Application of artificial neural networks in breast and colorectal surgery

• Main Author: Ramesh, Aswatha Narayana Murthy
• Other Authors: Drew, Philip
• Published: University of Hull 2008
• Subjects: 610; Medicine
• Online Access: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.503731

Accurate prediction of a clinical event in an individual patient is extremely useful, as treatment can be appropriately tailored to that individual, avoiding either over or under treatment. More importantly interventions based on prediction in a group of patients is no longer acceptable to an individual patient, as they increasingly demand to know what is likely to happen to them. Artificial neural networks (ANNs) are a nonlinear regression method capable of accurately predicting outcome in an individual patient. We identified two scenarios where accurate prediction in individual patient can greatly influence their management. In the first study, we investigated the ability of ANNs to predict axillary lymph node metastasis in patients with breast cancer and compared with traditional logistic regression analysis. In addition, the ability of ANNs to generalise across institutions was studied and their ability to predict in individual patient was explored. In the second study, the predictive capability of ANNs was explored to predict anastomotic leak following left sided large bowel anastomosis and compared with traditional logistic regression analysis. Multi-layer perceptron ANNs utilising a backpropagation learning algorithm were designed and then trained and validated on breast and colorectal cancer databases. ANNs were superior to logistic regression method in predicting lymph node metastasis in breast cancer. Furthermore, the previously trained ANNs were able to provide accurate predictions across an independent institution and in individual patients. ANNs achieved similar accuracy as logistic regression analysis in predicting anastomotic leak but were able to accurately predict leak in individual patients. ANNs have the potential to be used as predictive tool in individual patients in a variety of clinical setting.