Demographics

In total, 70 individuals were invited to participate. The response rate among midwives was 52.9% (18/34) and 63.9% among other individuals (23/36). Forty individuals completed the REDCap survey, while 41 participated in IDIs or FGDs, indicating that one person did not complete the survey.

Among the survey respondents, most were healthcare providers, with 42.5% midwives/nurses and 20% physicians (Table 2). Over 70% of participants had more than 10 years of experience in the obstetrics field. While approximately two-thirds were LMIC residents—specifically from Kenya, Uganda, Nigeria, Burkina Faso and Zambia—those that did not currently live in an LMIC have worked extensively in these contexts. Three-quarters were familiar with POCUS (i.e., had first-hand experience or were conceptually familiar), and 55% had prior knowledge of efforts in the AI-enabled ultrasound space before participating in this study.

The following sections present quantitative and qualitative results related to perceptions of standard and AI-enabled POCUS (Domains 2 and 3, respectively – Table 1) to compare and contrast opinions of the existing and emerging technology.

Priority AI capabilities

Respondents were asked to rate via a Likert scale the importance of select maternal and fetal assessments that would be most helpful for AI-enabled POCUS to automatically screen for in a basic emergency obstetric and neonatal care (BEmONC) facility (Table 3). Fetal heart rate/viability, multiple gestation and placental location were the three most highly ranked conditions—the latter two were included in the described prototype which was introduced after completion of the survey.

Qualitatively, the prototype capabilities presented to respondents was seen as acceptable:

“That’s some of the biggest risk which we identified: lie or the presentation is one of the risks, and number of fetuses is another risk, and the placenta location is another big risk, which we identified among several others.” – Researcher 2, IDI, LMIC

While only 70% of respondents felt gestational age – another prototype feature and part of the WHO recommendation—was very important, an additional 22.5% of respondents felt it was somewhat important. Gestational age was flagged during conversation as an important component:

“I think this artificial intelligence will help us more because, first of all, it has the gestational age. Most of our mothers, a big percentage, are not very sure of their dates, so it’s going to really help us.” – Midwife 10, FGD, LMIC

However, getting women into ANC earlier was a key consideration:

“Our colleagues think this is really critical to have gestational age […] but we have to get women in earlier in order to get the most accurate dating.” – Funder/Nurse-midwife 1, FGD, HIC

Several respondents flagged detection of congenital anomalies as both important and concerning, particularly in LMIC contexts. While ethical concerns were raised about screening for conditions that healthcare providers and systems may not be prepared to manage, others argued for the importance of screening for congenital anomalies to allow clients the option of termination if desired.

“I think that [detecting congenital anomalies with AI] is also a slippery slope […] I think it would be really hard if providers had access to that information – that’s a whole different type of counseling.” – Policy-maker/Physician 1, IDI, HIC

“[With POCUS], the mother gets benefited by knowing that she’s going to deliver a live baby… At the same time, also knowing any deformity or congenital abnormality that the baby is having, so that she decides before the time for delivery.” – Midwife 2, FGD, LMIC

This sentiment was countered with several respondents emphasizing that both standard and AI-enabled POCUS, particularly in the hands of midwives, should be introduced as a screening tool rather than a diagnostic tool.

“My thinking as an obstetrician, […] if there’s any doubt, then that will be an indication to send this patient [to a] radiologist to rule out the abnormality that you are worried about. So, I think AI will be vital for screening, but we shall need a detailed scan just for confirmatory tests to make sure we reduce on the issue of over treatment.” – OBGYN/Researcher 2, FGD, LMIC

Potential impact on ANC quality, services and clinical outcomes

ANC utilization and experience of care

Survey respondents were asked about their perceptions regarding how standard POCUS and AI-enabled POCUS might impact ANC utilization and experience of care. When comparing their survey responses, there was overall strong agreement that both technologies could improve ANC attendance (75% standard and 65% AI), though there was less agreement with AI-enabled POCUS. Fewer respondents felt that AI-enabled POCUS would increase trust between providers and women compared to standard POCUS (60% and 82.5%, respectively). Approximately half of respondents largely agreed that neither technology would lead to clients deciding to forego other ANC necessities in order to pay for a scan. Across these questions, there was slightly more uncertainty with AI-enabled POCUS (Fig. 1).

