Automated alerts in obstetrics

Abstract

We thank Drs Klumpner, Kountanis, and Tremper for their interest in and response to our article. Their recent report 1 highlights several issues that merit continued attention by investigators seeking to improve patient outcomes using predictive analytics. First, their report is one of the few that provide temporal data on the performance characteristics of alerts. Such temporal data are not just needed to compare systems-they are an essential ingredient in the development of clinician workflows needed to respond to an alert. Second-and this is a deficiency shared by our model 2 -the positive predictive value (PPV) of their alerts (5.1%), while clearly superior to that of a manually assigned score (3.3%), is still fairly low compared with what is now possible in nonlaboring adults, which is in the range of 10 to 20%. 3 Finally, their report highlights an important challenge in obstetrics-small numbers. Although they tested their model in a population with a much higher baseline risk than ours (postpartum hemorrhage prevalence of 1.5% compared with 0.3% in our population), they only had 120 cases. Given that, in developed nations, most laboring women are healthy, low numbers will remain a continuing challenge.Although we employed different methods, our reports are in fundamental agreement: improvements in the early detection of impending or ongoing obstetrical complications using only manually calculated scores are unlikely. Given the low numbers (ie, weak signals), it is unlikely that simple combinations of vital signs can achieve what constitutes a "triple aim" in predictive analytics: high sensitivity, high PPV, and plenty of lead time. Only models or systems that achieve these 3 aims can be used to answer a much more important question: does early detection lead to improved patient outcomes?