{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T09:54:25Z","timestamp":1790848465139,"version":"4.1.0"},"update-to":[{"DOI":"10.12688\/wellcomeopenres.26477.1","type":"new_version","label":"New version","source":"publisher","updated":{"date-parts":[[2026,9,21]],"date-time":"2026-09-21T00:00:00Z","timestamp":1789948800000}}],"reference-count":24,"publisher":"F1000 Research Ltd","license":[{"start":{"date-parts":[[2026,9,21]],"date-time":"2026-09-21T00:00:00Z","timestamp":1789948800000},"content-version":"tdm","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"funder":[{"DOI":"10.13039\/100010269","name":"Wellcome Trust","doi-asserted-by":"publisher","award":["227486"],"award-info":[{"award-number":["227486"]}],"id":[{"id":"10.13039\/100010269","id-type":"DOI","asserted-by":"publisher"}]}],"content-domain":{"domain":["wellcomeopenresearch.org"],"crossmark-restriction":false},"short-container-title":["Wellcome Open Res"],"abstract":"<ns3:p>Background Antimicrobial resistance (AMR) is a growing global public health threat, particularly in low- and middle-income countries (LMICs), where delayed antimicrobial susceptibility testing (AST) and limited diagnostic capacity complicate timely treatment and antimicrobial stewardship (AMS). Although large-scale surveillance programmes routinely collect antimicrobial susceptibility data, these datasets remain underutilised for predictive analytics. Methods We developed a surveillance-driven machine learning (ML) framework using isolate-level data from African sites participating in the Pfizer Antimicrobial Testing Leadership and Surveillance (ATLAS) programme between 2019 and 2023. Seven antibiotic-specific Extreme Gradient Boosting (XGBoost) models were developed to predict antimicrobial susceptibility using routinely collected microbiological, demographic, clinical, and geographical metadata. Model development included structured data preprocessing, RandomOverSampler-based class balancing restricted to the training data, hyperparameter optimisation using stratified cross-validation, and evaluation on a held-out test dataset. A rule-based MIC interpretation system and interactive dashboard were developed to demonstrate implementation of the analytical workflow. Results The antibiotic-specific models demonstrated moderate predictive performance, with test accuracies ranging from 57% to 76%. Ceftazidime-Avibactam achieved the highest test accuracy (76%), followed by Gentamicin (65%) and Imipenem (64%), while Amikacin showed the lowest performance (57%). Feature-importance analysis identified bacterial species as consistently among the most influential predictors. In contrast, bacterial family, country of isolate collection, specimen source, clinical specialty, and demographic characteristics contributed to varying degrees across antibiotics. Performance was generally stronger for the more frequently represented susceptible class than for intermediate and resistant isolates. Conclusion Routinely collected AMR surveillance data can support antibiotic-specific ML predictions without requiring genomic sequencing or detailed patient-level clinical information. This study provides a proof-of-concept framework for surveillance-driven predictive analytics that could complement conventional AMR surveillance and AMS. External validation, calibration, prospective clinical evaluation, and implementation studies are required before routine deployment.<\/ns3:p>","DOI":"10.12688\/wellcomeopenres.26477.2","type":"journal-article","created":{"date-parts":[[2026,9,21]],"date-time":"2026-09-21T15:14:07Z","timestamp":1790003647000},"page":"393","update-policy":"https:\/\/doi.org\/10.12688\/wellcomeopenres.crossmark-policy","source":"Crossref","is-referenced-by-count":0,"title":["Surveillance-Driven Machine Learning for Prediction of Antimicrobial Susceptibility: An Explainable Modeling Framework using the Pfizer ATLAS Dataset (2004 - 2023)"],"prefix":"10.12688","volume":"11","author":[{"ORCID":"https:\/\/orcid.org\/0009-0003-8551-9134","authenticated-orcid":false,"given":"Raphael","family":"Mutua","sequence":"first","affiliation":[{"name":"Center for Microbiology Research Nairobi Nairobi County 54840-00200 Kenya Medical Research Institute"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"David","family":"Gichohi","sequence":"additional","affiliation":[{"name":"Department of Computer Science Nyeri Nyeri County 657-10100 Dedan Kimathi University of Technology"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0000-0002-3518-1142","authenticated-orcid":false,"given":"Samuel K.","family":"Ndegwa","sequence":"additional","affiliation":[{"name":"Center for Microbiology Research Nairobi Nairobi County 54840-00200 Kenya Medical Research