|
Abstract:
|
Our previous work explored that our regularized modified neuro-fuzzy classifier (R-mGNNF) effectively predicts the outcome of longitudinal behavioral intervention data using demographic and intervention attributes. Now we extend the R-mGNNF 's capacity to perform the semi-supervised learning to predict the chronic disease outcomes using latent dietary trajectory patterns and available attributes. We propose an optimization method built upon SVM but tailored to our model to determine the decision boundary. Real-world dietary data (e.g., Biomarkers, latent diet quality trajectory patterns and relevant attributes) and simulation are used to identify the regularization parameter and associated fuzzifiers to improve the generalizability of R-mGNNF for chronic disease outcome prediction. We evaluate the performance of R-mGNNF against neural networks and their variants, fuzzy rule-based models, and generalized linear mixed models by using MSE as a metric. We use 60% of the data as training, 20% for validation and the remining for testing of our model, in addition to cross-validation. Our preliminary results indicate the proposed method minimizes MSE and computational time.
|
