Excluding Geographic Variables Does Not Fix Regional Bias in Machine Learning Antimicrobial Resistance Prediction: Analysis of 77,548 Isolates Across 132 Countries

Excluding Geographic Variables Does Not Fix Regional Bias in Machine Learning Antimicrobial Resistance Prediction: Analysis of 77,548 Isolates Across 132 Countries | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Excluding Geographic Variables Does Not Fix Regional Bias in Machine Learning Antimicrobial Resistance Prediction: Analysis of 77,548 Isolates Across 132 Countries Hayden Farquhar This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8772201/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Background Machine learning models for antimicrobial resistance (AMR) prediction exhibit geographic performance disparities, with models trained on high-income country data underperforming in low- and middle-income settings. A seemingly straightforward solution is to exclude geographic variables from models. We tested whether this approach eliminates regional bias. Methods We analysed 77,548 bacterial isolates from the BV-BRC database across two cohorts. The Primary Cohort (n = 39,859 Escherichia coli from 132 countries) quantified regional ciprofloxacin resistance prevalence. The Genomic Cohort (n = 37,689 E. coli with fluoroquinolone resistance gene annotations) tested whether a model trained exclusively on genomic features—with geographic variables explicitly excluded—would produce equitable sensitivity across regions. We evaluated sensitivity disparities at multiple classification thresholds. Results Despite excluding all geographic variables, the genomic model produced significantly different prediction scores by region (ANOVA F = 4.99, p = 1.45×10⁻⁴). At a threshold of 0.30, sensitivity ranged from 61.4% (Oceania) to 81.0% (Africa)—a 19.6 percentage point disparity. No threshold achieved sensitivity variation below 10 percentage points across all regions. The underlying cause: resistance gene prevalence itself varies geographically (qnr genes: 1.0% North America vs 6.8% Asia, p < 0.001), meaning any model using these biologically relevant features will inherit geographic structure. Conclusions Excluding geographic variables does not fix regional bias in AMR prediction because the bias is encoded in the underlying biology, not the model's feature set. Resistance genes vary geographically due to antimicrobial pressure, horizontal gene transfer, and clonal expansion. These findings demonstrate that geographic fairness requires region-specific models or thresholds, not simply removing location data. We recommend mandatory geographic stratification in model evaluation and recognition of geography as a protected attribute in medical AI fairness frameworks. antimicrobial resistance machine learning algorithmic fairness geographic bias health equity Escherichia coli fluoroquinolone resistance Figures Figure 1 Figure 2 Figure 3 Figure 4 INTRODUCTION Machine learning models for antimicrobial resistance (AMR) prediction face a well-documented problem: they perform worse in some geographic regions than others. 1 – 3 Models trained predominantly on isolates from high-income countries show degraded accuracy when applied to low- and middle-income settings, where the burden of resistant infections is highest. 4 This geographic performance gap is not merely a technical inconvenience—it risks entrenching health disparities by providing inferior diagnostic support to populations who need it most. The intuitive solution is straightforward: remove geographic variables from the model. If a model cannot "see" where an isolate originated, the reasoning goes, it cannot discriminate based on geography. This approach aligns with established fairness strategies in other domains, where excluding protected attributes (race, sex, age) has been proposed as a path to equitable prediction. 5 We tested whether this intuitive solution works for AMR prediction. Using a large multi-country dataset, we trained a model using only fluoroquinolone resistance genes—mechanistically relevant genomic features—with geographic variables explicitly excluded. We then evaluated whether the model achieved equitable sensitivity across regions. Our findings reveal a fundamental problem: excluding geographic variables does not eliminate geographic bias. The bias persists because it is encoded in the underlying biology. Resistance gene prevalence varies across regions due to differences in antimicrobial prescribing, horizontal gene transfer dynamics, and clonal expansion patterns. 6 – 8 Any model that uses these biologically relevant features will inherit their geographic structure—regardless of whether "region" appears in the feature set. This finding has immediate implications for how we develop, evaluate, and regulate AMR prediction tools. It suggests that the path to geographic fairness runs not through feature exclusion, but through region-specific models, region-aware thresholds, and mandatory geographic stratification in model evaluation. The Geographic Bias Problem in AMR Prediction Antimicrobial resistance caused an estimated 4.95 million deaths globally in 2019, with mortality rates in Western sub-Saharan Africa more than four times higher than in Australasia. 4 Machine learning approaches for AMR prediction have proliferated rapidly, with meta-analyses reporting pooled AUCs of 0.78–0.82. 9–11 However, the geographic distribution of training data raises critical fairness concerns. Major genomic databases substantially overrepresent isolates from the United States and Western Europe. 12 – 13 The landmark Valavarasu et al. 2025 study explicitly acknowledged "significant underrepresentation of data from Sub-Saharan Africa" and "bias towards high-income countries." 11 Kim et al.'s systematic review concluded: "An ML model trained on samples from one source may be ineffective in predicting the resistance phenotype for isolates from different environments, and universal ML models for AMR may not be feasible." 10 The most compelling evidence comes from Nsubuga et al.'s 2024 cross-continental validation study. 1 Models trained on English isolates showed dramatic performance degradation on African data: ciprofloxacin prediction accuracy dropped from 87% to 50%, and cefotaxime prediction plummeted from 92% to 45% when applied to Ugandan, Nigerian, and Tanzanian isolates. Paradoxically, ampicillin prediction inverted from 58% to 94%—suggesting models may learn geographic signatures rather than true resistance determinants. Why Feature Exclusion Seems Like a Solution The algorithmic fairness literature offers a seemingly relevant precedent. Obermeyer et al.'s foundational 2019 study demonstrated that a healthcare cost-prediction algorithm systematically disadvantaged Black patients—not because race was an explicit feature, but because the algorithm used healthcare costs as a proxy for health needs. 14 Removing or adjusting such proxy variables improved equity. Seyyed-Kalantari et al. showed similar patterns in chest X-ray AI, where technical factors created artifactual prediction shifts that could be addressed through recalibration. 15 This logic suggests that if we train AMR models using only mechanistically relevant features—the actual genes that confer resistance—rather than geographic proxies, we might achieve equitable prediction. The resistance genes are the "true" signal; geography is merely a confounder. While algorithmic fairness researchers have long understood that excluding protected attributes does not guarantee equitable outcomes, this principle has not been empirically demonstrated in AMR prediction—nor has it penetrated clinical or regulatory guidance for medical AI deployment. We hypothesised that this reasoning fails for AMR prediction because resistance gene prevalence itself varies geographically. The genes are not independent of geography; they carry geographic information. If so, excluding geographic variables would be ineffective—the bias would persist through the genomic features themselves. Study Objectives Our objectives were to: (1) quantify geographic variation in E. coli ciprofloxacin resistance across 132 countries; (2) test whether a model trained exclusively on genomic features—with geographic variables excluded—produces equitable sensitivity across regions; (3) identify the mechanism underlying any persistent disparities; and (4) evaluate the clinical and regulatory implications of our findings. METHODS Data Source This retrospective cross-sectional study analysed antimicrobial susceptibility data from the Bacterial and Viral Bioinformatics Resource Center (BV-BRC), a comprehensive genomic database containing over 2 million bacterial genomes with associated antimicrobial resistance phenotypes. 12 Data were accessed via the BV-BRC API in December 2024. Primary Cohort: Geographic Prevalence Analysis The Primary Cohort comprised E. coli isolates with ciprofloxacin susceptibility data and geographic metadata. Genome metadata (n = 63,702) and ciprofloxacin susceptibility records (n = 141,537) were extracted via paginated API queries. After filtering to isolates with definitive resistant or susceptible phenotypes and excluding Antarctica (n = 1), the final dataset comprised 39,859 isolates from 132 countries across six geographic regions (Africa, Asia, Europe, North America, Oceania, South America), classified using BV-BRC's geographic_group schema. Country-level income classification followed World Bank 2024 categories. Genomic Cohort: Feature Exclusion Analysis The Genomic Cohort tested whether excluding geographic variables eliminates regional bias. We queried specialty gene annotations for all E. coli genomes with ciprofloxacin susceptibility phenotypes and geographic metadata. Binary features were extracted for 10 established fluoroquinolone resistance determinants: quinolone resistance-determining region (QRDR) targets (gyrA, gyrB, parC, parE); plasmid-mediated quinolone resistance (PMQR) genes (qnrA/B/C/D/S); efflux pump components (acrAB, tolC, marRAB, oqxAB); and the aminoglycoside acetyltransferase variant aac(6')-Ib-cr. 6–8 Critically, geographic region was explicitly excluded from the feature set. The model received only genomic information—the same information a laboratory-based prediction system would have access to. The final Genomic Cohort comprised 37,689 isolates with resistance gene annotations. Model Training and Evaluation A logistic regression model was trained using only the 10 genomic features as predictors. The dataset was split 70%/30% for training/testing, stratified by resistance phenotype. We tested whether prediction scores differed by region using one-way ANOVA, and evaluated sensitivity disparities across regions at multiple classification thresholds (0.25, 0.30, 0.35, 0.40). Chi-squared tests assessed the association between genomic feature prevalence and geographic region. Fairness Definition We define fairness primarily in terms of equal sensitivity (true positive rates) across geographic regions. This choice reflects clinical consequences: a missed resistance prediction leads directly to inappropriate empiric therapy and potential treatment failure—consequences disproportionately severe in resource-limited settings with fewer second-line options. 