Cervical dilatation over time is a poor predictor of severe adverse birth outcomes: a diagnostic accuracy study

Objective: To assess the accuracy of the World Health Organization (WHO) partograph alert line and other candidate predictors in the identification of women at risk of developing severe adverse birth outcomes. Design: A facility-based, multicentre, prospective cohort study. Setting: Thirteen maternity hospitals located in Nigeria and Uganda. Population: A total of 9995 women with spontaneous onset of labour presenting at cervical dilatation of ≤6 cm or undergoing induction of labour. Methods: Research assistants collected data on sociodemographic, anthropometric, obstetric, and medical characteristics of study participants at hospital admission, multiple assessments during labour, and interventions during labour and childbirth. The alert line and action line, intrapartum monitoring parameters, and customised labour curves were assessed using sensitivity, specificity, positive and negative likelihood ratios, diagnostic odds ratio, and the J statistic. Outcomes: Severe adverse birth outcomes. Results: The rate of severe adverse birth outcomes was 2.2% (223 women with severe adverse birth outcomes), the rate of augmentation of labour was 35.1% (3506 women), and the caesarean section rate was 13.2% (1323 women). Forty-nine percent of women in labour crossed the alert line (4163/8489). All reference labour curves had a diagnostic odds ratio ranging from 1.29 to 1.60. The J statistic was less than 10% for all reference curves. Conclusions: Our findings suggest that labour is an extremely variable phenomenon, and the assessment of cervical dilatation over time is a poor predictor of severe adverse birth outcomes. The validity of a partograph alert line based on the ‘one-centimetre per hour’ rule should be re-evaluated. Tweetable abstract: The alert line in check: results from a WHO study. The BOLD project included quantitative, qualitative, and service‐design research conducted in Nigeria and Uganda. The methodological details of the BOLD project have been described elsewhere.13, 14 This analysis is based on the quantitative component, a facility‐based, multicentre, prospective cohort study. In brief, this study included women admitted for vaginal birth with single live fetuses during the early first stage of labour across 13 hospitals in both countries. Women with spontaneous onset of labour presenting at cervical dilatation of ≤6 cm and women undergoing induction of labour took part in the study. Women with multiple pregnancies, women with pregnancies with gestational ages of less than 34 weeks 0 days, women choosing elective caesarean section, and women who were incapable of giving consent because of labour distress or obstetric emergencies at arrival were excluded. Participating institutions had a minimum of 1000 deliveries per year, with stable access to caesarean section, augmentation of labour, and assisted vaginal birth. Midwives, obstetricians, or obstetric residents provided intrapartum health care to women in labour. Doptones were used to assess fetal vital status at hospital admission and for intermittent monitoring through labour and childbirth. Labour management protocol, as well as the number and timing of pelvic examinations, were not standardised across participating institutions. None of the institutions subscribed to the active management of labour protocol during the study period. Although the partograph was a standard element of medical records in all participating health facilities, its prospective application to guide labour management during the study period varied widely across the hospitals. Eligible women were recruited into the study between December 2014 and November 2015. From the medical records, trained research nurses prospectively extracted detailed information on the sociodemographic, anthropometric, obstetric, and medical characteristics of the study participants at hospital admission, multiple assessments during labour monitoring, interventions performed throughout the first and second stages of labour, and maternal and neonatal labour outcomes. Attending staff were approached to complement medical records data when needed. Data collection was limited to hospital stay of the mother and baby, and there was no post‐hospital discharge follow‐up. The current analysis was based on information on maternal baseline and admission characteristics, repeated assessments of cervical dilatation versus time, and maternal and neonatal outcome data. Severe adverse birth outcomes were defined as the occurrence of any of the following: stillbirths, intra‐hospital early neonatal deaths, neonatal use of anticonvulsants, neonatal cardiopulmonary resuscitation, Apgar score of <6 at 5 minutes, uterine rupture, and maternal death or organ dysfunction with dystocia. Details of the sample size calculation are provided in the supporting information (Box S1). Simple frequencies and proportions were used to describe the characteristics of the study population. Sensitivity, specificity, positive and negative likelihood ratios, diagnostic odds ratios, and the J statistic (Youden's index), with 95% confidence intervals, were used to estimate