Findings from IDI/FGDs provided some nuance to these survey data. For example, several respondents believed that AI-enabled POCUS could be seen as less engaging to clients particularly if a clear image of the fetus was not readily available. This could impact ANC utilization, demand and provider–client trust as compared to standard POCUS:

“Ultimately, a mother wants to see as much as possible of what their baby looks like, and so restricting the images that the nurse gets out of these blind sweeps, I’d say to some extent, [is] a bit limiting to what the mother might want to be seeing and what the nurse might want to be sharing with the mother.” – Implementer 2, IDI, LMIC

“And they will be asking us questions like, ‘So that is the only thing that you’ve done and you’re telling me everything is okay? I’ve not seen you looking for the head.’ […] But for this new thing, we are not going to be showing them and it’s just going to be displaying itself there. I think they’ll be asking us a lot of questions…” – Midwife 3, FGD, LMIC

Respondents noted that the time spent with a client during a standard POCUS scan was critical for human connection and that AI-enabled POCUS may compromise this.

“For me, I think it’ll [AI-enabled POCUS] reduce physical contact with the patient – doctor-patient relationship may be minimal. […] if it is minimal, sometimes you are not able to explain well, the condition of the patient to the client. And they may feel they are not getting enough information also.” – Midwife 8, FGD, LMIC

Quality and content of ANC services

The majority (85%) of respondents agreed that standard POCUS can strengthen health care providers’ ability and confidence in making appropriate clinical decisions; however, 70% felt AI-enabled POCUS would build confidence (Fig. 2). Thus, there was more uncertainty with AI-enabled POCUS in reinforcing confidence compared to standard POCUS, a sentiment that was reinforced by qualitative findings.

Perceptions of standard and AI-enabled POCUS on client services (n = 40)

Some providers tended to express enthusiasm for AI, believing it could enhance accuracy and usability:

“Those simple sweeps [that are] reproducible and reduce chance of making error is a very good way to go. If those sweeps done correctly can give you, at least those very basic outcomes that you want to know: the presentation, the placental position, and so forth, it’ll be useful at least to make that quick decision at a point of care.” – OBGYN/Researcher 3, FGD, LMIC

“Yeah, that one will be just a very good help because it is not going to disturb us – like previously, we used to look for the presentation and you move with the probe all over the mother’s abdomen looking for one thing, but now this time around we are just going to use it three times and the things will just display. It’s just going to be very easy for us.” – Midwife 3, FGD, LMIC

On the other hand, some respondents expressed concern that clinical acumen may diminish with increased reliance on AI and negatively affect clinical decision-making and confidence.

“Good midwifery skills are all about hands on bellies, and all about interaction with the woman and all about having a dialogue. […] I just worry a little bit that if you then put this great machine in the middle of all that, then the emphasis is on the machine. And that if the machine’s broken, how will the health worker have maintained their skills and be able to do just hands on belly to work out where the baby is and potentially what size it is. […] But there are sets of skills that midwives and OBs develop over time which are critically important to maintain.” – Funder, IDI, HIC

There were more mixed perceptions around whether it would decrease time available for other services and how it would impact provider workload with 30–40% either strongly agreeing or strongly disagreeing with the statement (Fig. 2). A few respondents raised the issue that introduction of POCUS could replace time spent completing another critical ANC components:

“What are you not gonna do? I feel like in this whole prioritization conversation, nobody says, ‘well, if you’re prioritizing something new, you have to deprioritize something’, and we never, ever acknowledge that. So, it means those things get dropped randomly, which means that if the midwife is now prioritizing her time on AI ultrasounds, that she can do, what is she not doing? And if we don’t provide guidance on that, she’s gonna drop whatever the hardest thing to do is, and that might be the one thing that would save more lives […].” – Funder/Nurse-midwife 2, FGD, HIC

While some said poorly remunerated nurses and heavy workloads could de-motivate scanning, others mentioned that AI-enabled ultrasound could mitigate this by decreasing the amount of scanning time needed per client. Several midwife respondents acknowledged the workload, but this sentiment was outweighed by their excitement for the new technology and its potential to allow for more equitable access to ultrasound.

“Of course the workload will there, but I will [prefer it] this way – I am very okay to do it.” – Midwife 3, FGD, LMIC

“You go to the next client, you may conduct the ultrasound to as many clients as possible. Unlike the [standard] POCUS, we sometimes do five to eight [scans]. But for this one [AI-enabled POCUS], you may conduct the ultrasound to almost every client because the results are very fast. I think it’s a game changer also to me.” – Midwife 9, FGD, LMIC

Referrals and clinical outcomes

The majority of survey respondents felt both technologies could improve appropriate referrals (85% and 72.5% for standard and AI, respectively) and neonatal outcomes (85% standard and 72.5% AI) – and to a lesser degree, maternal outcomes (Fig. 3).

