Predicting outcomes of gastric endoscopic submucosal dissection using a Bayesian approach: a step for individualized risk assessment

• Main Authors: Diogo Libânio, Mário Dinis-Ribeiro, Pedro Pimentel-Nunes, Cláudia Camila Dias, Pedro Pereira Rodrigues
• Format: Article
• Language: English
• Published: Georg Thieme Verlag KG 2017-06-01
• Series: Endoscopy International Open
• Online Access: http://www.thieme-connect.de/DOI/DOI?10.1055/s-0043-106576

Background and study aims Efficacy and adverse events probabilities influence decisions regarding the best options to manage patients with gastric superficial lesions. We aimed at developing a Bayesian model to individualize the prediction of outcomes after gastric endoscopic submucosal dissection (ESD). Patients and methods Data from 245 gastric ESD were collected, including patient and lesion factors. The two endpoints were curative resection and post-procedural bleeding (PPB). Logistic regression and Bayesian networks were built for each outcome; their predictive value was evaluated in-sample and validated through leave-one-out and cross-validation. Clinical decision support was enhanced by the definition of risk matrices, direct use of Bayesian inference software and by a developed online platform. Results ESD was curative in 85.3 % and PPB occurred in 7.7 % of patients. In univariate analysis, male sex, ASA status, carcinoma histology, polypoid or depressed morphology, and lesion size ≥ 20 mm were associated with non-curative resection, while ASA status, antithrombotics and lesion size ≥ 20 mm were associated with PPB. Naïve Bayesian models presented AUROCs of ~80 % in the derivation cohort and ≥ 74 % in cross-validation for both outcomes. Risk matrices were computed, showing that lesions with cancer at biopsies, ≥ 20 mm, proximal or in the middle third, and polypoid are more prone to non-curative resection. PPB risk was < 5 % in lesions < 20 mm in the absence of antithrombotics. Conclusions The derived Bayesian model presented good discriminative power in the prediction of ESD outcomes and can be used to predict individualized probabilities, improving patient information and supporting clinical and management decisions.