Introduction

Endoscopic submucosal dissection (ESD) is a widely used minimally invasive treatment for early gastric cancer (EGC). Post-ESD bleeding, which occurs in 3.1–6.5% of cases, is the most common complication^1,2. With Japan's aging population, more patients with EGC are receiving antithrombotic agents, increasing the risk of bleeding. Accurate prediction of post-ESD bleeding facilitates the implementation of effective preventive measures.
Machine learning (ML) models can predict clinical outcomes and aid in identifying crucial factors³. Thus, a predictive model based on pre- and intra-operative data may prove useful for predicting post-ESD bleeding and elucidating its key contributing factors.

Aims & Methods

This study aimed to develop an ML predictive model for post-ESD bleeding in patients with EGC, using a large sample dataset obtained from a multicenter study, and to investigate the key factors associated with post-ESD bleeding.
This multicenter study included patients who underwent ESD for EGC at Niigata University Hospital and Murakami General Hospital between 2012–2023. Patient data such as age, sex, development of post-ESD bleeding, comorbidities, and intake of antithrombotic agents (i.e., direct oral anticoagulants [DOACs], warfarin, or antiplatelets) were pulled from medical records. The lesion locations and specimen sizes were also recorded. Clinical characteristics were compared between the patients without post-ESD bleeding (control group) and those with bleeding (bleeding group). The factors associated with post-ESD bleeding were investigated using a logistic regression model. A predictive ML model was then constructed using pre- and intra-operative data. An extreme gradient-boosted tree classifier, a commonly used ML algorithm, was used to construct the model. To train it, the data were randomly partitioned so that 80% were used for training and 20% for testing. The area under the curve (AUC) was calculated for each model using the test dataset. Shapley additive explanations values, a quantitative method to evaluate the contribution of each variable to the model, were calculated to determine the key variables related to post-ESD bleeding.

Results

A total of 1,084 consecutive patients who underwent ESD for EGC were included, of whom 63 (5.8%) developed post-ESD bleeding. The median age [interquartile range] was significantly higher in the bleeding group than in the control one (77.0 [71.0–81.0] vs. 75.0 [70.0–79.0], p=0.040), although the female ratio was similar (22.2% vs. 25.0%, p=0.623). The frequencies of DOACs (39.7% vs. 9.0%, p<0.001), warfarin (27.0% vs. 4.8%, P <0.001), and P2Y12 receptor antagonist use (12.7% vs. 2.8%, p<0.001) were significantly higher in the bleeding group.
Multivariate logistic regression analysis showed significant associations between post-ESD bleeding and DOACs (odds ratio [OR]: 10.50, 95% confidence interval [CI]: 5.48–20.00, p<0.001), warfarin (OR: 10.70, 95% CI: 4.57–24.80, p<0.001), and P2Y12 receptor antagonist use (OR: 4.09, 95% CI: 1.55–10.80, p<0.001). Our ML model with all variables was able to predict post-ESD bleeding with an accuracy of 93.3% and exhibited an AUC of 0.83 (95% CI: 0.76–0.91). The use of DOACs and warfarin were the top two contributing factors in the ML model.

Conclusion

Post-ESD bleeding was predicted with high accuracy using an ML model, revealing that DOACs and warfarin represent key factors related to post-ESD bleeding. For patients predicted to experience bleeding, meticulous hemostasis measures such as post-ESD ulcer closures may be necessary.

References

1. Yoshio, T. et al. Effect of direct oral anticoagulants on the risk of delayed bleeding after gastric endoscopic submucosal dissection. Dig. Endosc. 29, 686–694 (2017).
2. Nagami, Y. et al. Antithrombotics increase bleeding after endoscopic submucosal dissection for gastric cancer: Nationwide propensity score analysis. Dig. Endosc. 34, 974–983 (2022).
3. Takahashi, K. et al. Achalasia phenotypes and prediction of per‐oral endoscopic myotomy outcomes using machine learning. Dig. Endosc. (2023) doi:10.1111/den.14714.