ARTIFICIAL NEURAL NETWORK FOR THE PREDICTION OF MORTALITY IN PATIENTS PRESENTED WITH NON-VARICEAL UPPER GASTROINTESTINAL BLEEDING

Aims Upper gastrointestinal bleeding (UGIB) represents a common cause of gastroenterological admission and usually requires risk stratification for level of care determination as well as a rapid decision management. The aim of our study is to assess the use of an artificial neuronal network (ANN) that may help predict mortality in patients which present with non-variceal UGIB.

Methods All patients admitted with non-variceal UGIB within the Gastroenterology Department of Craiova County Hospital between 1^st of January 2017 and 31^st December 2019 were included in our study. We performed a patient analysis on 914 patients by using the Rockall, AIM65 and the Glasgow-Blatchford score. A two layers ANN of the endoscopic scores was developed using Python 3.10.0 and tried to provide a higher prediction of patient's mortality. The neural network was validated on the patients admitted in 2019. Each nod was assessed by using random weights of each parameter, which are further adjustable according to the prediction errors. The second layer consisted of 7 nods.

Results Our ANN was able to predict the mortality rate of patients non-variceal UGIB better than the three scores taken separately with an accuracy>95%. The second iteration revealed a specificity of 0.96 and a sensibility of 0.76 which was higher than the used scores alone.

Conclusions Using the three UGIB scores in ANN may perform better for patient’s mortality assessment in a non- variceal setting. However, there is a need of external validation in external patient’s population to be validated.

Publication History

Article published online:
14 April 2022