Institute"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"ORCID":"https:\/\/orcid.org\/0009-0008-2927-6402","authenticated-orcid":false,"given":"Rahma O.","family":"Golicha","sequence":"additional","affiliation":[{"name":"Center for Microbiology Research Nairobi Nairobi County 54840-00200 Kenya Medical Research Institute"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Frida","family":"Njeru","sequence":"additional","affiliation":[{"name":"Center for Microbiology Research Nairobi Nairobi County 54840-00200 Kenya Medical Research Institute"}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Benson","family":"Kituku","sequence":"additional","affiliation":[{"name":"Department of Computer Science Nyeri Nyeri County 657-10100 Dedan Kimathi University of Technology"}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"2560","published-online":{"date-parts":[[2026,9,21]]},"reference":[{"key":"ref1","doi-asserted-by":"publisher","DOI":"10.70389\/PJPH.100023","article-title":"Emerging Trends in Antimicrobial Resistance: A Global Review of Surveillance, Challenges, and Strategies.","author":"R Ahmed","year":"2025","journal-title":"Premier J. Public Health."},{"key":"ref23","doi-asserted-by":"publisher","first-page":"e0319460","DOI":"10.1371\/journal.pone.0319460","article-title":"Machine learning for predicting antimicrobial resistance in critical and high-priority pathogens: A systematic review considering antimicrobial susceptibility tests in real-world healthcare settings.","volume":"20","author":"C Ardila","year":"2025 Feb 25","journal-title":"PLoS One."},{"key":"ref2","doi-asserted-by":"publisher","first-page":"3941-3956","DOI":"10.1080\/13501763.2024.2410919","article-title":"More than medicine: Antimicrobial resistance (AMR) as a social and political challenge that can be overcome.","volume":"31","author":"E Baekkeskov","year":"2024","journal-title":"J. Eur. Publ. Policy."},{"key":"ref3","doi-asserted-by":"publisher","first-page":"7-12","DOI":"10.1016\/j.ejim.2022.01.029","article-title":"Optimizing antibiotic therapies to reduce the risk of bacterial resistance.","volume":"99","author":"S Bassetti","year":"2022","journal-title":"Eur. J. Intern. Med."},{"key":"ref4","doi-asserted-by":"publisher","DOI":"10.48550\/arXiv.2208.04683","article-title":"Applying data technologies to combat AMR: Current status, challenges, and opportunities on the way forward.","author":"L Chindelevitch","year":"2022","journal-title":"(arXiv:2208.04683). arXiv."},{"key":"ref6","doi-asserted-by":"publisher","DOI":"10.3390\/microorganisms11041050","article-title":"Antimicrobial Resistance in the Global Health Network: Known Unknowns and Challenges for Efficient Responses in the 21st Century.","volume":"11","author":"T Coque","year":"2023","journal-title":"Microorganisms."},{"key":"ref26","doi-asserted-by":"publisher","first-page":"214-221","DOI":"10.4258\/hir.2021.27.3.214","article-title":"Machine Learning for Antibiotic Resistance Prediction: A Prototype Using Off-the-Shelf Techniques and Entry-Level Data to Guide Empiric Antimicrobial Therapy.","volume":"27","author":"G Feretzakis","year":"2021 Jul","journal-title":"Healthc. Inform. Res."},{"key":"ref7","doi-asserted-by":"publisher","first-page":"e2400259","DOI":"10.1002\/jobm.202400259","article-title":"Understanding the Evolution and Transmission Dynamics of Antibiotic Resistance Genes: A Comprehensive Review.","volume":"64","author":"A Hossain","year":"2024","journal-title":"J. Basic Microbiol."},{"key":"ref8","doi-asserted-by":"publisher","first-page":"55-74","DOI":"10.1146\/annurev-publhealth-051920-093141","article-title":"Innovations in Public Health Surveillance for Emerging Infections.","volume":"44","author":"P Jia","year":"2023","journal-title":"Annu. Rev. Public Health."},{"key":"ref9","doi-asserted-by":"publisher","first-page":"3589-3609","DOI":"10.2147\/IDR.S342753","article-title":"Antimicrobial Resistance Rates and Surveillance in Sub-Saharan Africa: Where Are We Now?","volume":"15","author":"S Kariuki","year":"2022","journal-title":"Infect. Drug Resist."},{"key":"ref10","doi-asserted-by":"publisher","first-page":"e0017921-e0017921","DOI":"10.1128\/cmr.00179-21","article-title":"Machine Learning for Antimicrobial Resistance Prediction: Current Practice, Limitations, and Clinical Perspective.","volume":"35","author":"J Kim","year":"2022","journal-title":"Clin. Microbiol. Rev."},{"key":"ref22","doi-asserted-by":"publisher","first-page":"140","DOI":"10.12688\/wellcomeopenres.15477.2","article-title":"Feasibility of informing syndrome-level empiric antibiotic recommendations using publicly available antibiotic resistance datasets","volume":"4","author":"Q Leclerc","year":"2020","journal-title":"Wellcome Open Res."