Perceptions of standard and AI-enabled POCUS on referrals and clinical outcomes (n = 40)

Qualitatively, respondents emphasized that without strengthened and functional referral systems, it is unlikely that either standard or AI-enabled POCUS will have any impact on clinical outcomes.

“We have to take like a giant step back. […] There needs to be forward thinking of the referral pathway. […] Because, what I’ve seen is that introduction of point of care ultrasound by itself, even in my experience, is never the thing that works by itself. […] I’m one of the biggest proponents of POCUS, but even more so, I’m one of the biggest proponents of a reminder that POCUS is just a tool in the larger toolbox of clinical tools that we have, including your physical exam and your clinical gestalt of a patient’s disease process.” – Researcher/Physician 3, IDI, HIC

“When you have added the ultrasound, what do the women do? Where do they go? So I think that we need to really consider how this is embedded within the system and the modifications that need to happen within the clinical pathway for it to be feasibly and sustainably integrated.” – Funder/Policy-maker, IDI, LMIC

“The purpose of the ultrasound is to identify potential problems and then deal with them appropriately. That’s all part of the intervention. So going back to referrals – I think that has to be front and center. You identify and then you have to know what to do with what you have identified. That’s not a small part of it either.” – Radiologist/Physician, IDI, HIC

Some respondents expressed concern that referrals would increase overall, potentially straining systems and depleting family resources.

“If someone gets an ultrasound and they misinterpret something and it leads them to refer a patient that they otherwise wouldn’t have referred […], the process of referring someone is a pretty resource intensive one. When you’re coming from a little clinic, and now all of a sudden you have to put someone in an ambulance or you have to tell them you have to mobilize your own resources to go to a hospital that’s like, you know, 70 kilometers away to be seen for this thing; they have to go mobilize their own money, a lot of people don’t have insurance, so you have to get your own transportation, you have to go pay for the extra test, pay for the extra consultation, worry a lot [about] what could be happening to me, only for them to go and find, ‘Oh no, you’re actually okay’. I think that’s a potential source of harm.” – Researcher/Physician 2, IDI, LMIC

In addition to increasing inappropriate referrals, another potential unintended consequence related to health outcomes was inappropriate clinical management. Several respondents voiced concerns about liability risks involved with missed diagnoses or mismanagement due to algorithmic limitations or inaccuracy:

“If for example, it makes a wrong impression, and I go in and intervene and I’m in error, do I blame the clinician or you blame the machine? [With] the other one [standard POCUS], the person was the one interpreting the picture and saying, ‘I think from my training it is this’, [but] now the [AI] machine has given you that this baby is distressed [then when you go] and deliver, you deliver a preterm baby. That could give you challenges: ‘the machine told me that it was distressed, but it was not actually’.” – OBGYN/Researcher 4, FGD, LMIC

Considerations for implementation and health system integration

Target user and potential for task sharing

There was general agreement among respondents that midwives should be the primary target-user of AI-enabled POCUS. In the survey (respondents could select up to 3 provider cadres), midwives and nurses were selected most frequently (97.2%), followed by other doctors (55.6%), and OBGYNs (52.8%).

While the central role of midwives was echoed qualitatively, several others also emphasized the need to ensure doctors/OBGYNs were equally trained.

“I think that’s a great idea because, both in private and in public, midwives are the ones who spend a substantial amount of time with the patients. […] So, strengthening their capacity and empowering them to do point of care ultrasound, I think that would be very welcome to improve outcomes.” – OBGYN/Researcher 3, FGD, LMIC

“If for example, the midwife is able to pick it [AI-enabled POCUS], [then], for me to embrace it, to understand what she has referred to me a patient with certain condition, then I need to also get used to it, such that I’m able to be in the same boat with her. […Otherwise, I’m] subjecting the patient to another scan.” – OBGYN/Researcher 4, FGD, LMIC

However, some respondents specifically called out AI integration as being more likely to result in professional displacement.

