Abstract
Background
Delayed post-polypectomy bleeding (DPPB) is a common adverse event after endoscopic mucosal resection (EMR) of colorectal polyps. However, the role of high-risk adenoma (HRA) in DPPB has been rarely studied, as prior work focused on polyp number, age, and comorbidities; its size-specific effect remains unclear.
Methods
Conducted at a single center, this retrospective cohort study enrolled 4,035 patients undergoing EMR for colorectal polyps (October 2016–March 2025) after excluding inflammatory bowel disease, familial polyposis, incomplete records, and non-EMR techniques. The HRA–DPPB association was evaluated with sequential multivariable logistic regression across three nested adjustment models, with interactions tested against high-risk serrated polyp (HRSP) and polyp size (≤1 cm vs. >1 cm). A simplified clinical risk score derived by the Sullivan method was internally validated through an optimism bootstrap that rebuilt the score on every resample, assessing discrimination (area under the curve, AUC) and calibration.
Results
Of 4,035 patients, 106 (2.63%) developed DPPB. HRA was independently associated with DPPB (adjusted odds ratio [OR] 2.90, 95% confidence interval [CI] 1.74–4.85), and the estimate strengthened rather than weakened with sequential adjustment (crude OR 2.49). Pathology-stratified analysis showed a dose–response pattern—HRA only (OR 2.90), HRSP only (OR 2.27), both (OR 5.54)—with an additive interaction that trended positive without reaching significance [relative excess risk due to interaction (RERI) 1.37, 95% CI −2.65 to 5.38]. The HRA × polyp size interaction (P = 0.005) was driven by the overlap between the HRA size criterion (≥1 cm) and the size-based grouping: when HRA was redefined by its non-size criteria alone (high-grade dysplasia, villous component, or ≥3 adenomas), the interaction disappeared (P = 0.95), leaving histology (OR ≈ 1.5–2.1) and size (OR 7.4 at 2 cm vs. 0.8 cm) as independent predictors of bleeding. The simplified score [age, sex, polyp size, count, body mass index (BMI)] achieved an optimism-corrected AUC of 0.729 with adequate calibration (Hosmer–Lemeshow P = 0.222; corrected slope 0.923), stratifying patients into low (0.73%), medium (2.03%), and high (8.84%) risk groups.
Conclusion
HRA is an independent risk factor for DPPB after EMR, and high-risk histology predicts bleeding across the size spectrum, complementing a strong, independent effect of polyp size. A simplified clinical score stratifies risk adequately, but as a single-center, internally validated instrument it awaits external confirmation before clinical adoption.
Keywords: delayed post-polypectomy bleeding, endoscopic mucosal resection, high-risk adenoma, interaction, risk score, size-stratified analysis
1. Introduction
Colorectal cancer (CRC) ranks third in global incidence and second in cancer-related mortality (1). Colorectal adenomas progress toward carcinoma through the conventional adenoma–carcinoma sequence, whereas roughly 15%–30% of CRCs arise from serrated polyps (2, 3). Endoscopic mucosal resection (EMR) is now the mainstay of colorectal polyp treatment, offering minimal trauma and rapid recovery, yet it carries complications such as bleeding, perforation, and post polypectomy coagulation syndrome (4, 5). Among these, delayed post-polypectomy bleeding (DPPB) is the most frequent, occurring in 6%–7% of large lesions (6). Defined conventionally as hemorrhage within 30 days of polypectomy, DPPB prolongs hospital stay, increases cost, and may prove life-threatening in severe cases (7). Identifying its risk factors early and strengthening the care of susceptible patients are therefore of clear clinical value.
Earlier studies have linked DPPB to several factors, including polyp size, location, and anticoagulant use (8). Whether high-risk adenoma (HRA) independently predisposes to DPPB, however, has received little attention (9). HRA is defined by a diameter ≥10 mm, high-grade dysplasia, a villous component, or the presence of three or more adenomas; because size is embedded in this definition, polyp size may confound the HRA–bleeding association or modify its effect (10). Beyond size, HRA also captures villous architecture, high-grade dysplasia, and overall adenoma burden—features with distinct vascular and architectural properties that may bear on bleeding independently of lesion size. Evidence on the HRA–DPPB relationship remains scarce, and no study has yet examined how HRA interacts with polyp size in driving DPPB.
A simple, reliable tool for predicting DPPB is also lacking. Nomogram-based models have been proposed for risk assessment, yet their bedside application remains cumbersome (11). A simplified risk score can fill this gap by translating a complex regression model into an integer point table that clinicians can tally quickly to gauge post-polypectomy bleeding risk.
This study therefore set out to (1) determine whether HRA is associated with DPPB; (2) probe the HRA–polyp size interaction through pathology- and size-stratified analyses; and (3) derive a simplified clinical risk score from the identified predictors and evaluate its discrimination and calibration using bootstrap internal validation.
2. Materials and methods
2.1. Study design and population
In this retrospective cohort study, we enrolled consecutive patients who underwent EMR for colorectal polyps at Peking University International Hospital between 1 October 2016 and 31 March 2025. The inclusion criteria were as follows: (1) age 18–85 years; (2) endoscopically confirmed colorectal polyps; (3) signed procedural informed consent; and (4) absence of contraindications to endoscopic therapy. The exclusion criteria were inflammatory bowel disease, familial adenomatous polyposis, incomplete clinical or pathological records, and treatment by endoscopic submucosal dissection or cold snare polypectomy, which preserves a homogeneous EMR cohort; the numbers excluded under each criterion are reported in the Results (Section 3.1, Figure 1). Each record was verified manually for resection technique.
Figure 1.
Patient flow diagram.
The Peking University International Hospital Ethics Committee approved the protocol on October 24, 2025 (approval 2025-KY-0101-01) and waived informed consent given the retrospective design.
2.2. Data collection and variable definitions
We collected data across three domains. Baseline characteristics comprised age, sex, BMI, hypertension, diabetes, coronary heart disease, history of stroke, liver cirrhosis, anticoagulant or antiplatelet use, and blood-pressure control among hypertensive patients. Polyp- and procedure-related variables included polyp number, maximum polyp diameter, HRA, high-risk serrated polyp (HRSP), resection technique (verified manually), polyp morphology and location, prophylactic clipping, intraprocedural bleeding, endoscopist seniority (≥10 years senior vs. <10 years junior), bowel-preparation quality [Aronchick or Boston Bowel Preparation Scale (BBPS)], and sedation type (unsedated vs. sedated) (12, 13). Laboratory measures were platelet count and estimated glomerular filtration rate.
Periprocedural antithrombotic management followed institutional practice in consultation with the prescribing specialty: warfarin was typically interrupted 5 days before EMR, with the INR checked before the procedure and bridged with low-molecular-weight heparin in patients judged to have high thrombotic risk; non-vitamin K antagonist oral anticoagulants were typically interrupted at least 48 h pre-procedure based on renal function without bridging; antiplatelet agents were continued or interrupted according to individual thrombotic risk, as bridging is not applicable to antiplatelet therapy; agents were resumed after hemostasis was secured, with the timing individualized according to each patient's thrombotic and bleeding risk.
All EMR procedures followed a standardized electrosurgical protocol on an ERBE VIO unit (ERBE Elektromedizin, Tübingen, Germany): Endocut Q mode (blended cut–coagulation) with Effect 3, Duration 1, and Interval 6 (peak voltage 770 Vp) was used for resection, and forced coagulation at Effect 2 and 40 W (peak voltage 1,100 Vp) for hemostasis. Submucosal injection was performed with normal saline through a 23–25 G injection needle inserted obliquely into the submucosa at an angle of 30–45°, with the volume adjusted to lesion size until adequate lift-off was achieved. These settings remained unchanged throughout the study period.
HRA was defined as any adenoma ≥1 cm in diameter, harboring a villous component or high-grade dysplasia, or the presence of three or more adenomas. HRSP was defined as a serrated lesion that was either a sessile serrated polyp ≥1 cm or showed dysplasia (10). DPPB was defined as any bleeding event within 30 days after the procedure that prompted an outpatient or emergency visit or repeat endoscopic intervention. Events were identified through retrospective review of the hospital electronic record—captured from discharge diagnoses and readmission records throughout the 30-day window—without active patient follow-up, so admissions to other hospitals and minor self-limited bleeding may have been missed.
Prophylactic clipping was performed at the discretion of the operating endoscopist; no standardized institutional protocol governed clip placement during the study period. Clip status (any clip placed during the procedure) was ascertained from the free-text procedure reports (available for 4,004 of 4,035 procedures); lesion-level clip attribution and the number of clips were not uniformly documented and were not analyzed.
