Classification of gastrointestinal bleeding data

摘要

Acute gastrointestinal bleeding (GIB) is an increasing healthcare problem due to rising NSAID (non-steroidal anti-inflammatory drugs) use in an aging population. In the emergency room (ER), the ER physician can misdiagnose a GIB patient at least 50% of the time. While it is best for a gastroenterologist to diagnose GIB patients, it is not feasible due to time and cost constraints. Classification models can be used to assist the ER physician to diagnose GIB patients more efficiently and effectively, targeting scarce healthcare resources to those who need it the most.;Currently, there have not been models developed which can predict all three sources of bleeding simultaneously (upper, middle, and lower bleeding). Eight classification models were trained and tested by performing ten repetitions of ten-fold cross validation on a 192 patient dataset. The classification models considered were: artificial neural network, boosting, k-nearest neighbor, linear discriminant analysis, logistic regression, random forest, shrunken centroid, and support vector machine. The four response variables classified were: source of bleeding, need for urgent resuscitation, need for urgent endoscopy, and disposition. Performance was assessed by accuracy and balance of sensitivity and specificity. The top three models (random forest, support vector machine, and artificial neural network) were externally validated. It was determined that random forest performed the best overall.;The Rockall and Blatchford scores have been previously developed for upper GIB patients. The random forest model was found to be comparable to these scores for upper GIB patients. In addition, simulation studies were done to compare the eight classification models and to compare to the results obtained with the actual GIB data. Simulated GIB data that was unbalanced versus balanced and correlated versus independent was considered, with accuracy and balance of sensitivity and specificity being the performance measures of the models. Random forest was again seen to be the best performing model. An online tool was developed for a user-friendly interface that physicians and nurses can utilize. This online tool will be utilized in future studies in the hope this tool or something similar can be adopted for routine use in caring for GIB patients.