16 We consider sensitivity disparities exceeding 10 percentage points to be clinically significant. Statistical Analysis Regional resistance prevalence was calculated with 95% confidence intervals using bootstrap resampling (n = 1,000). Between-region comparisons used chi-squared tests with Bonferroni correction for multiple comparisons (15 pairwise comparisons, adjusted α = 0.0033). Effect sizes were quantified using odds ratios with 95% confidence intervals. Analyses were performed using Python 3.10 with scipy 1.11 and statsmodels 0.14. Multi-Organism Validation To assess generalisability, we analysed geographic patterns in Klebsiella pneumoniae and Staphylococcus aureus ciprofloxacin resistance using the same methodology. RESULTS Primary Finding: Excluding Geographic Variables Does Not Eliminate Regional Bias The logistic regression model, trained exclusively on 10 fluoroquinolone resistance genes with geographic variables explicitly excluded, produced prediction scores that varied significantly by region (one-way ANOVA: F = 4.99, p = 1.45×10⁻⁴). Mean prediction scores ranged from 0.299 (Oceania) to 0.312 (Asia). This regional variation in prediction scores translated directly into sensitivity disparities at any fixed threshold. At a classification threshold of 0.30, sensitivity ranged from 61.4% (Oceania, n = 83 resistant cases) to 81.0% (Africa, n = 174 resistant cases)—a 19.6 percentage point disparity (Table 1 ). This means that a resistant isolate from Oceania has a 20% lower probability of being correctly identified compared to a resistant isolate from Africa, despite using the same model and threshold. No single threshold achieved equitable performance. At threshold 0.25, the sensitivity gap narrowed to 12.8 percentage points but remained clinically significant. At threshold 0.40, the gap compressed to 4.5 percentage points—but only because nearly all cases were missed across all regions (> 3,000 cases missed). No operating point achieved both acceptable case detection and sensitivity variation below 10 percentage points (Table 1 , Fig. 1 ). Table 1 Genomic Model Sensitivity by Threshold and Region Threshold Africa Asia Europe N. America Oceania S. America Gap (pp) 0.25 87.4% 87.5% 86.0% 87.4% 74.7% 84.1% 12.8 0.30 81.0% 80.9% 78.7% 79.4% 61.4% 76.8% 19.6 0.35 12.1% 14.4% 18.1% 16.1% 10.8% 13.9% 7.2 0.40 9.2% 10.7% 12.9% 11.0% 8.4% 10.6% 4.5 Logistic regression model trained on 10 fluoroquinolone resistance gene features with geographic variables excluded. N = 11,307 test isolates. pp=percentage points. Mechanism: Resistance Gene Prevalence Varies Geographically The persistent regional bias has a clear biological explanation: the genomic features themselves carry geographic information. Resistance gene prevalence varied significantly across regions for all tested determinants (χ² tests, all p < 0.001) (Table 2 ). The most striking example is plasmid-mediated quinolone resistance (PMQR) genes. Qnr gene prevalence ranged from 1.0% in North America to 6.8% in Asia—a nearly 7-fold difference. This pattern reflects the documented epidemiology of PMQR dissemination: these resistance elements emerged in Asia and have spread globally at varying rates depending on antimicrobial selection pressure, horizontal gene transfer dynamics, and clonal expansion. 7 QRDR target genes showed similar geographic structure: gyrA prevalence ranged from 19.5% (North America) to 26.6% (Asia). Efflux pump genes (acrAB) ranged from 33.3% (North America) to 45.4% (Africa). Because these genes are mechanistically relevant—they actually confer resistance—any well-calibrated model must produce higher predicted probabilities in regions where they are more prevalent. The geographic bias is not an artifact of the model; it is encoded in the biology. Table 2 Genomic Feature Prevalence by Geographic Region Region N Resistance gyrA parC qnr acrAB Asia 13,331 45.9% 26.6% 19.0% 6.8% 44.0% North America 9,911 16.2% 19.5% 12.9% 1.0% 33.3% Europe 9,166 27.1% 23.2% 16.4% 2.8% 40.9% Africa 2,560 23.6% 26.1% 16.7% 4.1% 45.4% South America 1,534 32.7% 23.9% 17.4% 5.4% 39.0% Oceania 1,187 23.8% 25.3% 15.9% 1.3% 44.3% All genomic features significantly associated with region (χ² tests, all p < 0.001). qnr includes qnrA/B/C/D/S variants. Geographic Context: Underlying Resistance Prevalence To contextualise the genomic findings, we quantified regional ciprofloxacin resistance prevalence in the Primary Cohort (n = 39,859 isolates from 132 countries). Resistance prevalence ranged from 16.8% (95% CI: 16.1–17.5%) in North America to 44.1% (95% CI: 43.3–44.9%) in Asia—a 27.3 percentage point absolute difference (Table 3 ). The odds of resistance in Asia were 3.90 times higher than in North America (95% CI: 3.67–4.15; p < 0.001). Of 15 pairwise regional comparisons, 13 remained statistically significant after Bonferroni correction. Table 3 E. coli Ciprofloxacin Resistance by Geographic Region (Primary Cohort) Region N Prevalence 95% CI OR (95% CI)* Asia 14,157 44.1% 43.3–44.9% 3.90 (3.67–4.15) South America 1,625 31.6% 29.2–33.7% 2.28 (2.03–2.56) Europe 9,817 26.7% 25.8–27.6% 1.80 (1.68–1.93) Africa 2,701 23.5% 22.0-25.2% 1.52 (1.37–1.68) Oceania 1,220 23.4% 21.1–25.8% 1.51 (1.31–1.75) North America 10,339 16.8% 16.1–17.5% Reference *Odds ratio compared to North America (reference). All comparisons p < 0.001. The Threshold Selection Problem To illustrate why this geographic prevalence heterogeneity creates an unavoidable threshold problem, we conducted a simulation analysis. We trained a Random Forest classifier using only geographic region as a predictor. This deliberately simplified model demonstrates a limiting case: when the only predictive feature is region, a well-calibrated model outputs the regional prevalence as its predicted probability. At any single global threshold, regions partition into discrete groups: those above the threshold (all isolates classified as resistant) and those below (all classified as susceptible). At a threshold of 0.30 (approximating global prevalence), only Asia and South America exceed the threshold. All 1,594 resistant isolates from Europe, Africa, Oceania, and North America—44.3% of total resistant cases—would be classified as susceptible (Table 4 ). This simulation is not a recommended modelling strategy; it demonstrates the mathematical constraint that any model correctly calibrated to regional base rates faces when a single threshold is applied to populations with heterogeneous prevalence. Table 4 Simulation: Impact of Global Threshold Selection on Regional Classification Threshold Regions Predicted Resistant Regions Predicted Susceptible Cases Missed 0.20 Asia, S.Am, Europe, Africa, Oceania North America 518 (14.4%) 0.25 Asia, S.Am, Europe Africa, Oceania, N.Am 788 (21.9%) 0.30 Asia, South America Europe, Africa, Oceania, N.Am 1,594 (44.3%) 0.35 Asia S.Am, Europe, Africa, Oceania, N.Am 1,777 (49.4%) 0.45 None All regions 3,600 (100%) Resistant cases missed calculated from test set (n = 11,918; 3,600 resistant). Simulation uses regional prevalence as predictions to illustrate the threshold selection problem. Multi-Organism Validation Analysis of K. pneumoniae and S. aureus confirmed that geographic heterogeneity is not specific to E. coli or ciprofloxacin. K. pneumoniae showed the largest disparity: 70.1% resistance in South America versus 29.9% in Oceania (40.2 percentage point gap). S. aureus showed 37.0% resistance in Asia versus 15.6% in Oceania (21.4 percentage point gap). Importantly, different organisms showed different regional rankings: E. coli highest in Asia, K. pneumoniae highest in South America. This pattern indicates that observed disparities reflect organism-specific epidemiology rather than simple "high-resistance regions" (Fig. 4 ). Data Representation Inversely Related to Resistance Burden Analysis by World Bank income classification revealed an inverse relationship between data representation and resistance burden. High-income countries contributed 58.5% of isolates but had 19.9% resistance prevalence. Lower-middle-income countries contributed only 6.8% of isolates but had 53.7% resistance prevalence. Models trained on available data are thus optimised for settings where resistance is lowest, while populations bearing the highest AMR burden are most poorly represented. DISCUSSION Principal Findings This study demonstrates that excluding geographic variables does not fix regional bias in AMR prediction models. A model trained exclusively on fluoroquinolone resistance genes—with geographic variables explicitly excluded—still produced 19.6 percentage point sensitivity disparities across regions. No classification threshold achieved both acceptable case detection and equitable performance. The mechanism is clear: resistance gene prevalence itself varies geographically. Qnr genes are nearly 7 times more prevalent in Asia than North America. Because these genes are mechanistically relevant—they actually confer resistance—any model using them must produce geographically varying predictions. The bias is encoded in the biology, not the feature set. This finding has important implications. It means that the standard fairness strategy of excluding protected attributes cannot work for AMR prediction. Geographic information permeates the biologically relevant features. The only paths to equity are region-specific models, region-aware thresholds, or fundamentally different approaches to global AMR prediction. Clinical Implications Consider a patient presenting with suspected urinary tract infection requiring empiric antibiotic selection. An AI-assisted decision support tool predicts fluoroquinolone resistance probability. Our findings show that at any fixed threshold, patients in low-prevalence regions will more frequently receive "susceptible" predictions. When resistance is actually present, these patients experience treatment failure. The clinical consequences are asymmetric. A patient incorrectly classified as susceptible receives ineffective antibiotics—leading to treatment failure, prolonged illness, and potentially death. A patient incorrectly classified as resistant receives broader-spectrum therapy that remains clinically effective, though suboptimal for antimicrobial stewardship. In AMR contexts, false negatives cause more immediate harm than false positives. Critically, the regions with highest false negative rates (lowest sensitivity) are often those with the highest true resistance prevalence. Patients in Asia—where 44% of isolates are resistant—face the highest absolute burden of both resistance and model underperformance. This compounds existing health inequities. Relationship to Algorithmic Fairness Theory Our findings instantiate the fairness impossibility theorems established by Kleinberg et al. and Chouldechova in theoretical contexts. 