The state-of-the-art monitoring systems for critical care measure vital signs and generate alerts based on the logic of general patient population models, but they lack the capabilities of accurately correlating physiological data with clinical events and of adapting to individual patient's characteristics that do not fit the population models. This research examines the feasibility of developing patient-specific alarm algorithms in real time at the bedside and evaluates the potential of these algorithms in helping improve patient monitoring. Modular components that facilitate real-time development of alarm algorithms were added to a system that simultaneously collects physiological data and clinical annotations at the bedside. At a pediatric intensive care unit (ICU), classification trees and neural networks for generating clinical alarms were trained for individual patients. These algorithms were evaluated immediately after training on subsequently collected data. The implemented system was capable of training and evaluating patient-specific algorithms in a consistent manner in real time at the bedside. The performance of patient-specific alarm algorithms improved as training data increased. Neural networks with eight hours of training data on average achieved a sensitivity of 0.96, a specificity of 0.99, a positive predictive value of 0.79, and an accuracy of 0.99; these figures were 0.84, 0.98, 0.72, and 0.98 respectively for the classification trees. These results suggest that real-time development of patient-specific alarm algorithms is feasible using machine learning techniques.