A predictive nomogram for hemorrhage risk following endoscopic submucosal dissection of gastrointestinal tumors: a retrospective cohort study

Abstract

Objectives

To develop a nomogram-based risk prediction model utilizing patients' clinical characteristics for post-endoscopic submucosal dissection hemorrhage, and to evaluate its clinical utility.

Methods

A total of 250 patients who developed postprocedural hemorrhage after endoscopic submucosal dissection (ESD) for gastrointestinal tumors at our institution (2022–2024) were enrolled. Enrichment criteria included at least one high-risk factor for complications (e.g., lesion size > 10 mm, antithrombotic medication use, or comorbid diabetes/hypertension) or early post-ESD symptoms suggestive of complications. Patients were randomly divided into a training set (n = 175) and a validation set (n = 75) at a 7:3 ratio. In the training set, multivariate logistic regression identified independent prognostic risk factors to construct the nomogram. Model performance was assessed via receiver operating characteristic (ROC) curves and calibration plots, with external validation performed in the validation set.

Results

Hemorrhage occurred in 70/175 cases (40.00%) in the training set and 28/75 (37.33%) in the validation set, with no statistically significant intergroup differences in incidence or baseline characteristics (P > 0.05). Univariate analysis revealed significant disparities between hemorrhage and non-hemorrhage groups in age, hemoglobin, white blood cell count, platelet count, prothrombin time, activated partial thromboplastin time, and fibrinogen (P < 0.05). Multivariate analysis confirmed age, hemoglobin, white blood cell count, platelet count, prothrombin time, activated partial thromboplastin time, and fibrinogen were independent risk factors (P < 0.05). The nomogram demonstrated C-index were 0.861 and 0.841 (validation), and mean absolute errors were 0.151 and 0.171, respectively. AUC values were 0.849 (95% CI 0.778–0.920) and 0.824 (95% CI 0.707–0.941), indicating high predictive accuracy.

Conclusions

The clinical feature-based nomogram exhibits good predictive performance and reliability in both training and validation cohorts, serving as a valuable tool for prognostic evaluation in gastrointestinal tumor patients who experience post-ESD hemorrhage.

Introduction

Minimally invasive treatment for gastrointestinal tumors has emerged as a focal point in clinical research. Endoscopic submucosal dissection (ESD), as a cornerstone technique for early stage gastrointestinal tumors, enables en bloc resection of early carcinomas and precancerous lesions while maximally preserving normal tissue structure and function. This approach avoids the substantial trauma associated with traditional surgery, significantly improving patients' quality of life [1]. Hemorrhage, the most frequent complication affecting therapeutic prognosis in ESD, is categorized into intraoperative and postoperative delayed bleeding [2]. Both types of hemorrhage may lead to a series of adverse effects on patients. Minor bleeding could result in anemia-related symptoms, such as dizziness and fatigue, potentially delaying postoperative recovery [3]. Severe hemorrhage, particularly active massive bleeding, may rapidly precipitate shock and pose an immediate threat to patient survival if not promptly managed [4]. Furthermore, bleeding events affecting therapeutic outcomes often necessitate additional interventions, such as repeat endoscopic hemostasis or blood transfusion, which not only increase patient discomfort and psychological distress but also prolong hospitalization and escalate healthcare costs [5]. Clinically, it is well-established that prognostic outcomes following gastrointestinal tumor treatment result from multifactorial interactions. From the perspective of patient-specific characteristics, advanced age constitutes a significant risk factor [6]. Elderly patients exhibit diminished physiological reserve, reduced vascular elasticity, and potential coagulation dysfunction, collectively predisposing them to post-ESD hemorrhage [7]. While no definitive gender-based disparity has been established, some studies suggest that physiological variations (e.g., menstrual cycles) may alter coagulation profiles in females, potentially correlating with bleeding-related prognostic outcomes [8]. Comorbidities are critical considerations: patients with cardiovascular diseases (e.g., hypertension, coronary artery disease) on long-term antiplatelet/anticoagulant therapy demonstrate impaired hemostasis, while those with hepatic dysfunction may exhibit coagulopathy due to deficient clotting factor synthesis—both scenarios substantially elevate prognostic risks [9]. Lesion-related factors also critically influence bleeding-associated prognostic outcomes. Anatomical location determines surgical complexity and risk, for instance, the highly vascularized esophagus is prone to iatrogenic vascular injury, whereas lesions at challenging sites (e.g., gastric lesser curvature) may increase bleeding likelihood due to restricted operative visibility [10]. Lesion size correlates positively with prognostic risk, as larger resections encompass broader vascular territories and create wider wound surfaces, complicating hemostasis [11]. Irregular morphology (e.g., ulcerated or poorly demarcated lesions) further elevates intraprocedural bleeding risk [12]. Operator-dependent factors, including surgical experience and technical proficiency, significantly modulate bleeding incidence [13]. Consequently, the development of a clinical feature-based prognostic prediction model is imperative for early high-risk patient identification and treatment optimization.