17 – 18 Their proofs demonstrate that when base rates differ between groups, no classifier can simultaneously achieve calibration, equal false positive rates, and equal false negative rates. The 27.3 percentage point prevalence gap between Asia and North America creates conditions where mathematical impossibility becomes clinical inevitability. Our contribution is not rediscovering these impossibility results, but demonstrating their practical inevitability in AMR prediction—and showing that the intuitive fix (excluding geography) does not escape the problem. This extends prior work on medical AI fairness. Liu et al.'s 2025 scoping review found that geography was examined in only 2.4% of 467 medical AI fairness studies. 19 Even when examined, geography was treated as a data source characteristic rather than a protected attribute requiring fairness evaluation. Our findings suggest this neglect is consequential. Implications for Model Development and Deployment Our findings suggest three approaches for achieving more equitable AMR prediction: 1. Regional models : Train and deploy separate models for distinct geographic contexts. This avoids the threshold selection problem by allowing each region to have its own appropriately-calibrated model. 2. Region-aware global models : If a single global model is required, incorporate geographic features explicitly and apply region-specific classification thresholds. This requires transparent documentation of regional performance characteristics. 3. Mandatory geographic fairness reporting : All AMR prediction models should report performance stratified by continent, country income level, and data representation. Models should not be deployed in regions where they were not validated. The counterintuitive implication is that achieving geographic fairness may require including geographic information, not excluding it—using region-specific thresholds that account for local prevalence. Regulatory Implications Current regulatory frameworks do not adequately address geographic fairness in medical AI. FDA guidance emphasises demographic subgroup performance reporting, but only 13% of FDA-cleared AI/ML devices report demographic subgroup performance, and geographic representation is not tracked. 20 – 21 The EU AI Act requires that training datasets "take into account... the characteristics or elements that are particular to the specific geographical, contextual, behavioural or functional setting," but implementation guidance remains under development. 22 WHO guidance acknowledges that "systems trained primarily on HIC data may not perform well for individuals in low- and middle-income settings" but provides no specific metrics or requirements. 23 – 24 No AMR-specific AI regulatory framework exists despite antimicrobial resistance ranking among WHO's top ten global health threats. We submit that deploying a single-threshold AMR prediction model globally without regional validation constitutes a foreseeable and preventable inequity. Regulatory frameworks should require geographic subgroup performance reporting with the same rigour currently applied to demographic attributes. Geography as a Protected Attribute A potential objection to treating geography as a protected attribute is that geography is theoretically modifiable—patients can relocate. However, at the point of clinical care, a patient's geographic context is effectively immutable: the infection has already occurred, the local resistance patterns have already shaped the pathogen's genotype, and the healthcare system must make treatment decisions based on available information. This parallels the treatment of other "contextual" protected attributes in fairness frameworks. Socioeconomic status is theoretically modifiable but fixed at the point of clinical encounter. 14 Insurance status determines access at the moment care is needed. 15 Geographic fairness follows the same logic: regardless of whether a patient could theoretically relocate, the prediction must be fair given their current geographic context. Strengths and Limitations Strengths include the large sample size (77,548 isolates), comprehensive geographic coverage (132 countries), direct test of the feature exclusion hypothesis using real genomic data, multi-organism validation, and identification of the underlying mechanism (geographic variation in resistance gene prevalence). Limitations include reliance on a single database (BV-BRC) with known geographic biases toward high-income countries. The genomic model achieved modest AUC (0.567) because it detects gene presence rather than specific resistance-conferring mutations; however, this limitation strengthens our argument, as higher-performing models would show even stronger geographic calibration shifts. Race/ethnicity data were not available, preventing analysis of how geographic and racial disparities interact. CONCLUSIONS Excluding geographic variables does not fix regional bias in AMR prediction. We demonstrate that a model trained exclusively on fluoroquinolone resistance genes—with geography explicitly excluded—still produced 19.6 percentage point sensitivity disparities across regions. The bias persists because resistance gene prevalence itself varies geographically, reflecting differences in antimicrobial pressure, horizontal gene transfer, and clonal expansion. These findings have immediate implications for AMR model development, deployment, and regulation. The intuitive fairness strategy of removing protected attributes does not work when the protected attribute is encoded in the biologically relevant features. Achieving geographic equity requires region-specific models or thresholds, mandatory geographic stratification in model evaluation, and recognition of geography as a protected attribute in medical AI fairness frameworks. As machine learning tools for AMR prediction move toward clinical deployment, we must ensure they do not systematically disadvantage patients in regions that already bear the highest burden of resistant infections. Abbreviations AMR: Antimicrobial resistance; ANOVA: Analysis of variance; AUC: Area under the receiver operating characteristic curve; BV-BRC: Bacterial and Viral Bioinformatics Resource Center; CI: Confidence interval; E. coli: Escherichia coli; K. pneumoniae: Klebsiella pneumoniae; OR: Odds ratio; PMQR: Plasmid-mediated quinolone resistance; QRDR: Quinolone resistance-determining region; S. aureus: Staphylococcus aureus Declarations Ethics approval and consent to participate Not applicable. This study used de-identified, publicly available data from the BV-BRC database and did not involve human participants, human tissue, or animals. Consent for publication Not applicable. Availability of data and materials The BV-BRC database is publicly available at https://www.bv-brc.org/. Processed datasets used in this analysis are available from the corresponding author upon reasonable request. Code availability Analysis code is available at: https://github.com/hayden-farquhar/amr-geographic-fairness Acknowledgements This study used data from the Bacterial and Viral Bioinformatics Resource Center (BV-BRC), which is funded by the National Institute of Allergy and Infectious Diseases (NIAID). We acknowledge the researchers and public health laboratories worldwide who contributed isolate data to BV-BRC, without which this analysis would not have been possible. AI-assisted tools were used during manuscript preparation for literature synthesis and editorial refinement. The author takes full responsibility for all content. Authors' contributions HF: Conceptualisation, Methodology, Software, Formal Analysis, Investigation, Data Curation, Writing – Original Draft, Writing – Review & Editing, Visualisation. The author read and approved the final manuscript. Competing interests The author declares that they have no competing interests. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. References Nsubuga M, et al. Generalizability of machine learning in predicting antimicrobial resistance in E. coli: A cross-continental validation study. BMC Genomics. 2024;25. 10.1186/s12864-024-10214-4 . Lv J, Wang X. Machine learning for antimicrobial resistance prediction: An updated systematic review. Technol Health Care. 2024. 10.3233/THC-240119 . Lewis JM, et al. Genomic epidemiology and antimicrobial resistance in Escherichia coli in Malawi. Microb Genom. 2023;9(6). 10.1099/mgen.0.001035 . Antimicrobial Resistance Collaborators. Global burden of bacterial antimicrobial resistance in 2019: a systematic analysis. Lancet. 2022;399(10325):629–55. Chouldechova A. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data. 2017;5(2):153–63. Hooper DC, Jacoby GA. Mechanisms of drug resistance: quinolone resistance. Ann N Y Acad Sci. 2015;1354:12–31. Robicsek A, Jacoby GA, Hooper DC. The worldwide emergence of plasmid-mediated quinolone resistance. Lancet Infect Dis. 2006;6(10):629–40. Redgrave LS, Sutton SB, Webber MA, Piddock LJ. Fluoroquinolone resistance: mechanisms, impact on bacteria, and role in evolutionary success. Trends Microbiol. 2014;22(8):438–45. Tang C, et al. Machine learning versus traditional approaches for antimicrobial resistance prediction: A systematic review and meta-analysis. Int J Antimicrob Agents. 2022;60(5):106684. Kim J, Greenberg DE, et al. Machine learning for antimicrobial resistance prediction: Current practice, limitations, and clinical perspective. Clin Microbiol Rev. 2022;35(3):e00179–21. Valavarasu R, Ren Y, et al. Genomic and phenotypic prediction of antimicrobial resistance using machine learning. Nat Sci Rep. 2025. 10.1038/s41598-025-14078-w . Davis JJ, et al. The PATRIC Bioinformatics Resource Center: Expanding data and analysis capabilities. Nucleic Acids Res. 2020;48(D1):D525–34. Pfizer. ATLAS Antimicrobial Surveillance Database. Nat Commun. 2022;13:2817. Obermeyer Z, Powers B, et al. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447–53. Seyyed-Kalantari L, et al. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med. 2021;27(12):2176–82. Celi LA, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities—A global review. PLOS Digit Health. 2022;1(3):e0000022. Kleinberg J, Mullainathan S, Raghavan M. Inherent trade-offs in the fair determination of risk scores. Proceedings of Innovations in Theoretical Computer Science (ITCS). 2017. Chouldechova A. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data. 2017;5(2):153–63. Liu M, et al. Algorithmic fairness in AI for healthcare: A scoping review of 467 studies. npj Digit Med. 2025;8:360. Muralidharan V et al. Demographic representation in FDA-approved AI/ML medical devices. npj Digit Med. 2024. FDA. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. Draft Guidance; January 2025. European Parliament. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). Official J Eur Union. 2024. World Health Organization. Ethics and Governance of Artificial Intelligence for Health. Geneva: WHO. 2021. ISBN:9789240029200. World Health Organization. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. Geneva: WHO. 2024. ISBN:9789240084759. Additional Declarations No competing interests reported. