Exploring multidrug resistance patterns in community-acquired E. coli urinary tract infections with machine learning

BackgroundWhile associations of antibiotic resistance traits are not random in multidrug-resistant (MDR) bacteria, clinically relevant resistance patterns remain relatively underexplored. This study used machine learning, specifically association-set mining, to explore resistance associations within E. coli isolates from community-acquired urinary tract infections (UTIs). MethodsWe analysed antibiograms of community-acquired E. coli UTI isolates collected from 2018 to 2022 by Frances national surveillance system. Association-set mining was applied separately to extended-spectrum beta-lactamase-producing E. coli (ESBL-EC) and non-ESBL-EC. MDR patterns that had expected support (reflecting pattern frequency) and conditional lift (reflecting association strength) higher than expected by chance (p-value[≤]0.05) were used to construct resistance networks, and analysed according to time, age and gender. FindingsThe number of isolates increased from 360 287 in 2018 (10 150 ESBL-EC, 350 137 non-ESBL-EC) to 629 017 in 2022 (18 663 ESBL-EC, 610 354 non-ESBL-EC). More MDR patterns were selected in ESBL-EC than non-ESBL-EC (2022: 1770 vs 93 patterns), with higher respective network densities (2022: 0.230 vs 0.074). Fluoroquinolone, third-generation cephalosporin and penicillin resistances were strongly associated in ESBL-EC. The median densities of resistance association networks increased from 2018 to 2022 in both ESBL-EC (0.238 to 0.302, p-value=0.06, Pearson test) and non-ESBL-EC (0.074 to 0.100, p-value=0.04). Across all years, median network densities were higher in men than women in both ESBL-EC (2022: 0.305 vs 0.276) and non-ESBL-EC (2022: 0.128 vs 0.094); they were also higher in individuals over 65 years old than under 65 in ESBL-EC (2022: 0.289 vs 0.275) and non-ESBL-EC (2022: 0.103 vs 0.088). InterpretationThese findings, which show increasing MDR associations, especially in men and older individuals, highlight the importance of ongoing resistance surveillance to understand the future evolution of resistance patterns. FundingThis work received funding from the French government through the National Research Agency project COMBINE ANR-22-PAMR-0003. Research in contextO_ST_ABSEvidence before this studyC_ST_ABSWe searched Pubmed for previously published articles without any date or language restrictions using the search terms (multiresistan* OR "multidrug-resistan*") AND ("data mining" OR "machine learning" OR "artificial intelligence") AND (pattern* OR associat*). We found three studies that used machine learning to identify multiresistance patterns in various pathogens (chicken-associated Escherichia coli, human-associated Staphylococcus aureus and cattle-associated Salmonella enterica) in the United States. However, to our knowledge, no machine-learning studies to date have explored multiresistance patterns in human-associated Enterobacterales, especially within European contexts. Added value of this studyOur study provided a novel and detailed analysis of multiresistance patterns in community-acquired E. coli urinary tract infection collected from a French national surveillance system. Our findings confirmed that association-set mining is effective for identifying resistance associations in antibiotic resistance surveillance data. We explored the temporal evolution of resistance associations, gender-specific and age-specific differences, which to our knowledge, had not been previously analysed. Implications of all the available evidenceOur results suggest a temporal increase of resistance associations in community-acquired E. coli UTI and identify key patterns in different subpopulations. In the context of rising antibiotic resistance, optimizing the use of current medications is crucial, as few new antibiotics have been developed in the past two decades. With further research, this work could provide insight for targeted antibiotic stewardship strategies.