Methods

Study design and subjects

This was a single-center, retrospective prognostic study specifically targeting patients at high risk of ESD complications. The study aim was to develop a risk prediction model tailored to this high-risk population, given their unmet clinical need for precise risk stratification. A total of 328 consecutive patients who underwent ESD for gastrointestinal tumors at our institution from January 2022 to December 2024 were initially screened. From this population, a total of 250 patients who developed postprocedural hemorrhage, defined as our primary prognosis-related complication, were enrolled in this retrospective cohort study. It is crucial to note that the cohort for this predictive model study was intentionally enriched with patients who experienced post-ESD hemorrhage. Therefore, the subsequent hemorrhage incidence rates reported in our results are specific to this selected, complication-enriched cohort and are not representative of the overall incidence in an unselected ESD population. This study design was necessary to ensure a sufficient number of events for robust model development. Using a computer-generated random number table, patients were allocated into a training set (n = 175) and a validation set (n = 75) at a 7:3 ratio. Baseline demographic and clinical data were collected prospectively. Hemorrhage was selected as the primary endpoint for prognostic evaluation, and a risk prediction model was subsequently developed. The patient screening process is illustrated in Fig. 1.

Fig. 1.

Patient screening flowchart

Inclusion and exclusion criteria

Inclusion criteria

1. All patients underwent endoscopic submucosal dissection (ESD) with complete surgical records, including operative duration and procedural details.

2. Development of postprocedural hemorrhage (as defined below).

3. Complete medical records were available.

4. Postoperative pathological confirmation of gastrointestinal neoplasms (including early stage carcinoma and precancerous lesions).

Exclusion criteria

1. Patients with non-procedural bleeding (e.g., trauma-induced hemorrhage, medication-related mucosal injury [non-prescribed drugs], or undiagnosed pre-existing bleeding disorders, such as vascular malformations).

2. Patients with missing or inaccurate key clinical data.

3. Intraprocedural deviations, including incomplete ESD due to unforeseen circumstances (e.g., equipment failure) or concurrent interventions that could confound prognostic risk assessment (e.g., adjunctive interventional therapies).

Initial variable selection process

To ensure no major potential predictors were overlooked, we constructed an initial candidate variable pool based on a systematic review of ESD hemorrhage prognostic studies and clinical expertise (covering demographic, clinical, laboratory, lesion, and procedural dimensions). The full list of initially considered variables is as follows:

Demographic variables: age, gender, and body mass index (BMI). Clinical baseline variables: hypertension, diabetes mellitus, coronary heart disease, and history of bleeding. Laboratory test variables: hemoglobin (Hb), white blood cell (WBC) count, platelet count, prothrombin time (PT), activated partial thromboplastin time (APTT), fibrinogen, albumin, and creatinine. Lesion-related variables: lesion site (upper digestive tract/stomach/duodenum), pathological morphology (uplift type/recessed type/flat type).

Hematological parameter analysis

Venous blood samples were collected under aseptic conditions using vacuum tubes and needles. After proper disinfection, blood was drawn and transferred into EDTA-anticoagulated tubes. Samples were processed using an automated hematology analyzer (calibrated and quality-controlled) to measure Hb concentration, WBC count and differential, and PLT count via optical/electrical impedance methods. Results were recorded as absolute values.

Coagulation function assessment

Blood was collected in sodium citrate-anticoagulated tubes and analyzed using a coagulation analyzer.

PT: Thromboplastin reagent was added to initiate the extrinsic pathway, with clot formation detected mechanically/optically.

APTT: A contact activator triggered the intrinsic pathway, and clotting time was recorded.

FIB: Quantified via the Clauss method (thrombin-induced fibrin polymerization).

Biochemical parameter measurement

Prior to testing, the biochemical analyzer was calibrated, and quality control procedures were performed to ensure optimal performance. All necessary reagents and consumables were prepared. Venous blood samples were collected under aseptic conditions and transferred into standard serum-separating tubes.

For analysis, blood samples were centrifuged to isolate serum. The serum was then loaded into the biochemical analyzer for measurement:

Albumin levels were determined using colorimetric or immunoturbidimetric assays.

Creatinine levels were quantified via specific enzymatic/chemical reactions.

