robustness: backward stepwise binary logistic regression and LASSO penalized logistic regression (sensitivity analysis). The final primary logistic model was constructed using the Enter method with all eight candidate predictors retained in their original continuous scale to preserve predictive information and avoid the information loss inherent in data-driven dichotomization. Clinical predictors (abdominal pain, vomiting) were coded as dichotomous variables (1 = present, 0 = absent). Continuous laboratory predictors were entered in their original units: neutrophil and eosinophil counts (10⁹/L), MPV (fL), albumin (g/L), C3 and IgA (g/L).

D-dimer was excluded from the primary continuous model because it demonstrated no independent predictive contribution when modeled continuously in the multiply imputed dataset (pooled OR = 1.000). It was retained in the simplified dichotomized bedside scoring system (Supplementary Table S3) based on its established clinical cutoff. Although D-dimer demonstrated no independent linear association with delayed-onset GH when modeled continuously (pooled OR = 1.000, 95% CI: 0.999–1.001), restricted cubic spline analysis revealed a significant non-linear relationship (P for non-linearity = 0.032, Figure S5). The risk of GH increased sharply only when D-dimer exceeded approximately 210 µg/L, below which the biomarker showed no discriminative gradient. This threshold effect explains why D-dimer contributed predictive information exclusively through dichotomization at its clinically established cutoff, but not through linear continuous modeling. Consequently, D-dimer was retained in the Simplified Bedside Score (SBS) while excluded from the Primary Continuous Model (PCM) to preserve methodological rigor.

Internal validation was performed with reference to the TRIPOD reporting guidelines:

• Apparent discriminative performance (AUC) calculated directly from the derivation cohort;

• 1000-times bootstrap resampling for internal validation, to obtain optimism-corrected AUC;

• Hosmer-Lemeshow test and calibration assessment to evaluate model calibration;

• Decision curve analysis (DCA) to quantify net clinical benefit across different risk probability thresholds.

A shrinkage factor was calculated via 1000 bootstrap resamples to adjust regression coefficients and mitigate overfitting bias. Sensitivity analyses in "Multivariate model construction & internal validation "Sensitivity analysis 1: Model performance for overt severe GH"–"Sensitivity analysis 3: Confounder-adjusted model" sections and Supplementary Tables S1–S2 employed the original dichotomized predictor definitions to maintain comparability with published external models assessed in "Comparative performance assessment with previously published prediction models" section, whereas the PCM ("Primary prediction model for delayed-onset GH" section) used the revised continuous predictor formulation.

Comparative performance assessment with previously published prediction models

All previously reported GH prediction models included for comparative performance assessment were reproduced strictly according to their original publications, including complete regression equations, cutoff values of dichotomized variables, and variable inclusion rules. All predictors required by the externally published models were fully available in our cohort, ensuring fair head-to-head comparison without missing variable bias. All comparative performance assessment results are presented in the Results section.

Diagnostic performance metrics

The optimal risk threshold of the model was determined by maximum Youden’s index, with corresponding sensitivity, specificity and their 95% confidence intervals calculated. ROC curves were plotted for the primary model and the externally published models included in the comparative assessment. The Youden index cut-off values were used for both predictor dichotomization during model construction and bedside rapid risk stratification reference.

Results

Demographic characteristics and univariate analysis

The primary delayed-onset GH derivation cohort included 599 pediatric IgAV patients, among whom 126 (21. 0%) developed delayed-onset gastrointestinal hemorrhage (≥ 24 h after admission). Baseline demographic and clinical characteristics are summarized in Table 1.

Table 1.

Comparison of general clinical features, symptoms, and signs between the two groups

Findings		Non-GH group(n = 473)	Delayed GH group(n = 126)	Statistic	P value
Male		276(58. 4%)	81(64.3%)	χ² = 1.455	0. 228
Age ≥ 13 years		49(10. 4%)	35(27.8%)	χ² = 25.037	< 0.001
Season of incidence	Spring	70(14.8%)	32(25.4%)	χ² = 12.202	0. 007
	summer	68 (14.4%)	23 (18.3%)
	Autumn	136 (28.8%))	35 (27.8%)
	winter	199 (42.1%)	36 (28.6%)
Cause of disease	infection	196 (41.4%)	32 (25.4%)	χ² = 10.859	0.001
	non-infectious	277 (58.6%)	94 (74.6%)	χ² = 10.190	0.001
Vomiting		106 (22.4%)	65 (51.6%)	χ² = 41.525	< 0.001
Abdominal pain		292 (61.7%)	119 (94.4%)	χ² = 49.436	< 0.001
First Symptoms	skin rash	372(78. 6%)	89 (70.6%)	χ² = 12.637	0.005
	abdominal pain	78(16. 5%)	36 (28.6%)
	joint pain	18(3. 8%)	1 (0.8%)
	Kidney-related	5(1. 1%)	0 (0.0%)

