Magazine article AI Magazine

A Machine-Learning Approach to the Detection of Fetal Hypoxia during Labor and Delivery

Magazine article AI Magazine

A Machine-Learning Approach to the Detection of Fetal Hypoxia during Labor and Delivery

Article excerpt

The lifelong disability that can result from oxygen deprivation during childbirth is rare but devastating for families, clinicians, and the health-care system. Between 1 and 7 in 1000 fetuses experience oxygen deprivation during labor that is severe enough to cause fetal death or brain injury (Saphier et al. 1998); the range of this estimate reflects considerable regional variation and some clinical debate on the definition of brain injury. The main source of information used by clinicians to assess the fetal state during labor is cardiotocography (CTG), which measures maternal uterine pressure (UP) and fetal heart rate (FHR). These signal are routinely recorded during labor, using monitors of the type presented in figure 1.

Clinicians look at these signals and use visual pattern recognition and their prior experience to decide whether the fetus is in distress and to pick an appropriate course of action (such as performing a Cesarean section). However, there is great variability among physicians in terms of how they perform this task (Parer et al. 2006). Furthermore, because significant hypoxia is rare, false alarms are common, leading physicians to disregard truly abnormal signals. Indeed, approximately 50 percent of birth-related brain injuries are deemed preventable, with incorrect CTG interpretation leading the list of causes (Draper et al. 2002; Freeman, Garite, and Nageotte 2003). The social costs of such errors are massive: intrapartum care generates the most frequent malpractice claims and the greatest liability costs of all medical specialties (Saphier et al. 1998). Thus, there is great motivation to find better methods to discriminate between healthy and hypoxic conditions.


In this article, we summarize our recent work on a novel approach to this problem, which relies heavily on machine-learning methods; a more detailed account of the methods is presented in two biomedical journal publications (Warrick et al. 2009; 2010) as well as in a Ph.D. dissertation (Warrick 2010). Although the system we present Was designed specifically for labor monitoring, we believe that the general steps we took can inform other AI medical monitoring systems, as well as, more generally, applications for time-series prediction and analysis.

We built an automated detector of fetal distress by using data from normal and pathological cases. We had access to a unique database, which contains labor monitoring data from an unusually large number of births; a significant number of the cases are pathological examples (well above the natural frequency of occurrence of such problems). All the data has been collected under clinical conditions; as a result, it is very noisy. To handle this problem, we modeled the fetal heart rate signal through several components. The parameters of these models, which have been learned from data, are then used to build a classifier for a given time period. Because the state of the fetus can change during labor, classification is performed repeatedly on data segments of limited duration. A majority vote of recent labels determines if and when a fetus is considered pathological.

The article is organized as follows. First, we give some background on the problem and type of data used. Then, we describe our general approach. We present empirical results and a discussion of the main findings, as well as the next steps towards clinical deployment.


Clinicians' interpretation of intrapartum CTG signals relies on the temporary decreases in FHR (FHR decelerations) in response to uterine contractions. FHR decelerations are due mainly to two contraction-induced events: (1) umbilical-cord compression and (2) a decrease in oxygen delivery through an impaired utero-placental unit. There is general consensus that deceleration depth, frequency, and timing with respect to contractions are indicators of both the insult and the ability of the fetus to withstand it. …

Search by... Author
Show... All Results Primary Sources Peer-reviewed


An unknown error has occurred. Please click the button below to reload the page. If the problem persists, please try again in a little while.