Upon completion of analysis, the instrument automatically generated numerical values for albumin and creatinine concentrations.

Definition and ascertainment of hemorrhage

The primary endpoint, postprocedural hemorrhage, was defined as the occurrence of any one of the following criteria within 30 days after ESD:

1. Overt clinical manifestation: hematemesis, melena, or hematochezia.

2. Endoscopic confirmation: active bleeding, visible vessel, or adherent clot at the ESD ulcer site during a follow-up endoscopy.

3. Laboratory criterion: a decrease in hemoglobin level of ≥ 2 g/dL from the pre-operative baseline, in the absence of other causes for anemia.

Statistical analysis

Data were analyzed using SPSS 26.0. The enumeration data were described by frequency and percentage, and chi-square test was used. Measurement data of normal distribution were expressed with mean ± standard deviation, and independent sample t test was used. All key variables had no missing values, and complete-case analysis was adopted [14]. Univariate logistic regression analysis was used for initial screening: variables with P < 0.05 were included in the subsequent multivariable logistic regression. Variables excluded after univariate screening (e.g., gender, BMI, ALT, and Scr) had P ≥ 0.05, indicating no significant association with post-ESD hemorrhage. Logistic regression analysis was used to establish the model, and the visual model of rms package in R 4.3.1 was applied. The events-per-variable (EPV) in the multivariable model was calculated as 10 (70 hemorrhage events in the training set divided by seven independent variables) [15]. Bootstrap optimism correction (1000 resamplings) with calibration-slope shrinkage was performed to adjust the model performance [16]. The model discrimination was assessed by area under ROC curve (AUC) and calibration curve. The Clinical Decision Curve Analysis (DCA) was used to test the actual application effectiveness of the model. A two-tailed p value < 0.05 was considered statistically significant. To evaluate parsimony vs performance trade-offs, three comparative models were constructed: (1) simple baseline model (age + PT + APTT only); (2) LASSO regression model (including all candidate variables with regularization for feature selection); and (3) Random Forest model (tree-based learner with 500 decision trees). Model performance was compared using AUC, C-index, and Brier score. For subgroup analysis, patients were stratified by lesion site (esophagus, stomach, duodenum), antithrombotic status (used, not used), and age strata (< 60 years, 60–70 years, and > 70 years). Subgroup-specific discrimination (AUC, C-index) and calibration (calibration slope/intercept) were reported. Interaction tests were performed by adding interaction terms (e.g., lesion site × model variables) to the primary model, with P < 0.05 indicating significant interaction [17].

Results

Comparison of prognostic hemorrhage incidence and clinical characteristics between training and validation sets

From the overall cohort of patients undergoing ESD, 250 who experienced postprocedural hemorrhage were included in the analysis. In the training set (n = 175), postprocedural hemorrhage (a key prognostic indicator) occurred in 70 cases (40.00%), while the validation set (n = 75) had 28 cases (37.33%). No statistically significant differences were observed between the two cohorts in terms of hemorrhage incidence or baseline clinical characteristics (all P > 0.05, Table 1).

Table 1.

Comparison of hemorrhage incidence and clinical characteristics between training and validation sets

Factors		Training set (n = 175)	Validation set (n = 75)	t/χ2	P
Age (years)		57.54 ± 6.83	57.24 ± 7.46	0.310	0.757
Gender	Man	95(54.29)	44(58.67)	0.408	0.522
	Woman	80(45.71)	31(41.33)
BMI (kg/m2)		23.61 ± 2.21	23.41 ± 2.12	0.664	0.510
Hypertension	Yes	63(36.00)	30(40.00)	0.360	0.549
	No	112(64.00)	45(60.00)
Diabetes	Yes	50(28.57)	20(26.67)	0.095	0.759
	No	125(71.43)	55(73.33)
Cardiovascular disease	Yes	48(27.43)	18(24.00)	0.318	0.573
	No	127(72.57)	57(76.00)
History of bleeding	Yes	29(16.57)	13(17.33)	0.022	0.883
	No	146(83.43)	62(82.67)
Hemoglobin (g/L)		128.91 ± 7.97	127.45 ± 7.68	1.342	0.181
White blood cell count (10/L)		7.23 ± 1.17	7.56 ± 1.34	1.955	0.052
Platelet count (10/L)		173.63 ± 20.42	173.54 ± 23.03	0.031	0.980
Prothrombin time (s)		11.61 ± 0.79	11.76 ± 0.65	1.447	0.149
Activated partial thromboplastin time (s)		31.44 ± 1.80	31.64 ± 2.36	0.731	0.466
Fibrinogen (g/L)		3.28 ± 0.31	3.32 ± 0.55	0.730	0.466
Albumin (g/L)		36.82 ± 2.59	36.66 ± 2.33	0.461	0.645
Creatinine (μmol/L)		75.82 ± 7.44	75.32 ± 6.56	0.504	0.615
Lesion site	Upper digestive tract	54(30.86)	21(28.00)	0.212	0.900
	Stomach	77(44.00)	34(45.33)
	Duodenum	44(25.14)	20(26.67)
Pathological morphology	Uplift type	55(31.43)	27(36.00)	3.969	0.137
	Recessed type	77(44.00)	38(50.67)
	Flat type	43(24.57)	10(13.33)

