Elena Maria Tosca (1), Simone Milanesi (2), Marta Colaneri (3,4), Giuseppe De Nicolao (1,5)
(1) Electrical, Computer, and Biomedical Engineering, University of Pavia, Pavia, Italy, (2) Department of Mathematics, University of Pavia, Pavia, Italy, (3) Infectious Diseases and Immunopathology, Department of Clinical Sciences, Università di Milano, L. Sacco Hospital, Milan, Italy, (4) Centre for Multidisciplinary Research in Health Science (MACH), University of Milano, Milano, Italy (5) Division of Infectious Diseases I, Fondazione IRCCS Policlinico San Matteo, Pavia, Italy.
Introduction: Antibiotic resistance poses a significant global health challenge [1], necessitating robust surveillance methods to monitor its prevalence and trends across different regions. According to [2], “Setting up country-level surveillance of resistance and consumption is vital for understanding the impact of AMR and to reduce the spread of resistant pathogens”.
Objectives:
- To develop an appropriate diagnostic tool for statistically monitoring antibiotic resistance at country level, as reported in the WHO GLASS dashboard;
- to define a model of natural variability of antibiotic resistance that considers the number of performed tests and provides alarm limits for public health purposes;
- to obtain the distribution of antibiotic resistance across countries, highlighting possible clusters of countries sharing common features.
Methods: Antimicrobial resistance data for year 2020 relative to the WHO European region were acquired from the GLASS dashboard, which presents global antimicrobial consumption (AMC) and resistance (AMR) data for countries, territories, and areas (CTAs) that were enrolled in GLASS (Global Antimicrobial Resistance and Use Surveillance System) [3]. For each country the resistance percentage of different bacterial pathogens (Acinetobacter spp., Escherichia coli, Klebsiella pneumoniae, Salmonella spp., Streptococcus pneumoniae) to several antibiotics (Cephalosporins and Carbapenems) is presented. For each pathogen-antibiotic pair, the GLASS dashboard provides the number of Antibiotic Susceptibility Tests (AST) and the corresponding resistance percentage. In order to characterize the statistical variability of antibiotic resistance as a function of AST numerosity, we propose utilizing the funnel plot methodology [4], a validated approach that compares the performance of units having different sizes and provides probabilistic alarm limits. In particular, the inter-unit variability is described either by multiplicative overdispersion or by a random effect model [5]. To address the limitations of standard funnel plots, two new multi-funnel models are developed, based either on multiplicative overdispersion or the introduction of a Gaussian Mixture Model (GMM) random effect, needed to account for some multimodal distributions of antibiotic resistance observed in the dashboard.
Results: Some pathogen-antibiotic pairs were well-modeled by standard funnel plots, such as in the case of (Escherichia coli, Cefotaxime): the antibiotic resistance percentage was estimated as theta = 9.5%, with the multiplicative overdispersion model, and theta = 10.2%, with omega = 1.4% for the mixed effect model. In both cases, Italy with its 26.2% resistance was the only country outside the upper alarm limit, in accordance with [6]. Differently, in several cases, the standard model is outperformed by the multi-funnel one. An example is given by the (Acinetobacter spp., Meropenem) pair where countries form two clusters. The average resistance in North and Central Europe countries is 5%, while average resistance in South and East Europe is estimated as large as 80%, for both the multiplicative and mixed effect models. Also the Klebsiella resistance data typically require a multi-funnel model. For the (Klebsiella, Ceftadizime) pair the average resistance in North and Central Europe is 10% against 55% in South and East Europe. In the mixed effect model, the first and second components of the GMM have omega_1=0.2% and omega_2=1.4%, corresponding to a much larger variability of resistance in the South and East Europe cluster.
Conclusions: These findings show the effectiveness of funnel models for the statistical characterization of the variability of antibiotic resistance across European countries. In some cases, variability is adequately explained by a single random effect accounting for country variability, but in several other cases the variability calls for the use of the new multi-funnel GMM model. In this latter case, countries can be clustered based on the mixture components. A first cluster, characterized by relatively low resistance and small inter-country variability seems to correspond to North and Centre Europe, while a second cluster, associated with South and East Europe is characterized by much larger resistance and inter-country variability.
References:
[1] World Health Organization. Antimicrobial resistance surveillance in Europe 2022–2020 data. World Health Organization. Regional Office for Europe, 2022.
[2] Kumar. Antimicrobial Resistance: A Top Ten Global Public Health Threat. N.p. Print.
[3] https://worldhealthorg.shinyapps.io/glass-dashboard/_w_b064ed35/#!/amr
[4] Spiegelhalter, D. J. Funnel plots for comparing institutional performance. Stat. Med. 24(8), 1185–1202 (2005).
[5] Spiegelhalter, David J. “Handling over-dispersion of performance indicators.” BMJ Quality & Safety 14.5 (2005): 347-351.
[6] https://www.ecdc.europa.eu/en/publications-data/ecdc-country-visit-italy-discuss-antimicrobial-resistance-issues
Reference: PAGE 32 (2024) Abstr 10935 [www.page-meeting.org/?abstract=10935]
Poster: Methodology - Other topics