2.3. Statistical analysis
Continuous variables were expressed as mean ± standard deviation and compared with Student's t-test; categorical variables were summarized as frequencies and proportions, n (%), and compared using the χ2 test or Fisher's exact test. To assess the independent association of HRA with DPPB, we built three prespecified sequential (nested) logistic regression models: Model 1 included HRA alone (unadjusted); Model 2 additionally adjusted for age, sex, and BMI; Model 3 further adjusted for diabetes, hypertension, anticoagulant or antiplatelet use, and renal impairment (CKD stages 2–5). All covariates were entered by forced entry without stepwise selection based on univariable P-values. Where a covariate showed complete separation within a stratum (zero events in one cell), standard maximum-likelihood estimation becomes unbounded; in that setting we used Firth (1993) penalized likelihood regression, which yields finite estimates with profile-likelihood confidence intervals. Calendar year was assessed in a sensitivity adjustment to account for secular changes in endoscopic practice. To probe whether the inverse crude associations of diabetes and antithrombotic use with DPPB reflected differential process of care, prophylactic clip receipt (ascertainable in 4,004/4,035 procedures) was compared across comorbidity groups and additionally modeled as a binary outcome with lesion characteristics and comorbidities as predictors.
In the pathology-stratified analysis, patients were classified into four groups—neither HRA nor HRSP, HRA only, HRSP only, and coexisting HRA and HRSP—and both crude and adjusted odds ratios (ORs) were computed. Additive interaction indices (relative excess risk due to interaction [RERI], attributable proportion [AP], and synergy index [S]) and a multiplicative interaction term evaluated the joint effect of HRA and HRSP. For the size-stratified analysis (≤1 cm vs. > 1 cm), logistic regression adjusting for HRA and the covariates above—excluding polyp size and number—was fitted within each stratum to test whether the HRA effect varied with size, and a multiplicative interaction term (HRA × size category) assessed the significance of this modification. We assessed multicollinearity using variance inflation factors (VIF).
Variable selection was prospectively guided by clinical accessibility, prior literature, and univariable results, retaining five independent predictors: age (years), sex, maximum polyp diameter (cm), polyp count, and BMI (kg/m2). To corroborate that this clinically prespecified set was not an artifact of post-hoc selection, we additionally performed a Lasso (L1, 10-fold cross-validation) analysis over a twelve-candidate pool. Pathology (HRA) was deliberately omitted because the score is intended as a point-of-care intraprocedural tool, whereas HRA becomes available only after pathological review. Each variable was categorized at clinically meaningful cut-points: age as ≥60, 50–59, or <50; sex as female or male; maximum diameter as <1.0, 1.0–1.9, or ≥2.0 cm; polyp count as 1, 2–3, or ≥4; and BMI as ≥24 or <24 kg/m2.
Discrimination was quantified by the area under the receiver operating characteristic curve (AUC); the DeLong test compared the score with the full model. For internal validation we used a Harrell optimism bootstrap that rebuilds the entire score on every resample—multivariable refit, coefficient-to-integer rounding (B = 0.40), and score-model refit—and corrected both the AUC and the calibration slope. Calibration was appraised with the Hosmer–Lemeshow test, calibration intercept and slope, and the Brier score. Total scores stratified patients into low-, medium-, and high-risk groups at the empirical tertiles. All tests were two-sided, with P < 0.05 deemed significant. Analyses were performed in R, version 4.3.1 (R Foundation for Statistical Computing, Vienna, Austria).
3. Results
3.1. Patient selection and baseline characteristics
Within the analytic set, all twelve key variables (age, sex, BMI, maximum polyp diameter, polyp count, HRA, HRSP, platelet count, CKD stage, diabetes, hypertension, and anticoagulant/antiplatelet use) had 0.00% missing, so complete-case analysis could not introduce missing-data bias (Supplementary Table S6). Of 4,356 patients who underwent EMR during the study period, 4,035 were retained for analysis after excluding 68 with inflammatory bowel disease, 10 with familial polyposis, and 243 with incomplete records (Figure 1). DPPB occurred in 106 (2.63%), while the remaining 3,929 (97.37%) did not bleed. Baseline characteristics of the two groups are compared in Table 1. Patients who bled were younger (51.14 ± 12.76 vs. 56.81 ± 11.71 years, P < 0.001), more often male (78.30% vs. 62.10%, P < 0.001), and less likely to have diabetes (6.60% vs. 14.81%, P = 0.018) or to be taking anticoagulants or antiplatelets (5.66% vs. 12.14%, P = 0.043). Their polyps were also more numerous (≥6 lesions: 15.09% vs. 8.02%, P < 0.001), larger (maximum diameter 1.41 ± 0.58 vs. 1.11 ± 0.52 cm, P < 0.001), and more likely to be HRA (82.08% vs. 64.77%, P < 0.001) or HRSP (16.98% vs. 8.83%, P = 0.004). BMI, hypertension, CKD stage, platelet count, endoscopist seniority, bowel-preparation score, and sedation type did not differ between the groups.
Table 1.
Baseline characteristics of study patients.
| Variable | Total (n = 4,035) | Non-bleeding (n = 3,929) | Bleeding (n = 106) | P value |
|---|---|---|---|---|
| Age, mean ± SD | 56.67 ± 11.77 | 56.81 ± 11.71 | 51.14 ± 12.76 | <0.001 |
| BMI, mean ± SD | 25.14 ± 3.45 | 25.15 ± 3.45 | 24.77 ± 3.38 | 0.272 |
| Gender, n (%) | <0.001 | |||
| Male | 2,523 (62.53) | 2,440 (62.10) | 83 (78.30) | |
| Female | 1,512 (37.47) | 1,489 (37.90) | 23 (21.70) | |
| Hypertension, n (%) | 1,416 (35.09) | 1,380 (35.12) | 36 (33.96) | 0.805 |
| Blood pressure control (among HTN), n (%) | 0.428 | |||
| Controlled | 838 (59.18) | 819 (59.35) | 19 (52.78) | |
| Uncontrolled | 578 (40.82) | 561 (40.65) | 17 (47.22) | |
| Diabetes mellitus, n (%) | 589 (14.60) | 582 (14.81) | 7 (6.60) | 0.018 |
| Coronary heart disease, n (%) | 255 (6.32) | 251 (6.39) | 4 (3.77) | 0.275 |
| History of cerebral infarction, n (%) | 125 (3.10) | 123 (3.13) | 2 (1.89) | 0.773 |
| Liver cirrhosis, n (%) | 16 (0.40) | 15 (0.38) | 1 (0.94) | 0.347 |
| CKD stage 2, n (%) | 987 (24.46) | 969 (24.66) | 18 (16.98) | 0.069 |
| CKD stage 3–5, n (%) | 127 (3.15) | 126 (3.21) | 1 (0.94) | 0.262 |
| Anticoagulant/antiplatelet use, n (%) | 483 (11.97) | 477 (12.14) | 6 (5.66) | 0.043 |
| Platelet count, mean ± SD | 228.40 ± 56.75 | 228.50 ± 56.66 | 224.81 ± 60.21 | 0.510 |
| Number of polyps, n (%) | <0.001 | |||
| <3 | 2,758 (68.35) | 2,705 (68.85) | 53 (50.00) | |
| 3–5 | 946 (23.44) | 909 (23.14) | 37 (34.91) | |
| ≥6 | 331 (8.20) | 315 (8.02) | 16 (15.09) | |
| Maximum polyp diameter (cm), mean ± SD | 1.12 ± 0.52 | 1.11 ± 0.52 | 1.41 ± 0.58 | <0.001 |
| High-risk adenoma, n (%) | 2,632 (65.23) | 2,545 (64.77) | 87 (82.08) | <0.001 |
| High-risk serrated polyp, n (%) | 365 (9.05) | 347 (8.83) | 18 (16.98) | 0.004 |
| Endoscopist experience, n (%) | 0.823 | |||
| Senior (≥10 years) | 873 (21.64) | 851 (21.66) | 22 (20.75) | |
| Junior (<10 years) | 3,162 (78.36) | 3,078 (78.34) | 84 (79.25) | |
| Bowel preparation scale, n (%) | 0.560 | |||
| Aronchick | 1,787 (44.29) | 1,743 (44.36) | 44 (41.51) | |
| BBPS | 2,248 (55.71) | 2,186 (55.64) | 62 (58.49) | |
| Aronchick score, n (%) | 0.056 | |||
| Score 1 | 160 (8.95) | 156 (8.95) | 4 (9.09) | |
| Score 2 | 947 (52.99) | 919 (52.73) | 28 (63.64) | |
| Score 3 | 592 (33.13) | 585 (33.56) | 7 (15.91) | |
| Score 4 | 85 (4.76) | 80 (4.59) | 5 (11.36) | |
| Score 5 | 3 (0.17) | 3 (0.17) | 0 (0.00) | |
| BBPS total score, mean ± SD | 7.24 ± 1.27 | 7.24 ± 1.27 | 7.19 ± 1.20 | 0.786 |
| Colonoscopy type, n (%) | 0.408 | |||
| Conventional | 568 (14.08) | 556 (14.15) | 12 (11.32) | |
| Painless (sedated) | 3,467 (85.92) | 3,373 (85.85) | 94 (88.68) |
Continuous variables: mean ± SD (Student's t-test). Categorical variables: n (%) (Chi–square test or Fisher's exact test).
