Enumerating the economic cost of antimicrobial resistance per antibiotic consumed to inform the evaluation of interventions affecting their use

Human antimicrobial consumption, whether or not clinically warranted, is associated with propagation of antimicrobial resistance (AMR) [1, 2]. This and other key drivers of AMR are listed in Fig. 1, notably widespread antibiotic use prophylactically and as growth promoters in agriculture [3].

Treatment of resistant infections is associated with higher costs for second line drugs, additional investigations, and longer hospitalisation [4]. Other indirect costs associated with AMR include productivity losses due to excess morbidity and premature mortality. These costs can be conceptualised as a negative externality to antimicrobial consumption accrued by all members of society, which are not reflected in the market price of antimicrobials [5, 6].

In addition to curative use in infectious diseases, antimicrobials are widely used presumptively, in mass treatment programmes (anti-helminths, antimalarials), and as prophylactics in surgical procedures and alongside immunocompromising treatments [2, 7]. Many other healthcare interventions such as vaccinations, diagnostics, and treatments for infectious diseases affect antimicrobial consumption, and consequently increase or decrease the risks of AMR. Economic evaluations of such interventions, however, have failed to internalise the potential costs of AMR into the analyses, leaving policymakers to intuitively consider these alongside more tangible costs and benefits in the evaluation [4, 8]. This can result in uninformed decision making, as the cost of AMR is likely to be under- or over-estimated by policymakers, if it is considered at all [4, 8, 9].

In 1996 Coast et al. argued that the omission of the cost of AMR in economic evaluation is partly explained by the challenges to its quantification [4], with extensive uncertainties surrounding resistance mechanisms, paucity and poor quality of relevant data, and other methodological challenges [5, 10]. The (mis)perception that the impact of AMR will only be felt in future years might also deter analysts from including them in the evaluation, assuming policymakers operate with a myopic view of health gains and costs. As confirmed in a recent review, very few attempts have since been made to quantify the externality of AMR [11].

Policymakers and key stakeholders, however, appear increasingly concerned with AMR, with unprecedented funding being allocated to interventions to mitigate its impact. In late 2016 the UN General Assembly held a special meeting on the topic, passing a unanimous resolution from Member States committing to adopt such measures [12]. Without enumerating the cost of AMR per antimicrobial consumed, it will be difficult to determine the allocative efficiency of these investments, and particularly so in low/middle income countries (LMICs) with more tangible causes of ill-health to invest in.

Therefore, despite the challenges, there is a clear need for costing the negative externality of AMR that can be affixed to the consumption of antimicrobials. The rare occasions where this has been done indicate the importance of such efforts. In a German hospital setting, for example, the use of a single defined daily dose of a second or third generation cephalosporin was associated with €5 and €15 respectively in costs of AMR [6]. The current analysis produced a menu of economic costs of AMR per antibiotic consumed for a variety drug classes, stratified into LMICs and high-income country settings. The output can be applied in future economic evaluations of interventions that involve or affect antibiotic consumption.

Evidence-based policy draws on economic evaluation to allocate resources most efficiently [29], but this is entirely dependent on the inclusion of all pertinent costs and benefits associated with interventions under consideration. This is, to our knowledge, a first attempt at estimating the costs of AMR per antibiotic consumed by drug class and across national income brackets. We chose simple and transparent methods and restricted our assessment to the current burden of AMR, rather than more uncertain future projections, and to tangible factors including only direct medical costs and productivity losses due to AMR attributable deaths. Even within this restrictive framework there is much uncertainty surrounding interactions between antibiotic consumption, development of resistance, and its economic implications, but our underlying assumptions and parameter estimates were conservative.

The cost per SU of antibiotic differed between the US and Thailand for several reasons. First, the burden of AMR is considerably higher in Thailand, with a total of 28 AMR associated deaths per 100,000 as compared with 4.6 per 100,000 in the US (Table 1). Furthermore, the two countries had different epidemiological profiles, such as a higher burden of Acinetobacter associated mortality in Thailand as compared with the dominance of MRSA in the US. There were also notable differences in the cost data between the two countries; as the unit costs per infection for Thailand were only available from hospital settings, they tended to be higher than those in the US, which included both hospital and community settings. Other factors contributing to this difference are the higher GDP per capita and lower per capita consumption of antibiotics in the US.

The costs of AMR for drug classes also varied widely, driven primarily by the degree to which they were assumed to propagate resistance in the selected infections; NSPs were assumed to drive resistance only in S. aureus, while cephalosporins were implicated in resistance in all pathogens. The costs per full course of antibiotics were mostly determined by the number of SU per course, which for glycopeptides is high - a full course of vancomycin being 56 SU (four daily over 14 days) as compared with three daily units for a course of azithromycin (Additional file 1: Table S1).