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8772201","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":585977315,"identity":"bc672682-e17f-4579-812f-af6052c8a3f4","order_by":0,"name":"Hayden Farquhar","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCklEQVRIiWNgGAWjYBACAyCWAGJ+EIcZiOXYGJgbGBjYCGuRbIBqMWZjYCRRS2IDIS3m7GcMb3zcwSDBP/vwNumCmnvpfewHGx/zlDHI84sdwKrFsifH2HLmGQYJiXNpZdIzjhXntvEkNhvznGMwnDk7AbvDDuSYSfO2MdQxnOExk+ZhS8htk2BsA4kkGNzGoeX8GzPpv20MEvJgLf8S0tkIarkBtIURqMUApIW3LSGBCC3Pii172yQkDM+wFVvz9iUYgvxiOOecBG6/nE/eeONnm42E3Bnmjbd5viXIy7cfPvjgTZmNPL80di0MDBywqAHHERxI4FAOAuwP4DbiUTUKRsEoGAUjGQAAlPtS+Lqtq/cAAAAASUVORK5CYII=","orcid":"","institution":"Independent Researcher","correspondingAuthor":true,"prefix":"","firstName":"Hayden","middleName":"","lastName":"Farquhar","suffix":""}],"badges":[],"createdAt":"2026-02-03 07:09:33","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8772201/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8772201/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":101942657,"identity":"e1c5b483-567d-4941-a50d-49410e32a830","added_by":"auto","created_at":"2026-02-05 09:31:03","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":447063,"visible":true,"origin":"","legend":"\u003cp\u003eGenomic Model Analysis Demonstrates Persistent Regional Bias Despite Excluding Geographic Variables. (A) Prediction score distributions by region—despite excluding geography from features, distributions shift by region (ANOVA p=1.45×10⁻⁴). (B) Regional calibration shift showing mean predicted probability with 95% CIs. (C) Sensitivity varies by region at any threshold, with Oceania consistently lowest. (D) The underlying cause: fluoroquinolone resistance gene prevalence varies significantly by region, with qnr genes showing nearly 7-fold variation between North America (1.0%) and Asia (6.8%). N=37,689 isolates.\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8772201/v1/90ed3207dcee80790c48394a.png"},{"id":102294753,"identity":"07ff6218-8ffa-437f-915d-14c38ac22479","added_by":"auto","created_at":"2026-02-10 09:55:54","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":420102,"visible":true,"origin":"","legend":"\u003cp\u003eGeographic and Income-Based Disparities in E. coli Ciprofloxacin Resistance. (A) Resistance rates by region and antibiotic, showing geographic variation is not specific to ciprofloxacin. (B) Resistance prevalence by World Bank income classification, demonstrating an inverse relationship between data contribution and resistance burden—high-income countries contribute most data but have lowest resistance. (C) Magnitude of geographic disparity exceeds 20 percentage points for all antibiotics tested. (D) The threshold selection trade-off: any global threshold creates either high miss rates (red) or high false positive rates (blue). N=39,859 isolates from 132 countries.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8772201/v1/efeb71522a2a2a1e7eda077c.png"},{"id":101942658,"identity":"8af86af6-9244-451a-9138-5af21ff1bd4f","added_by":"auto","created_at":"2026-02-05 09:31:03","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":622293,"visible":true,"origin":"","legend":"\u003cp\u003eThe Threshold Selection Impossibility (Genomic Validation Model). (A) Sensitivity curves for each geographic region from the genomic model trained on fluoroquinolone resistance genes. No single threshold achieves equal sensitivity across all regions. At threshold 0.30, sensitivity ranges from 61.4% (Oceania) to 81.0% (Africa)—a 19.6 percentage point gap. (B) The threshold selection dilemma: lower thresholds improve equity (12.8pp disparity at 0.25) while missing fewer cases, but higher thresholds paradoxically reduce disparity (4.5pp at 0.40) only because nearly all cases are missed across all regions. Green dashed line indicates the 10 percentage point clinical significance threshold. N=11,307 test isolates.\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8772201/v1/75063011d81dd5910918f324.png"},{"id":101942659,"identity":"bf0fa336-a4fd-40e0-a88e-57be2901affc","added_by":"auto","created_at":"2026-02-05 09:31:03","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":588120,"visible":true,"origin":"","legend":"\u003cp\u003eMulti-Organism Geographic Disparities in Ciprofloxacin Resistance. (A) Resistance rates by organism and region. (B) All three pathogens show geographic disparities exceeding 20 percentage points. (C) Different organisms show different regional rankings—E. coli highest in Asia, K. pneumoniae highest in South America—ruling out simple \"high-resistance regions\" explanations. This indicates organism-specific epidemiology driven by differential antimicrobial pressure, clonal expansion, and horizontal gene transfer. Total N=87,118 isolates across three species.\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8772201/v1/75b9834fc523bc5688fc56d2.png"},{"id":104397420,"identity":"f3b7a1d1-96df-42df-88db-ddd3a044e89d","added_by":"auto","created_at":"2026-03-11 11:47:52","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":3207458,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8772201/v1/d34e83ac-d29c-498c-85e1-1f696ae8a5b4.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Excluding Geographic Variables Does Not Fix Regional Bias in Machine Learning Antimicrobial Resistance Prediction: Analysis of 77,548 Isolates Across 132 Countries","fulltext":[{"header":"INTRODUCTION","content":"\u003cp\u003eMachine learning models for antimicrobial resistance (AMR) prediction face a well-documented problem: they perform worse in some geographic regions than others.\u003csup\u003e\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e Models trained predominantly on isolates from high-income countries show degraded accuracy when applied to low- and middle-income settings, where the burden of resistant infections is highest.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e This geographic performance gap is not merely a technical inconvenience\u0026mdash;it risks entrenching health disparities by providing inferior diagnostic support to populations who need it most.\u003c/p\u003e \u003cp\u003eThe intuitive solution is straightforward: remove geographic variables from the model. If a model cannot \"see\" where an isolate originated, the reasoning goes, it cannot discriminate based on geography. This approach aligns with established fairness strategies in other domains, where excluding protected attributes (race, sex, age) has been proposed as a path to equitable prediction.\u003csup\u003e\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eWe tested whether this intuitive solution works for AMR prediction. Using a large multi-country dataset, we trained a model using only fluoroquinolone resistance genes\u0026mdash;mechanistically relevant genomic features\u0026mdash;with geographic variables explicitly excluded. We then evaluated whether the model achieved equitable sensitivity across regions.\u003c/p\u003e \u003cp\u003eOur findings reveal a fundamental problem: excluding geographic variables does not eliminate geographic bias. The bias persists because it is encoded in the underlying biology. Resistance gene prevalence varies across regions due to differences in antimicrobial prescribing, horizontal gene transfer dynamics, and clonal expansion patterns.\u003csup\u003e\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e Any model that uses these biologically relevant features will inherit their geographic structure\u0026mdash;regardless of whether \"region\" appears in the feature set.\u003c/p\u003e \u003cp\u003eThis finding has immediate implications for how we develop, evaluate, and regulate AMR prediction tools. It suggests that the path to geographic fairness runs not through feature exclusion, but through region-specific models, region-aware thresholds, and mandatory geographic stratification in model evaluation.\u003c/p\u003e\n\u003ch3\u003eThe Geographic Bias Problem in AMR Prediction\u003c/h3\u003e\n\u003cp\u003eAntimicrobial resistance caused an estimated 4.95\u0026nbsp;million deaths globally in 2019, with mortality rates in Western sub-Saharan Africa more than four times higher than in Australasia.\u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e\u003c/sup\u003e Machine learning approaches for AMR prediction have proliferated rapidly, with meta-analyses reporting pooled AUCs of 0.78\u0026ndash;0.82.\u003csup\u003e9\u0026ndash;11\u003c/sup\u003e However, the geographic distribution of training data raises critical fairness concerns.\u003c/p\u003e \u003cp\u003eMajor genomic databases substantially overrepresent isolates from the United States and Western Europe.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e The landmark Valavarasu et al. 2025 study explicitly acknowledged \"significant underrepresentation of data from Sub-Saharan Africa\" and \"bias towards high-income countries.\"\u003csup\u003e11\u003c/sup\u003e Kim et al.'s systematic review concluded: \"An ML model trained on samples from one source may be ineffective in predicting the resistance phenotype for isolates from different environments, and universal ML models for AMR may not be feasible.\"\u003csup\u003e10\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThe most compelling evidence comes from Nsubuga et al.'s 2024 cross-continental validation study.\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e Models trained on English isolates showed dramatic performance degradation on African data: ciprofloxacin prediction accuracy dropped from 87% to 50%, and cefotaxime prediction plummeted from 92% to 45% when applied to Ugandan, Nigerian, and Tanzanian isolates. Paradoxically, ampicillin prediction inverted from 58% to 94%\u0026mdash;suggesting models may learn geographic signatures rather than true resistance determinants.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eWhy Feature Exclusion Seems Like a Solution\u003c/h2\u003e \u003cp\u003eThe algorithmic fairness literature offers a seemingly relevant precedent. Obermeyer et al.'s foundational 2019 study demonstrated that a healthcare cost-prediction algorithm systematically disadvantaged Black patients\u0026mdash;not because race was an explicit feature, but because the algorithm used healthcare costs as a proxy for health needs.\u003csup\u003e\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e Removing or adjusting such proxy variables improved equity. Seyyed-Kalantari et al. showed similar patterns in chest X-ray AI, where technical factors created artifactual prediction shifts that could be addressed through recalibration.\u003csup\u003e\u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eThis logic suggests that if we train AMR models using only mechanistically relevant features\u0026mdash;the actual genes that confer resistance\u0026mdash;rather than geographic proxies, we might achieve equitable prediction. The resistance genes are the \"true\" signal; geography is merely a confounder.