CKD, chronic kidney disease; BMI, body mass index; BBPS, Boston Bowel Preparation Scale.
Among the 483 patients using antithrombotic agents, the agents comprised aspirin (n = 375, including 17 on dual antiplatelet therapy), P2Y12 inhibitors (n = 84), non-vitamin K antagonist oral anticoagulants (n = 14), warfarin (n = 5), and other agents (n = 23, predominantly indobufen and traditional Chinese medicines with antiplatelet activity), categories are not mutually exclusive. Notably, bridging-eligible patients (warfarin users) constituted only 0.12% of the cohort.
3.2. Univariable analysis
Univariable logistic regression (Table 2) linked DPPB to age (OR 0.96 per year, P < 0.001), male sex (OR 2.20, P < 0.001), polyp count (OR 1.10 per additional lesion, P < 0.001), maximum polyp diameter (OR 1.86 per cm, P < 0.001), HRA (OR 2.49, P < 0.001), and HRSP (OR 2.11, P = 0.005). Diabetes (OR 0.41, P = 0.022) and anticoagulant or antiplatelet use (OR 0.43, P = 0.049) emerged as protective in this unadjusted analysis, a pattern likely reflecting selection bias or clinical intervention. To test whether these inverse associations reflected differential process of care, we compared prophylactic clipping rates across comorbidity groups: clipping did not differ by diabetes (80.7% vs. 79.1%, P = 0.41) or antithrombotic use (79.5% vs. 79.3%, P = 0.98), and clip receipt was independently associated with lesion size >1 cm (aOR 2.82) and HRA (aOR 1.42) but with neither comorbidity (diabetes aOR 1.06, P = 0.60; antithrombotic use aOR 0.91, P = 0.45).
Table 2.
Univariable logistic regression analysis for risk factors of delayed post-polypectomy bleeding.
| Variable | OR (95% CI) | P value |
|---|---|---|
| Renal insufficiency (CKD 2–5) | 0.57 (0.34–0.93) | 0.026 |
| Age (per year) | 0.96 (0.95–0.98) | <0.001 |
| Male sex | 2.20 (1.38–3.51) | <0.001 |
| BMI (per kg/m2) | 0.97 (0.91–1.03) | 0.272 |
| Hypertension | 0.95 (0.63–1.43) | 0.805 |
| Diabetes mellitus | 0.41 (0.19–0.88) | 0.022 |
| Coronary heart disease | 0.57 (0.21–1.57) | 0.281 |
| Cerebral infarction | 0.60 (0.15–2.44) | 0.471 |
| Liver cirrhosis | 2.49 (0.33–18.99) | 0.380 |
| Anticoagulant/antiplatelet use | 0.43 (0.19–1.00) | 0.049 |
| Platelet (per 109/L) | 1.00 (1.00–1.00) | 0.509 |
| Polyp count (per 1) | 1.10 (1.04–1.17) | <0.001 |
| Maximum polyp size (per cm) | 1.86 (1.48–2.32) | <0.001 |
| High-risk adenoma | 2.49 (1.51–4.11) | <0.001 |
| High-risk serrated polyp | 2.11 (1.26–3.55) | 0.005 |
| Junior endoscopist | 1.06 (0.66–1.70) | 0.823 |
| Painless colonoscopy | 1.29 (0.70–2.37) | 0.410 |
CKD, chronic kidney disease; BMI, body mass index.
3.3. Independent association of HRA with DPPB: sequential adjustment models
Sequential multivariable logistic regression (Table 3) revealed a robust HRA–DPPB association across all three models: unadjusted Model 1 (OR 2.49, 95% CI 1.51–4.11, P < 0.001), Model 2 adjusting for age, sex, and BMI (OR 2.97, 95% CI 1.78–4.97, P < 0.001), and the fully adjusted Model 3 (OR 2.90, 95% CI 1.74–4.85, P < 0.001). Counterintuitively, the HRA odds ratio rose rather than fell after adjustment for age and other confounders (2.49 → 2.97).
Table 3.
Univariable and sequential multivariable logistic regression analyses of factors associated with delayed post-polypectomy bleeding.
| Variable | Univariable OR (95% CI) | Univariable P | Model 1 OR (95% CI) | Model 1 P | Model 2 OR (95% CI) | Model 2 P | Model 3 OR (95% CI) | Model 3 P |
|---|---|---|---|---|---|---|---|---|
| High-risk adenoma | 2.49 (1.51–4.11) | <0.001 | 2.49 (1.51–4.11) | <0.001 | 2.97 (1.78–4.97) | <0.001 | 2.90 (1.74–4.85) | <0.001 |
| Age (per year) | 0.96 (0.95–0.98) | <0.001 | — | — | 0.95 (0.94–0.97) | <0.001 | 0.96 (0.94–0.97) | <0.001 |
| Male sex | 2.20 (1.38–3.51) | <0.001 | — | — | 2.14 (1.32–3.47) | 0.002 | 2.16 (1.33–3.51) | 0.002 |
| BMI (per kg/m2) | 0.97 (0.91–1.03) | 0.272 | — | — | 0.94 (0.89–1.00) | 0.038 | 0.93 (0.88–0.99) | 0.025 |
| Diabetes mellitus | 0.41 (0.19–0.88) | 0.022 | — | — | — | — | 0.52 (0.23–1.15) | 0.104 |
| Hypertension | 0.95 (0.63–1.43) | 0.805 | — | — | — | — | 1.51 (0.96–2.39) | 0.077 |
| Anticoagulant/antiplatelet | 0.43 (0.19–1.00) | 0.049 | — | — | — | — | 0.62 (0.26–1.46) | 0.273 |
| CKD 2–5 | 0.57 (0.34–0.93) | 0.026 | — | — | — | — | 0.85 (0.50–1.46) | 0.562 |
We constructed three sequential multivariable logistic regression models. Model 1 included only the primary variable of interest (HRA). Model 2 additionally adjusted for age, sex, and BMI. Model 3 further adjusted for diabetes mellitus, hypertension, anticoagulant/antiplatelet use, and CKD 2–5. Polyp size and count were not adjusted in primary models (see stratified analysis for natural control by size).
OR, odds ratio; CI, confidence interval; BMI, body mass index; CKD, chronic kidney disease.
Within Model 3, male sex was an additional independent risk factor (OR 2.16, P = 0.002), whereas age (OR 0.96 per year, P < 0.001) and BMI (OR 0.93 per kg/m2, P = 0.025) were protective. Adding polyp size and count to Model 3 attenuated the HRA estimate from 2.90 to 2.12 (95% CI 1.25–3.62, P = 0.006), which retained significance (Supplementary Table S1). No harmful collinearity was detected (HRA VIF 1.05 in the primary model and 1.23 after adding size and count; all variables < 2; Supplementary Table S5). Size and count therefore partly confound the HRA–DPPB link. This estimate sits alongside the component-decomposition result (3.02 → 1.71, P = 0.06; same table), in which size alone is added to a different covariate set; the two analyses bracket the HRA effect between a fully adjusted estimate that stays significant (2.12) and a size-only decomposition that does not (1.71), indicating that size accounts for a substantial share, though not all, of the association.
