Neuro-fuzzy modeling: An accurate and interpretable method for predicting bladder cancer progression

Purpose: New methods are required to improve the prediction of cancer progression as traditional statistical tests have limited accuracy. Accurate predictions would allow physicians to offer specific treatment according to individual patient risk. While predictive improvements are obtained using ANN, the hidden nature of these networks prevents insight and has hindered their widespread implementation. NFM is an alternate form of artificial intelligence using fuzzy logic (which is a multivalued logic which provides reasoning under uncertainty). By defuzzification the NFM rule base becomes transparent to overcome the black box nature of ANN. Materials and Methods: Combinations of clinicopathological (tumor stage and grade, patient age, gender, and smoking status) and molecular (immunohistochemical expression of p53 and methylation status of 11 loci) data from 117 patients were used to develop and compare predictive models of tumor progression using NFM, ANN and LR. Results: NFM (88% to 100% sensitivity, 97% to 100% specificity and 94% to 100% accuracy) predicted the presence and timing of cancer progression more accurately than ANN (81% to 87%, 95% to 100% and 89% to 90%, p = 0.002) and LR (3%, 61% to 72% and 47% to 53%, p = 0.00005). NFM was able to interrogate the clinicopathological and molecular data, and select the most important parameters (age, grade, stage, smoking, methylation) for progression prediction. Conclusions: Intelligent systems and molecular biomarkers improved the accuracy of cancer progression predictions. NFM appeared superior to ANN in terms of accuracy, sensitivity, specificity and transparency. The use of NFM in routine clinical practice warrants further validation.