A cross-sectional, descriptive, qualitative study was conducted in two public health facilities in Kilifi county in coastal Kenya. The two health facilities are the rural Rabai Health Centre (primary care facility) and the larger, urban Mariakani Sub-county Hospital (secondary care facility). Both facilities would later participate in the PRECISE (Pregnancy Care Integrating Translational Science, Everywhere) pregnancy cohort study [20]. At the time of data collection, enrollment to the PRECISE cohort had not yet started.

ANC, routine delivery care, and emergency care for pregnancy complications are provided primarily by nurses/nurse-midwives with support by clinical officers (nonphysician clinicians) at both facilities. At Mariakani Hospital, doctors and an obstetrician/gynecologist provide specialist services for high-risk pregnancies. Ultrasound services are not available routinely, but can be undertaken at Mariakani Hospital (and other private facilities) upon referral for pregnancy complications or uncertainties regarding GA.

We sought to enroll two main groups of participants: (1) HCWs directly involved in the provision of services for pregnant women, as well as managers and health administrators; and (2) community members, represented by pregnant women participating in ANC and their family members (partners, as well as the pregnant woman’s parents and parents-in-law).

HCWs were purposely sampled to cover providers at the ANC clinic, maternity, outpatient, and radiology (including ultrasound) departments. Health administrators were also purposely sampled to ensure inclusion of facility and subcounty managers overseeing reproductive health services. Pregnant women were approached by research assistants when they presented for routine ANC, and participating women could invite their partners or parents.

Data were collected between March and May 2019 by two Kenyan researchers: a social scientist (PMM) and a maternal health researcher and obstetrician (AK), who were assisted by two trained local research assistants who took notes during the sessions. Researchers were familiar with the local setting and the Kenyan health care system, and were fluent in both English and Swahili. The research assistants were also fluent in Mijikenda. None of the data collectors were involved in the participants’ clinical care; however, some HCWs had previous interactions with the two researchers as part of PRECISE study preparations.

In-depth interviews (IDIs) with HCWs and focus group discussions (FGDs) with pregnant women and their families were conducted in person in private areas of the health facility, away from the clinical areas. We developed a semistructured interview guide, which was piloted on two HCWs at Rabai Health Centre and revised prior to the subsequent interviews. The topic guide began with simple assessment of prior exposure to computers, smartphones, and obstetric ultrasound, followed by a discussion on existing methods of assessing GA and the potential value to pregnant women and HCWs.

The study was started before clinical implementation of TraCer. A video demonstrating its use during routine ANC was shown to participants. The 5-minute video, recorded at one of the facilities in Swahili, showed a nurse using TraCer with a pregnant woman who was unsure of her LMP. In the video, the nurse shows the mother the image of the fetal heartbeat and the head of the baby on the tablet screen, reports the GA, and then gives a date for the next clinic visit.

Participants were encouraged to voice their thoughts and ask any questions during and immediately after watching the video. Participants were asked what they liked or disliked about TraCer as seen in the video, whether TraCer could be introduced to their health facility, how confident they would be in its findings, any outcomes (positive or negative) they expected with its introduction, the type of provider they thought could use TraCer, and whether they would recommend it to other pregnant women and health facilities.

IDIs were conducted in the participants’ language of choice. HCWs preferred English, whereas pregnant women and their families preferred Swahili. All sessions were audio-recorded with permission and field notes were taken during each session. After each IDI and FGD, the research team debriefed to update field notes, discussed revisions and additional probes to the topic guide, and assessed data saturation.

All recordings were transcribed verbatim in the language of the interview and translated to English (where applicable) by research assistants. A sample of transcripts was compared with the recordings to ensure accuracy. NVivo 12 (QSR International, Melbourne, Australia) was used to manage the transcripts, and to facilitate coding and collaborative data analysis. The data analysis team comprised three Kenyan researchers familiar with the study site and local languages (AK, PMM, GMM), including two who had participated in the data collection (AK, PMM) and two experienced Canadian social scientists (MWK, MV). The data analysis team first familiarized themselves with the transcripts. Employing a directed content analysis approach [21], the coding framework was developed deductively from the research question based on pre-existing definitions of acceptability and feasibility [22] that were modified to fit our study. Acceptability was assessed according to the perceptions of the appropriateness of TraCer to participant needs, preferences, and sociocultural norms, along with factors that would influence willingness to use the device. Feasibility was assessed according to perceptions on whether TraCer could be implemented in the study health facilities and factors required for its successful implementation.

