Medical predictive analytics offers the possibility to translate large amounts of data into patient-specific insights that support clinical decision-making and potentially improve patient outcomes. Indeed, many systems have been developed that estimate a patients' risk for a disease or a future medical event. However, none of these algorithms are routinely used in practice and this is often attributed to poor calibration, limited external validity, and inadequate workflow integration. In this talk, we introduce an algorithm that overcomes these limitations. Based on the general principles of super learning, we designed a framework for building and ensembling a library of pre-specified machine learning algorithms that are 1) regularly updated with batches of new data, and 2) learn from within individual time series as well as across patients. Applicability of this algorithm is demonstrated with a real data example involving forecasting acute hypotensive episodes, which are among the most frequent and critical events that occur in intensive care units. We show that this algorithm adapts over time to build trust with the clinician, and optimize patient-specific forecasts.