Abstract:
|
Accurately predicting the outcomes of online randomized controlled trials (RCT) can facilitate the implementation of actual interventions. Complex models are less preferred in real clinical studies, especially those without tractability or interpretability. For example, due to its black-box nature, neural networks are less used in RCT outcome prediction. Our previous work studied a Regularized modified-generalized-neuro-fuzzy (R-mGNNF) classifier and examined its utility in predicting the outcome of a longitudinal behavioral RCT. Our preliminary results indicate the R-mGNNF exhibits higher accuracy compared to similar fuzzy logic-based classifiers. There has been no uniform agreement on validation metrics for neuro-fuzzy classifiers, to our knowledge. Since accuracy has been criticized as not being a sufficient metric to be used in different scenarios, we propose a visualization-aided validation technique to evaluate neuro fuzzy classifiers. This method also incorporates the idea of confusion entropy but based on fuzzy set theory, to evaluate multi-class classifiers. Using our real and simulated data, we systematically evaluate our proposed metric against other well known metrics.
|