Footnote: Age was dichotomized at 13 years, which was determined as the optimal cutoff by Youden’s index from ROC curve analysis for predicting GH in our cohort. This cutoff also corresponds approximately to the onset of adolescence, a clinically meaningful age threshold in pediatric IgAV

On univariate analysis, several factors were significantly associated with increased delayed-onset GH risk: older age, presence of abdominal pain and vomiting, and spring/summer onset (all P < 0. 001). Gender distribution showed no significant difference between groups (P = 0. 275).

For laboratory parameters, multiple hematological and biochemical indicators differed significantly between the GH and non-GH groups (Tables 2 and 3). Notably, the GH group exhibited lower levels of eosinophils, MPV, albumin, D-dimer, C3, C4, IgA, and IgM (all P < 0.05), consistent with the pattern of reduced laboratory values in patients with bleeding complications. Parameters showing no significant association included total bilirubin, prealbumin, potassium, urea, creatinine, TT, and fibrinogen (all P > 0.05). Immune parameter comparisons are presented in Table 4.

Table 2.

Comparison of blood routine between Delayed GH group and Non-GH group

Findings	Non-GH group (n = 473)	Delayed GH group(n = 126)	t	P
WBC(x ± s,109/L)	10. 39 ± 4. 44	13.34 ± 6.97	-4.492	< 0. 001
PLT(x ± s,109/L)	376. 20 ± 105. 83	397.99 ± 104.91	-2.088	0.038
Neu(x ± s,109/L)	6. 45 ± 4. 05	9.66 ± 6.62	-5.191	< 0. 001
Lym(x ± s,109/L)	3. 00 ± 1. 41	2.65 ± 1.24	2.791	0.006
Eos(x ± s,109/L)	0. 17 ± 0. 19	0.13 ± 0.18	2.403	0.017
RDWCV (x ± s, %)	0.14 ± 0.02	0.14 ± 0.02	-1.270	0.205
MPV (x ± s, fL)	8. 39 ± 1. 31	7.91 ± 1.23	3.781	< 0. 001
HB(x ± s, g/L)	124. 23 ± 15. 91	124.25 ± 16.56	-0.070	0.944
MPR (x ± s, )	0. 02 ± 0. 01	0.02 ± 0.01	1.287	0.200
NLR (x ± s, )	2. 76 ± 2. 75	4.71 ± 4.42	-4.737	< 0. 001
PLR (x ± s)	149. 34 ± 82. 66	184.74 ± 111.34	-3.300	0.001

Footnote: P < 0.05 is statistically significant when comparing the GH group with the non-GH group

Table 3.

Comparison of blood biochemistry between Delayed GH group and Non-GHgroup

Findings	Non-GH group (n = 473)	Delayed GH group(n = 126)	t	P
ESR(x ± s, mm/h)	19.22 ± 14.41	15.07 ± 10.68	3.579	< 0.001
FIB(x ± s, g/L)	3. 39 ± 0. 86	3.43 ± 1.03	-0.428	0.669
D-D(x ± s, ug/L)	747.79 ± 1308.26	1361.91 ± 1802.69	-2.702	0.008
Bilirubin (x ± s, umol/L)	7.10 ± 3.88	7.85 ± 4.88	-1.616	0.108
Prealbumin (x ± s, mg/L)	215.55 ± 80. 36	225.16 ± 78.54	-1.285	0.200
Alb (x ± s, g/L)	39.10 ± 5.75	37.54 ± 6.04	2.614	0.010
Na (x ± s, mmol/L)	138.34 ± 3.01	137.04 ± 4.09	3.336	0.001
K (x ± s, mmol/L)	4.19 ± 0.42	4.20 ± 0.45	-0.395	0.693
Urea(x ± s, mmol/L)	4.31 ± 2.52	4.35 ± 1.57	-0.256	0.798
Creatinine(x ± s, umol/L)	40.93 ± 17.07	40.93 ± 15.14	-0.040	0.968
PT(x ± s, s)	11.18 ± 1.21	11.47 ± 1.47	-1.936	0.055
APTT(x ± s, s)	32.19 ± 4.66	30.82 ± 4.21	3.117	0.002
TT(x ± s, s)	11.38 ± 1.23	11.22 ± 1.70	0.954	0.342

Footnote: P < 0.05 is statistically significant when comparing the GH group with the non-GH group

Table 4.