the diagnostic accuracy of the alert line and the action line in the identification of women who would develop a severe adverse birth outcome.15, 16, 17, 18 We used the true‐positive rate (i.e. sensitivity) and the false‐positive rate (i.e. 1 – specificity) to graphically represent the diagnostic accuracy of the partograph parameters in the receiver operating characteristic (ROC) space.19 Each point estimate in the ROC space represents a classification result for binary parameters, and the interpretation of the ROC space is similar to the ROC curve: optimal results are associated with high true‐positive rates combined with low false‐positive rates. The J statistic summarises the performance of a binary classifier,16 and also expresses the proportion of ideal performance of a diagnostic test (Box S2). The supporting information provides additional details related to the calculation and interpretation of these statistics (Tables S1–S3). The alert line and the action line are classifiers currently applied to all women, regardless of their obstetric characteristics (e.g. nulliparous, multiparous, spontaneous or induced labour, or previous caesarean section). We hypothesised that cervical dilatation curves customised according to the obstetric characteristics of the population could have a better accuracy than the generic alert and action lines. The study population was stratified into mutually exclusive, totally inclusive obstetric groups according to the 10‐group Robson classification:20 group 1 (nulliparous, single cephalic pregnancy, 37 weeks of gestation or more, with spontaneous onset of labour); group 2 (nulliparous women, single cephalic pregnancy, 37 weeks of gestation or more, with induced onset of labour); group 3 (multiparous women without previous caesarean section, with single cephalic pregnancy, 37 weeks of gestation or more, with spontaneous onset of labour); group 4 (multiparous women without previous caesarean section, with single cephalic pregnancy, 37 weeks of gestation or more, with induced onset of labour); group 5 (all multiparous women with at least one previous caesarean section, single cephalic pregnancy, at 37 weeks of gestation or more); and group 10 (all women with singleton cephalic preterm pregnancy at less than 37 weeks of gestation at childbirth). As a result of the eligibility criteria, this study has no women from group 8 (multiple pregnancies) or with caesarean section before labour. Women with non‐cephalic presentations (groups 6, 7, and 9) were grouped together. Groups 1–5 and 10, were further divided according to the use of augmentation of labour (present or absent), totalling 12 subgroups. Using data from women who did not have any severe adverse birth outcome, customised labour curves were generated for each of these 12 subgroups. Data from women pertaining to groups 6, 7, and 9 were not used to generate customised curves because of the small numbers involved. The customised cervical dilatation curves were created using a multi‐state Markov model,21, 22 which represented the cervical dilation pattern through intermediate states from 2 cm to 10 cm, and childbirth by selected percentiles and obstetric group (i.e. one labour curve for each obstetric group and selected percentile). In this model, each centimetre of cervical dilatation represented an intermediate state, and childbirth was the final ‘absorbing’ state. The model was generated as a progressive unidirectional labour‐to‐childbirth model, and the time of state change was determined by a set of transition intensities. The transition intensity represents the instantaneous likelihood of moving from one state to another, and is generated as part of the multi‐state Markov model. For each one of the 12 obstetric subgroups, the multi‐state Markov model generated labour curves representing the progress of labour in women that was either faster or at the 50, 60, 70, 80, 90, and 95th percentiles. Once the percentile curves were generated for each obstetric subgroup of women without severe adverse birth outcomes, women were classified as having crossed or not having crossed each of the percentile curves of their relevant obstetric subgroup. The study population was then consolidated and all women who crossed their relevant 50th percentile curves were grouped together (i.e. women in which labour progressed more slowly than the customised 50th percentile curve). Similarly, women were classified as having labour that progressed either slower or faster/equal to the relevant 60, 70, 80, 90, and 95th percentiles. We estimated the accuracies of the customised percentile curves in the identification of women who would develop a severe adverse birth outcome, by comparing women with labour progress that was slower than the specific percentile with women in which labour progressed faster or equal to that percentile. Sensitivity, specificity, positive and negative likelihood ratios, diagnostic odds ratios, with 95% confidence intervals, the J statistic, and ROC space plotting were used to estimate the accuracy of the percentile curves in the identification of women who would develop a severe adverse birth outcome. Statistical analyses were carried in r and Microsoft excel (2010).23