3.4. Pathology-stratified analysis: joint effects of HRA and HRSP
Pathology-stratified analysis (Table 4) took patients with neither HRA nor HRSP as the reference (bleeding rate 1.23%). HRA alone carried a rate of 3.04% (adjusted OR 2.90, 95% CI 1.66–5.07, P < 0.001), HRSP alone 2.97% (adjusted OR 2.27, 95% CI 0.64–8.01, P = 0.202), and coexisting HRA and HRSP the highest rate, 5.68% (adjusted OR 5.54, 95% CI 2.67–11.49, P < 0.001). Additive interaction testing yielded an adjusted RERI of 1.37 (95% CI −2.65 to 5.38), S of 1.43, and AP of 0.25 (Table 5). The intervals are wide and cross zero, in part because the HRSP-only group contained only three events, so these indices support a trend toward positive additive interaction rather than a statistically significant one; nominally, a quarter of the risk in the co-presence group is attributable to the interaction. On the multiplicative scale the interaction OR was 0.84 (95% CI 0.21–3.36, P = 0.805), indicating independent effects. Each high-risk pathology thus raised DPPB risk on its own, and the two together posed the greatest hazard (Supplementary Table S4).
Table 4.
Pathology-stratified analysis of post-EMR bleeding risk.
| Pathology group | n | Bleeding n (%) | Crude OR (95% CI) | Crude P | Adjusted OR* (95% CI) | Adjusted P* |
|---|---|---|---|---|---|---|
| Neither HRA nor HRSP | 1,302 | 16 (1.23) | 1.00 | — | 1.00 | — |
| HRA only | 2,368 | 72 (3.04) | 2.52 (1.46–4.35) | <0.001 | 2.90 (1.66–5.07) | <0.001 |
| HRSP only | 101 | 3 (2.97) | 2.46 (0.70–8.59) | 0.158 | 2.27 (0.64–8.01) | 0.202 |
| Both HRA + HRSP | 264 | 15 (5.68) | 4.84 (2.36–9.92) | <0.001 | 5.54 (2.67–11.49) | <0.001 |
*Adjusted for age, sex, BMI, diabetes, hypertension, anticoagulant/antiplatelet use, and CKD stages 2–5 (not adjusted for polyp size or count).
Table 5.
Interaction between HRA and HRSP.
| Measure | Crude | Adjusted* |
|---|---|---|
| RERI (Relative Excess Risk due to Interaction) | 0.86 | 1.37 (−2.65 to 5.38) |
| AP (Attributable Proportion due to Interaction) | 0.18 | 0.25 (−0.39 to 0.89) |
| S (Synergy Index) | 1.29 | 1.43 (−0.13 to 2.99) |
| Multiplicative interaction OR (95% CI) | — | 0.84 (0.21–3.36) |
| Multiplicative interaction P | — | 0.805 |
*Adjusted for age, sex, BMI, diabetes, hypertension, anticoagulant/antiplatelet use, and CKD 2–5 (not adjusted for polyp size/count).
RERI > 0 indicates positive additive interaction; AP > 0 indicates proportion attributable to interaction; S > 1 indicates synergism.
HRA, high-risk adenoma; HRSP, high-risk serrated polyp; RERI, relative excess risk due to interaction; AP, attributable proportion; S, synergy index.
3.5. Stratification by polyp size: effect modification of HRA
Size-stratified analysis (Table 6) disclosed a significant HRA × size interaction (interaction OR 0.22, 95% CI 0.08–0.63, P = 0.005). Among polyps ≤1 cm (n = 2,621; 38 bleeds), HRA multiplied DPPB risk 3.23-fold (OR 3.23, 95% CI 1.50–6.97, P = 0.003), whereas among polyps >1 cm (n = 1,414; 68 bleeds) no residual association was observed (OR 0.80, 95% CI 0.38–1.69, P = 0.561)—an estimate reflecting near-universal exposure in this stratum, since 1,269/1,414 patients (90%) were HRA-positive through the size criterion alone. Within the ≤1 cm stratum no diabetic patient bled, producing complete separation for diabetes under standard estimation; Firth penalized likelihood returned a finite estimate (OR 0.082, profile-likelihood CI 0.001–0.584, P = 0.065) (Supplementary Table S3), read as a selection artifact rather than a genuine protective effect of diabetes.
Table 6.
Stratified analysis of post-EMR bleeding risk by polyp size.
| Variable | Polyp ≤1 cm OR (95% CI) | P | Polyp >1 cm OR (95% CI) | P |
|---|---|---|---|---|
| High-risk adenoma | 3.23 (1.50–6.97) | 0.003 | 0.80 (0.38–1.69) | 0.561 |
| Age (per year) | 0.97 (0.94–1.00) | 0.083 | 0.95 (0.93–0.97) | <0.001 |
| Male sex | 1.77 (0.82–3.86) | 0.148 | 2.78 (1.48–5.21) | 0.001 |
| BMI (per kg/m2) | 0.98 (0.89–1.08) | 0.692 | 0.91 (0.84–0.98) | 0.018 |
| Diabetes mellitus | 0.082 (0.001–0.584)† | 0.065 | 0.94 (0.40–2.18) | 0.878 |
| Hypertension | 1.69 (0.81–3.54) | 0.164 | 1.38 (0.76–2.49) | 0.287 |
| Anticoagulant/antiplatelet | 0.30 (0.04–2.24) | 0.240 | 0.72 (0.27–1.92) | 0.506 |
| Renal insufficiency (CKD 2–5) | 0.84 (0.35–2.01) | 0.688 | 0.84 (0.42–1.68) | 0.622 |
HRA × Polyp Size Interaction: interaction OR = 0.22 (0.08–0.63), P = 0.005.
Polyp size and count were not included as covariates in stratified models (natural control within strata).
N = 2,621 (38 events) in ≤1 cm stratum; N = 1,414 (68 events) in >1 cm stratum.
Estimated by Firth penalized likelihood with profile-likelihood confidence intervals (complete separation: no diabetic patient bled within the ≤1 cm stratum).
Because the HRA definition itself incorporates a≥10 mm criterion, the exposure and the stratifying threshold share the 1 cm boundary; we therefore repeated the analysis with this definitional overlap removed. Within the ≤1 cm stratum, lesions recorded as exactly 1.0 cm—the only subgroup that can simultaneously satisfy the HRA size criterion and remain classified as small—accounted for 1,175/2,621 patients (45%) and 26/38 bleeding events (68%), and recorded sizes showed strong digit preference (1.0 cm in 1,175 patients, vs. 33 at 0.9 cm and 6 at 1.1 cm), so the boundary group is best read as a rounding mixture of true 9–11 mm lesions. In a boundary-decomposition model within this stratum, an exactly-1.0 cm indicator entered together with a size-independent histology exposure (villous component, high-grade dysplasia, or ≥3 adenomas) assigned the apparent small-polyp effect to the boundary lesions themselves (adjusted OR 2.53, 95% CI 1.32–4.87, P = 0.005) rather than to histology (adjusted OR 1.52, 95% CI 0.78–2.94, P = 0.219). Under this non-size definition, the ≤1 cm estimate attenuated from 3.23 to 1.66 (95% CI 0.78–3.34, P = 0.136), the >1 cm stratum regained an informative contrast (adjusted OR 2.04, 95% CI 1.21–3.44, P = 0.008), and the exposure × size interaction disappeared (interaction OR 1.03, P = 0.95). Modelling size as a restricted cubic spline likewise showed no interaction for either the original HRA (joint P = 0.48) or the non-size definition (joint P = 0.85): histology predicted bleeding with a broadly stable adjusted OR of approximately 1.5–2.1 across 0.5–2.0 cm, whereas size itself showed a steep independent dose–response (vs. 0.8 cm: adjusted OR 1.65 at 1.0 cm, 2.55 at 1.2 cm, 4.35 at 1.5 cm, and 7.43 at 2.0 cm). The interaction observed in Table 6 is therefore definition-dependent rather than biological: size and histology each carry independent bleeding-risk information across the size spectrum.
Decomposing HRA into its components further examined the contribution of size to the HRA effect (Supplementary Table S1). Across the full cohort the HRA odds ratio fell from 3.0 to 1.71 (P = 0.06) once polyp size was added in the component-decomposition model, yet within ≤1 cm polyps HRA retained an independent effect (adjusted OR 3.21, 95% CI 1.48–6.98, P = 0.003), and a polyp count ≥3 predicted bleeding independently in both strata (≤1 cm adjusted OR 2.50; >1 cm 2.28). These decomposition estimates differ slightly from the main-analysis values (univariable HRA OR 2.49; size-stratified ≤1 cm HRA OR 3.23 in Table 6) because the component model adjusts for a different covariate set. The bleeding load of multiple polyps is therefore distinct from the histology captured by HRA. When intraprocedural variables extracted from the free-text reports were added to the model, pedunculated (Ip) morphology emerged as the independent predictor (adjusted OR 2.04, 95% CI 1.35–3.08, P = 0.0007), and prophylactic clipping carried an adjusted OR of 2.72 (95% CI 1.50–4.95)—a direction best explained by confounding by indication: in the absence of a protocolized clipping policy, clip placement tracked the endoscopist's real-time perception of bleeding risk (clip use 92.5% among patients who developed DPPB vs. 79.1% among those who did not; approximately 92% in the >1 cm stratum). Size >1 cm remained the dominant driver (adjusted OR 2.09) (Supplementary Table S2). Calendar year showed no significant temporal trend across 2016–2025 (adjusted OR 1.08 per year, P = 0.12; Supplementary Table S13).