Very few attempts have been made to quantify the cost of AMR per antibiotic consumed and internalise them in evaluations of interventions that involve or affect the use of antimicrobials. A recent study by Oppong et al. was one of the first attempts to do so in an evaluation focusing on antibiotic treatment of respiratory infections, demonstrating the decisive impact this had on outcomes [30]. Their estimate for the cost of AMR, however, assumed that resistance is driven exclusively by human antimicrobial consumption and that consumption of all drug classes contribute to resistance in all pathogens equally. It also ignored the considerable differences in the burden of resistance across countries, as apparent in the much higher burden of AMR in Thailand compared with that in the US. An earlier study evaluating the cost-effectiveness of malaria rapid tests used a similarly crude estimate for the cost of antimalarial resistance, also showing the large impact this had in swaying results and conclusions [31]. Elbasha, building on previous work by Phelps [32] estimated the deadweight loss of resistance due to overtreatment and found a higher cost of AMR of $35 (2003) per course of amoxicillin in the US context [33].

Several studies have explored the correlation between antimicrobial consumption and resistance [34‐36]. The correlation coefficients in the current study are smaller than prior estimates. For example, the coefficient for resistance in E. coli in this analysis was 0.27 (Table 4) in comparison to 0.74 from Goossens et al. [34]. This could be explained by the latter using 14 European countries in contrast to 44 countries from different regions in our study, and more abundant data for European countries that enabled correlating between the consumption and resistance of specific drugs, rather than drug classes as done here. The smaller coefficients imply a conservative assessment of the cost of AMR attributable to human antibiotic consumption.

Kaier et al. derived measures of association between antibiotic consumption and resistance from a time-series analysis using a multivariate regression model with different drug classes [37]. This would be a better approach for calculating the RMf, rather than the ecological associations used here. We were restricted, however, by having only 10 years of consumption data and even sparser and more heterogeneous resistance data.

There were many assumptions and limitations in the analysis (see Additional file 1: Table S2). One key limitation was the inclusion of a limited number of organisms, while consumption of the same antibiotics could drive resistance in other organisms with additional costs. The Thai estimates also focused only on the burden of AMR within hospital settings, excluding the possible excess burden in primary care and the community. These and other listed limitations result in a conservative estimate of the economic costs of AMR in our model.

Taking the human capital approach to productivity losses implies much higher estimates than would have been derived using friction costs; given the context of this analysis, trying to capture the full societal costs of AMR, this was deemed appropriate. This is essentially equivalent to the widespread use of GDP/capita as a proxy for the ceiling ratio in cost-effectiveness analyses to classify interventions as cost-effective.

The direct medical costs assigned to resistant infections were derived very differently in each country; the US estimates were taken from a recent study providing a national estimate of the incremental healthcare cost of treating millions of patients with antibiotic sensitive and resistant infections [15]. The Thai estimates used rudimentary costing methods, largely relying on expert opinion to estimate the cost of antibiotics required to treat resistant infections.

The selection of drug classes implicated in propagation of resistance in the respective organisms were based on limited available evidence [24]. This might explain some apparent anomalies, like the relatively low costs for NSPs, which were assumed to drive resistance only in S. aureus. Another reason for this anomaly relates to the entire framework of the analysis, whereby the cost of AMR is approximated from its current (or recent) estimated burden, rather than projections of what will happen if resistance to last line drugs, such as carbapenem, were to spread, for which there are alarming early indications. Such an approach is arguably more relevant than focusing on the present burden of AMR, but it requires many more strong and contestable assumptions.

The data on consumption and resistance levels used to derive the RMf were limited to 10 years and a causal relationship was assumed. For many pathogens and types of infections, however, this is not realistic as increasing resistance could alter consumption patterns as patients and physicians adapt their behaviour in order to provide the best possible treatment in a changing environment of resistance and therefore counteract the assumed dose-response relationship.

These rudimentary estimates for the economic cost of AMR per antibiotic consumed could be improved upon in several ways in future work as better data become available. In addition to addressing the above limitations, the link between human antibiotic consumption and resistance can be disaggregated into hospital vs. community use. The model can be further extended to other organisms including parasites and viruses and their varying distribution in different health sectors and geographical locations (global/regional/country/hospital/community).

The estimates of the economic costs of AMR per antibiotic consumed in this analysis were high. Incorporation of such estimates in economic evaluation of interventions that affect the use of antibiotics will better portray their true costs and benefits, and could act as a catalyst for more efficient deployment of interventions to mitigate the burden of AMR. We highlight the limitations of the analysis to emphasise the need for further development of the methods, and point to the notable differences in the costs of AMR per antibiotic consumed between the two countries and within the different drug classes to encourage their adaptation to other settings as relevant data become available.