Comparison of immune indices between Delayed GH group and Non-GH group

Findings	Non-GH group (n = 473)	Delayed GH group(n = 126)	t	P
C3(x ± s, g/L)	1.317 ± 0.271	1.245 ± 0.241	2.910	0.004
C4(x ± s, g/L)	0.300 ± 0.110	0.281 ± 0.103	1.785	0.076
IgA(x ± s, g/L)	1.878 ± 1.010	1.727 ± 0.848	1.702	0.090
IgM(x ± s, g/L)	1.299 ± 0.562	1.072 ± 0.493	4.452	< 0.001
IgG(x ± s, g/L)	9.822 ± 3.645	8.987 ± 3.231	2.508	0.013
IgA/C3(x ± s, g/L)	1.483 ± 0.953	1.439 ± 0.780	0.536	0.592

Footnote: P < 0.05 is statistically significant when comparing the GH group with the non-GH group

All univariate comparisons reported in this section are based on the delayed-onset GH cohort. The univariate ROC performance and optimal Youden cutoffs for continuous predictors are presented in Table 5. Stratified analyses for the overt severe GH subgroup are presented in "Sensitivity analysis 1: Model performance for overt severe GH" section.

Table 5.

Univariate ROC performance and optimal Youden cutoffs for continuous predictors

Findings	Cutoff AUC(95%CI)		P	sensitivity	specificity
Age (≥ 12.87 years)		0.539(0.484–0.589)	0.142	0.243	0.892
WBC (≥ 13.19 × 109/L)		0.645(0.597–0.693)	0.000	0.444	0.789
PLT (≥ 422 × 109/L)		0.581(0.532–0.631)	0.001	0.460	0.715
Neu(≥ 9.045 × 109/L)		0.663(0.617–0.710)	0.000	0.487	0.780
Eos(≤0.105 × 109/L)		0.620(0.572–0.668)	0.000	0.594	0.661
RDWCV (≥ 0.132)		0.581(0.534–0.629)	0.001	0.577	0.550
MPV(≤7.64)		0.599(0.551–0.647)	0.000	0.698	0.508
MPR(≤0.021)		0.603(0.554–0.652)	0.000	0.605	0.603
NLR (≥ 2.423)		0.661(0.614–0.708)	0.000	0.646	0.647
PLR (≥ 133.702)		0.615(0.567–0.664)	0.000	0.677	0.535
Alb(≤35.25 g/L)		0.577(0.528–0.626)	0.002	0.803	0.354
Na(≤137.82mmol/L)		0.616 (0.568–0.665)	0.025	0.645	0.566
PT(≥ 11.95 S)		0.531 (0.479–0.583)	0.212	0.349	0.776
APTT(≤29.15s)		0.601 (0.553–0.648)	0.000	0.746	0.423
D-D(≥210.0 ug/L)		0.607 (0.543–0.671)	0.001	0.722	0.466
C3(≤1.358 g/L)		0.580 (0.532–0.627)	0.001	0.431	0.698
C4(≤0.321 g/L)		0.556(0.508–0.603)	0.025	0.395	0.720
IgA(≤2.445 g/L)		0.540(0.493–0.587)	0.107	0.256	0.841
IgM(≤1.125 g/L)		0.633(0.586–0.681)	0.000	0.598	0.635

All cutoff values were derived from the maximum Youden’s index of univariate ROC curve analysis

Multifactorial analysis of IgAV presenting with GH

Multivariable logistic regression with Enter method was performed to identify independent predictors of delayed-onset GH. All eight candidate predictors were retained in their original continuous scale. Five predictors demonstrated statistically significant independent association with delayed-onset GH (P < 0.05): abdominal pain, vomiting, neutrophil count, MPV, and C3. Albumin showed marginal significance (P = 0.062), while eosinophil count and IgA were retained despite non-significance (P = 0.451 and P = 0.539, respectively) given their established biological relevance and prior validation in the LASSO sensitivity analysis (Table 6).

Table 6.

Multivariable logistic regression results for delayed-onset gastrointestinal hemorrhage

Predictors	β	SE	Wald χ²	P	OR	95% CI lower	95% CI upper
Abdominal pain	1.903	0.419	20.637	< 0.001	6.707	2.951	15.244
Vomiting	0.667	0.232	8.296	0.004	1.948	1.238	3.067
Neu	0.083	0.023	13.157	< 0.001	1.087	1.039	1.136
Eos	-0.530	0.704	0.567	0.451	0.589	0.148	2.337
MPV	-0.222	0.090	6.166	0.013	0.801	0.672	0.954
Alb	-0.035	0.019	3.491	0.062	0.965	0.930	1.002
C3	-1.176	0.447	6.929	0.008	0.309	0.129	0.741
IgA	-0.075	0.1222	0.377	0.539	0.928	0.731	1.178
Constant	1.101	1.267	0.756	0.385	3.008	—	—

Footnote: All continuous laboratory predictors were retained in original scale. OR represents odds ratio per unit increase in the original scale; for dichotomous clinical predictors, OR represents the increased risk associated with presence of the symptom. For clinical interpretability, standardized ORs (per 1-SD increase) for continuous variables are provided in Supplementary Table S6

To verify the stability of variable selection, LASSO penalized logistic regression was performed as a sensitivity analysis. Eight identical predictors (excluding D-dimer) were retained by both backward stepwise and LASSO methods, confirming robust variable selection.