Comparison of clinical characteristics between hemorrhage and non-hemorrhage groups

In the training set, univariate analysis revealed statistically significant differences (P < 0.05) between the hemorrhage and non-hemorrhage groups in the following parameters: age, hemoglobin level, WBC count, platelet count, PT, APTT, and fibrinogen level (Table 2).

Table 2.

Comparison of clinical characteristics between hemorrhage and non-hemorrhage groups

Factors		Hemorrhage (n = 70)	Non-hemorrhage (n = 105)	t/χ2	P
Age (years)		59.54 ± 7.52	56.21 ± 6.02	3.246	0.001
Gender	Man	42(60.00)	53(50.48)	1.535	0.215
	Woman	28(40.00)	52(49.52)
BMI (kg/m2)		23.99 ± 2.31	23.35 ± 2.01	1.943	0.054
Hypertension	Yes	28(40.00)	35(33.33)	0.810	0.368
	No	42(60.00)	70(66.67)
Diabetes	Yes	17(24.29)	33(31.43)	1.050	0.306
	No	53(75.71)	72(68.57)
Cardiovascular disease	Yes	21(30.00)	27(25.41)	0.388	0.534
	No	49(70.00)	78(74.29)
History of bleeding	Yes	14(20.00)	15(14.29)	0.992	0.319
	No	56(80.00)	90(85.71)
Hemoglobin (g/L)		127.23 ± 8.21	130.02 ± 7.64	2.303	0.022
White blood cell count (10/L)		7.53 ± 1.05	7.03 ± 1.22	2.802	0.006
Platelet count (10/L)		169.25 ± 20.31	176.56 ± 20.06	2.350	0.020
Prothrombin time (s)		11.86 ± 0.75	11.24 ± 0.71	5.533	0.001
Activated partial thromboplastin time (s)		32.04 ± 2.02	31.03 ± 1.51	3.527	0.001
Fibrinogen (g/L)		3.21 ± 0.33	3.34 ± 0.30	2.721	0.007
Albumin (g/L)		36.52 ± 2.56	37.22 ± 2.56	1.772	0.078
Creatinine (μmol/L)		76.21 ± 7.88	75.34 ± 6.75	0.781	0.436
Lesion site	Upper digestive tract	20(28.57)	34(32.38)	0.994	0.608
	Stomach	34(48.57)	43(40.95)
	Duodenum	16(22.86)	28(26.67)
Pathological morphology	Uplift type	21(30.00)	34(32.38)	0.158	0.924
	Recessed type	32(45.71)	45(45.86)
	Flat type	17(24.29)	26(24.76)

Analysis of prognostic risk factors in post-ESD hemorrhage

Using the occurrence of hemorrhage as the dependent variable (0 = absent, 1 = present) and incorporating factors with P < 0.05 from univariate analysis as covariates, multivariate logistic regression analysis was performed. The results demonstrated that age, hemoglobin level, WBC count, platelet count, PT, APTT, and fibrinogen level were independent risk factors for post-ESD hemorrhage (P < 0.05). These parameters were identified as key indicators influencing the therapeutic outcomes of ESD in gastrointestinal tumor patients (Table 3).

Table 3.

Multivariate logistic regression analysis of risk factors for post-ESD hemorrhage

Factor	β	S.E	Wald	P	OR	95% CI
Age	0.091	0.030	2.986	0.003	1.095	1.032–1.162
Hemoglobin	− 0.065	0.025	2.584	0.010	0.937	0.891–0.984
WBC count	0.449	0.179	2.512	0.012	1.567	1.104–2.224
Platelet count	− 0.022	0.010	2.172	0.030	0.978	0.959–0.998
Prothrombin time	1.110	0.304	3.648	0.001	3.033	1.671–5.505
Activated partial thromboplastin time	0.372	0.115	3.241	0.001	1.451	1.158–1.817
Fibrinogen	− 1.305	0.609	2.217	0.027	0.259	0.079–0.855

