This study, set in the adult medical wards and ICU of a tertiary referral hospital in Kigali, Rwanda, found high rates of antimicrobial resistance, particularly among Gram-negative organisms. Critically, however, 98% of the patients were already antibiotic-exposed at the time of culture sampling, introducing a potential selection bias towards isolation of resistant organisms and highlighting the limitations of routine laboratory data for antimicrobial resistance (AMR) surveillance in this setting. There were significant delays in culture taking, with a median duration of antibiotic exposure of 3 days prior to culture sampling. Unsurprisingly, discordant antibiotic therapy was significantly associated with in-hospital mortality.

The rates of resistance presented here are broadly consistent with those reported previously in Rwanda. A retrospective study from 2017 to 2018, which looked at 341 bacteraemias from the same referral hospital, found an identical rate of ceftriaxone resistance of 73% amongst the tested Gram-negative organisms (97 of 132 isolates) [6]. Worryingly, however, the rates of imipenem resistance were much higher in our cohort compared to this cohort from only 4 years earlier (15% vs. 1%). By contrast, MRSA rates were higher in the earlier cohort, with a prevalence of 33% (23 out of 69 isolates), and there were 3 cases of vancomycin-resistant Staphylococcus aureus (VRSA; 3.7%, 3 out of 78 tested isolates). This contrasts with no MRSA or VRSA seen in our study. Of note, however, the earlier cohort was hospital-wide, included only bloodstream infections and 25% of the cohort were patients under 14 years of age. Importantly, their data were only laboratory-based and so there was no corresponding clinical information about antibiotic exposure at the time of culture.

Another study from 2019 at the same institution enrolled 647 adult patients with suspected infection and financed microbiological cultures for them according to the presumed site of infection [7]. They identified 323 positive cultures, with a similar predominance of Gram-negative organisms (88% compared to 82% in our cohort). They also found a similar rate of ceftriaxone resistance amongst Gram-negatives (76% vs. 73% in our cohort), but lower rates of carbapenem resistance (4% compared to 15% in our cohort). They were able to perform limited testing for ESBL-production and found this to be the main mechanism of cephalosporin resistance (71% of the 92 tested Enterobacteriaceae). Similar to Habyarimana et al., they found rates of MRSA of 32%. Sixty-five percent of their patients had been exposed to antibiotics in the 30 days prior to enrolment but they did not collect data on the nature, timing or duration of that exposure, nor on how many patients received antibiotics after enrolment but prior to culture.

A retrospective study from a different referral hospital in Kigali examined resistance patterns amongst 5296 laboratory isolates from 2009 to 2013 [9]. They found lower rates of resistance compared to our study: 64% of E. coli and 75% of Klebsiella pneumoniae isolates were resistant to co-amoxiclav (compared 97.3% and 96.9% respectively in our cohort), and only 25% of E. coli were resistant to cefuroxime (compared to 81% resistance to ceftriaxone in our cohort). Similar to our study, they found low rates of MRSA and VRE (only 2.2% and 0.6% respectively). Again, however, as a laboratory-based study, they did not report on antibiotic exposure of patients prior to these cultures being taken.

There appears to be a clear temporal trend towards worsening AMR amongst Gram-negative organisms isolated in referral hospitals in Rwanda, with a worrying increase in carbapenem resistance in our cohort. Indeed, an older study from 2009 found only 38% ceftriaxone resistance amongst inpatient E. coli isolates from urine (compared to 81% in our cohort) [15]. Although testing capacity for resistance mechanisms is limited, it is likely that extended-spectrum beta-lactamase (ESBL) production accounts for a significant proportion of this worsening resistance. A study looking at carriage of ESBL-producing Enterobacteriaceae amongst patients and caregivers in a referral hospital in southern Rwanda identified high rates of ESBL-carriage (50% of patients and 37% of caregivers at admission, increasing to 65% and 47% at discharge) [8].

Nonetheless, our study highlights the significant challenge of using routine laboratory samples to infer the broader burden of AMR in our setting. The majority of existing literature in Rwanda, and indeed many low-income countries, provides little or no data on the clinical context of the cultures being analysed. By linking clinical and laboratory data, we found that 98% of our patients had received antibiotic therapy prior to the culture being taken, with 66% specifically exposed to ceftriaxone and a median prior exposure of 3 days. In this context, it is not surprising to find a high level of ESBL-producing organisms amongst culture results. Patients with susceptible infections are either not being cultured (due to rapid clinical improvement with empiric therapy) or their cultures are more likely to be negative due to prior effective antibiotic exposure (with the exception perhaps of those with a deep-focus of infection and inadequate source control). This selection bias means that resistant infections may be over-represented among positive laboratory cultures, particularly compared to high-income settings where microbiological culture is more routinely performed prior to antibiotic administration and at lower-level health facilities. This phenomenon is increasingly recognised as a limitation of passive, laboratory-based AMR surveillance [2, 16, 17]. In our institution, availability of blood culture bottles is inconsistent and doctors have to check with the laboratory before requesting. This discourages routine requesting of blood cultures unless a patient is deteriorating on first-line therapy.

Evidence-based antimicrobial stewardship programmes and national antimicrobial guidelines rely entirely on the quality of data used to inform them. The existing data from Rwanda would support the need for a carbapenem and/or amikacin for all unwell patients admitted to teaching hospitals in Rwanda with a possible Gram-negative infection. However, our findings on the delay to culture and significant selection bias from prior antibiotic exposure should give policymakers pause for thought. The risk of over-estimating AMR is significant, particularly in the context of the already increasing prevalence of carbapenem resistance seen in our cohort. There is a need for enhanced AMR surveillance, culturing antibiotic-naïve patients at all levels of healthcare (community, primary and secondary, not just in tertiary referral hospitals). This would not only provide data for better-informed national policies, but also support clinicians in narrowing the spectrum of antibiotic therapy in those with susceptible infections that are not being isolated in current practice.

Finally, the reported prevalence of MRSA in Rwanda varies considerably, ranging from 0 to 2% in our study and Carroll et al., to 32–33% in Habyarimana et al. and Sutherland et al., and even 82% in one study from 2013 [6, 7, 9, 10]. Whilst variation in patient cohort selection may account for some differences, it is possible that rates of MRSA may be miscalculated due to faulty laboratory materials or over-estimated where mixed staphylococcal cultures are not correctly identified (disc diffusion tests may be misread as resistant in mixed infections as coagulase-negative Staphylococci are usually methicillin-resistant). Further work is needed to delineate this issue, as most patients do not receive empiric MRSA coverage in Rwanda and the introduction of such practice would pose a significant risk in the absence of therapeutic drug monitoring for glycopeptides and could drive further resistance.

Our study has several limitations, most notably being single-centre, single-specialty and relying on a single non-automated laboratory for culture interpretation. Nonetheless, these real-world clinical culture results are all that are available to policymakers in Rwanda and the detailed clinical and prescribing data presented here highlight the challenges of relying on such data to infer the true prevalence of AMR.