3.6. Development and validation of the clinical risk score
The Lasso analysis over a twelve-candidate pool retained all twelve candidates at the optimal penalty (C = 3.92); the five clinically prespecified predictors were all retained (5/5 overlap), with age, polyp size, and polyp count entering the coefficient path first, corroborating that the clinically chosen set was not an artifact of post-hoc selection (Supplementary Table S7). The multivariable regression spanned eight non-reference categories across five predictors, with effect directions matching clinical expectation—younger age, male sex, larger diameter, higher polyp count, and lower BMI each raised bleeding risk. The strongest contributors were a maximum diameter ≥2.0 cm (OR 7.480, 95% CI 3.599–15.549), followed by 1.0–1.9 cm (OR 3.676, 95% CI 1.980–6.825) and age <50 years (OR 3.173, 95% CI 1.966–5.120). Following Sullivan and colleagues, we scaled the regression coefficients to integers with a constant B = 0.40, producing a 0–13-point score (Table 7); the largest single contribution came from a diameter ≥2.0 cm (5 points) and the smallest from age 50–59 years, 2–3 polyps, and BMI <24 kg/m2 (1 point each).
Table 7.
Sullivan score assignment table for the 5-variable simplified risk score.
| Variable | Category | β (logit)a | OR (95% CI) | W = β × (X_i − X_ref)b | Pointsc |
|---|---|---|---|---|---|
| Age (years) | ≥60 | 0 | 1.00 | 0.0000 | 0 |
| 50–59 | 0.4606 | 1.585 (0.944–2.660) | 0.4606 | 1 | |
| <50 | 1.1546 | 3.173 (1.966–5.120) | 1.1546 | 3 | |
| Sex | Female | 0 | 1.00 | 0.0000 | 0 |
| Male | 0.7109 | 2.036 (1.258–3.294) | 0.7109 | 2 | |
| Maximum polyp diameter (cm) | <1.0 | 0 | 1.00 | 0.0000 | 0 |
| 1.0–1.9 | 1.3018 | 3.676 (1.980–6.825) | 1.3018 | 3 | |
| ≥2.0 | 2.0123 | 7.480 (3.599–15.549) | 2.0123 | 5 | |
| Polyp count (n) | 1 | 0 | 1.00 | 0.0000 | 0 |
| 2–3 | 0.3978 | 1.489 (0.908–2.442) | 0.3978 | 1 | |
| ≥4 | 0.9935 | 2.701 (1.581–4.615) | 0.9935 | 2 | |
| BMI (kg/m2) | ≥24 | 0 | 1.00 | 0.0000 | 0 |
| <24 | 0.4444 | 1.560 (1.044–2.330) | 0.4444 | 1 | |
| Intercept (β₀) | — | −6.3577 | — | — | — |
β coefficients from the multivariable logistic regression on categorized predictors (reference categories set to β = 0).
For categorical predictors, X_i = 1 for the index category and X_ref = 0 for the reference category; hence W_i = β_i. All W≥0, guaranteeing a non-negative integer score.
Points = round(W_i/B), where B = 0.40 (e^0.40 ≈ 1.49; each 1-point increment ≈ OR 1.49 ≈ + 50% risk). Total score range = 0–13. Reference-group baseline risk = 0.17%. Risk formula: P(bleeding) = 1/(1 + e^(6.358 − 0.40 × Total Score)).
Patients were then grouped by total score into low (0–3), medium (4–8), and high (9–13) risk (Table 8). The low-risk group comprised 820 patients (20.3%) with 6 bleeds (0.73%); the medium-risk group 2,706 (67.1%) with 55 bleeds (2.03%); and the high-risk group 509 (12.6%) with 45 bleeds (8.84%). Although only 12.6% of the cohort, the high-risk group concentrated 42.5% of all bleeding events and carried roughly 12 times the risk of the low-risk group. Predicted bleeding rates tracked observed rates closely across score levels (Figure 2), with Wilson 95% confidence intervals largely overlapping the ideal diagonal.
Table 8.
Risk stratification according to the simplified score (N = 4,035; 106 events).
| Risk group | Score range | N | % of cohort | Bleeding events, n | Bleeding rate (%) |
|---|---|---|---|---|---|
| Low risk | 0–3 | 820 | 20.3 | 6 | 0.73 |
| Medium risk | 4–8 | 2,706 | 67.1 | 55 | 2.03 |
| High risk | 9–13 | 509 | 12.6 | 45 | 8.84 |
| Total | 0–13 | 4,035 | 100.0 | 106 | 2.63 |
The high-risk group (12.6% of the cohort) accounted for 42.5% of all bleeding events, with an absolute rate approximately 12 × that of the low-risk group (8.84% vs. 0.73%). The medium-risk group captured 51.9% of events; the low-risk group captured 5.7%.
Figure 2.
Calibration of the 5-variable simplified risk score. Blue circles represent observed bleeding probabilities at each score level (size proportional to N), with 95% Wilson confidence intervals shown as vertical bars. The red curve represents a Loess smoother. The dashed gray line indicates ideal calibration (y = x). Calibration intercept = 0.080; calibration slope = 1.001; Hosmer–Lemeshow χ2(7) = 9.45, P = 0.222; Brier score = 0.0248; AUC = 0.7447 (95% CI 0.6962–0.7931).
Discrimination and calibration of the simplified score and the full five-variable model are compared in Tables 8, 9 and Figures 3, 4. The AUC was 0.7447 (95% CI 0.6962–0.7931) for the score vs. 0.7479 (95% CI 0.6995–0.7963) for the full model, an insignificant difference (ΔAUC = −0.0032, DeLong Z = −0.975, P = 0.329). The optimism bootstrap, which rebuilt the score on every resample, returned a corrected AUC of 0.729 (optimism 0.016) and a corrected calibration slope of 0.923 (Supplementary Table S10). Calibration was otherwise satisfactory: Hosmer–Lemeshow χ2(7) = 9.445, P = 0.222; calibration intercept 0.080 and apparent slope 1.001; Brier score 0.0248. Reclassification gains were marginal (NRI 0.041, IDI 0.0026). The AIC values (simplified 899.35; full 911.94) are reported for completeness only (Supplementary Table S11); the two models are not meaningfully ranked by AIC, because the score is a rounded restriction of the full model and the apparent ΔAIC of 12.59 reflects the parameter-count penalty rather than a fit advantage (likelihood-ratio χ2 = 1.42, df = 7, P = 0.985).
Table 9.
Performance metrics of the simplified score vs. the full 5-variable logistic regression model.
| Domain | Metric | Simplified score (0–13) | Full 5-variable modela |
|---|---|---|---|
| Discrimination | AUC (C-statistic) | 0.7447 | 0.7479 |
| 95% CI | (0.6962–0.7931) | (0.6995–0.7963) | |
| Bootstrap-corrected AUCb | 0.7289 | — | |
| Optimism | 0.0157 | — | |
| DeLong test (Z; P) | — | −0.975; 0.329 | |
| Calibration | Hosmer–Lemeshow χ2 (df) | 9.445 (7) | 8.814 (8) |
| H–L P-value | 0.222 | 0.358 | |
| Calibration interceptc | 0.080 | 0.000 | |
| Calibration slopec | 1.001 (apparent)/0.923 (corrected) | 1.000 | |
| Brier score | 0.0248 | 0.0248 | |
| Reclassificationd | NRI (overall) | — | 0.041 |
| NRI (events) | — | 0.076 | |
| NRI (non-events) | — | −0.034 | |
| IDI | — | 0.0026 | |
| Overall fit | AIC | 899.35 | 911.94 |
| ΔAIC | — | 12.59 (parameter-count penalty, not a fit advantage) | |
| No. of parameters (k) | 2 | 9 |
Interpretation: ΔAUC = 0.003 (DeLong P = 0.329), comparable H–L P (>0.20), and identical Brier scores indicate the simplified score retains essentially all discriminative and calibrated performance of the full model while being more parsimonious (2 vs. 9 parameters). The AIC values are reported for completeness only; the two models are not meaningfully ranked by AIC, because the score is a rounded restriction of the full model and the apparent ΔAIC reflects the parameter-count penalty rather than a fit advantage (likelihood-ratio χ2 = 1.42, df = 7, P = 0.985).