\u003c/p\u003e \u003cp\u003eWhile algorithmic fairness researchers have long understood that excluding protected attributes does not guarantee equitable outcomes, this principle has not been empirically demonstrated in AMR prediction\u0026mdash;nor has it penetrated clinical or regulatory guidance for medical AI deployment.\u003c/p\u003e \u003cp\u003eWe hypothesised that this reasoning fails for AMR prediction because resistance gene prevalence itself varies geographically. The genes are not independent of geography; they carry geographic information. If so, excluding geographic variables would be ineffective\u0026mdash;the bias would persist through the genomic features themselves.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eStudy Objectives\u003c/h3\u003e\n\u003cp\u003eOur objectives were to: (1) quantify geographic variation in E. coli ciprofloxacin resistance across 132 countries; (2) test whether a model trained exclusively on genomic features\u0026mdash;with geographic variables excluded\u0026mdash;produces equitable sensitivity across regions; (3) identify the mechanism underlying any persistent disparities; and (4) evaluate the clinical and regulatory implications of our findings.\u003c/p\u003e"},{"header":"METHODS","content":"\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003eData Source\u003c/h2\u003e \u003cp\u003eThis retrospective cross-sectional study analysed antimicrobial susceptibility data from the Bacterial and Viral Bioinformatics Resource Center (BV-BRC), a comprehensive genomic database containing over 2\u0026nbsp;million bacterial genomes with associated antimicrobial resistance phenotypes.\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e Data were accessed via the BV-BRC API in December 2024.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003ePrimary Cohort: Geographic Prevalence Analysis\u003c/h3\u003e\n\u003cp\u003eThe Primary Cohort comprised E. coli isolates with ciprofloxacin susceptibility data and geographic metadata. Genome metadata (n\u0026thinsp;=\u0026thinsp;63,702) and ciprofloxacin susceptibility records (n\u0026thinsp;=\u0026thinsp;141,537) were extracted via paginated API queries. After filtering to isolates with definitive resistant or susceptible phenotypes and excluding Antarctica (n\u0026thinsp;=\u0026thinsp;1), the final dataset comprised 39,859 isolates from 132 countries across six geographic regions (Africa, Asia, Europe, North America, Oceania, South America), classified using BV-BRC's geographic_group schema. Country-level income classification followed World Bank 2024 categories.\u003c/p\u003e \u003cdiv id=\"Sec8\" class=\"Section2\"\u003e \u003ch2\u003eGenomic Cohort: Feature Exclusion Analysis\u003c/h2\u003e \u003cp\u003eThe Genomic Cohort tested whether excluding geographic variables eliminates regional bias. We queried specialty gene annotations for all E. coli genomes with ciprofloxacin susceptibility phenotypes and geographic metadata. Binary features were extracted for 10 established fluoroquinolone resistance determinants: quinolone resistance-determining region (QRDR) targets (gyrA, gyrB, parC, parE); plasmid-mediated quinolone resistance (PMQR) genes (qnrA/B/C/D/S); efflux pump components (acrAB, tolC, marRAB, oqxAB); and the aminoglycoside acetyltransferase variant aac(6')-Ib-cr.\u003csup\u003e6\u0026ndash;8\u003c/sup\u003e\u003c/p\u003e \u003cp\u003e \u003cb\u003eCritically, geographic region was explicitly excluded from the feature set.\u003c/b\u003e The model received only genomic information\u0026mdash;the same information a laboratory-based prediction system would have access to. The final Genomic Cohort comprised 37,689 isolates with resistance gene annotations.\u003c/p\u003e \u003c/div\u003e\n\u003ch3\u003eModel Training and Evaluation\u003c/h3\u003e\n\u003cp\u003eA logistic regression model was trained using only the 10 genomic features as predictors. The dataset was split 70%/30% for training/testing, stratified by resistance phenotype. We tested whether prediction scores differed by region using one-way ANOVA, and evaluated sensitivity disparities across regions at multiple classification thresholds (0.25, 0.30, 0.35, 0.40). Chi-squared tests assessed the association between genomic feature prevalence and geographic region.\u003c/p\u003e\n\u003ch3\u003eFairness Definition\u003c/h3\u003e\n\u003cp\u003eWe define fairness primarily in terms of equal sensitivity (true positive rates) across geographic regions. This choice reflects clinical consequences: a missed resistance prediction leads directly to inappropriate empiric therapy and potential treatment failure\u0026mdash;consequences disproportionately severe in resource-limited settings with fewer second-line options.\u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e We consider sensitivity disparities exceeding 10 percentage points to be clinically significant.\u003c/p\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003eStatistical Analysis\u003c/h2\u003e \u003cp\u003eRegional resistance prevalence was calculated with 95% confidence intervals using bootstrap resampling (n\u0026thinsp;=\u0026thinsp;1,000). Between-region comparisons used chi-squared tests with Bonferroni correction for multiple comparisons (15 pairwise comparisons, adjusted α\u0026thinsp;=\u0026thinsp;0.0033). Effect sizes were quantified using odds ratios with 95% confidence intervals. Analyses were performed using Python 3.10 with scipy 1.11 and statsmodels 0.14.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003eMulti-Organism Validation\u003c/h2\u003e \u003cp\u003eTo assess generalisability, we analysed geographic patterns in Klebsiella pneumoniae and Staphylococcus aureus ciprofloxacin resistance using the same methodology.\u003c/p\u003e \u003c/div\u003e"},{"header":"RESULTS","content":"\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003ePrimary Finding: Excluding Geographic Variables Does Not Eliminate Regional Bias\u003c/h2\u003e \u003cp\u003eThe logistic regression model, trained exclusively on 10 fluoroquinolone resistance genes with geographic variables explicitly excluded, produced prediction scores that varied significantly by region (one-way ANOVA: F\u0026thinsp;=\u0026thinsp;4.99, p\u0026thinsp;=\u0026thinsp;1.45\u0026times;10⁻⁴). Mean prediction scores ranged from 0.299 (Oceania) to 0.312 (Asia).\u003c/p\u003e \u003cp\u003eThis regional variation in prediction scores translated directly into sensitivity disparities at any fixed threshold. At a classification threshold of 0.30, sensitivity ranged from 61.4% (Oceania, n\u0026thinsp;=\u0026thinsp;83 resistant cases) to 81.0% (Africa, n\u0026thinsp;=\u0026thinsp;174 resistant cases)\u0026mdash;a \u003cb\u003e19.6 percentage point disparity\u003c/b\u003e (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). This means that a resistant isolate from Oceania has a 20% lower probability of being correctly identified compared to a resistant isolate from Africa, despite using the same model and threshold.\u003c/p\u003e \u003cp\u003eNo single threshold achieved equitable performance. At threshold 0.25, the sensitivity gap narrowed to 12.8 percentage points but remained clinically significant. At threshold 0.40, the gap compressed to 4.5 percentage points\u0026mdash;but only because nearly all cases were missed across all regions (\u0026gt;\u0026thinsp;3,000 cases missed). No operating point achieved both acceptable case detection and sensitivity variation below 10 percentage points (Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eGenomic Model Sensitivity by Threshold and Region\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThreshold\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAfrica\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAsia\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eEurope\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eN. America\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eOceania\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eS. America\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eGap (pp)\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e87.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e87.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e86.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e87.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e74.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e84.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e12.8\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e81.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e80.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e78.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e79.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e61.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e76.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e19.6\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e12.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e14.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e18.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e10.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e13.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e7.2\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.40\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e10.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e12.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e11.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e8.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e10.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c8\"\u003e \u003cp\u003e4.5\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eLogistic regression model trained on 10 fluoroquinolone resistance gene features with geographic variables excluded. N\u0026thinsp;=\u0026thinsp;11,307 test isolates. pp=percentage points.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003eMechanism: Resistance Gene Prevalence Varies Geographically\u003c/h2\u003e \u003cp\u003eThe persistent regional bias has a clear biological explanation: the genomic features themselves carry geographic information. Resistance gene prevalence varied significantly across regions for all tested determinants (χ\u0026sup2; tests, all p\u0026thinsp;\u0026lt;\u0026thinsp;0.001) (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe most striking example is plasmid-mediated quinolone resistance (PMQR) genes. Qnr gene prevalence ranged from 1.0% in North America to 6.8% in Asia\u0026mdash;a nearly 7-fold difference. This pattern reflects the documented epidemiology of PMQR dissemination: these resistance elements emerged in Asia and have spread globally at varying rates depending on antimicrobial selection pressure, horizontal gene transfer dynamics, and clonal expansion.\u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e \u003cp\u003eQRDR target genes showed similar geographic structure: gyrA prevalence ranged from 19.5% (North America) to 26.6% (Asia). Efflux pump genes (acrAB) ranged from 33.3% (North America) to 45.4% (Africa).\u003c/p\u003e \u003cp\u003eBecause these genes are mechanistically relevant\u0026mdash;they actually confer resistance\u0026mdash;any well-calibrated model must produce higher predicted probabilities in regions where they are more prevalent. The geographic bias is not an artifact of the model; it is encoded in the biology.