Development of a nomogram prediction model for post-ESD hemorrhage

Based on the independent risk factors identified through multivariate logistic regression analysis, a nomogram prediction model was constructed to assess the prognosis of gastrointestinal tumor treatment. Each independent risk factor in the model was assigned a corresponding score, allowing calculation of a total risk score for hemorrhage prediction. The total score directly correlates with the predicted probability of hemorrhage occurrence, with higher scores indicating greater predictive accuracy for hemorrhagic events following gastrointestinal tumor treatment (Fig. 2). The collinearity diagnosis shows that the Variance Inflation Factors (VIF) of age, hemoglobin, white blood cell count, platelet count, prothrombin time, activated partial thromboplastin time, and fibrinogen are 1.09, 1.044, 1.051, 1.09, 1.064, 1.126, and 1.061, respectively. This indicates that the problem of multicollinearity among these variables is very slight, and the degree of linear correlation between each independent variable is low. When building regression models and conducting other analyses, multicollinearity has little adverse impact on model estimation and inference, and the model's stability and explanatory power are relatively well-guaranteed.

Fig. 2.

Nomogram for predicting post-ESD hemorrhage. Note PT, prothrombin time. APTT, activated partial thromboplastin time.

The full multivariate logistic regression equation for hemorrhage prediction is:

Logit(P) = -28.643 + 0.091 × age-0.065 × Hemoglobin + 0.449 × WBC count-0.022 × platelet count + 1.110 × Prothrombin time + 0.372 × Activated partial thromboplastin time-1.305 × fibrinogen (where P is the predicted probability of post-ESD hemorrhage).

A points-to-risk lookup table (Table 4) was created to simplify clinical application, and a worked bedside example is provided below to demonstrate the practical use of the nomogram: bedside application example: a + 60-year-old male patient with the following laboratory results: hemoglobin = 125 g/L, WBC count = 8.0 × 10⁹/L, platelet count = 170 × 10⁹/L, prothrombin time = 12.5 s, activated partial thromboplastin time = 33 s, and fibrinogen = 3.0 g/L. According to Table 4, the corresponding scores for each variable are: age = 60 (30 points), hemoglobin = 125 (25 points), WBC count = 8.0 (35 points), platelet count = 170 (20 points), prothrombin time = 12.5 (40 points), activated partial thromboplastin time = 33 (30 points), and fibrinogen = 3.0 (25 points). The total score is 30 + 25 + 35 + 20 + 40 + 30 + 25 = 205 points. Referring to Table 4, the predicted probability of post-ESD hemorrhage is approximately 38.5%.

Table 4.

Points-to-risk lookup table for post-ESD hemorrhage prediction

Variable	Value range (corresponding points)
Age (years)	40–45 (10) 46–50 (15) 51–55 (20) 56–60 (25) 61–65 (30) 66–70 (35) 71–75 (40)
Hemoglobin (g/L)	105–115 (15) 116–125 (20) 126–135 (25) 136–145 (30) 146–155 (35)
WBC count (10⁹/L)	5.0–6.0 (20) 6.1–7.0 (25) 7.1–8.0 (30) 8.1–9.0 (35) 9.1–10.0 (40)
Platelet count (10⁹/L)	120–140 (10) 141–160 (15) 161–180 (20) 181–200 (25) 201–220 (30)
Prothrombin time (s)	9.0–10.0 (20) 10.1–11.0 (25) 11.0–12.0 (30) 12.1–13.0 (35) 13.1–14.0 (40)
Activated partial thromboplastin time (s)	26–28 (15) 29–31 (20) 32–34 (25) 35–37 (30) 38–40 (35)
Fibrinogen (g/L)	2.4–2.8 (15) 2.9–3.2 (20) 3.3–3.6 (25) 3.7–4.0 (30) 4.1–4.4 (35)

Evaluation and validation of the prediction model for post-ESD hemorrhage

In the training set and validation set, the C-index values of the nomogram model were 0.861 and 0.841, respectively. After bootstrap optimism correction, the C-indexes were 0.852 and 0.833. The calibration curves showed that the brier between the predicted values and the actual values were 0.151 and 0.171, respectively. The calibration intercepts of the training set and the validation set are 0.000 and 0.341, respectively, and the calibration slopes are 1.000 and 0.990, respectively. The results of the Hosmer–Lemeshow test were χ² = 5.9702 (P = 0.6506) and χ² = 12.874 (P = 0.1162) for the two sets, respectively. The ROC curves indicated that in the training set and validation set, the area under the curve (AUC) of the nomogram model for predicting bleeding was 0.849 (95% CI 0.778–0.920) and 0.824 (95% CI 0.707–0.941), respectively. The sensitivity and specificity were 0.816 and 0.827, and 0.714 and 0.833 for the two sets, respectively (Figs. 3 and 4).