Full model = multivariable logistic regression with the same 5 categorized predictors.
500-resample bootstrap (Steyerberg/Harrell heuristic).
Logistic refit of observed outcome on logit of predicted probability (ideal: intercept = 0, slope = 1).
Reclassification metrics using the simplified score as the “old” model and the full regression as the “new” model, with a 5% risk-threshold boundary.
Figure 3.
Calibration plot of the simplified risk score with bootstrap internal validation. Blue circles with error bars represent decile observed probabilities with 95% Wilson confidence intervals. The red solid curve and dashed logistic refit line depict apparent calibration. The green curve represents bias-corrected calibration (500 bootstrap resamples). The darkened reference line indicates ideal calibration (y = x). Bootstrap-corrected AUC = 0.7289.
Figure 4.
ROC curves of the simplified score vs. full model with bootstrap internal validation. The dark blue curve depicts the simplified score (AUC = 0.7447, 95% CI 0.6962–0.7931), with the shaded band representing 95% CI (500 bootstrap). The red curve depicts the full 5-variable model (AUC = 0.7479, 95% CI 0.6995–0.7963). The bias-corrected simplified score AUC is 0.7289 (dashed blue line). The gray diagonal line represents no discrimination. DeLong test: Z = −0.975, P = 0.329.
4. Discussion
Drawing on 4,035 EMR-treated patients, this study systematically appraised the HRA–DPPB relationship and arrived at three principal findings. First, HRA independently predicts DPPB (adjusted OR 2.90), and the association strengthened rather than weakened as major confounders were added sequentially, a descriptive pattern consistent with negative confounding by age (younger patients bled more often), offered without causal attribution to the adjustment itself. Second, size and histology carry independent, complementary risk information rather than interacting: non-size histology predicted bleeding across the size spectrum (adjusted OR ≈ 1.5–2.1, no effect modification), while polyp size itself exerted a steep, independent dose–response (up to OR 7.4 at 2 cm). Third, a score built from five simple variables discriminates DPPB risk effectively—a 12-fold gradient between high- and low-risk groups—and shows good discrimination and calibration after bootstrap validation.
The adenoma–DPPB link has seldom been examined; our three sequential models demonstrate that the HRA effect is stable (9, 14). The HRA odds ratio climbed from 2.49 to 2.97 once age was adjusted for, because younger patients in this cohort bled more often than older ones—a direction reported in several earlier series as well (15). This pattern reads more coherently as selection bias than as biology. Younger patients undergo colonoscopy for symptoms, a family history of neoplasia, or a known polyp rather than for screening, assembling a higher-risk subset (indication bias); at the other extreme, patients aged ≥80 who reach colonoscopy are disproportionately fit (healthy-survivor bias). Restricted cubic splines (four knots) detected no significant nonlinearity for age (P = 0.18) or BMI (P = 0.37), so the linear estimates are retained (Supplementary Table S12).
The stratified pattern—an elevated point estimate within ≤1 cm polyps (OR 3.23) and a null estimate beyond 1 cm (OR 0.80)—initially suggested that the HRA effect varies with polyp size, but four complementary analyses indicate that the pattern is definition-dependent rather than biological. The mechanism lies at the 1 cm boundary itself, the one point shared by the HRA definition and the stratification: lesions recorded as exactly 1.0 cm accounted for 45% of the ≤1 cm stratum (1,175/2,621) and 68% of its bleeding events (26/38), and recorded sizes cluster strongly at this value (1,175 patients at 1.0 cm vs. 33 at 0.9 cm and 6 at 1.1 cm), making the boundary group a rounding mixture of true 9–11 mm lesions. Within the ≤1 cm stratum, entering the exactly-1.0 cm indicator and a size-independent histology exposure in the same model assigned the apparent small-polyp effect to the boundary lesions themselves (adjusted OR 2.53, 95% CI 1.32–4.87, P = 0.005) rather than to histology (adjusted OR 1.52, 95% CI 0.78–2.94, P = 0.219). The null estimate beyond 1 cm, in turn, reflected exposure saturation—1,269/1,414 patients (90%) were HRA-positive through the size criterion alone, leaving only 145 referents—rather than irrelevance of histology, which regained an informative contrast under the non-size definition (adjusted OR 2.04, 95% CI 1.21–3.44, P = 0.008); the interaction vanished under either remedy (non-size exposure, P = 0.95; continuous size modelling, P = 0.48–0.85). Consistent with size and histology carrying independent information, the HRA odds ratio falls from 3.0 to 1.71 (P = 0.06) once size is added in the component-decomposition model across the full cohort, while a polyp count ≥3 predicts bleeding independently in both strata (≤1 cm adjusted OR 2.50; >1 cm 2.28) (Supplementary Table S1). The sensitivity analysis restricted to lesions strictly smaller than 1.0 cm still rests on 12 events, with point estimates in the same direction under either definition (original HRA OR 3.05; non-size definition 2.96) but profile CIs including 1 (Supplementary Table S14), so a histology signal in truly diminutive lesions can be neither confirmed nor excluded. The interaction is therefore best read as a definitional artifact; its practical message survives in modified form—small polyps do not carry uniformly low bleeding risk, chiefly because lesions at the 1 cm boundary bleed far more often than smaller ones (2.2% vs. 0.83%), with at most a suggestive histology contribution below 1.0 cm (16).
Pathology stratification showed the highest DPPB risk when HRA and HRSP coexisted (adjusted OR 5.54; bleeding rate 5.68%), with adjusted additive indices of RERI 1.37 (95% CI −2.65 to 5.38), S 1.43, and AP 0.25. These values are consistent with a trend toward positive additive interaction, although the wide intervals—driven by the three events in the HRSP-only group—temper the strength of the conclusion; nominally, a quarter of the risk in the co-presence group is attributable to the interaction. The multiplicative term was non-significant (OR 0.84, P = 0.805), showing no departure from independent multiplicative effects. Even if the two factors act through biologically separate channels, their combination can push absolute bleeding risk beyond the sum of the individual effects. HRA may heighten bleeding because villous adenomas carry richer microvascular structure, and high-grade dysplasia entails architectural disarray and neovascularization, both of which leave the post-EMR wound prone to oozing or arterial hemorrhage (17). HRSP may operate differently: serrated lesions often grow flat or laterally and cluster in the right colon, a region where several studies record higher post-polypectomy bleeding rates (18).
The multiplicative test asks whether the joint effect exceeds the product of the individual effects, whereas the additive test asks whether the absolute risk difference surpasses the sum of the individual risk differences. Here the multiplicative interaction OR of 0.84 (P = 0.805) indicates the joint effect stays within the range expected of a product, hinting that HRA and HRSP may partly overlap in their causal pathways. The additive indices (RERI 1.37, AP 0.25), though estimated on sparse events, nominally place 25% of the excess absolute risk among patients exposed to both high-risk polyp types on synergy between them. Reference patients (neither HRA nor HRSP) bled at 1.23% vs. 5.68% in the co-presence group, an absolute difference of 4.45 percentage points; per 1,000 co-presence patients roughly 57 bleed vs. 12 in the reference group, leaving about 45 excess events attributable to the dual exposure. The standalone HRSP estimate rests on sparse events (3 bleeds among 101 patients), but this does not weaken the highly significant co-presence effect (P < 0.001). Patients carrying both high-risk adenoma and high-risk serrated features are, in absolute terms, a very-high-risk subgroup for DPPB; whether measures such as prophylactic hemostatic clipping, extended post-procedural observation, or closer outpatient follow-up (19) should be adopted for them is a consideration to be tested prospectively, not a recommendation supported by this single-center retrospective association.
BMI emerged as protective against DPPB (OR 0.93 per kg/m2, P = 0.025), each unit lowering bleeding risk by 7%. Such a direction has seldom been reported and merits discussion, though the effect is modest: the 7% reduction per unit carries a upper confidence limit (0.99), so its clinical import may be limited. The inverse association can still be read in three ways.
The first is an obesity paradox: across critical illnesses such as heart failure, end-stage renal disease, and respiratory failure, overweight or obese patients often fare better (20). This pattern is ascribed to a subclinical procoagulant state (20). In this state, higher BMI tracks with elevated fibrinogen, heightened platelet activity, and inhibited fibrinolysis—a thrombotic tilt that, in a bleeding context, may act protectively (21).