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eGenomic Feature Prevalence by Geographic Region\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eResistance\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003egyrA\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eparC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eqnr\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eacrAB\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAsia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e13,331\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e45.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e26.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e19.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e6.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e44.0%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNorth America\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9,911\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e19.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e12.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.0%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e33.3%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEurope\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9,166\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e27.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e23.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e2.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e40.9%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAfrica\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,560\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e23.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e26.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e16.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e4.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e45.4%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSouth America\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,534\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e32.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e23.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e17.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e5.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e39.0%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOceania\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,187\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e23.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e15.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e1.3%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e44.3%\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eAll genomic features significantly associated with region (χ\u0026sup2; tests, all p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). qnr includes qnrA/B/C/D/S variants.\u003c/em\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec16\" class=\"Section2\"\u003e \u003ch2\u003eGeographic Context: Underlying Resistance Prevalence\u003c/h2\u003e \u003cp\u003eTo contextualise the genomic findings, we quantified regional ciprofloxacin resistance prevalence in the Primary Cohort (n\u0026thinsp;=\u0026thinsp;39,859 isolates from 132 countries). Resistance prevalence ranged from 16.8% (95% CI: 16.1\u0026ndash;17.5%) in North America to 44.1% (95% CI: 43.3\u0026ndash;44.9%) in Asia\u0026mdash;a 27.3 percentage point absolute difference (Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe odds of resistance in Asia were 3.90 times higher than in North America (95% CI: 3.67\u0026ndash;4.15; p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Of 15 pairwise regional comparisons, 13 remained statistically significant after Bonferroni correction.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eE. coli Ciprofloxacin Resistance by Geographic Region (Primary Cohort)\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"5\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRegion\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eN\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrevalence\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e95% CI\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eOR (95% CI)*\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAsia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e14,157\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e44.1%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e43.3\u0026ndash;44.9%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e3.90 (3.67\u0026ndash;4.15)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eSouth America\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,625\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e31.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e29.2\u0026ndash;33.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e2.28 (2.03\u0026ndash;2.56)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eEurope\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e9,817\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e26.7%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e25.8\u0026ndash;27.6%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.80 (1.68\u0026ndash;1.93)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAfrica\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e2,701\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e23.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e22.0-25.2%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.52 (1.37\u0026ndash;1.68)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eOceania\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e1,220\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e23.4%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e21.1\u0026ndash;25.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e1.51 (1.31\u0026ndash;1.75)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNorth America\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e10,339\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e16.8%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e16.1\u0026ndash;17.5%\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003eReference\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003e*Odds ratio compared to North America (reference). All comparisons p\u0026thinsp;\u0026lt;\u0026thinsp;0.001.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003eThe Threshold Selection Problem\u003c/h2\u003e \u003cp\u003eTo illustrate why this geographic prevalence heterogeneity creates an unavoidable threshold problem, we conducted a simulation analysis. We trained a Random Forest classifier using only geographic region as a predictor. This deliberately simplified model demonstrates a limiting case: when the only predictive feature is region, a well-calibrated model outputs the regional prevalence as its predicted probability.\u003c/p\u003e \u003cp\u003eAt any single global threshold, regions partition into discrete groups: those above the threshold (all isolates classified as resistant) and those below (all classified as susceptible). At a threshold of 0.30 (approximating global prevalence), only Asia and South America exceed the threshold. All 1,594 resistant isolates from Europe, Africa, Oceania, and North America\u0026mdash;44.3% of total resistant cases\u0026mdash;would be classified as susceptible (Table\u0026nbsp;\u003cspan refid=\"Tab4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThis simulation is not a recommended modelling strategy; it demonstrates the mathematical constraint that any model correctly calibrated to regional base rates faces when a single threshold is applied to populations with heterogeneous prevalence.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab4\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eSimulation: Impact of Global Threshold Selection on Regional Classification\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"4\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eThreshold\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eRegions Predicted Resistant\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003eRegions Predicted Susceptible\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eCases Missed\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAsia, S.Am, Europe, Africa, Oceania\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eNorth America\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e518 (14.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAsia, S.Am, Europe\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAfrica, Oceania, N.Am\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e788 (21.9%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.30\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAsia, South America\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eEurope, Africa, Oceania, N.Am\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,594 (44.3%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAsia\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eS.Am, Europe, Africa, Oceania, N.Am\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e1,777 (49.4%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e0.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003eNone\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003eAll regions\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e3,600 (100%)\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003cem\u003eResistant cases missed calculated from test set (n\u0026thinsp;=\u0026thinsp;11,918; 3,600 resistant). Simulation uses regional prevalence as predictions to illustrate the threshold selection problem.\u003c/em\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003eMulti-Organism Validation\u003c/h2\u003e \u003cp\u003eAnalysis of K. pneumoniae and S. aureus confirmed that geographic heterogeneity is not specific to E. coli or ciprofloxacin. K. pneumoniae showed the largest disparity: 70.1% resistance in South America versus 29.9% in Oceania (40.2 percentage point gap). S. aureus showed 37.0% resistance in Asia versus 15.6% in Oceania (21.4 percentage point gap).\u003c/p\u003e \u003cp\u003eImportantly, different organisms showed different regional rankings: E. coli highest in Asia, K. pneumoniae highest in South America. This pattern indicates that observed disparities reflect organism-specific epidemiology rather than simple \"high-resistance regions\" (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003eData Representation Inversely Related to Resistance Burden\u003c/h2\u003e \u003cp\u003eAnalysis by World Bank income classification revealed an inverse relationship between data representation and resistance burden. High-income countries contributed 58.5% of isolates but had 19.9% resistance prevalence. Lower-middle-income countries contributed only 6.8% of isolates but had 53.7% resistance prevalence. Models trained on available data are thus optimised for settings where resistance is lowest, while populations bearing the highest AMR burden are most poorly represented.