“So, if everybody and anybody can scan [because of access to AI], then the potential is that people are going to be displaced. And in LMICs, people are looking for jobs. So we don’t want professional conflicts arising out of this.” – Funder/Policy-maker, IDI, LMIC

Relevant to both standard and AI-enabled POCUS, many respondents agreed that provider protections are critical as task sharing of new technologies emerge:

“I think in terms of training, regulation is not very clear. […] So I think there is a need for a policy review to ensure that if it is going to be task shifting, then those people are allowed, but also protected of litigation, it’s becoming quite common. […] So I think definitely regulation has to be important; I know sometimes technology comes and these things are rushed because there is a need, but I think looking into that aspect is also important to make sure that there’s a policy that is generally agreeable and it define clearly: what one can do and what one can’t do. In fact, with the increase in technology, this is becoming quite an important space.” – Researcher/Physician 4, IDI, LMIC

Introduction outside of the health facility and in the community was met with more heterogenous response. Approximately one-quarter 27.8% of survey respondents felt that community health workers (CHWs) should be trained to use AI-enabled POCUS. Many recognized that CHWs are often the first touchpoint with women, while others voiced concerns related to CHWs’ limited obstetric clinical training and competency to communicate obstetric findings.

“We are ignoring the fact that women come into contact with the health system in the community. So, we keep trying to incentivize women […] to come to the facility early. And it really doesn’t work. […] On the other hand, these community health promoters are seeing those women […]. Let’s be very open minded about that first contact, and don’t force women to change their behavior. But just leverage what’s already there.” – Funder/Nurse-midwife 2, FGD, HIC

Others believed community-based scanning might de-motivate women from seeking care.

“The community health promoters are service providers, but they are not well trained in [healthcare]. Their work is to bring us clients as early as possible, like those who are pregnant they should start clinic early. […] But it [AI-enabled POCUS] is a technical thing which can be used by a provider who can explain more to the client what is happening to the baby.” – Midwife 9, FGD, LMIC

AI-POCUS training

In terms of training, many believed that AI-enabled POCUS could reduce training requirements compared to standard POCUS training given the ease of blind sweeps. Approximately two-thirds of respondents (67.5%) believed that standard POCUS required intensive training compared to 42.5% for AI-enabled POCUS (Fig. 4).

“I think that training may not be very complicated, or maybe may not take a lot of time, rather, because the way I see it, it may be easier than [standard] POCUS.” – Midwife 9, FGD, LMIC

Perceptions of health system integration of standard and AI-enabled POCUS (n = 40)

However, several respondents emphasized that concomitant training in POCUS and image interpretation remained critical.

“But, what knowledge does this type of technology add to me as a provider, because it is just like giving me everything; I’m not even thinking. That may be my worry – a situation where we don’t have so much providers who have been trained on ultrasound. Maybe what I would suggest even at the deployment: training on the analog then to the AI, so that at least somebody knows how to interpret images so that even when now I go to this magic gadget that will interpret everything, at least I have something that I’ve learned from that session.” – OBGYN/Researcher 1, FGD, LMIC

“And maybe [with] the AI, it may be easier for them [midwives] to do it maybe in a faster time to be able to get the findings. Except that I believe that training is very important because having the background of POCUS would also be very helpful even as they use the AI […] I think it’s also good to understand, how will the midwife confirm that this machine has told me the right thing?” – Nurse-educator/Researcher, FGD, LMIC

At least one participant saw AI-enabled POCUS as an opportunity to enhancing existing training programs and focus on strengthening clinical-decision making:

“Because of the fact that the training itself of the use of ultrasound might be a little bit more facilitated by AI, might leave a little bit more time to focus on the ‘what do you do next’ part, which often, to be very honest, isn’t always focused on in point of care obstetric ultrasound classes because people are very, very focused on learning the ultrasound and not learning the clinical algorithm, which is ironic because you need to know the second part of that when you’re learning the first part.” – Researcher/Physician 3, IDI, HIC

Resources, data systems and facility infrastructure

Similar to standard POCUS, issues around device maintenance, availability of consumables, machine cost, misuse and overuse (e.g., fetal sex determination, over-charging clients for unnecessary scans, undue anxiety among clients) were mentioned in IDIs/FGDs. In the survey data, there was little to no difference between how either technology might increase ultrasound misuse or overuse with approximately 40% strongly disagreeing with this statement (Fig. 4); however, respondents offered some ideas for how AI might change this, such as disabling identification of fetal sex through “digital diapers.”