},{"key":"ref12","doi-asserted-by":"publisher","first-page":"385-390","DOI":"10.2147\/CEOR.S130725","article-title":"A cost analysis of a broad-spectrum antibiotic therapy in the empirical treatment of health care-associated infections in cirrhotic patients.","volume":"9","author":"C Lucidi","year":"2017","journal-title":"Clinicoecon. Outcomes Res."},{"key":"ref24","doi-asserted-by":"publisher","first-page":"2865-2882","DOI":"10.3233\/THC-240119","article-title":"Machine learning-based antibiotic resistance prediction models: An updated systematic review and meta-analysis.","volume":"32","author":"G Lv","year":"2024","journal-title":"Technol. Health Care."},{"key":"ref13","doi-asserted-by":"publisher","first-page":"629-655","DOI":"10.1016\/S0140-6736(21)02724-0","article-title":"Global burden of bacterial antimicrobial resistance in 2019: A systematic analysis.","volume":"399","author":"C Murray","year":"2022","journal-title":"Lancet."},{"key":"ref14","doi-asserted-by":"publisher","first-page":"103743","DOI":"10.1016\/j.eclinm.2025.103743","article-title":"Machine learning for predicting clinical outcomes of hospitalised children: A systematic review of applications in low- and middle-income countries.","volume":"91","author":"W Nkhono","year":"2026","journal-title":"EClinicalMedicine."},{"key":"ref15","doi-asserted-by":"publisher","first-page":"461-474","DOI":"10.56557\/jomahr\/2025\/v10i210043","article-title":"Real-Time Genomic Surveillance for Multidrug-Resistant Pathogens in Low-Resource Settings: A Narrative Review of Challenges, Innovations, and Implementation Frameworks.","volume":"10","author":"M Olorunkosebi","year":"2025","journal-title":"J. Med. Health Res."},{"key":"ref16","doi-asserted-by":"publisher","first-page":"1330-1351","DOI":"10.1002\/bcp.70011","article-title":"Factors influencing physicians\u2019 antimicrobial prescribing decisions: A systematic review of qualitative studies.","volume":"91","author":"S Reali","year":"2025","journal-title":"Br. J. Clin. Pharmacol."},{"key":"ref17","doi-asserted-by":"publisher","first-page":"325-334","DOI":"10.1093\/bioinformatics\/btab681","article-title":"Prediction of antimicrobial resistance based on whole-genome sequencing and machine learning.","volume":"38","author":"Y Ren","year":"2022","journal-title":"Bioinformatics."},{"key":"ref18","doi-asserted-by":"publisher","first-page":"147-160","DOI":"10.1080\/14787210.2021.1951705","article-title":"Antimicrobial resistance in low- and middle-income countries: Current status and future directions.","volume":"20","author":"G Sulis","year":"2022","journal-title":"Expert Rev. Anti-Infect. Ther."},{"key":"ref19","doi-asserted-by":"publisher","first-page":"1795-1805","DOI":"10.7150\/thno.19217","article-title":"Current and emerging techniques for antibiotic susceptibility tests.","volume":"7","author":"K Syal","year":"2017","journal-title":"Theranostics."},{"key":"ref25","doi-asserted-by":"publisher","first-page":"30509","DOI":"10.1038\/s41598-025-14078-w","article-title":"Prediction of antibiotic resistance from antibiotic susceptibility testing results from surveillance data using machine learning.","volume":"15","author":"S Valavarasu","year":"2025","journal-title":"Sci. Rep."},{"key":"ref21","doi-asserted-by":"publisher","DOI":"10.3390\/antibiotics14060543","article-title":"What Are the Drivers Triggering Antimicrobial Resistance Emergence and Spread? Outlook from a One Health Perspective.","volume":"14","author":"Z Ye","year":"2025","journal-title":"Antibiotics."},{"key":"ref27","doi-asserted-by":"publisher","first-page":"1143-1152","DOI":"10.1038\/s41591-019-0503-6","article-title":"Personal clinical history predicts antibiotic resistance of urinary tract infections.","volume":"25","author":"I Yelin","year":"2019","journal-title":"Nat. Med."}],"container-title":["Wellcome Open Research"],"original-title":[],"language":"en","link":[{"URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2\/xml","content-type":"application\/xml","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2\/pdf","content-type":"application\/pdf","content-version":"vor","intended-application":"text-mining"},{"URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2\/iparadigms","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,10,1]],"date-time":"2026-10-01T08:59:10Z","timestamp":1790845150000},"score":1,"resource":{"primary":{"URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,9,21]]},"references-count":24,"URL":"https:\/\/doi.org\/10.12688\/wellcomeopenres.26477.2","relation":{"has-version":[{"id-type":"doi","id":"10.12688\/wellcomeopenres.26477.1","asserted-by":"subject"}],"has-review":[{"id-type":"doi","id":"10.21956\/wellcomeopenres.30447.r169686","asserted-by":"subject"},{"id-type":"doi","id":"10.21956\/wellcomeopenres.30447.r169686","asserted-by":"object"}]},"ISSN":["2398-502X"],"issn-type":[{"value":"2398-502X","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,9,21]]},"assertion":[{"value":"Indexed","URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2#article-reports","order":0,"name":"referee-status","label":"Referee