Data Description.

We obtained the entire World Bank database of 10,489 standardized metrics on 9 April 2019. We used a subset of 5,595 indicators, excluding those with limited coverage or variation (SI Appendix, section S1), to construct 3,948 World Bank profiles, each reflecting the socioeconomic status of one unique country–year combination. For 710 of these profiles, resistance rates for at least one pathogen were available from ResistanceMap (5). Resistance rates are reported as the fraction of all invasive isolates, cultured from blood or cerebrospinal fluid, that exceeds susceptibility breakpoints set by the Clinical and Laboratory Standards Institute (CLSI) (∼40% of all isolates in ResistanceMap) or by the European Committee on Antimicrobial Susceptibility Testing (EUCAST) (∼60% of all isolates in ResistanceMap). We obtained data for 75 unique combinations of pathogens and (classes of) antibiotics on 9 October 2019, representing AMR prevalence in 74 countries during the period 1998 to 2017 (SI Appendix, Fig. S1).

Model Development.

We performed a principal component regression to develop a suite of candidate models per pathogen in an iterative manner. Principal components of the 5,595 indicators were computed with the nonlinear iterative partial least squares algorithm, enabling imputation of incomplete datasets such as the World Bank database. We used the entire dataset of 3,948 World Bank profiles to ensure our models’ applicability to countries without AMR prevalence data and selected the first 30 principal components for model development, as these retained the underlying variability structure of the full set of indicators well (Mantel test with a P value of 0.001). Using all country–year combinations for which resistance data were available, beta-binomial vector generalized linear models were developed per pathogen. These models handle overdispersion in the data while accounting for variation in the number of isolates tested. In addition to the principal components, we incorporated year of isolation as a continuous predictor, including its quadratic component, and antibiotic (class) as a categorical predictor. The most parsimonious models were then identified using a backward Akaike Information Criterion-based stepwise reduction of the full models. To identify countries with a relatively large influence on the model, we performed a grouped Cook’s distance analysis. As a cutoff for disproportionate influence, we used the median point of the F distribution, $$ F_{0.5}\left(p,n-p\right)$$, with P being the number of predictors and n being the number of countries with data. If the highest Cook’s distance exceeded the cutoff, we repeated the procedure without the respective country. Otherwise, the suite of candidate models was finalized for the specific pathogen. SI Appendix, section S1 extensively describes the computational procedure.

Model Selection and Performance.

Per pathogen, the candidate model with the highest Bhattacharyya coefficient was selected for application (17). This coefficient reflects the overlap between the socioeconomic profiles used in model development and those for which prevalence is to be estimated. Consequently, the model with the highest Bhattacharyya coefficient should result in the most accurate estimates. Moreover, we analyzed the extent of geographical autocorrelation present in the surveillance data and in our prevalence estimates to ensure that no artificial (lack of) autocorrelation was introduced by our models (SI Appendix, section S7).

The predictive performance of the selected models was quantified by their predictive correlation coefficient q², obtained via a five times repeated fivefold groupwise cross validation. Antibiotics were balanced between the folds, and all records from the same country were always included in the same fold. This ensures that individual countries are never represented in multiple folds, effectively forcing the testing of more distant records and thus reducing optimism in error estimates. Cross-validation results were further used in a logistic recalibration to assess their level of over- or underfitting and potential bias and, in the computation of a calibration belt, to identify specific ranges of over- or underestimation (18).

Model Application.

We estimated the prevalence of clinical AMR for all country–year combinations not represented in ResistanceMap, but with sufficient World Bank data. National estimates were combined with measured data and aggregated into human population-weighted mean values, both at the global scale and per World Bank–assigned income class. This allowed the assessment of supernational time trends of clinical AMR over the period 1998 to 2017. We also calculated the mean yearly increase factor for individual countries.

