Clinical Antibiotic Resistance Patterns Across 70 Countries

We sought global patterns of antibiotic resistant pathogenic bacteria within the AMR Research Initiative database, Atlas. This consists of 6.5M clinical minimal inhibitory concentrations (MICs) observed in 70 countries in 633k patients between 2004 and 2017. Stratifying MICs according to pathogens (P), antibiotics (A) and countries (C), we found that the frequency of resistance was higher in Atlas than other publicly available databases. We determined global MIC distributions and, after showing they are coherent between years, we predicted MIC changes for 43 pathogens and 827 pathogen-antibiotic (PAs) pairings that exhibit significant resistance dynamics, including MIC increases and even decreases. However, many MIC distributions are multi-modal and some PA pairs exhibit sudden changes in MIC. We therefore analysed Atlas after replacing the clinical classification of pathogens into ‘susceptible’, ‘intermediate’ and ‘resistant’ with an information-optimal, cluster-based classifier to determine subpopulations with differential resistance that we denote S and R. Accordingly, S and R clusters for different PA pairs exhibit signatures of stabilising, directional and disruptive selection because their respective MICs can have different dynamics. Finally, we discuss clinical applications of a (R, dR/dt) ‘phase plane’ whereby the MIC of R is regressed against change in MIC (dR/dt), a methodology we use to detect PA pairs at risk of developing clinical resistance.

marginally, while the R-MIC is decreasing. Disruptive selection is the most common of these cases and R MIC increases are typically greater than the respective global mean change ( Figure 3C). 142 Clinical Observations 143 To detail these dynamics we introduce two phase planes: first, a 2-dimensional depiction of (R, dR/dt) 144 ( Figure 4A) and, second, a plot of (dS/dt, dR/dt) ( Figure S11). The latter is consistent with disruptive, 145 purifying and directional selection ( Figure S11) but the former is more useful: it highlights where the 146 most clinically relevant cluster, R, is now and where it is heading. Figure 4B                doripenem resistance is increasing from sub-to super-clinical whereas meropenem appears stable.
The R cluster of imipenem is slowly decreasing according to a regression (red dashes) but, in fact, this has merged with the S sub-population as it achieves super-clinical dosages.

424
Given a time series consisting of a number of MICs, n 1 , . . . , n T say, we want to test the null hypothesis, 425 H 0 , that all sampled points come from the same distribution against the alternative hypothesis, H 1 , 426 that there is a single changepoint in mean, µ, at timepoint τ : Assuming that the sample is normally distributed and that the variance, σ 2 , does not change across the sample, the log-likelihood ratio between the two hypotheses is: . We then find the value of τ that maximizes R τ , and we define We accept the existence of a changepoint, at timeτ , if G is larger than a 428 critical value, λ * , defined by the Bayesian Information Criterion 52 λ * = 1 2 log T . Some countries contribute more data than others, even after accounting for population size. c) More MIC data has been added to the Atlas database in recent times with a noticeable increase around 2012; this may be due to more countries having more data to contribute to the programme through time. d) An analogous comment applies to bacterial species that increase in number through time.

Spectral Analysis of Correlation Matrices
Given the lack of replication in clinical MIC assays, it is important to ask whether MIC timeseries from 431 the Atlas database are noise-like or are they formed from coherent signals drawn from multiple patients 432 with small noise? We therefore use a test for MIC coherency between years using the following null Given this, we now define functionals τ and τ N that can be used as the basis of a test that compares correlation matrices for PA pairs independently of the value of N over which the MIC data stream has been gathered. As X = L + I + L T , where L is a lower-diagonal matrix with entries between 0 and 1, then we are interested in whether the mean square off-diagonal entry in L, l ij , is 0, 1, or neither. Now, this mean square, noting tr(X) = tr(I) = N , is where τ lies between 0 and 1 2 .

463
Or, for a fixed N , we can test the total lower diagonal diagonal square entries of L, namely Others with more data are poorly ranked because they have block structures consistent with high year-year correlations followed by sudden changes in MIC distribution.