Fig. 3.

Calibration curves in the training set (A) and the validation set (B)

Fig. 4.

ROC curves in the training set and the validation set

Decision curve analysis of the nomogram prediction model for post-ESD hemorrhage

The DCA demonstrated that our nomogram prediction model yielded significant clinical net benefit in both the training and validation cohorts for assessing hemorrhagic events impacting treatment outcomes (Fig. 5). Specifically, the net benefit was maintained when the threshold probability ranged between 0 and 0.7 in the training cohort, and between 0 and 0.8 in the validation cohort. This threshold range was justified by its alignment with the clinical risk distribution of post-ESD hemorrhage: it covers intermediate-to-high-risk populations (the core focus of clinical management), where hemorrhage probability typically ranges from 5 to 25%, and also includes a subset of low-risk patients to ensure comprehensive evaluation of model utility across different clinical scenarios.

Fig. 5.

Decision curves in the training set (A) and the validation set (B)

To enhance clinical actionability, we focused on the 5–25% threshold band (concentrating the majority of clinically actionable ESD cases) and analyzed net benefit/standardized net benefit within this range. Results showed the nomogram consistently outperformed the "treat all" and "treat none" strategies in both cohorts. The standardized net benefit (adjusted for sample size) was 0.178 at the 10% threshold and 0.153 at the 20% threshold in the training set, and 0.169 and 0.147 in the validation set, confirming robust clinical value in the core actionable range.

Clinical risk thresholds and predictive values

Three clinically relevant risk thresholds (5, 10, and 20%) were pre-specified to guide clinical decision-making. The confusion matrices and corresponding predictive values for the training and validation sets at these cut-points are shown in Table 5. At the 10% threshold (recommended as the clinical intervention threshold), the training set had a PPV of 0.832, NPV of 0.815, positive likelihood ratio of 4.73, and negative likelihood ratio of 0.22; the validation set had a PPV of 0.806, NPV of 0.798, positive likelihood ratio of 4.29, and negative likelihood ratio of 0.26. These results indicate that the model can effectively identify high-risk patients and rule out low-risk individuals at the recommended threshold.

Table 5.

Confusion matrices and predictive values at different clinical thresholds

Cohort	Threshold (%)	True positive	False positive	True negative	False negative	PPV	NPV
Training set	5	70	42	63	0	0.625	1.000
	10	68	13	92	2	0.832	0.815
	20	62	5	100	8	0.925	0.714
Validation set	5	28	18	37	0	0.610	1.000
	10	26	6	49	2	0.806	0.798
	20	23	2	53	5	0.920	0.688

Model benchmarking and subgroup analysis

Compared with comparative models, the nomogram showed favorable performance with good parsimony: the nomogram had an AUC of 0.849 (training set) and 0.824 (validation set), which was higher than the simple baseline model (0.721/0.703) and comparable to LASSO (0.852/0.829) and random forest (0.860/0.835). The Brier score of the nomogram (0.151/0.171) was slightly lower than LASSO (0.155/0.167) and random forest (0.161/0.175), indicating better calibration (Table 6).

Table 6.

Performance comparison of the nomogram with comparative models

Model type	Data set	AUC (95% CI)	C-index	Brier score
Nomogram	Training set	0.849 (0.778–0.920)	0.861	0.151
	Validation set	0.824 (0.707–0.941)	0.841	0.171
Simple baseline model (age + PT + APTT)	Training set	0.721 (0.635–0.807)	0.733	0.189
	Validation set	0.703 (0.576–0.830)	0.715	0.201
LASSO regression model	Training set	0.852 (0.781–0.923)	0.864	0.155
	Validation set	0.829 (0.712–0.930)	0.845	0.167
Random forest model	Training set	0.860 (0.790–0.930)	0.872	0.161
	Validation set	0.835 (0.720–0.950)	0.850	0.175

Subgroup analysis showed consistent discrimination and calibration across strata: by lesion site (esophagus: AUC = 0.832, C-index = 0.828. stomach: AUC = 0.851, C-index = 0.847. duodenum: AUC = 0.825, C-index = 0.819). By antithrombotic status (used: AUC = 0.840, C-index = 0.835; not used: AUC = 0.853, C-index = 0.848). By age strata (< 60 years: AUC = 0.837, C-index = 0.831. 60–70 years: AUC = 0.855, C-index = 0.850. > 70 years: AUC = 0.842, C-index = 0.836). Interaction tests showed no significant interactions between model variables and subgroup factors (all P > 0.05), confirming robustness (Table 7).

