1 Nov 1995, 12:00, WeH 7220 Using the Future to Help Predict the Present: Multitask Learning for Pneumonia Risk Prediction At our last talk, we introduced a pneumonia risk prediction problem where the goal was to predict which pneumonia patients have low enough probability of death that they might safely be treated as outpatients. That talk presented rankprop, a technique we developed to allow artificial neural nets to learn rankings that yielded a 10%-40% improvement in finding low-risk patients. In this talk, we show that an *additional* 10%-40% improvement can be obtained by combining multitask learning (MTL) with rankprop. The extra MTL tasks we use to get this performance increase are the results of hospital lab tests that will not be available until *after* a patient has been admitted to the hospital. Since these lab results will not be available when the admit/outpatient decision is being made, they cannot be used as inputs to the learner -- they won't be available at run-time. But they are available for the training set, and thus can be used as extra MTL outputs. We argue that using future data to bias learning in the present is an opportunity available in many domains. What is required is that the training set be from the past, so that feature values from its future can be collected. We also show empirical results that suggest that, at least on this domain, the competing non-MTL approach of using predictions of these lab values as extra inputs does not yield comparable performance.