Introduction The majority of patients with acute lower gastrointestinal bleeding (ALGIB) have a low risk of requiring intervention, rebleeding or death. Nevertheless in routine clinical practice most are admitted to hospital for observation and endoscopy increasing cost of care. There are no risk scores routinely used in clinical practice which differentiate high risk patients who should be admitted to hospital from those who could be managed as outpatients. British Society of Gastroenterology/Scottish intercollegiate guidelines network (SIGN) have published expert opinion based criteria for non-admission but the accuracy of these is unclear.
Methods The aim of this study was to compare an artificial neural network's (ANN) performance in distinguishing high-risk from low-risk patients with ALGIB to SIGN guidelines (six clinical variables) and BLEED score (five clinical variables). Data were collected retrospectively from patients with ALGIB who were admitted to the emergency department of a teaching hospital between 2007 and 2010 (n=174). A multi-layered perceptron ANN model using back propagation and logistic activation function with hidden nodes to make a prediction was constructed from 35 clinical and laboratory variables. The ANN was trained and validated internally using leave-one-out method. The primary composite end point was the need for intervention, rebleeding or death. Sensitvity, specificity, predictive values and accuracy were calculated to compare the performance of the scores in predicting the composite end point.
Results Overall demographics and outcome of the 174 patients identified with ALGIB were: mean age 68 year (range 16–99), male:female 1:1, rebleeding rate (16.1% n=28), 30 day in hospital mortality (2.3% n=4). The most common diagnoses were diverticular disease (36%), haemorrhoids (10%) and colorectal carcinoma (10%). Twenty-three patients (13%) required intervention; endoscopic therapy (n=7), angiographic embolisation (n=8), or surgery (n=8). Notably, only four (2.3%) patients satisfied the SIGN criteria for non-admission. Predictive scores for each tool were: ANN (sensitivity 50%, specificity 83%, PPV 44%, NPV 83%), BLEED (sensitivity 67%, specificity 44%, PPV 28%, NPV 81%) and SIGN (sensitivity 100%, specificity 3%, PPV 25%, NPV 100%). The ANN performed significantly better in predicting the composite outcome (accuracy 0.76, 95% CI 0.70 to 0.83) compared with BLEED (0.49, 95% CI 0.42 to 0.57) and SIGN (0.26, 95% CI 0.20 to 0.33) scores.
Conclusion A non-endoscopic based artificial neural network model was more accurate than published guidelines/scores in predicting an adverse outcome in patients with ALGIB.
Competing interests None declared.