Donabedian (1988) distinguished three aspects of care that can be assessed: structure, process and outcome. Outcomes of postpartum haemorrhage such as hysterectomy or death are too rare on a single ward to reveal much, and they depend on case mix. Process, meaning what was done and when, is more informative and more directly within the control of staff. Criterion-based audit compares the process of care recorded in case notes against a short list of explicit, evidence-based criteria, and it has been applied to obstetric emergencies in settings with few resources because it requires little more than records and a trained reviewer (Graham, 2009). The World Health Organization includes it among the recommended approaches for reviewing maternity care (WHO, 2004).

Checklists reduce reliance on memory in tasks where steps are known but easily omitted under pressure (Hales and Pronovost, 2006), and the surgical safety checklist demonstrated their effect on outcomes across diverse hospitals (Haynes et al., 2009). In obstetrics, the consensus bundle on obstetric haemorrhage of the National Partnership for Maternal Safety organised care into readiness, recognition and prevention, response, and reporting and systems learning (Main et al., 2015). A World Health Organization technical consultation defined a first response bundle and a bundle for refractory bleeding (Althabe et al., 2020). The E-MOTIVE trial then tested early detection with a calibrated drape together with a bundle of uterine massage, oxytocic drugs, tranexamic acid, intravenous fluids, examination and escalation, and found a lower risk of the composite of severe haemorrhage, laparotomy or death from bleeding than with usual care (Gallos et al., 2023). These sources supply the content from which audit criteria can be drawn.

The Cochrane review of audit and feedback found that it generally produces small to moderate improvements in professional practice, with larger effects when baseline performance is low, when feedback comes from a supervisor or colleague, when it is given more than once and in both verbal and written form, and when it includes explicit targets and an action plan (Ivers et al., 2012). An audit that ends with a report filed in an office should not be expected to change anything. The method below builds feedback and re-audit into the design for that reason.

A small multidisciplinary group, including midwives, obstetricians, an anaesthetist and a member of the blood bank staff, selects the criteria. Each criterion should be supported by guidance that the unit already accepts, should be verifiable from the record, and should refer to something staff can change. The criteria are grouped into the six domains set out in Table 1. The group also sets the standard for each, meaning the proportion of cases in which the criterion ought to be met. For core first-response actions the standard will normally be all cases; for criteria with legitimate exceptions it may be lower, and the exceptions should be written down.

Cases are identified from the birth register using the unit's own threshold for recording a haemorrhage. All cases in a defined period are preferable to a selection, because selection invites the choice of memorable cases. If the number is too large, a random or systematic sample is drawn from the list. The audit should state how cases were found, since bleeds that were never recognised will not be in the register, and under-recognition is itself a failure the audit cannot see directly. Visual estimation is known to underestimate larger losses (Bose et al., 2006), so a ward that relies on it will register fewer cases than one that measures.

Two abstractors who were not involved in the care review each record independently for the first batch of cases. For every criterion they mark met, not met, or not applicable, and a fourth option, not recorded, when the notes are silent. Keeping not recorded separate from not met is essential, because the two call for different remedies. Where the record gives clock times, the abstractors note the time of birth, recognition, call for help, first uterotonic, tranexamic acid and arrival of senior staff, from which intervals can be derived. A short written guide defines each criterion and gives worked examples of borderline entries.

Agreement between the two abstractors is then calculated for each criterion using the kappa statistic, which corrects for agreement expected by chance (Cohen, 1960) and is commonly interpreted against the descriptive bands suggested by Landis and Koch (1977). Criteria with poor agreement are redefined or dropped before the remaining cases are abstracted by a single reviewer.

Results are summarised by criterion, as the proportion of applicable cases in which it was met, and set beside the agreed standard. A single overall score per case is tempting and should be avoided, because it hides which actions are being missed and implies that all criteria carry equal weight. Intervals are summarised by their median and range. The proportion of items not recorded is reported for each criterion as a finding in its own right. The analysis is descriptive, and comparisons between shifts or staff groups should be made only where numbers are adequate and never in a form that identifies individuals.

Findings are presented to the staff who provide the care, in a meeting and in a one-page written summary, by someone they regard as a colleague. Following the evidence on effective feedback, the summary names the two or three criteria furthest from standard, proposes a target for each and records who will do what (Ivers et al., 2012). Typical responses include a haemorrhage box with drugs and a prompt card, a pre-printed record sheet with time fields, drills, or a change in how blood loss is measured.

The same checklist is applied to a new sample after the changes have had time to take effect. Criteria, definitions and sampling are kept constant so that the two rounds can be compared. If the work is written up, the SQUIRE guidelines for reporting quality improvement provide a suitable structure (Ogrinc et al., 2016).