Neural network analysis of clinicopathological and molecular prognostic markers in bladder cancer

Introduction: The objective of this study was to retrospectively evaluate the ability of an artificial neural network (ANN) to predict bladder cancer recurrence and stage progression in patients with Ta/Tl bladder cancer, and 12 month cancer-specific survival, in a group of patients with T2-T4 bladder cancer. Patients and methods: Data were extracted from a database and analysed using a NeuralWorks Professional IIIPlus software package using the radial basis function algorithm. The input neural data for the prediction of stage progression in Ta/Tl bladder cancers (n = 105) included tumour stage, grade, size, multiplicity, male/female and epidermal growth factor receptor status. For the prediction of tumour recurrence within 6 months for Ta/Tl tumours (n= 56) and 12 month cancer-specific survival for T2T4 tumours (n = 40), additional neural inputs were; smoking habit, histology of mucosal biopsies, presence of carcinoma in situ, tumour metaplasia, tumour architecture/site, c-erbB-2 and p53 status. The network was trained on randomly selected patients and validated on the remaining exemplars. Results: The ANN analysis of the data gave (%); Sensitivity Specificity Accuracy Stage progression (Ta/T1) 70 82 80 Recurrence within 70 80 75 6 months (Ta/T1) 12 month cancer-specific 100 80 82 survival (T2-T4) When restricting the testing subset to patients with T1G3 tumours in relation to stage progression, the sensitivity of the ANN analysis ncreased to 100%. with a specificity of 78% and an accuracy of 82%. Conclusions: ANNs have been shown to be a useful adjunct in predicting outcomes in patients with bladder cancer.