The second concerns nutritional reserve and tissue repair: patients with higher BMI typically carry greater reserves and may synthesize protein and regenerate tissue more readily, so the post-polypectomy wound heals faster and the bleeding window narrows (22). The third is operator behavior: larger patients are often presumed to be technically harder, prompting endoscopists toward more aggressive prevention—more prophylactic clips, longer intraprocedural observation (23). The BMI–DPPB link is therefore unlikely to reflect a single mechanism but rather the combined action of coagulation, reserve, and operator bias, a question prospective studies should resolve.
The inverse pattern extends beyond age and BMI. Diabetes mellitus, renal impairment, and antithrombotic use each trended protective in the unadjusted analysis (OR 0.41, 0.57, and 0.43, respectively) yet attenuated toward the null under multivariable adjustment (Model 3: diabetes OR 0.52, P = 0.104; anticoagulant or antiplatelet use OR 0.62, P = 0.273; CKD stages 2–5 OR 0.85, P = 0.562). These inverse crude associations most plausibly reflect a combination of patient- and lesion-case-mix confounding and unmeasured periprocedural risk mitigation—chiefly antithrombotic agent interruption—rather than a biological shield conferred by the comorbidity itself; differential prophylactic clipping was specifically excluded as a contributing channel, since clipping rates did not differ by diabetes (80.7% vs. 79.1%) or antithrombotic use (79.5% vs. 79.3%), and clip receipt was driven by lesion risk rather than comorbidity. Consistent with risk-adapted management, antithrombotic-treated patients underwent resection of larger lesions (mean maximum diameter 1.21 vs. 1.11 cm, P < 0.001) yet developed DPPB less often (5.66% of bleeders vs. 12.14% of non-bleeders, P = 0.043). Each association is read descriptively, without causal attribution.
The simplified score offers clear practical advantages. It rests on only five variables—age, sex, BMI, polyp size, and polyp count—each available before or during the procedure, so the endoscopist can stratify risk and decide on prophylaxis before leaving the table, turning risk assessment into a real-time bedside judgment. Pathology (high-risk adenoma) was deliberately left out because it is unavailable until after the procedure; a Lasso analysis over a twelve-candidate pool retained all twelve candidates at the optimal penalty, with the five clinically chosen predictors all among them and age, size, and count entering the coefficient path first; adding HRA to the score left discrimination unchanged (ΔAUC 0.000; Supplementary Table S8), confirming that its information is already captured by size and count. The score also dispenses with complex comorbidity or lifestyle data. Earlier predictive models have incorporated diabetes, hypertension, anticoagulation, smoking, and alcohol intake, information that demands detailed history-taking and electronic-record review, inflating the time cost of data collection and degrading completeness and accuracy through reliance on patient recall and chart quality (11, 24). The five variables chosen here are objective, readily captured, and immune to subjective reporting bias, and they require no prior records, reducing the chance of missing data. The score needs no device or online tool: an integer point table that any clinician can use, with zero training barrier and strong portability.
The three-tier stratification is intended to inform rather than dictate management, and its boundaries are robust rather than arbitrary: the integer cuts track the empirical 33rd and 67th score percentiles, and sensitivity analyses across median split, quartile, and a 5% risk-threshold rule produced comparable gradients (high-to-low rate ratios 4.8–12.1), so the layering does not hinge on a single cutpoint (Supplementary Table S9). The gradient itself is descriptive: low-risk patients (0.73%) bled in fewer than 1 in 100 cases, whereas high-risk patients (8.84%) bled at roughly 12 times the low-group rate. How such risk information might eventually guide decisions on prophylaxis, observation, resection strategy, scheduling, or follow-up is a matter for prospective multicenter study; the present single-center retrospective association does not warrant prescribing routine prophylactic clipping, extended observation, staged resection, or day-case discharge for any risk tier.
Several limitations warrant mention. First, the retrospective design cannot rule out unmeasured confounding; intraprocedural variables (polyp morphology, prophylactic clipping, intraprocedural bleeding) were extracted from the free-text reports and entered the adjusted models. Moreover, prophylactic clipping was discretionary and unprotocolized, precluding adjustment for a structured clipping indication. En-bloc vs. piecemeal technique, defect size, antithrombotic interruption or resumption, the duration of post-procedure observation or inpatient admission, and periprocedural glycemic control were not captured in structured fields and remain unmeasured confounders. Second, DPPB was ascertained through retrospective review of the hospital electronic record without active 30-day patient contact, so admissions to other hospitals and minor self-limited bleeding may have been missed, and precise time-to-bleed was not uniformly recorded. Third, the analysis is patient-level: in patients with multiple lesions the largest diameter, total count, and HRA/HRSP status are summarized per patient without attributing the bleed to a specific lesion, which limits lesion-level inference. Fourth, the 106 bleeding events constrain power in subgroup analyses—only 3 events occurred in the HRSP-only group, and some covariates in the ≤1 cm stratum were estimated imprecisely, so those results call for cautious interpretation. Fifth, the score has undergone only internal bootstrap validation; its discrimination (AUC ≈ 0.73) is moderate, the optimism-corrected calibration slope is 0.923, and its generalizability is untested in an external cohort, which multicenter prospective studies will need to establish. Sixth, because the HRA definition embeds size (≥1 cm) and adenoma count (≥3), the HRA × size interaction must be interpreted with methodological care: a lesion of exactly 1.0 cm satisfies the HRA size criterion yet falls within the ≤1 cm stratum, so the ≤1 cm HRA estimate is largely carried by these boundary lesions. Further analyses showed that this interaction is definition-dependent—it disappeared when the size criterion was removed from the exposure (P = 0.95) or when size was modelled continuously (P = 0.48–0.85)—and that recorded sizes cluster at the 1.0 cm boundary (1,175 patients vs. 33 at 0.9 cm and 6 at 1.1 cm), a digit preference that makes the boundary group itself a rounding mixture of true 9–11 mm lesions. Within-stratum comparison was intended to control size by design, but the shared boundary leaves stratification itself vulnerable to this overlap, so continuous modelling of size is the more robust approach; restricting the ≤1 cm stratum to lesions strictly smaller than 1.0 cm left only 12 bleeding events and an HRA point estimate (OR 3.05) that was no longer significant, so any histology signal in truly diminutive lesions remains suggestive rather than conclusive. Seventh, calendar year was included as a covariate and showed no significant temporal trend across 2016–2025 (adjusted OR 1.08 per year, P = 0.12; Supplementary Table S13), yet unmeasured secular changes in practice may persist. Eighth, this is a single-center study, and external generalizability remains to be confirmed.
In summary, HRA is an independent risk factor for DPPB after EMR, and high-risk histology predicts bleeding across the size spectrum, while polyp size is itself a strong, independent risk factor. Polyp pathology should therefore enter bleeding-risk assessment on equal footing with size. A simplified five-variable score showed adequate discrimination and calibration on internal validation but, as a single-center instrument, awaits external confirmation.
Acknowledgments
The authors thank all endoscopists and nursing staff of the Gastroenterology Endoscopy Center at Peking University International Hospital for their support in data collection.
Funding Statement
The author(s) declared that financial support was not received for this work and/or its publication.
Footnotes
Edited by: Ravi Kumar Sharma, Chandigarh University, India
Reviewed by: Brett J. Hoskins, Indiana University School of Medicine, United States
Tengfei Hao, The Seventh Affiliated Hospital of Sun Yat-sen University, China
Data availability statement
The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.
Ethics statement
The studies involving humans were approved by Peking University International Hospital Biomedical Ethics Committee. The studies were conducted in accordance with the local legislation and institutional requirements. The ethics committee/institutional review board waived the requirement of written informed consent for participation from the participants or the participants' legal guardians/next of kin because this study is a retrospective cohort study that involved only the review of existing medical records and routine clinical data. No additional interventions, procedures, or direct patient contact were performed for the purposes of this research. All data were fully de-identified prior to analysis, and patient confidentiality was strictly maintained. Given the retrospective design, the large sample size, and the extended study period (October 2016 to March 2025), obtaining written informed consent from all included patients was impracticable and would have posed an undue burden. The study presented no more than minimal risk to participants. The study protocol was reviewed and approved by the Institutional Review Board (IRB)/Ethics Committee of Peking University International Hospital, which formally waived the requirement for written informed consent [approval number: 2025-KY-0101-01]. The research was conducted in accordance with the Declaration of Helsinki and applicable institutional and national guidelines for retrospective studies.
