Semi-Supervised Approach to Predictive Analysis Using Temporal Data

redicting a target event from temporal data using supervised learning alone presents a number of chal-lenges. It assumes that members falling into the same class have similar historical characteristics, which is a too strong an assump-tion. Additionally, it can be difficult for the algorithm to underline the differences from a large volume of data and multitude of temporal projections. In such situations, a combination of supervised and unsupervised learning proved to be superior in performance as compared to supervised learning alone. In the proposed methodology, we develop feature vectors of temporal events that are subsequently split into groups by similarity of spatio-temporal characteristics using a clustering algorithm. We then apply a supervised learning methodology to predict the class within each of these subpopulations. We show a dramatic improvement in predictive power of this joint methodology as compared to supervised learning alone. The case study that we use to demonstrate the methodology utilizes medical claims data to predict a patient’s short-term risk of myocardial infarction. In particular, we identify groups of people with temporal diagnostic patterns associated with a high-risk of myocardial infarction in the coming three months. We use these patterns as a profile reference for assessing the state of new patients. We demonstrate that the newly developed combined approach yields improved predictions for myocardial infarction over using classification alone.