Major themes and subthemes were explored related to acceptability and feasibility. The coding framework (see Multimedia Appendix 1) was tested on three transcripts to refine and ensure agreement between coders. Transcripts were then divided between coders for analysis. Emergent common and divergent patterns of responses between participants were explored through discussion within the team. Factors that were considered included differences in the site characteristics (urban vs rural, level of facility, access to ultrasound), HCW characteristics (skill level/cadre and prior experience of ultrasound), and community member characteristics (age, gender).

The study obtained ethical approval from Aga Khan University Institutional Ethical Research Committee (2018_REC_47), King’s College London (Ref HR-17/18-7855), and University of British Columbia (H18-02828). All participants provided individual written informed consent prior to research activities. Confidentiality and safe storage of the data were ensured through deidentification of transcripts and electronic storage in password-protected devices accessible only to members of the research team.

In total, we conducted 18 IDIs and 6 FGDs involving 52 community members; the IDIs lasted 27-64 minutes and the FGDs lasted 34-77 minutes. In the IDIs, nurses represented the largest group interviewed, and other cadres were clinical officers, doctors, sonographers, and public health officials (Table 1). Seven of these HCWs also had administrative responsibilities. The 52 members engaged in the FGDs included 31 in three FGDs at Rabai Health Centre and 21 in three FGDs at Mariakani Hospital. Overall, FGDs engaged 19 pregnant women, 15 partners (all male), and 18 parents (all mothers or mothers-in-law). Fewer than 20% (9/52) of pregnant women or their families had prior experience of ultrasound (Table 2).

In this study, we assessed a low-cost, portable, and artificial intelligence (AI)-enabled ultrasound device in two Kenyan settings. Overall, the device concept was found to be highly acceptable to women, their families, and HCWs, and implementation at health care facilities was felt to be feasible. The introduction of the device was predicted to contribute to reduced anxiety around the stage of pregnancy and fetal well-being, increase ANC clinic attendance, increase confidence by women in HCWs, and potential time- and cost-savings. It was also predicted to boost HCWs’ professional self-image. The preferred users of TraCer were ANC nurses. There was a clear message that effective implementation requires adequate consideration on training, sustainability, and consistency of service. In addition, addressing potential misconceptions is a clear outcome of this work.

Various features of TraCer contributed to its acceptance. In particular, the value of seeing an image of the baby and the ability to see fetal heart activity were appreciated and strongly linked to trust of the device’s findings/results. The added value of “seeing” has been highlighted in other studies [23]. Device features such as the small size, portability, and absence of connecting cords were cited as strengths.

Previous studies conducted in sub-Saharan Africa highlight the positive attitude toward new technological innovations in health [24-26]. Despite little previous experience, HCWs showed high levels of acceptance toward new technologies owing to expectations that new tools would be easy to use, make work easier, and take less time [25,26]. This is also true for studies evaluating ultrasound devices [23,27,28]. Our study is unique in that it assessed a tool using AI to automate analysis, and that the tool is intended for a narrow use-case of GA assessment. Despite these features, the positive attitude toward the tool and the willingness to have it introduced was shared. The use of the specific device was viewed favorably and equated with quality, with HCWs expressing eagerness to learn its use. In our study, HCWs expressed confidence in automation and felt that the results would be more reliable due to reduction of human error and absence of interference from providers. This is an important finding relevant to future technologies incorporating AI in medical devices in these settings.

Some hesitancy, due to novelty, was also expressed, indicating the need for robust testing of performance and the effect of implementation over time. This is an important finding for implementation, demonstrating that immediate acceptance of new tools upon introduction cannot be expected. It is likely that the experience during early phases of implementation may influence eventual acceptability of novel tools.

There is increasing use of low-cost portable ultrasound in low-resource settings [29]; however, the majority of studies have been conducted in large tertiary hospitals with only a few conducted in rural facilities. Our study demonstrated higher acceptance of TraCer in the rural site. It is possible that this is due to the lack of any ultrasound service in that setting (Rabai) and the challenges associated with referral to the urban site. This unequal distribution of ultrasound services is common in sub-Saharan Africa [13], and sites for future implementation of TraCer and similar technologies should be carefully selected with this in mind. In contrast, the higher hesitancy in the urban site (Mariakani) could be a reflection of the higher existing technical skills and better understanding of potential pitfalls. Thus, HCWs with higher skill levels and prior experience with pregnancy ultrasound had more questions regarding accuracy and recommended more intensive training. They should therefore be provided with robust technical information and given ample time for training and testing.