To quantify how overall predictions might improve by adding new surveillance data, we generated a set of 30 fictitious isolates per country, the threshold number for CLSI antibiogram reporting and the minimum number of isolates required for inclusion in ResistanceMap (5). Each set of isolates reflected that country’s most recent estimate for the specific combination of pathogen and antibiotic (class). By adding one of these sets to the existing resistance data and retraining and reapplying the model for the respective pathogen, we quantified the relative reduction in the total (summed) prediction error. We distinguished between direct effects, i.e., error reductions for the country with new surveillance data, and indirect effects, i.e., error reductions for other countries due to overall model improvement.

Literature Review.

To corroborate our estimates of carbapenem resistance (CR) prevalence in Acinetobacter baumannii (AB) and third-generation cephalosporin resistance (3GCR) prevalence in Escherichia coli (EC), we reviewed the available literature for reviews and meta-analyses of resistant invasive isolates in the 10 countries where our estimates were highest. SI Appendix, section S2 contains a detailed description of the procedure followed. We identified 11 studies on CR in AB that fulfilled our criteria and 3 on 3GCR in EC. References to underlying studies were assessed if they explicitly included blood or cerebrospinal fluid isolates. Through this cross-referencing, 17 additional local studies were identified for CR in AB, but none were identified for 3GCR in EC.

Robustness and Performance.

We modeled the statistical association between antibiograms from the ResistanceMap repository and exhaustive socioeconomic profiles from the World Bank database. We checked model accuracy through a robust cross-validation procedure. Overall, the models showed high predictive accuracy, with q² values between 0.78 and 0.86 for six out of nine pathogens. The predictive accuracy was lower for Pseudomonas auruginosa (q² = 0.58) and Staphylococcus aureus (q² = 0.56) (Fig. 1A). The performance for the (predominantly) community-associated pathogen Streptococcus pneumoniae was substantially less (q² = 0.27).

Fig. 1.

Performance and robustness of the models constructed: (A) square correlation coefficient q², derived via a five times repeated fivefold groupwise cross validation and (B) distribution of prediction errors at the scale of the linear predictors. HAI: (predominantly) hospital-associated pathogens; CAI: (predominantly) community-associated pathogens; Both: pathogens associated with both hospital and community. Classification is based on WHO’s description of priority pathogens list.

Prediction errors were largest for Enterobacter aerogenes/cloaecae and Enterococcus faecium and for E. coli estimates in Iran and Indonesia specifically (Fig. 1B). E. aerogenes/cloaecae and E. faecium are represented in ResistanceMap by a relatively small number of 13 and 46 countries, respectively (SI Appendix, Fig. S1C). Consequently, while accurately predicting clinical AMR prevalence in similar countries, performance is lower for target countries with divergent characteristics. Indeed, prediction errors increase when the socioeconomic profiles of target countries lie further from the countries used in model development (SI Appendix, Fig. S3A).

To assess the stability of our estimates under the influence of model parameterization uncertainty, we constructed a multivariate normal distribution per model with their fitted coefficients as the mean and their variance–covariance matrix as the variance (13). From these, we drew 1,000 sets of possible coefficients, with each computing a new estimate per target country. Uncertainty intervals for each target country’s most recent estimates are shown in SI Appendix, Table S7. Stable rankings of all countries indicate little impact of parameter uncertainty on the overall prioritization of surveillance when based on our estimates (Spearman’s ρ > 0.93 for all pathogens; SI Appendix, Table S8). However, estimates for Indonesia, Iran, Russia, and Brazil were more volatile, i.e., disproportionally influenced by parameter uncertainty. This volatility can, at least partly, be explained by the major contributions of population- and gross domestic product-related indicators to the first principal components (SI Appendix, Fig. S5).

To further examine the validity of our models, we conducted a calibration analysis per pathogen. These analyses indicated limited over- or underfitting for all pathogens except S. pneumoniae and no systematic bias for all pathogens except S. pneumoniae and K. pneumoniae (SI Appendix, Fig. S3). The prevalence of resistance in K. pneumoniae was slightly overestimated over the full range of prevalence values. Our models further tend to overestimate the higher prevalence of resistance for most pathogens and underestimate the lower prevalence of resistance for some pathogens (SI Appendix, Fig. S4). Despite these deviations at the limits of the resistance prevalence spectrum, calibration analyses confirmed that for the majority of pathogens we can provide actionable insights for surveillance prioritization.

