Abstract
With the increasing adoption of machine learning in digital trials for pretrial planning and participant management, the amount of data generated from each large longitudinal trial has consistently been increasing over recent years. This warrants innovation in methodologies and frameworks to handle these large data while retaining the interpretability of the resulting complex models. Specifically, neural networks have proven to be universal approximators and powerful in learning and prediction. However, their black-box nature limits their use in outcome studies of longitudinal digital trials. In contrast, fuzzy systems imitate human-like knowledge representation and explanation abilities yet lack a systematic methodology for their design. The proposed hybrid intelligent system, iPredict, encapsulates three novel components: a novel regularized modified Generalized Neuro-Fuzzy classifier (R-mGNNF), a novel regularization approach based on a support vector machine (SVM) to handle overfitting of the model, and a novel validation approach for the proposed system. R-mGNNF is a variant of adaptive artificial neural networks (ANN) that builds upon the theories of fuzzy logic and neural networks to complement their weaknesses. R-mGNNF is designed as a five-layer regularized neuro-fuzzy that tunes the fuzzifiers of the fuzzy rules in contrast to traditional neural networks. It is used to predict the outcome of longitudinal randomized control trials (RCT) and observational studies (OS) using demographic and background information about the participant. R-mGNNF was trained, tested, and validated using five real longitudinal trial studies and numerical analyses. Along with real longitudinal trial study data, iPredict was evaluated using simulated data generated with parameters and covariance matrices from these trials. We evaluate the performance of iPredict using five real world datasets obtained from four dietary intervention studies and one observation study and compare it to well-known kernel-based algorithms such as support vector regression (SVR) and relevant vector regression (RVR) for regression tasks and modified Generalized Neuro-Fuzzy classifiers and least square support vector machines for classification tasks. We also describe a preprocessing and data cleaning sub-section and present results for hyperparameter tuning using regularization parameters and number of membership functions. In addition, we use a suite of evaluation metrics to assess the performance of the proposed model, including true positive rate, false positive rate, precision, and F-measure. Our results show that iPredict outperforms other models in terms of RMSE and accuracy for regression tasks and better class separation for classification tasks.