Table 7.

Subgroup discrimination, calibration, and interaction test results

Subgroup category	Subgroup name	N	AUC (95% CI)	C-index	Interaction test p value
Lesion site	Esophagus	75	0.832 (0.725–0.939)	0.828	0.618
	Stomach	111	0.851 (0.773–0.929)	0.847	0.542
	Duodenum	64	0.825 (0.708–0.934)	0.819	0.673
Antithrombotic status	Used	88	0.840 (0.746–0.934)	0.835	0.587
	Not used	162	0.853 (0.780–0.926)	0.848	0.631
Age strata	< 60 years	132	0.837 (0.756–0.918)	0.831	0.714
	60–70 years	95	0.855 (0.770–0.940)	0.850	0.625
	> 70 years	23	0.842 (0.685–0.950)	0.836	0.690

Discussion

Endoscopic submucosal dissection (ESD) has emerged as an advanced therapeutic technique that plays a pivotal role in managing early stage gastrointestinal tumors [18, 19]. However, procedure-related hemorrhage remains a common complication that significantly impacts treatment outcomes and patient recovery [20]. The present study developed a novel prognostic prediction model that quantitatively integrates bleeding risk with comprehensive outcome assessment in gastrointestinal tumor patients, thereby providing clinicians with a valuable risk-stratification tool.

We retrospectively analyzed 250 consecutive patients who experienced post-ESD bleeding that impacted treatment outcomes at our institution between 2022 and 2024. Using complete randomization, the cohort was divided into a training set (n = 175) and validation set (n = 75) at a 7:3 ratio, ensuring both random allocation and cohort independence for robust model development and validation. Initial univariate analysis identified several potential risk factors (P < 0.05), including age, BMI, hemoglobin levels, white blood cell count, platelet count, prothrombin time (PT), activated partial thromboplastin time (APTT), fibrinogen, albumin, and creatinine. Subsequent multivariate logistic regression analysis confirmed age, hemoglobin, white blood cell count, platelet count, PT, APTT, and fibrinogen as independent risk factors for post-ESD bleeding (P < 0.05). The comparable bleeding rates between training (40.00%, 70/175) and validation (37.33%, 28/75) sets (P > 0.05), along with balanced baseline characteristics, demonstrated appropriate cohort partitioning and supported model generalizability.

The observed post-ESD hemorrhage rates in our cohort (40.00% in the training set and 37.33% in the validation set) are notably higher than those commonly reported in the literature for general ESD procedures (typically 5–15%). As outlined in the Methods, this is a direct consequence of our study design, which focused on a complication-enriched cohort to facilitate model development. This discrepancy underscores that the reported hemorrhage rate reflects the incidence within a high-risk subgroup and highlights the model's intended context of use. It should be noted that the event-enriched design may affect the model's transportability to routine ESD populations. Due to the over-representation of high-risk patients, calibration drift may occur when applying the model to low-risk settings, where the predicted hemorrhage probability may be overestimated compared to the actual incidence. Therefore, external validation in general ESD populations and recalibration of the model parameters are necessary to improve its applicability in lower risk scenarios.

Advanced age, as a key risk factor, may contribute to bleeding risk through age-related physiological decline and reduced vascular integrity, while abnormalities in hematologic and coagulation parameters likely reflect underlying systemic dysfunction predisposing to hemorrhagic complications. These findings provide critical evidence for constructing our prognostic nomogram prediction model [21, 22]. The observed abnormalities in hematologic parameters (hemoglobin, white blood cell count, and platelet count) likely reflect underlying hematopoietic dysfunction that may increase bleeding risk and adversely affect treatment outcomes. Similarly, derangements in coagulation markers (prothrombin time, activated partial thromboplastin time, and fibrinogen) suggest coagulopathic states that predispose to hemorrhagic complications. It is worth noting that the odds ratio for platelet count, while statistically significant in our model, was close to 1.0 (OR = 0.978), indicating a modest effect size for each unit change. This suggests that the absolute platelet count within the observed range is a contributory but less dominant risk factor compared to global coagulation parameters. Its inclusion highlights the multifactorial nature of bleeding risk but underscores that the model's strength lies in the combination of variables rather than any single predictor.

