Anticipatory intervention, rather than a reactionary response to the deterioration of critically ill children, is ideal. Even for experienced clinicians, however, it is challenging to differentiate who among high-risk patients are at highest probability of decompensation, especially among patients with single-ventricle physiology. Making this prediction with certainty and foresight that afford time for preventive intervention has not been achievable thus far. In this issue of the Journal, Ruiz and colleagues 1 approach this dilemma by creating prediction models that are based on clinician knowledge and machine learning to attempt to anticipate critical events as early as 8 hours in advance. The study suggests that routinely collected data can be leveraged to build an early warning system, and that models centered on a combination of machine learning and expert knowledge are superior to models reliant exclusively on expert knowledge.Although the findings of Ruiz and colleagues 1 are of interest, various aspects of the study design and interpretation merit thoughtful reflection. Although a single-ventricle population was targeted, the study cohort may not be representative because of the number of patients housed outside a cardiac intensive care unit, the distribution of congenital heart disease lesions, and the timeline and frequency of adverse events. Specifically, the majority of the events occurred greater than 7 days after surgical palliation, which is incongruent with literature, suggesting that the highest risk of decompensation occurs during the immediate postoperative period. 2,3 The event frequency also appears lower than expected. Furthermore, a critical event was defined solely through medical records; however, many procedures may have been elective rather than truly emergency in nature. For example, many extracorporeal membrane oxygenation cannulations were not preceded by either cardiopulmonary resuscitation or intubation, 4 and only 5% of intubations were followed by further deterioration. In terms of variables, the data were not longitudinal, may have been outdated, and are not all