\u003c/p\u003e \u003c/div\u003e"},{"header":"DISCUSSION","content":"\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\n \u003ch2\u003ePrincipal Findings\u003c/h2\u003e\n \u003cp\u003eThis study demonstrates that excluding geographic variables does not fix regional bias in AMR prediction models. A model trained exclusively on fluoroquinolone resistance genes\u0026mdash;with geographic variables explicitly excluded\u0026mdash;still produced 19.6 percentage point sensitivity disparities across regions. No classification threshold achieved both acceptable case detection and equitable performance.\u003c/p\u003e\n \u003cp\u003eThe mechanism is clear: resistance gene prevalence itself varies geographically. Qnr genes are nearly 7 times more prevalent in Asia than North America. Because these genes are mechanistically relevant\u0026mdash;they actually confer resistance\u0026mdash;any model using them must produce geographically varying predictions. The bias is encoded in the biology, not the feature set.\u003c/p\u003e\n \u003cp\u003eThis finding has important implications. It means that the standard fairness strategy of excluding protected attributes cannot work for AMR prediction. Geographic information permeates the biologically relevant features. The only paths to equity are region-specific models, region-aware thresholds, or fundamentally different approaches to global AMR prediction.\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\n \u003ch2\u003eClinical Implications\u003c/h2\u003e\n \u003cp\u003eConsider a patient presenting with suspected urinary tract infection requiring empiric antibiotic selection. An AI-assisted decision support tool predicts fluoroquinolone resistance probability. Our findings show that at any fixed threshold, patients in low-prevalence regions will more frequently receive \u0026quot;susceptible\u0026quot; predictions. When resistance is actually present, these patients experience treatment failure.\u003c/p\u003e\n \u003cp\u003eThe clinical consequences are asymmetric. A patient incorrectly classified as susceptible receives ineffective antibiotics\u0026mdash;leading to treatment failure, prolonged illness, and potentially death. A patient incorrectly classified as resistant receives broader-spectrum therapy that remains clinically effective, though suboptimal for antimicrobial stewardship. In AMR contexts, false negatives cause more immediate harm than false positives.\u003c/p\u003e\n \u003cp\u003eCritically, the regions with highest false negative rates (lowest sensitivity) are often those with the highest true resistance prevalence. Patients in Asia\u0026mdash;where 44% of isolates are resistant\u0026mdash;face the highest absolute burden of both resistance and model underperformance. This compounds existing health inequities.\u003c/p\u003e\n \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\n \u003ch2\u003eRelationship to Algorithmic Fairness Theory\u003c/h2\u003e\n \u003cp\u003eOur findings instantiate the fairness impossibility theorems established by Kleinberg et al. and Chouldechova in theoretical contexts.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e Their proofs demonstrate that when base rates differ between groups, no classifier can simultaneously achieve calibration, equal false positive rates, and equal false negative rates.\u003c/p\u003e\n \u003cp\u003eThe 27.3 percentage point prevalence gap between Asia and North America creates conditions where mathematical impossibility becomes clinical inevitability. Our contribution is not rediscovering these impossibility results, but demonstrating their \u003cem\u003epractical inevitability\u003c/em\u003e in AMR prediction\u0026mdash;and showing that the intuitive fix (excluding geography) does not escape the problem.\u003c/p\u003e\n \u003cp\u003eThis extends prior work on medical AI fairness. Liu et al.\u0026apos;s 2025 scoping review found that geography was examined in only 2.4% of 467 medical AI fairness studies.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e19\u003c/span\u003e\u003c/sup\u003e Even when examined, geography was treated as a data source characteristic rather than a protected attribute requiring fairness evaluation. Our findings suggest this neglect is consequential.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\n \u003ch2\u003eImplications for Model Development and Deployment\u003c/h2\u003e\n \u003cp\u003eOur findings suggest three approaches for achieving more equitable AMR prediction:\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e1. Regional models\u003c/strong\u003e: Train and deploy separate models for distinct geographic contexts. This avoids the threshold selection problem by allowing each region to have its own appropriately-calibrated model. \u003cspan\u003e\u003c/span\u003e\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e2. Region-aware global models\u003c/strong\u003e: If a single global model is required, incorporate geographic features explicitly and apply region-specific classification thresholds. This requires transparent documentation of regional performance characteristics.\u003c/p\u003e \u003cspan\u003e\n \u003cp\u003e\u003cstrong\u003e3. Mandatory geographic fairness reporting\u003c/strong\u003e: All AMR prediction models should report performance stratified by continent, country income level, and data representation. Models should not be deployed in regions where they were not validated.\u003c/p\u003e\n \u003c/span\u003e\n \u003cp\u003e\u003c/p\u003e\n \u003cp\u003eThe counterintuitive implication is that achieving geographic fairness may require \u003cem\u003eincluding\u003c/em\u003e geographic information, not excluding it\u0026mdash;using region-specific thresholds that account for local prevalence.\u003c/p\u003e\n \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e\n \u003ch2\u003eRegulatory Implications\u003c/h2\u003e\n \u003cp\u003eCurrent regulatory frameworks do not adequately address geographic fairness in medical AI. FDA guidance emphasises demographic subgroup performance reporting, but only 13% of FDA-cleared AI/ML devices report demographic subgroup performance, and geographic representation is not tracked.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e20\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e The EU AI Act requires that training datasets \u0026quot;take into account... the characteristics or elements that are particular to the specific geographical, contextual, behavioural or functional setting,\u0026quot; but implementation guidance remains under development.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e\u003c/p\u003e\n \u003cp\u003eWHO guidance acknowledges that \u0026quot;systems trained primarily on HIC data may not perform well for individuals in low- and middle-income settings\u0026quot; but provides no specific metrics or requirements.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e23\u003c/span\u003e\u0026ndash;\u003cspan class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e No AMR-specific AI regulatory framework exists despite antimicrobial resistance ranking among WHO\u0026apos;s top ten global health threats.\u003c/p\u003e\n \u003cp\u003eWe submit that deploying a single-threshold AMR prediction model globally without regional validation constitutes a foreseeable and preventable inequity. Regulatory frameworks should require geographic subgroup performance reporting with the same rigour currently applied to demographic attributes.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e\n \u003ch2\u003eGeography as a Protected Attribute\u003c/h2\u003e\n \u003cp\u003eA potential objection to treating geography as a protected attribute is that geography is theoretically modifiable\u0026mdash;patients can relocate. However, at the point of clinical care, a patient\u0026apos;s geographic context is effectively immutable: the infection has already occurred, the local resistance patterns have already shaped the pathogen\u0026apos;s genotype, and the healthcare system must make treatment decisions based on available information.\u003c/p\u003e\n \u003cp\u003eThis parallels the treatment of other \u0026quot;contextual\u0026quot; protected attributes in fairness frameworks. Socioeconomic status is theoretically modifiable but fixed at the point of clinical encounter.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e14\u003c/span\u003e\u003c/sup\u003e Insurance status determines access at the moment care is needed.\u003csup\u003e\u003cspan class=\"CitationRef\"\u003e15\u003c/span\u003e\u003c/sup\u003e Geographic fairness follows the same logic: regardless of whether a patient could theoretically relocate, the prediction must be fair given their current geographic context.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec27\" class=\"Section3\"\u003e\n \u003ch2\u003eStrengths and Limitations\u003c/h2\u003e\n \u003cp\u003eStrengths include the large sample size (77,548 isolates), comprehensive geographic coverage (132 countries), direct test of the feature exclusion hypothesis using real genomic data, multi-organism validation, and identification of the underlying mechanism (geographic variation in resistance gene prevalence).\u003c/p\u003e\n \u003cp\u003eLimitations include reliance on a single database (BV-BRC) with known geographic biases toward high-income countries. The genomic model achieved modest AUC (0.567) because it detects gene presence rather than specific resistance-conferring mutations; however, this limitation strengthens our argument, as higher-performing models would show even stronger geographic calibration shifts. Race/ethnicity data were not available, preventing analysis of how geographic and racial disparities interact.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e"},{"header":"CONCLUSIONS","content":"\u003cp\u003eExcluding geographic variables does not fix regional bias in AMR prediction. We demonstrate that a model trained exclusively on fluoroquinolone resistance genes\u0026mdash;with geography explicitly excluded\u0026mdash;still produced 19.6 percentage point sensitivity disparities across regions. The bias persists because resistance gene prevalence itself varies geographically, reflecting differences in antimicrobial pressure, horizontal gene transfer, and clonal expansion.\u003c/p\u003e \u003cp\u003eThese findings have immediate implications for AMR model development, deployment, and regulation. The intuitive fairness strategy of removing protected attributes does not work when the protected attribute is encoded in the biologically relevant features. Achieving geographic equity requires region-specific models or thresholds, mandatory geographic stratification in model evaluation, and recognition of geography as a protected attribute in medical AI fairness frameworks.\u003c/p\u003e \u003cp\u003eAs machine learning tools for AMR prediction move toward clinical deployment, we must ensure they do not systematically disadvantage patients in regions that already bear the highest burden of resistant infections.