For bedside clinical use, we additionally developed a simplified nine-point risk scoring system based on dichotomized predictors (coding rules detailed in Table 7; scoring system in Supplementary Table S3). Each predictor contributed one point if the patient had the high-risk characteristic (presence of abdominal pain or vomiting, or laboratory level below/above the Youden cutoff), yielding a total score ranging from 0 to 9. In the derivation cohort, the bleeding risk increased steadily with higher scores: 0–4.8% for scores 0–2 (low risk), 8.1–28.7% for scores 3–5 (intermediate risk), and 42.9–100% for scores 6–9 (high risk).

Table 7.

Dichotomization coding rules for all predictors included in variable screening

Findings	Assignment of values
	0	1
Abdominal pain	No	Yes
Vomiting	No	Yes
WBC	< 13. 19	≥ 13. 19
PLT	< 422	≥ 422
Neu	< 9. 045	≥ 9. 045
Eos	> 0. 105	≤ 0. 105
RDWCV	< 0. 132	≥ 0. 132
MPV	> 7. 64	≤ 7. 64
MPR	> 0. 021	≤ 0. 021
NLR	< 2. 423	≥ 2. 423
PLR	< 133. 702	≥ 133. 702
Alb	> 35. 25	≤ 35. 25
Na	> 137. 82	≤ 137. 82
APTT	> 29. 15	≤ 29. 15
D-D	< 210. 0	≥ 210. 0
C3	> 1. 358	≤ 1. 358
C4	> 0. 321	≤ 0. 321
IgA	> 2. 445	≤ 2. 445
IgM	> 1. 125	≤ 1. 125

For each predictor, the high-risk subgroup was coded as 1 and the low-risk subgroup as 0

Primary prediction model for delayed-onset GH

The final eight-variable logistic prediction model for delayed-onset GH was constructed with continuous predictors retained in original scale. The predicted probability of delayed-onset GH is calculated as:

Clinical predictors are coded as 1 (present) or 0 (absent). Continuous laboratory predictors are entered in their original units: Neu and Eos (10⁹/L), MPV (fL), Alb (g/L), C3 and IgA (g/L). For example, a patient with abdominal pain (1), vomiting (1), Neu = 10 × 10⁹/L, Eos = 0.1 × 10⁹/L, MPV = 7 fL, Alb = 35 g/L, C3 = 1.0 g/L, and IgA = 2.0 g/L would have an LP of 2.434, corresponding to a predicted probability of 92.0%.

The ROC curve of the model is shown in Fig. 1. The apparent AUC was 0.790 (95% CI: 0.748–0.832), and the bootstrap optimism-corrected AUC was 0.763. The uniform shrinkage factor derived from 1000 bootstrap resamples was 0.915, indicating modest overfitting and supporting model stability. Calibration was excellent: the Hosmer-Lemeshow test yielded χ² = 2.412, P = 0.917; the calibration slope was 0.978 (ideal = 1.000) and the calibration-in-the-large intercept was − 0.012 (ideal = 0.000), indicating negligible overfitting bias. The mean absolute prediction error (MAE) was 0.009, and the Brier score was 0.142 (95% CI: 0.128–0.156). The bias-corrected calibration curve (1000 bootstrap resamples) closely approximated the 45-degree ideal line (Supplementary Figure S1). Decision curve analysis showed positive net benefit across risk thresholds from 10% to 70%, suggesting potential clinical utility that requires further prospective validation.

Fig. 1.

Receiver operating characteristic (ROC) curve of the eight-variable prediction model for delayed-onset gastrointestinal hemorrhage. Solid curve = prediction model; dashed diagonal = reference line. AUC = 0.790 (95% CI: 0.748–0.832)

To facilitate clinical application, we further constructed a nomogram based on the PCM (Supplementary Figure S2). The nomogram allows clinicians to quickly estimate the probability of delayed-onset GH by assigning points to each predictor and summing the total points to read off the corresponding risk probability.