“In terms of the guarding against the misuse for fetal sex detection, and all the consequences of that, where such a limited visualization is desirable, but, on the other hand, the blind sweep paradigm is very restricted in terms of what it can detect.” – Implementer 2, IDI, LMIC

In the survey data, there was a slight difference in perceptions of how the technology might affect health system documentation with 70% of respondents believing that AI-enabled POCUS would streamline documentation compared to 62.5% for standard POCUS) (Fig. 4). However, many respondents raised larger issues around data privacy and storage, including conflicting interests between industry and governments:

“[Data storage] is going to become really, really important in the implementation because countries will be like, ‘I don’t want my data to go to these U.S. companies’ cloud’. And the more the more you know about it, even with regulations, what you think is anonymous, it can be so easily de-anonymous.” – Policy-maker/Physician 2, IDI, HIC

“How well can we localize some of this data […while still feeding] into the larger technology companies? Because obviously, I know my Minister of Health won’t have the capacity to develop some of these technologies – it needs either big corporations or big funders. But now it goes to the ethics part of AI in terms of how well the data is kept, how well all these policies -all these laws- that have been put in place, are being complied to.” – Implementer 1, IDI, LMIC

In terms of overall health system integration, many respondents emphasized that ultrasound was just one part of the continuum of care, underscoring the need to strengthen systems more broadly.

“…it can’t be called a game changer, per se, without other aspects like, strengthening the system, improving referral, and improving skills, improving availability of commodities, improving emergency obstetric care, but surprisingly also improving the welfare of women.” – Researcher/Physician 4, IDI, LMIC

“A lot of the work we’ve done […], is to better understand why some of the simpler interventions that are in our guidelines are still not being implemented. For example, why don’t we have a blood pressure cuff at every antenatal care clinic that functions? Why don’t we have a weighing scale? Why don’t we have calcium supplements? So, I think I’m a little bit concerned that if we jumped right to these technological solutions [AI-enabled POCUS], which I know have the potential for huge change, we also might lose on some of the really important interventions that we know are important. It might not be as attractive for providers or for women to pay for.” – Policy-maker/Physician 1, IDI, HIC

Evidence needed and research priorities

Respondents were asked to prioritize up to five research outcomes for AI-enabled POCUS (Fig. 5) from a list of 12 options. The top five outcomes included accuracy of diagnoses (77.5%), ANC quality (65%), early ANC attendance (50%), impact on referral (37.5%), and women’s experience of care (37.5%).

Priority research topics and outcomes related to AI-enabled POCUS (n = 40)

The importance of accuracy was conveyed by several respondents who cited potential inaccuracies stemming from a lack of diversity in training datasets. To mitigate bias, they emphasized the necessity of using representative and diverse datasets for algorithm training.

“My only question would be in terms of training these algorithms […] whether they’ll be consistent across the different populations […] so that at least whenever you get an outcome, it is an outcome that can actually help you make a decision, not an outcome that gets you into an error.”– OBGYN/Researcher 3, FGD, LMIC

“In India it’s generally known that a baby is small for gestational age, but may not be necessarily a preterm, there are more small babies. And then in Africa, there are bigger babies. So this algorithm has been tested on who? Does it work? Where does it work?” – Researcher/Physician 4, IDI, LMIC

Priorities related to ANC reflect the need to assess impact on ANC quality, early ANC attendance, and women’s experience of care during the pregnancy journey.

“I would want to, to answer the question: does it improve referrals that we’re getting over the analog? […] So on the part of the provider, do they think it’s better because it’s taking less time? […] We have challenges with human resources in our facilities. […] Does it improve the time they are taking to provide the service, but at the same time, not compromising quality of the work?” – Researcher 3, IDI, LMIC

“And how do we really assess what is good counseling and what does that that mean? So, I mean, if there’s an opportunity with all this excitement about AI ultrasound to really do some research around how do we ensure effective counseling around communicating what it means, and then how does the woman then interpret that and use that also for her own discussions with her, her family. […] I think there’s a lot of great questions that could be, that could be looked at more on some of the operational issues.” – Policy-maker/Physician 1, IDI, HIC

Measures related to maternal mortality and morbidity, stillbirth, neonatal mortality and morbidity were less prioritized (selected by 15–32% of the survey sample) than more proximal outcomes.

“I think we don’t need studies that look at, ‘Does it reduce maternal mortality? Does it reduce newborn mortality? […] We should look at earlier outcomes: ‘Is it diagnosing things correctly? Is it leading to correct referrals?’ I think that’s the kind of studies we should be designing. And of course, ‘What does it mean to implement it within a health system?’” – Policy-maker/Physician 2, IDI, HIC

“My point would be to think through what are these main causes of adverse maternal outcomes and birth outcomes. In our setting, we know it’s PPH, preeclampsia, sepsis, and so forth. Then see how we can integrate the tool towards mitigation of those big five. So, if we have that integrated within the tools, then we might see some level of reduction in the adverse outcomes. If not, then we’ll just still have the status quo.” – OBGYN/Researcher 3, FGD, LMIC