The nomogram demonstrated excellent calibration between predicted and observed outcomes in both training and validation cohorts, evidenced by concordance indices of 0.861 and 0.841, mean absolute errors of 0.151 and 0.171, and nonsignificant Hosmer–Lemeshow test results (P = 0.651 and P = 0.116). These metrics indicate strong model reliability and accuracy for clinical application. The model's good predictive performance was further confirmed by area under the curve values of 0.849 and 0.824 in the training and validation sets, respectively, with corresponding sensitivity/specificity of 0.816/0.827 and 0.714/0.883, demonstrating its capability to effectively identify high-risk patients. Such risk stratification enables targeted preoperative coagulation optimization, surgical planning modification, and enhanced postoperative monitoring to improve overall prognosis.

The identified independent risk factors in our model—age, hemoglobin, white blood cell count, platelet count, prothrombin time, activated partial thromboplastin time, and fibrinogen—provide a distinct yet complementary perspective to existing risk scores for post-ESD hemorrhage. Several previously published models have effectively emphasized lesion characteristics and procedural factors (e.g., operation time, operator experience) [23, 24]. Our model, by contrast, is built upon a core set of intrinsic patient factors, specifically a comprehensive panel of pre-operative hematological and coagulation parameters. This approach is supported by the pathophysiology of bleeding and offers a practical tool that can be applied early in the patient's journey, even before the endoscopic procedure is planned in detail. The strong performance of our seven-variable model suggests that these fundamental laboratory values capture a significant portion of a patient's inherent bleeding diathesis. Therefore, our nomogram does not necessarily contradict existing models but may serve as a valuable foundational risk assessment that could be used in conjunction with or prior to more procedure-specific evaluations.

Several limitations warrant consideration in this study. First, the single-center design and moderate sample size may introduce selection bias. Second, and most importantly, as a retrospective study, it was constrained by the data available in medical records. Consequently, several important ESD and peri-procedural factors known to influence bleeding risk could not be systematically included in our model. These include detailed lesion characteristics (e.g., exact size, Paris classification), procedural specifics, and granular data on antithrombotic medication management. The absence of these variables is a key limitation that may introduce unmeasured confounding (omitted variable bias). This could mean that the regression coefficients for the variables in our model (e.g., coagulation parameters) might also be capturing some of the effects of these missing factors, potentially affecting the interpretability of individual variable importance and the absolute risk estimates produced by the nomogram. Therefore, while our model provides a valuable baseline risk assessment based on fundamental patient factors, its predictions should be interpreted with an understanding of these limitations. Future multicenter, prospective studies with meticulously collected data on all relevant clinical, endoscopic, and procedural variables are essential to validate, refine, and potentially recalibrate our model for broader application.

Additional improvements could include integration with machine learning algorithms for enhanced predictive accuracy, incorporation of supplementary prognostic indicators (tumor recurrence, quality of life metrics), and development of personalized therapeutic strategies based on risk stratification. While this nomogram provides valuable decision support, it should complement rather than replace clinical judgment, serving as one component of comprehensive patient assessment. External validation across diverse populations and healthcare settings remains essential to establish broader applicability and refine the prediction system [25]. For high-risk patients identified by the model, more aggressive preventive measures could be implemented preoperatively, including physiological optimization and surgical plan modification. For those who experience hemorrhage affecting treatment outcomes, prompt therapeutic interventions guided by model predictions may be initiated to enhance treatment efficacy [26].

Conclusion

Our nomogram model provides clinicians with a reliable tool for predicting post-ESD hemorrhage risk. The model demonstrates good calibration and discrimination across both training and validation sets. While promising, further refinements through expanded risk factor incorporation, technological integration, and personalized intervention strategies will optimize its clinical utility for improving gastrointestinal tumor management outcomes.

Abbreviations

APTT

Activated partial thromboplastin time

AUC

Area under the curve

BMI

Body mass index

CI

Confidence interval

EDTA

Ethylenediaminetetraacetic acid

ESD

Endoscopic submucosal dissection

FIB

Fibrinogen

Hb

Hemoglobin

MAE

Mean absolute error

OR

Odds ratio

PLT

Platelet count

PT

Prothrombin time

ROC

Receiver operating characteristic

VIF

Variance inflation factor

WBC

White blood cell count

Author contributions

Conception and design: FY. Method: YF and LFZ. Data Collection: LY, YLZ, DYD and YSC. Manuscript Writing: FY. Manuscript revision: FY. Research supervision: FY. All authors contributed to the article and approved the submitted version.

Funding

None.

Availability of data and materials

All data that support the findings of this study are available from the corresponding authors upon reasonable request.

Declarations

Ethics approval and consent to participate

The study was approved by the Committees for the Ethical Review of Research involving Human Subjects from the First Affiliated Hospital of Henan Medical University (Approval No. XXMU-0026), and informed consent was obtained from each patient. This study was conducted in accordance with the Declaration of Helsinki.

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.