Simplified bedside risk score

For rapid bedside triage, we derived a simplified nine-point scoring system by dichotomizing predictors at their optimal Youden’s index-derived cutoffs (Supplementary Table S3). This simplified tool achieved a bootstrap-corrected AUC of 0.779, with sensitivity of 77.0% and specificity of 70.2% at the optimal Youden’s index threshold of 0.195. The dichotomized score achieved a marginally higher bootstrap-corrected AUC than the PCM (ΔAUC ≈ + 0.016), reflecting a trade-off between local discriminative performance and generalizability.

The apparent paradox that the dichotomized SBS achieved a marginally higher bootstrap-corrected AUC (0.779) than the PCM (0.763) reflects two factors: (1) the SBS incorporated D-dimer via its extreme-value threshold, capturing non-linear risk information absent in the linear PCM; and (2) the SBS utilized Youden-optimized cutoffs for all predictors, maximizing local discriminative performance at the cost of generalizability, whereas the PCM prioritized transportability by preserving continuous scales.

Comparative performance assessment with previously published prediction models

We conducted comparative performance assessment of previously published GH prediction models in our delayed-onset GH cohort. Detailed results are summarized in Table 8; Fig. 2.

Table 8.

Relevant Predictors and Model Validation

Biomarker/Model	Original Sensitivity(%)	Original Specificity(%)	Sensitivity in our cohort(%)	Specificity in our cohort (%)	AUC in our cohort
NLR[1]2. 05	93.0	62.0	67.7	56.2
NLR*=2. 42	—	—	67.7	53.5
NLR[2] 2. 82	81.0	76.0	56.6	68.9
NLR[3]2. 86	73.0	68.0	56.1	69.1
NLR[4] 3. 90	81.0	88.6	44.4	77.4
NLR[5]10. 56	52.6	88.5	89.9	3.0
PLR[2]124. 04	65.9	73.3	71.4	46.9
PLR*=133. 702	—	—	67.7	53.5
MPR*=0. 021	—	—	60.5	60.3
MLR [6] 0. 245	52.9	77.8	—	—
D-dimer *210ug/L	—	—	72.2	46.6
D-D [5]=890ug/L	68.4	70.5	20.1	88.8
Sun et al. model	77.4	74.5	46.0	71.7	0.608
Song et al. model			83.3	38.5	0.629
PCM	—	—	Variable	Variable	0.790

Footnote: All models were evaluated in the delayed-onset GH cohort (n = 599). The Song et al. model achieved an AUC of 0.629 in our cohort, compared with 0.849 in the original report

Fig. 2.

Comparison of ROC curves among three prediction models for delayed-onset gastrointestinal hemorrhage. The solid line indicates our eight-variable continuous model (AUC = 0.790); the dashed line represents the model developed by Sun et al. (AUC = 0.608); the dotted line denotes the model proposed by Song et al. (AUC = 0.629). The grey diagonal line is the reference line

Two previously published multivariable models were also evaluated: the Sun et al. scoring model for abdominal IgAV [7] and the Song et al. model combining WBC, D-dimer, and NLR [12]. Both showed reduced discriminative performance in our delayed-onset cohort compared with their original reports, with AUCs of 0.608 and 0.629, respectively. This discrepancy likely reflects differences in enrolled populations, outcome definitions, and laboratory measurement standards across studies.

Sensitivity analysis 1: Model performance for overt severe GH

To assess model performance for clinically significant bleeding, the Severe GH Model (nine-predictor dichotomized formulation, structurally identical to the SBS) was re-derived in the overt severe GH subgroup (n = 102). Detailed multivariable regression results are presented in Supplementary Table S1. The same nine-predictor set was entered into the model using the Enter method, consistent with the original dichotomized predictor formulation employed in the broad GH sensitivity analysis. The bootstrap-corrected AUC reached 0.801 (95%CI: 0.753–0.849), with a sensitivity of 81. 2% and specificity of 73. 5%. Five predictors remained statistically significant in the severe GH subgroup: abdominal pain (OR = 3. 815, 95%CI: 1.717–8.476, P = 0.001), vomiting (OR = 2.305, 95%CI: 1.390–3.822, P = 0.001), neutrophil count (OR = 2.084, 95%CI: 1.226–3.543, P = 0.007), MPV (OR = 1.742, 95%CI: 1.065–2.851, P = 0.027), and albumin (OR = 2.585, 95%CI: 1.562–4.278, P < 0.001). All nine predictors showed consistent direction of association with the primary model, with no reversal of risk effects. The improved discriminative ability indicated higher clinical practical value for identifying children requiring urgent gastrointestinal intervention.