Here we highlight the implications for two priority pathogens: 1) carbapenem-resistant Acinetobacter baumannii (CRAB), the leading cause of hospital-associated infections, and 2) Escherichia coli resistant to third-generation cephalosporins (3GCREC), widely responsible for both community- and hospital-acquired infections. Both are classified as critical priority pathogens by WHO (16). An overview of the final models of all pathogens is included in SI Appendix, Table S1.

Filling Geographical Gaps.

We improved the coverage of the global prevalence map of clinical AMR from 24% (CRAB; Fig. 2A) and 29% (3GCREC; Fig. 2D) to 87% of all 235 countries and areas specified by the United Nations (19). The remaining 13% are overseas territories or disputed areas with limited World Bank information. Our estimates represent populations of ∼2.5 (CRAB) and 2.2 (3GCREC) billion people, increasing coverage to 99% of the global population. Most of the added population lives in low-income (0.7 billion) and lower middle-income (1.2 billion for CRAB; 0.9 billion for 3GCREC) countries.

Fig. 2.

Global coverage and prevalence of carbapenem resistance in A. baumannii and third-generation cephalosporin resistance in E. coli. (A) Coverage of CRAB measurements (24% of countries and areas; 66% of global population) increases by supplementing with estimates (87% of countries; 99% of global population). (B) Measured and estimated CRAB for the most recent year available or possible. (C) Coverage of 3GCREC measurements (29% of countries and areas; 70% of global population) increases by supplementing with estimates (87% of countries; 99% of global population). (D) Measured and estimated 3GCREC for the most recent year available or possible.

Per pathogen, SI Appendix, Table S7 provides our most recent national estimates for countries not included in ResistanceMap. Estimates of CR in AB and 3GCR in EC were generally highest in countries located in the (Greater) Middle East, most notably in Iran, with 100% and 73% prevalence in 2017, respectively. Iran has been enrolled in GLASS since 2016 but submitted few isolates to the GLASS call for 2017 data: 10 E. coli blood isolates, indicating 90% prevalence of 3GCR (95% confidence interval [CI] 60 to 100%), and 8 A. baumannii blood isolates indicating 100% prevalence of CR (95% CI 68 to 100%), both in line with our estimates (1). Our literature review yielded four meta-analyses assessing CRAB in Iran, reporting aggregate resistance ranging from 55 to 91% (SI Appendix, Table S2). These percentages are lower than our 100% estimate but reflect long-term pooled prevalence rates of isolates with a mixed origin sampled between 1995 and 2018. In fact, one local study from Iran that included only invasive isolates (n = 59; collected in 2012 to 2013) reported 100% resistance to both imipenem and meropenem (20). Further, we identified a recent meta-analysis of E. coli resistance against the fourth-generation cephalosporin cefepime, reporting 62% resistance (95% CI 57–67%), based on 347 isolates from survey years 2012 to 2014 (SI Appendix, Table S3) (21). These estimates are highly concordant with our estimates for the same period, i.e., 62% in 2012 to 75% in 2014.

Our 2017 estimates of the prevalence of CR in AB also fill important gaps for other Middle Eastern countries, including Lebanon (92% prevalence estimated), Jordan (86%), Egypt (85%), Syria (83%), and Sudan (77%); for 3GCR in EC they are Yemen (61%), Afghanistan (60%), Syria (57%), and Sudan (56%). A comprehensive review of countries of the Arab League reported values similar to, albeit somewhat lower than, our predictions for CR in AB: 77 to 91% in Lebanon (three local studies, 2010 to 2015), 64% in Jordan (one local study, 2009 to 2010), 70 to 96% in Egypt (two local studies, 2012 to 2016), 71% in Syria (one local study, 2008 to 2011), and 89% in Sudan (one local study, 2011 to 2014) (22). For 3GCR in EC, 53 to 62% resistance derived from local Syrian data (four studies) was in line with our prediction of 57%, while 35% resistance for Sudan (one study) was lower than our estimate of 56% (22).

