The application of multi-model ensemble approach as a prognostic method to predict patient health status

Prognostic methods promise to improve patient healthcare if predictions of adverse disease and medical complications for each patient can be predicted in advance. Prognostics and prediction of patients' physiological health status are getting attention in medicine because they provide insight that can be used for medical interventions that prevent adverse medical complications. While various predictive analytics have been developed for detection and prediction of certain diseases, efforts to combine the predictive power of multiple algorithms have gone mostly unnoticed. This study proposes a prognostics engine using multiple models to predict patient physiological status. Given the diversity of clinical data and disease conditions, no single model can be the ideal prediction algorithm to cover all medical cases. Certain algorithms are more accurate than others depending on input data available, the type, amount and diversity of possible outcomes. In this study four different neural network algorithms were used for the prognostics engine and their accuracy on a dataset were compared. The study proposes using an ensemble of algorithms and an oracle, an overseer program to select the most accurate combination of the predictive models that is most suited for a particular disease prediction. The feasibility of this approach is tested using a clinical data set of 1,073 patient cases including 255 patients presented with Deep Vein Pulmonary Embolism. The study compared accuracy of five different schemas for constructing ensembles of various neural networks. The multiple schema approach combined with multi-model ensembles showed to improve accuracy of prediction for this case and promises to be a robust approach to other clinical prediction problems.