Introduction
South Africa has the highest use of health services across the continuum of care in sub-Saharan Africa. In the 2016 Demographic and Health Survey, 94% of women attended antenatal care from a skilled provider, 96% delivered in a health facility and 84% attended postnatal care within 2 days following birth.1 Despite increasing uptake of public sector health services, maternal and child mortality rates remain well above the Millennium Development Goal targets2 raising important questions about the underlying content and quality of care received.
Disrespect and abuse of women during childbirth—a subset of violence against women—has emerged as a key indicator of the overall quality of care and a barrier to improving maternal and child health outcomes.3 In South Africa, the importance of provider–patient relationships first gained traction over two decades ago with the 1998 publication of a qualitative study led by Dr Rachel Jewkes and colleagues, exploring the question of “why do nurses abuse patients?”.4 Through interviews with patients and staff, a complex picture of clinical neglect, verbal and physical abuse emerged, which suggested that mistreatment of women during childbirth had become commonplace in South Africa.4 Subsequent studies reinforce these findings, attributing mistreatment, in part, to a lack of accountability and action on the part of managers.3 5 6
In August of 2014, the National Department of Health (NDoH) launched the helpdesk as a social accountability mechanism for improving governance, allowing recipients of public sector services to hold providers and the NDoH accountable for the content and quality of care provided. Any individual attending public health facilities in South Africa can SMS (short message service) the helpdesk with complaints, compliments or questions. Importantly, the helpdesk is tied to MomConnect, a maternal messaging platform designed to support pregnancy and motherhood, leading to improved outcomes for South African women and their children.7 Thus, in addition to its broader functionality, the helpdesk is used by MomConnect users to opt out of messages or communicate important updates, such as the birth of a baby.
All incoming messages are sent directly to the NDoH, where the helpdesk, staffed by nurses, is physically located. Helpdesk messages are labelled and assigned to one of four full-time personnel for handling. Based on their content, responses to questions use one of 114 custom responses derived from frequently asked questions. In the event that none of the custom responses are appropriate, the woman is given a customised response, which often includes referral to her health facility. Processes for responding to complaints follow a lengthier process, which typically includes a response from the helpdesk to provincial representatives who then follow up on a case-by-case basis with district and facility authorities.
Since its launch in August of 2014, the helpdesk has received nearly 250 000 messages.7 However, little is known about the use of the helpdesk for reporting instances of violence against women, including mistreatment during pregnancy or childbirth. Understanding user engagement, coupled with the timeliness and appropriateness of the helpdesk’s response, to reported instances of violence that require an urgent or escalated response is vital for ensuring that it is responsive to population needs. At present, the management of messages has largely been manual, necessitating a gradual increase in the central-level personnel required to manage responses. As the helpdesk continues to expand its user base and move into new programme areas, efforts are needed to optimise response times and content and also to accommodate increases in message volume without overburdening the support staff. Where the prior paper in this series sought to describe user engagement with the helpdesk (ref series paper 5 on helpdesk), this paper outlines early efforts to understand and potentially enhance helpdesk performance through automated message triage using the handling of messages on mistreatment of women as a case study.
We begin by characterising the helpdesk response to incoming messages, with a focus on potential areas of improvement. Then, we explore the feasibility of an automated triage system, one which would sort and prioritise incoming messages, as a possible mechanism of improvement. Message handling is assessed in terms of response quality and timeliness among (1) a random sample of messages and (2) messages relating to mistreatment of women, as identified via expert-defined keywords. While all messages would ideally receive a prompt and appropriate response, the mistreatment of women is a high-importance topic warranting prioritisation should a triage system be implemented. After determining whether mistreatment can be identified by keyword, we further explore automated triage by training a naïve Bayes classifier to assign the message labels already used by NDoH staff. We hypothesise that (1) mistreatment of women can indeed be identified with keywords, (2) handling of messages reporting mistreatment is no better than the handling of other messages and (3) naïve Bayes will replicate less-common labels with high specificity, allowing them to be selectively identified to avoid overburdening NDoH staff. If confirmed, (1) and (2) would establish the need for triage, and (1) and (3) would demonstrate its feasibility.