Abstracts of the AGA Institute

Aims: Artificial neural networks (ANN) are computer programs used to identify complexrelations within data that cannot be detected with conventional linear-statistical analysis.The routine clinical predictions of need for lower gastrointestinal endoscopy have beenbased on population statistics with little meaning for individual patient. This results in largenumber of unnecessary colonoscopies. We aimed to develop a neural network algorithmwhich can accurately predict presence of pathology in patients attending routine outpatientclinics. Methods: 300 patients undergoing lower gastrointestinal endoscopy prospectivelycompleted a specifically developed questionnaire which included 40 variables based onclinical symptoms, signs, past and family history. Complete data sets of 50 percent of serieswere used to train the artificial neural network; the remaining 50 percent were used forinternal validation. The primary output used was a positive finding on the colonoscopy,including polyps, cancer, diverticular disease or colitis. Results: The outcome and pathologyreports of all the patients were obtained and assessed. Clear correlation between actual datavalue and artificial neural network value were found (r = 0.931; P = 0.0001). The predictiveaccuracy of neural network was 95% in the training group and was 89% (95% CI 84-96)in the validation set. This accuracy was significantly higher than the clinical accuracy (69%).Conclusions: We have shown that ANN is more accurate than standard statistics whenapplied to prediction in individual patients of need for lower gastrointestinal endoscopy.These results have obvious implications, with at least 20% resultant decrease in need forunnecessary lower gastrointestinal endoscopy. The logistic and economic impact with thisdevelopment is tremendous