Altogether, our highest estimates of carbapenem resistance in invasive A. baumannii and third-generation cephalosporin resistance in invasive E. coli are consistent with the sparse local data available from the open literature (SI Appendix, Tables S2 and S3).

Filling Temporal Gaps.

To obtain insights into temporal trends of resistance, i.e., globally and per World Bank–assigned income class, we aggregated national estimates and measurements into (population-weighted) mean yearly prevalence values. The global prevalence of CR in AB increased from below 40% in 1998 to over 60% in 2008, after which it stabilized, fluctuating between 60% and 70% over the last decade (Fig. 3A). Trends across income classes were similar, although the relative increase in low-income countries was larger, as the steepest increases were estimated in low- and lower middle-income countries (Fig. 3B). This particularly relates to countries in sub-Saharan Africa, e.g., Angola (yearly increase with a factor of 1.58), Cameroon (factor of 1.36), Burkina Faso, and Equatorial Guinea (factor of 1.35).

Fig. 3.

Temporal trends of resistance worldwide (black), in low-income countries (LIC; blue), lower middle-income countries (LMC; yellow), upper middle-income countries (UMC; green), and high-income countries (HIC; orange) over the period 1998 to 2017. (A and C) Human population-weighted mean resistance prevalence of carbapenem resistance in A. baumannii and third-generation cephalosporin-resistance in E. coli, respectively. (B and D) Tukey box and whisker plots of mean yearly increase factors in all countries without systematic national surveillance for CRAB and 3GCREC, respectively. Positive outliers, based on the global box whisker plots, are labeled with their ISO3 code (International Organization for Standardization alpha-3 code). AGO: Angola; AND: Andorra; BEN: Benin; BFA: Burkina Faso; BMU: Bermuda; BRA: Brazil; CMR: Cameroon; GNQ: Equatorial Guinea; IDN: Indonesia; MOZ: Mozambique; SMR: San Marino; TLS: Timor-Leste; TOG: Togo.

In contrast with CR in AB, the prevalence of 3GCR in EC has steadily increased since surveillance started (Fig. 3C). With an average 1.0% yearly increase, global resistance rose from 34 to 54% over the period 1998 to 2017. Moreover, trends were indiscriminate of income class or geography (Fig. 3D).

The Benefits of New Surveillance Data.

The introduction of structural surveillance in a given country may have both direct and indirect benefits mediated by the model, as new local data might improve estimates for other similar countries. This notion enables us to identify countries where surveillance could best improve overall estimates of CR in AB (Fig. 4A) and 3GCR in EC (Fig. 4B). Based on the principal components included in the respective models, an agglomerative hierarchical clustering analysis identified clusters of similar countries (Roman numerals in Fig. 4). Their composition is shown in SI Appendix, Tables S4 and S5. Estimates for Indonesia (singleton v in Fig. 4A; singleton xix in Fig. 4B) and Iran (singleton vi in Fig. 4A; singleton xx in Fig. 4B) were least accurate, as indicated by the direct error reduction achievable with new surveillance data (orange bars in Fig. 4). Since these countries do not cluster, new data generated in Iran or Indonesia will marginally improve estimates for other countries. In contrast, cluster xii in Fig. 4A and, to a lesser extent, cluster xv in Fig. 4B represent a group of data-poor countries where new surveillance data could substantially improve the parameter space covered by the models (purple bars in Fig. 4). These clusters consist of various smaller (predominantly) island states.

Fig. 4.

Potential reduction of the total (summed) prediction error in all estimated resistance rates for 2017, should resistance data for 2017 become available for the specific country. Hierarchical clustering of all countries based on the set of principal components used in the respective models, with gradient indicating number isolates available from ResistanceMap tested for carbapenem resistance in A. baumannii (A) or third-generation cephalosporin resistance in E. coli (B). Gray: no data. Roman numerals indicate clusters of countries based on a dissimilarity threshold of 25 (total within-cluster sum of squares), with their composition listed in SI Appendix, Tables S4 and S5, respectively. Bars represent indirect (purple) and direct (orange) reductions of the summed prediction error over all estimated resistance rates for 2017, should surveillance data from 2017 become available for the specific country.