Pre- and post-operative predictions of recurrence in patients with cancer of the oesophago-gastric junction using radial basis function artificial neural networks

Pre-operative data from 103 patients undergoing potentially curative resection of adenocarcinoma of the oesophago-gastric junction were collected prospectively and analysed by a radial basis function artificial neural network system. A separate neural structure was designed for use with selected additional post-operative parameters. Output variables for both systems were recurrence at 12, 18 and 24 months. Prediction specificities, using pre-operative data alone, were 60%, 62% and 60%, respectively. Inclusion of post-operative parameters improved the specificity to 72.7% at 12 months, 66% at 18 months and 62.2% at 24 months, while the sensitivity of prediction at 18 months reached 90% and at two years 93%. There was a strong correlation between the predictive value of pre- and post-operative findings in individual patients at each time period (correlation coefficient=0.6333, p<0.0001; 0.4873, p=0.0083; 0.266, p=0.0025). This study demonstrates that artificial neural networks are able to reliably predict, even with limited clinical pre-operative information, patients who are destined to fail when treated by surgery alone. This approach may have a role in assisting clinicians to achieve more appropriate selection of patients for surgery and neo-adjuvant therapy.