TOP 73 - Cross Institutional Validation of a Machine Learning Predictive Analytic Tool to Differentiate Bowel Necrosis from Intestinal Perforation

Publication Number: 4778.TOP 73

Allison C. Lure, Boston Children's Hospital, Chestnut Hill, MA, United States; Xinsong Du, Brigham and Women's Hospital/Harvard Medical School, Chelsea, MA, United States; Raechel Irons, Pediatrix Medical Group, Fort Myers, FL, United States; Jonathan L. Slaughter, Nationwide Children's Hospital and The Ohio State University, Columbus, OH, United States; Dominick Lemas, University of Florida College of Medicine, Gainesville, FL, United States; Daniel Ralph. Gipson, Norton Children’s Hospital / University of Louisville School of Medicine, Louisville, KY, United States; Diomel de la Cruz, UF Health Shands Children's Hospital, Gainesville, FL, United States; Josef Neu, University of Florida, GAINESVILLE, FL, United States

Background: While numerous machine learning models have been developed in recent years, most are derived from data at a single institution which hampers clinical implementation and generalizability. Previous attempts at cross-institutional model deployment have found the most success when retraining models on institution specific data, making multi-institutional deployment cumbersome. Integration of an institution-specific variable may obviate this requirement, but multi-institutional validation remains necessary. We previously built and validated random forest and ridge logistic regression machine learning models that were able to differentiate between bowel necrosis and intestinal perforation prior to surgery, with the hopes of improving clinical decision making and research, however this model was only validated at a single institution.

Objective: To determine the feasibility of creating and validating a multi-institutional novel machine learning predictive analytic tool capable of differentiating bowel necrosis from intestinal perforation in neonates.

Design/Methods: This is a retrospective study, with IRB exemption at all three participating hospital systems, inclusive of neonates born at ≤32 weeks gestation between 2011 and 2017, admitted to three level IV NICUs, who underwent abdominal drain placement or laparotomy for an intestinal injury and had bowel visualization. Neonates with an intestinal diagnosis other than bowel necrosis or perforation, such as atresia or malrotation, were excluded. Outcomes of bowel necrosis or intestinal perforation were determined by interoperative or autopsy findings. Predictors consisted of birth history, maternal history, medications, imaging findings, and laboratory values. Variables were limited to those available prior to or at the time of diagnosis, and before surgery proceeded to be clinically useful. Multiple machine learning models, including random forest, ridge regression, as well as other readily implementable approaches available in electronic health records, will be trained and validated on both single institutional and pooled data. We will test the model on different combinations of single institutional and pooled data and evaluate the area under the receiver operating curve. We will identify modifiable risk factors using interpretable machine learning models.