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eAMR: Antimicrobial resistance; ANOVA: Analysis of variance; AUC: Area under the receiver operating characteristic curve; BV-BRC: Bacterial and Viral Bioinformatics Resource Center; CI: Confidence interval; E. coli: Escherichia coli; K. pneumoniae: Klebsiella pneumoniae; OR: Odds ratio; PMQR: Plasmid-mediated quinolone resistance; QRDR: Quinolone resistance-determining region; S. aureus: Staphylococcus aureus\u003c/p\u003e\n"},{"header":"Declarations","content":"\u003ch2\u003eEthics approval and consent to participate\u003c/h2\u003e\n\u003cp\u003eNot applicable. This study used de-identified, publicly available data from the BV-BRC database and did not involve human participants, human tissue, or animals.\u003c/p\u003e\n\u003ch2\u003eConsent for publication\u003c/h2\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe BV-BRC database is publicly available at https://www.bv-brc.org/. Processed datasets used in this analysis are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCode availability\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAnalysis code is available at: https://github.com/hayden-farquhar/amr-geographic-fairness\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study used data from the Bacterial and Viral Bioinformatics Resource Center (BV-BRC), which is funded by the National Institute of Allergy and Infectious Diseases (NIAID). We acknowledge the researchers and public health laboratories worldwide who contributed isolate data to BV-BRC, without which this analysis would not have been possible.\u003c/p\u003e\n\u003cp\u003eAI-assisted tools were used during manuscript preparation for literature synthesis and editorial refinement. The author takes full responsibility for all content.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthors\u0026apos; contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHF: Conceptualisation, Methodology, Software, Formal Analysis, Investigation, Data Curation, Writing \u0026ndash; Original Draft, Writing \u0026ndash; Review \u0026amp; Editing, Visualisation. The author read and approved the final manuscript.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe author declares that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors.\u003c/p\u003e\n"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eNsubuga M, et al. Generalizability of machine learning in predicting antimicrobial resistance in E. coli: A cross-continental validation study. BMC Genomics. 2024;25. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1186/s12864-024-10214-4\u003c/span\u003e\u003cspan address=\"10.1186/s12864-024-10214-4\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLv J, Wang X. Machine learning for antimicrobial resistance prediction: An updated systematic review. Technol Health Care. 2024. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.3233/THC-240119\u003c/span\u003e\u003cspan address=\"10.3233/THC-240119\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLewis JM, et al. Genomic epidemiology and antimicrobial resistance in Escherichia coli in Malawi. Microb Genom. 2023;9(6). \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1099/mgen.0.001035\u003c/span\u003e\u003cspan address=\"10.1099/mgen.0.001035\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAntimicrobial Resistance Collaborators. Global burden of bacterial antimicrobial resistance in 2019: a systematic analysis. Lancet. 2022;399(10325):629\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChouldechova A. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data. 2017;5(2):153\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHooper DC, Jacoby GA. Mechanisms of drug resistance: quinolone resistance. Ann N Y Acad Sci. 2015;1354:12\u0026ndash;31.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRobicsek A, Jacoby GA, Hooper DC. The worldwide emergence of plasmid-mediated quinolone resistance. Lancet Infect Dis. 2006;6(10):629\u0026ndash;40.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRedgrave LS, Sutton SB, Webber MA, Piddock LJ. Fluoroquinolone resistance: mechanisms, impact on bacteria, and role in evolutionary success. Trends Microbiol. 2014;22(8):438\u0026ndash;45.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTang C, et al. Machine learning versus traditional approaches for antimicrobial resistance prediction: A systematic review and meta-analysis. Int J Antimicrob Agents. 2022;60(5):106684.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKim J, Greenberg DE, et al. Machine learning for antimicrobial resistance prediction: Current practice, limitations, and clinical perspective. Clin Microbiol Rev. 2022;35(3):e00179\u0026ndash;21.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eValavarasu R, Ren Y, et al. Genomic and phenotypic prediction of antimicrobial resistance using machine learning. Nat Sci Rep. 2025. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003e10.1038/s41598-025-14078-w\u003c/span\u003e\u003cspan address=\"10.1038/s41598-025-14078-w\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eDavis JJ, et al. The PATRIC Bioinformatics Resource Center: Expanding data and analysis capabilities. Nucleic Acids Res. 2020;48(D1):D525\u0026ndash;34.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePfizer. ATLAS Antimicrobial Surveillance Database. Nat Commun. 2022;13:2817.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eObermeyer Z, Powers B, et al. Dissecting racial bias in an algorithm used to manage the health of populations. Science. 2019;366(6464):447\u0026ndash;53.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSeyyed-Kalantari L, et al. Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nat Med. 2021;27(12):2176\u0026ndash;82.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCeli LA, et al. Sources of bias in artificial intelligence that perpetuate healthcare disparities\u0026mdash;A global review. PLOS Digit Health. 2022;1(3):e0000022.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKleinberg J, Mullainathan S, Raghavan M. Inherent trade-offs in the fair determination of risk scores. Proceedings of Innovations in Theoretical Computer Science (ITCS). 2017.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChouldechova A. Fair prediction with disparate impact: A study of bias in recidivism prediction instruments. Big Data. 2017;5(2):153\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu M, et al. Algorithmic fairness in AI for healthcare: A scoping review of 467 studies. npj Digit Med. 2025;8:360.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMuralidharan V et al. Demographic representation in FDA-approved AI/ML medical devices. npj Digit Med. 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFDA. Artificial Intelligence-Enabled Device Software Functions: Lifecycle Management and Marketing Submission Recommendations. Draft Guidance; January 2025.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEuropean Parliament. Regulation (EU) 2024/1689 laying down harmonised rules on artificial intelligence (AI Act). Official J Eur Union. 2024.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization. Ethics and Governance of Artificial Intelligence for Health. Geneva: WHO. 2021. ISBN:9789240029200.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWorld Health Organization. Ethics and Governance of Artificial Intelligence for Health: Guidance on Large Multi-Modal Models. Geneva: WHO. 2024. ISBN:9789240084759.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"antimicrobial resistance, machine learning, algorithmic fairness, geographic bias, health equity, Escherichia coli, fluoroquinolone resistance","lastPublishedDoi":"10.21203/rs.3.rs-8772201/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8772201/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eMachine learning models for antimicrobial resistance (AMR) prediction exhibit geographic performance disparities, with models trained on high-income country data underperforming in low- and middle-income settings. A seemingly straightforward solution is to exclude geographic variables from models. We tested whether this approach eliminates regional bias.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eWe analysed 77,548 bacterial isolates from the BV-BRC database across two cohorts. The Primary Cohort (n\u0026thinsp;=\u0026thinsp;39,859 Escherichia coli from 132 countries) quantified regional ciprofloxacin resistance prevalence. The Genomic Cohort (n\u0026thinsp;=\u0026thinsp;37,689 E. coli with fluoroquinolone resistance gene annotations) tested whether a model trained exclusively on genomic features\u0026mdash;with geographic variables explicitly excluded\u0026mdash;would produce equitable sensitivity across regions. We evaluated sensitivity disparities at multiple classification thresholds.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eDespite excluding all geographic variables, the genomic model produced significantly different prediction scores by region (ANOVA F\u0026thinsp;=\u0026thinsp;4.99, p\u0026thinsp;=\u0026thinsp;1.45\u0026times;10⁻⁴). At a threshold of 0.30, sensitivity ranged from 61.4% (Oceania) to 81.0% (Africa)\u0026mdash;a 19.6 percentage point disparity. No threshold achieved sensitivity variation below 10 percentage points across all regions. The underlying cause: resistance gene prevalence itself varies geographically (qnr genes: 1.0% North America vs 6.8% Asia, p\u0026thinsp;\u0026lt;\u0026thinsp;0.001), meaning any model using these biologically relevant features will inherit geographic structure.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e \u003cp\u003eExcluding geographic variables does not fix regional bias in AMR prediction because the bias is encoded in the underlying biology, not the model's feature set. Resistance genes vary geographically due to antimicrobial pressure, horizontal gene transfer, and clonal expansion. These findings demonstrate that geographic fairness requires region-specific models or thresholds, not simply removing location data. We recommend mandatory geographic stratification in model evaluation and recognition of geography as a protected attribute in medical AI fairness frameworks.\u003c/p\u003e","manuscriptTitle":"Excluding Geographic Variables Does Not Fix Regional Bias in Machine Learning Antimicrobial Resistance Prediction: Analysis of 77,548 Isolates Across 132 Countries","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-05 09:30:58","doi":"10.21203/rs.3.rs-8772201/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"f36f1663-78eb-42ee-a26f-95f5784f4976","owner":[],"postedDate":"February 5th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-02-22T16:39:21+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-05 09:30:58","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8772201","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8772201","identity":"rs-8772201","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}