Author contributions
YD: Supervision, Conceptualization, Writing – review & editing, Software, Writing – original draft, Investigation, Visualization, Project administration, Formal analysis, Validation, Data curation. HL: Writing – review & editing, Project administration, Investigation, Software, Data curation. XL: Project administration, Supervision, Writing – review & editing, Resources.
Conflict of interest
The author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Generative AI statement
The author(s) declared that generative AI was not used in the creation of this manuscript.
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Supplementary material
The Supplementary Material for this article can be found online at: https://www.frontiersin.org/articles/10.3389/fmed.2026.1913120/full#supplementary-material
References
- 1.Lee DJ, Parikh A, Sirohi B, Cao Y, Chan AT. Emerging trends in the global burden of colorectal cancer. Nat Rev Clin Oncol. (2026) 23(8):569–81. 10.1038/s41571-026-01149-8 [DOI] [PubMed] [Google Scholar]
- 2.Tang J, Lam GT, Brooks RD, Miles M, Useckaite Z, Johnson IRD, et al. Exploring the role of sporadic BRAF and KRAS mutations during colorectal cancer pathogenesis: a spotlight on the contribution of the endosome-lysosome system. Cancer Lett. (2024) 585:216639. 10.1016/j.canlet.2024.216639 [DOI] [PubMed] [Google Scholar]
- 3.Bleijenberg AGC, IJspeert JEG, Mulder JBG, Drillenburg P, Stel HV, Lodder EM, et al. The earliest events in BRAF-mutant colorectal cancer: exome sequencing of sessile serrated lesions with a tiny focus dysplasia or cancer reveals recurring mutations in two distinct progression pathways. J Pathol. (2022) 257:239–49. 10.1002/path.5881 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Copland AP, Kahi CJ, Ko CW, Ginsberg GG. AGA clinical practice update on appropriate and tailored polypectomy: expert review. Clin Gastroenterol Hepatol. (2024) 22:470–9. 10.1016/j.cgh.2023.10.012 [DOI] [PubMed] [Google Scholar]
- 5.Ortigão R, Weigt J, Afifi A, Libânio D. Cold versus hot polypectomy/endoscopic mucosal resection-a review of current evidence. United Eur Gastroent. (2021) 9:938–46. 10.1002/ueg2.12130 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Albéniz E, Montori S, Rodríguez de Santiago E, Lorenzo-Zúñiga V, Álvarez MA, Estremera-Arevalo F, et al. Preventing postendoscopic mucosal resection bleeding of large nonpedunculated colorectal lesions. Am J Gastroenterol. (2022) 117:1080–8. 10.14309/ajg.0000000000001819 [DOI] [PubMed] [Google Scholar]
- 7.Ferlitsch M, Moss A, Hassan C, Bhandari P, Dumonceau J-M, Paspatis G, et al. Colorectal polypectomy and endoscopic mucosal resection (EMR): european society of gastrointestinal endoscopy (ESGE) clinical guideline. Endoscopy. (2017) 49:270–97. 10.1055/s-0043-102569 [DOI] [PubMed] [Google Scholar]
- 8.Akizue N, Okimoto K, Matsumura T, Matsusaka K, Inaba Y, Horio R, et al. Utility of prophylactic clipping after colorectal cold snare polypectomy in patients on oral antithrombotic agents. Gastrointest Endosc. (2024) 99:826–9. 10.1016/j.gie.2024.01.016 [DOI] [PubMed] [Google Scholar]
- 9.Li Z, Yu F, Wang C, Du Z. Comparing efficacy and factors of postoperative bleeding in endoscopic mucosal resection vs coagulation for intestinal polyps. Med (Baltimore). (2023) 102:e34941. 10.1097/MD.0000000000034941 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Polychronidis G, He M, Vithayathil M, Knudsen MD, Wang K, Song M. Risk of colorectal neoplasia after removal of conventional adenomas and serrated polyps: a comprehensive evaluation of risk factors and surveillance use. Gut. (2024) 73:1675–83. 10.1136/gutjnl-2023-331729 [DOI] [PubMed] [Google Scholar]
- 11.Xu L, Zhang N, Zhang Y, Luo D, Lu H, Wang Y, et al. A dynamic online nomogram for predicts delayed postoperative bleeding after colorectal polyp surgery. Sci Rep-Uk. (2024) 14:19728. 10.1038/s41598-024-70635-9 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Lai EJ, Calderwood AH, Doros G, Fix OK, Jacobson BC. The Boston bowel preparation scale: a valid and reliable instrument for colonoscopy-oriented research. Gastrointest Endosc. (2009) 69:620–5. 10.1016/j.gie.2008.05.057 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Aronchick CA. Bowel preparation scale. Gastrointest Endosc. (2004) 60:1037–9. 10.1016/s0016-5107(04)02213-8 [DOI] [PubMed] [Google Scholar]
- 14.Bronsgeest K, Huisman JF, Langers A, Boonstra JJ, Schenk BE, De Vos Tot Nederveen Cappel WH, et al. Safety of endoscopic mucosal resection (EMR) of large non-pedunculated colorectal adenomas in the elderly. Int J Colorectal Dis. (2017) 32:1711–7. 10.1007/s00384-017-2892-7 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Park S-K, Seo JY, Lee M-G, Yang H-J, Jung YS, Choi KY, et al. Prospective analysis of delayed colorectal post-polypectomy bleeding. Surg Endosc. (2018) 32:3282–9. 10.1007/s00464-018-6048-9 [DOI] [PubMed] [Google Scholar]
- 16.Benazzato L, Zorzi M, Antonelli G, Guzzinati S, Hassan C, Fantin A. Colonoscopy-related adverse events and mortality in an Italian organized colorectal cancer screening program. Endoscopy. (2021) 53:501–8. 10.1055/a-1228-9225 [DOI] [PubMed] [Google Scholar]
- 17.Staton CA, Chetwood ASA, Cameron IC, Cross SS, Brown NJ, Reed MWR. The angiogenic switch occurs at the adenoma stage of the adenoma carcinoma sequence in colorectal cancer. Gut. (2007) 56:1426–32. 10.1136/gut.2007.125286 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Crockett SD, Khashab M, Rex DK, Grimm IS, Moyer MT, Rastogi A, et al. Clip closure does not reduce risk of bleeding after resection of large serrated polyps: results from a randomized trial. Clin Gastroenterol Hepatol. (2022) 20:1757–65. 10.1016/j.cgh.2021.12.036 [DOI] [PubMed] [Google Scholar]
- 19.Turan AS, Pohl H, Matsumoto M, Lee BS, Aizawa M, Desideri F, et al. The role of clips in preventing delayed bleeding after colorectal polyp resection: an individual patient data meta-analysis. Clin Gastroenterol Hepatol. (2022) 20:362–71. 10.1016/j.cgh.2021.05.012 [DOI] [PubMed] [Google Scholar]
- 20.Pan J, Yin J, Gan L, Xue J. Two-sided roles of adipose tissue: rethinking the obesity paradox in various human diseases from a new perspective. Obes Rev. (2023) 24:e13521. 10.1111/obr.13521 [DOI] [PubMed] [Google Scholar]
- 21.Andersson M, Thorell A, Henriksson P, Kallner A, Wallén H, Lindahl TL, et al. Bariatric surgery improves fibrinolysis in individuals with obesity with and without concomitant type 2 diabetes. Clin Nut. (2026) 58:106584. 10.1016/j.clnu.2026.106584 [DOI] [PubMed] [Google Scholar]
- 22.Heutlinger O, Acharya N, Tedesco A, Ramesh A, Smith B, Nguyen NT, et al. Nutritional optimization of the surgical patient: a narrative review. Adv Nut. (2025) 16:100351. 10.1016/j.advnut.2024.100351 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Sarno G, Reytor-González C, Frias-Toral E, Campuzano-Donoso M, Katsanos CS, Simancas-Racines D. Navigating the weight: the impact of obesity on gastrointestinal cancer surgery and strategies for improved outcomes. Semin Cancer Biol. (2025) 114:138–49. 10.1016/j.semcancer.2025.06.010 [DOI] [PubMed] [Google Scholar]
- 24.Albéniz E, Gimeno-García AZ, Fraile M, Ibáñez B, Guarner-Argente C, Alonso-Aguirre P, et al. Clinical validation of risk scoring systems to predict risk of delayed bleeding after EMR of large colorectal lesions. Gastrointest Endosc. (2020) 91:868–78. 10.1016/j.gie.2019.10.013 [DOI] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Data Availability Statement
The original contributions presented in the study are included in the article/Supplementary Material, further inquiries can be directed to the corresponding author.