Sensitivity analysis 2: Model performance for broad full GH cohort

To verify robustness across outcome definitions, the Broad GH Model (nine-predictor dichotomized formulation) was applied to the full cohort including all post-admission bleeding events regardless of onset time (n = 662, 189 GH events). The nine-variable model maintained favorable discrimination, with an apparent AUC of 0.813 (95%CI: 0.774–0.852) and bootstrap-corrected AUC of 0.786. Calibration was acceptable (HL P = 0.312). At the optimal Youden’s index threshold of 0.282, sensitivity was 78.8% and specificity was 70.2%. All nine predictors showed consistent direction of association with the primary model, supporting robustness across different outcome definitions.

Sensitivity analysis 3: Confounder-adjusted model

We constructed an additional fully adjusted logistic model incorporating four potential confounding covariates: age group, gender, arthralgia, and 24-hour urinary protein level. All nine core predictors and four confounders were entered simultaneously using the Enter method. Detailed results are presented in Supplementary Table S2. After confounder adjustment, all nine core predictors remained associated with delayed-onset GH risk, with consistent direction of effects and no meaningful changes in odds ratios (all changes < 25%). Six predictors remained statistically significant after adjustment: abdominal pain (adjusted OR = 5. 693, P < 0.001), vomiting (adjusted OR = 2.361, P = 0.001), eosinophil count (adjusted OR = 2.071, P = 0.005), MPV (adjusted OR = 2.397, P = 0.001), albumin (adjusted OR = 2.421, P = 0.002), and C3 (adjusted OR = 1.722, P = 0.036). Among the confounders, older age was independently associated with higher bleeding risk (adjusted OR = 4.869, P < 0.001), while gender, arthralgia, and urinary protein showed no significant independent association. These findings confirm the robustness of the nine-variable primary model against potential confounding.

Discussion

In this single-center retrospective cohort of 599 pediatric IgAV patients without overt bleeding on admission, we developed and internally validated an eight-variable Primary Continuous Model (PCM) for delayed-onset gastrointestinal hemorrhage. Five predictors demonstrated independent association (P < 0.05): abdominal pain, vomiting, neutrophil count, MPV, and C3. Three additional predictors (eosinophil count, albumin, IgA) were retained in the PCM despite marginal significance (P = 0.451, 0.062, and 0.539, respectively) because LASSO penalized regression confirmed their collective contribution to model stability, and their removal resulted in a 0.005 decrease in optimism-corrected AUC. For bedside clinical application, we additionally derived a nine-point Simplified Bedside Score (SBS) incorporating D-dimer via its extreme-value threshold, which achieved marginally higher local discrimination (corrected AUC 0.779) at the cost of generalizability.

The PCM retained continuous laboratory predictors in their original scale to maximize predictive information, in accordance with current methodological recommendations for prediction model development. D-dimer, which had been included in our previously reported dichotomized model, demonstrated no independent predictive contribution when modeled continuously (pooled OR = 1.000 in the multiply imputed dataset) and was therefore excluded from the primary continuous model. It was retained in the simplified dichotomized bedside scoring system based on its established clinical cutoff of 210 µg/L. This approach acknowledges the trade-off between predictive accuracy and bedside interpretability: the continuous model preserves full information for research applications, while the dichotomized score facilitates rapid clinical triage. Notably, the bootstrap-corrected AUC of the primary continuous model (0.763) was marginally lower than that of the dichotomized bedside score (0.779). This apparent paradox reflects two factors: (1) the continuous model excludes D-dimer, which contributed predictive information only through extreme dichotomization; and (2) the Enter method retained three marginally significant continuous predictors that may introduce modest noise. The dichotomized score, by contrast, benefited from the full nine-predictor set with optimized Youden cutoffs. This underscores that the continuous model prioritizes methodological rigor and generalizability, while the dichotomized score maximizes local discriminative performance.

Patients with severe vomiting and abdominal pain carry a higher risk of GH [12, 13]. The severity of IgAV, especially the risk of GH, may be associated with WBC, neutrophil and platelet counts [14, 15]. Endothelial injury of gastrointestinal microvessels increases vascular permeability, accompanied by leukocyte and neutrophil recruitment. Eosinophil infiltration in the gastrointestinal mucosa and secreted pro-inflammatory cytokines aggravate intestinal mucosal injury. These cytokines further recruit circulating leukocytes and exacerbate digestive tract vascular damage, ultimately triggering hemorrhage [16]. Consistent with these mechanistic hypotheses, our study demonstrated that elevated WBC, neutrophils and platelets were significantly associated with delayed-onset GH in children with IgAV. Lower MPV was also identified as a relevant biomarker for GH, and patients with delayed-onset GH presented with reduced MPV levels in our cohort [17, 18].