This study suggests that implementation of TraCer in this setting will be feasible, but several potential barriers to effective implementation should be considered. Previous studies have demonstrated the feasibility of the use of small portable (compact) ultrasound devices, particularly for obstetric applications [30]. Similar to other studies [15], HCWs highlighted the need for equipment maintenance to ensure sustainability and consistency of the service. Service interruptions may be disappointing to staff and pregnant women and may affect their attitudes toward the tool [25].

HCWs were confident that they would be able to perform the procedure if given adequate training, but their views on the desired nature and duration of the training were varied. Other studies have evaluated combinations of didactics courses, hands-on instruction, supervised scanning, and lectures [31,32]. The duration of training for obstetric point-of-care studies in one review [31] ranged from 3 days to 3 months. Findings from our study are not sufficient to inform the design of a training package for TraCer, which will be evaluated during implementation. Addition of a training module on basic troubleshooting and repair of the device may be beneficial.

Like in other settings, misconceptions around the use of the technological innovations was common [28], with frequent reports of overestimation of its diagnostic potential [28,33]. It is possible that part of the positive attitudes toward TraCer reported in this study may have been influenced by these misconceptions and overestimation of the tool’s scope and function. Providing false reassurance of fetal and maternal well-being could contribute to delays in care-seeking for women, or may lead to blame in case of an adverse pregnancy outcome. Utmost care should be exercised to avoid HCWs becoming less thorough during clinical assessments or changing referral thresholds for further care due to such false reassurance of fetal well-being [27,33]. Thus, the role and scope of TraCer should be emphasized for both pregnant women and HCWs. Referral pathways for formal ultrasound assessment should remain open and indications for this made clear.

Nevertheless, the presence of a visible fetal heart beat is an important measure of fetal well-being. HCWs perceived that using TraCer would make assessment of the fetal heartbeat faster and easier to operate than a Pinard fetoscope or Doppler fetal monitor. We suggest that messaging around fetal status should acknowledge the confirmation that the baby is alive at the time of the examination, but clarify that this does not give reassurance of well-being.

In view of the above findings, we propose that a fully validated and AI-driven TraCer device be used by nurses at the ANC clinic for assessment of all pregnant women. In this setting, TraCer would assess fetal viability and provide an accurate estimate of GA. Where there are indications for full obstetric ultrasound or other concerns for fetal well-being, women should be referred and encouraged to proceed for formal ultrasound assessment in line with current practice. With this approach, more women would have an accurate GA to inform subsequent decision-making and scheduling of effective ANC. Automated GA assessment using TraCer could reduce referrals for GA assessment only, giving higher-skilled ultrasound providers more time to focus on women needing specialist assessments. Such an approach has been suggested by other studies [13,34,35].

Our study has a number of strengths. Data were collected from various participant groups and evaluated in the setting intended for use (ie, low-resource settings in LMICs). To increase the generalizability of findings, data were obtained from two health care facilities with differences in location (periurban vs rural), level (primary vs secondary), and availability of ultrasound services. The research team comprised skilled qualitative researchers and experts in maternal health and pregnancy ultrasound. The data collection and analysis team included researchers familiar with the local language and context.

The use of video role-play has been successful in eliciting rich data on acceptability and feasibility of the tool prior to clinical implementation. This approach has allowed an early evaluation, ensuring that the initial views of users are incorporated into further tool development and planning for actual clinical implementation. The study has also helped to understand baseline user expectations that can be used to measure clinical implementation indicators.

Although useful, video role-play also has limitations, as it may create an inaccurate perception of proposed clinical implementation. In this case, study participants felt that use of the device took very little time, an impression that could be created by an edited video. To overcome this potential limitation, further assessments during active clinical implementation of the tool are important.

Some of the research team had had interactions with study participants (HCWs) prior to data collection as part of preparations for a larger research study. This is important as it could have introduced some social desirability bias, with these responders expressing a more favorable view toward TraCer. This was mitigated by selecting the interviewer who had had the least prior interaction with the participant.

We have shown that there is potential to implement AI-enabled ultrasound innovations in low-income settings, including a device that offers only a selected fetal assessment (in this case, GA). It is highly likely that the TraCer tool can be implemented in this and similar settings, and that users will find it valuable. Further device developments should ensure that the tool is simple and easy to use, and that results are obtained within a short time frame. Prior to clinical implementation, robust accuracy data must be provided and measures should be taken to ensure sustainability and consistency of the service. Clarity of messaging about the tool and its role in pregnancy is essential to prevent misconceptions and misuse. Our study points to further assessments required during the subsequent phases of clinical implementation, feasibility, acceptability, and device usability, as well as clinical validity and safety of the tool.