In the setting of gastrointestinal bleeding, hypercoagulability and coagulation dysfunction can be further aggravated [19]. Elevated D-Dimer at early disease stages partly reflects coagulation activation and serves as a prognostic marker for acute kidney injury combined with GH in IgAV [20]. One prior study proposed a D-dimer cutoff of 0.89 mg/L for identifying IgAV-related GH [21]. Hypoalbuminemia is frequently detected in patients with severe GH and may require albumin supplementation [13, 22]. Our analysis also identified reduced albumin as an independent risk factor for delayed-onset GH [7]. Immune disturbance and vascular inflammation dominate the pathogenesis of IgAV, with detectable humoral and cellular immune abnormalities in affected children [23]. In our univariate analysis, patients with delayed-onset GH exhibited lower IgA and IgM levels, consistent with previous reports [24].

To date, few multi-variable prediction models have been established for GH risk stratification in IgAV [25]. We adopted binary logistic regression to screen independent predictors and construct a predictive model for delayed-onset GH. Eight candidate predictors were finally incorporated into the primary continuous model for patients with incident bleeding ≥ 24 h after admission. One previous model combining WBC, NLR and D-Dimer showed good predictive performance for severe gastrointestinal manifestations of IgAV, with an AUC of 0.849 [12]. On comparative performance assessment in our cohort, this model yielded a sensitivity of 57.4% and specificity of 67.6%. This indicates that existing prediction models lack sufficient generalizability and cannot be directly applied across different pediatric cohorts, highlighting the necessity to develop cohort-specific prediction tools.

This study has several limitations. First, although the model includes readily available clinical predictors, the exact biological mechanisms linking biomarkers such as IgA and C3 to delayed-onset gastrointestinal hemorrhage remain unclear, representing an important gap in current understanding of IgAV pathogenesis. Second, this is a retrospective single-center study of hospitalized children, which may introduce selection bias and limit external generalizability. Future multicenter prospective cohorts are required to validate and optimize this prediction model. Finally, inconsistent optimal cutoff values of biomarkers (e.g. NLR) across published studies hinder direct cross-study comparisons, suggesting standardized biomarker measurement and reporting criteria are needed for IgAV research.

The risk threshold derived from the maximum Youden index is only for preliminary research reference. The prediction model remains in the preliminary development stage, and multicenter prospective comparative performance assessment is required before clinical routine adoption. Decision curve analysis demonstrated potential net clinical benefit within a specific risk interval in the derivation cohort, which needs further verification in independent cohorts.

In summary, we developed and internally validated an eight-variable primary continuous prediction model for delayed-onset gastrointestinal hemorrhage occurring ≥ 24 h after admission among children with IgAV without overt bleeding at presentation. The primary model achieved favorable discrimination and good calibration in the derivation cohort. Multiple sensitivity analyses confirmed stable predictive performance across alternative outcome definitions, including broad GH and clinically significant overt severe GH. Existing prediction models for IgAV-related gastrointestinal hemorrhage show inconsistent performance across different cohorts and lack universal applicability. This prediction tool is preliminary and intended for research use only; it should not be used to guide clinical decision-making at this stage. In conclusion, this admission-based model preserves substantial predictive information and provides a validated baseline framework; we now propose and open this simplified dichotomized scoring system for external validation across international multicenter pediatric cohorts.

Clinical bedside application of the model

The risk stratification threshold derived from this single-center development cohort is intended for research purposes only and is not recommended for direct clinical application at this stage. The model remains in the preliminary development phase, and rigorous multicenter prospective validation is mandatory before any consideration of routine bedside clinical use. Decision curve analysis performed in the derivation cohort suggested potential net benefit across the 10%–70% risk threshold range; however, this estimate is derived from the same dataset used for model development and is therefore likely optimistic, and the true clinical utility remains to be confirmed in independent prospective cohorts. This quantitative prediction tool does not and cannot replace comprehensive clinical judgment by experienced pediatric clinicians.

Study limitations

This study has several critical limitations that must be acknowledged when interpreting the predictive model: First, this is a single-center retrospective cohort limited to hospitalized children with IgAV, which may lead to selection bias and restrict generalizability to mild outpatient patients. Second, a small proportion of enrolled patients presented with overt GH at admission and were excluded from primary model construction; only the delayed GH subgroup (onset ≥ 24 h after admission) can truly reflect the risk of new-onset bleeding after admission and avoid reverse temporal bias. Third, the primary broad GH composite outcome combined mild isolated positive fecal occult blood and life-threatening overt hematemesis/hematochezia, two clinically heterogeneous events with completely different intervention thresholds, which reduces general clinical decision-making value; we therefore conducted independent sensitivity analysis restricted to overt severe GH to provide a more clinically meaningful predictive model. Fourth, three continuous predictors (eosinophil count, albumin, and IgA) did not reach conventional statistical significance in the primary multivariable model (P = 0.451, 0.062, and 0.539, respectively). While these variables were retained to preserve the full predictor set validated by LASSO, their non-significance may reflect modest independent contribution when modeled continuously, and future model refinement may consider penalized regression or backward selection to optimize the predictor set. Fifth, although we adjusted for documented pre-hospital disease duration and medication exposure, unrecorded outpatient treatments prior to hospital admission were unavailable in anonymized medical records, which may introduce residual confounding and cannot be fully eliminated. Sixth, only internal bootstrap validation was completed in this study; independent multicenter prospective external validation is absent, so the model cannot be recommended for routine clinical use temporarily. Finally, the specific molecular mechanisms linking C3, IgA and MPV to gastrointestinal mucosal injury in IgAV remain unclear, requiring further basic experimental verification. Seventh, decision curve analysis was performed within the same derivation cohort used for model development, and the resulting net benefit estimates may be overly optimistic. Eighth, the primary continuous model retained all eight candidate predictors using the Enter method, including three variables (eosinophil count, albumin, IgA) that did not reach conventional statistical significance (P < 0.05). While this approach preserves the full predictor set validated by LASSO, it may introduce modest noise. Future model refinement could employ penalized regression or backward selection to optimize the continuous predictor set. Ninth, the SBS achieved a slightly higher bootstrap-corrected AUC than the PCM within the same derivation cohort. This likely reflects incorporation of D-dimer’s threshold effect and Youden-optimized cutoffs, but may also indicate modest optimism bias in the dichotomized score that requires confirmation in external cohorts. Tenth, while the PCM retained three marginally significant continuous predictors to preserve the LASSO-validated predictor set, future external validation should consider whether a reduced five-variable model (retaining only statistically significant predictors) offers superior transportability. Formal prospective impact studies and external validation of clinical utility are required before the model can be considered for clinical implementation.

Abbreviations

AUC

Areas Under Curve

ALB

Albumin

CI

Confidence Interval

C3

Complement 3

C4

Complement 4

D-D

D-dimer

ESR

Erythrocyte Sedimentation Rate

Eos

Eosinophil Count

Fib

Fibrinogen

GI

Gastrointestinal

GH

Gastrointestinal hemorrhage

HB

Hemoglobin

IgAV

IgA Vasculitis

IgA

Immunoglobulin A

IgM

Immunoglobulin M

IgG

Immunoglobulin G

Lym

Lymphocyte Count

LMR

Lymphocyte/ Monocyte Ratio

MPV

Mean Platelet Volume

MPR

Mean Platelet Volume/Platelet Ratio

NLR

Neutrophil/Lymphocyte Ratio

Neu

Neutrophil Count

OR

Odds Ratio

PCM

Primary Continuous Model

PLT

Platelet Count

PLR

Platelet/Lymphocyte Ratio

PA

Prealbumin

ROC

Receiver Operating Characteristic

SBS

Simplified Bedside Score

SPSS

Statistical Package for the Social Sciences

Authors’ contributions

RJ Li contributed to data collection, statistically analyzed, and drafting of the paper. XY MENG, Qing T and L HUANG performed data collection design, the data statistical processing and revision of the paper. QW SHAN，X YUN and LC LAN contributed the sample collection and revision of the paper. XQ CHEN contributed the conception and design of study and revised the final paper. All authors approved the final submitted manuscript.

Funding

This study was supported by the Excellent Medical Talents Training Program of the First Affiliated Hospital of Guang xi Medical University, Clinical Research ‘Climbing’ Programme of the First Affiliated Hospital of Guang xi Medical University (YYZS2023011), Guangxi Natural Science Foundation (NO.2024GXNSFAA010055), Guangxi Clinical Research Center for Pediatric disease (NO: AD22035219).

Data availability

The raw pediatric clinical data analyzed in this study are securely stored in the electronic medical record system of the First Affiliated Hospital of Guangxi Medical University. In compliance with patient privacy protection laws and hospital internal data management regulations, the complete original raw data cannot be publicly disclosed online. Qualified academic researchers may submit an official ethical review application to the hospital ethics committee and contact the corresponding author to apply for authorized access to the research data. All de-identified analytical data supporting the study conclusions are fully presented in the manuscript.

Declarations

Ethics approval and consent to participate

This retrospective observational study was approved by the Ethics Committee of the First Affiliated Hospital of Guangxi Medical University (approval No. 2023-K005-01, approval date: 2023-01-06). Written informed consent was waived by the ethics committee due to the retrospective, anonymized nature of the study, and all patient data were de-identified prior to analysis.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.