status","group":{"name":"current-referee-status","label":"Current Referee Status"}},{"value":"10.21956\/wellcomeopenres.30447.r169686, Daniel Edbert, Atma Jaya Catholic University of Indonesia, Jakarta, 29 Sep 2026, version 2, indexed","URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2#referee-response-169686","order":0,"name":"referee-response-169686","label":"Referee Report","group":{"name":"article-reports","label":"Article Reports"}},{"value":"10.21956\/wellcomeopenres.29157.r159846, Timothy Miles Rawson, Imperial College London, London, 20 Jul 2026, version 1, 1 approved, 1 approved with reservations, 1 not approved","URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v1#referee-response-159846","order":1,"name":"referee-response-159846","label":"Referee Report","group":{"name":"article-reports","label":"Article Reports"}},{"value":"<b>Raphael Mutua<\/b>; \n<i>Posted: 30 Aug 2026<\/i>; We sincerely thank the reviewer for the careful and detailed assessment of our manuscript. We appreciate the reviewer\u2019s recognition of the importance of the research question and the potential value of using routinely collected antimicrobial resistance surveillance data for predictive modelling. We also acknowledge that the original manuscript did not provide sufficient methodological detail to permit a full assessment of the study\u2019s validity, reproducibility, and generalisability. In response, we have extensively revised the manuscript, with particular emphasis on the Methods, Results, Discussion, and Conclusion. The revised version now provides a clearer account of data selection, preprocessing, breakpoint interpretation, predictor variables, model development, internal validation, class-imbalance management, hyperparameter optimisation, performance evaluation, model interpretability, and study limitations. We have also moderated claims regarding clinical application and reframed the dashboard as a research prototype rather than a validated decision-support system. Our point-by-point responses are provided below. \n<b>Comment 1: Insufficient methodological reporting<\/b> \n<i>The Methods section lacks the detail required to understand, reproduce, or critically assess the study. Given that this is a prediction modelling study, the authors should follow established reporting standards such as the TRIPOD guidelines for predictive modelling studies.<\/i> \n<i>Specifically, the manuscript should clearly describe data extraction and inclusion\/exclusion criteria, organisms and antimicrobials included, data cleaning and preprocessing, missing data, feature engineering, model development, data partitioning, validation, performance metrics, and statistical analyses.<\/i> \n<b>Response<\/b> We agree that the original Methods section did not provide sufficient detail regarding the prediction-model development workflow. The Methods have now been comprehensively expanded and reorganised. The revised manuscript describes:  the retrospective study design and Pfizer ATLAS data source; restriction of the analysis to African isolates collected between 2019 and 2023; inclusion of organisms listed in the WHO Bacterial Priority Pathogens List; criteria used to select antibiotics for model development; exclusion of records with missing predictor or outcome information; exclusion of variables with extensive missingness or no informative observations; the complete set of candidate predictor variables; one-hot encoding of categorical predictors; label encoding of susceptibility outcomes; stratified 80:20 training and hold-out test partitioning; restriction of RandomOverSampler to the training dataset to prevent data leakage; hyperparameter optimisation using GridSearchCV and three-fold stratified cross-validation; the complete hyperparameter search grid and model-selection criterion; retraining of the selected model on the balanced training data; evaluation on an independent, unaltered test dataset; and the performance metrics and feature-importance approach used.  We also revised the organisation of the manuscript so that methodological information previously appearing only in the Results is now presented in the Methods. These changes were informed by established prediction-model reporting principles, including the need to distinguish clearly between model development, internal validation, and independent test evaluation.  \n<b>Comment 2: Lack of clarity regarding susceptibility testing methods<\/b> \n<i>The ATLAS dataset contains susceptibility data generated using different interpretive frameworks and laboratory methodologies over time. The manuscript does not adequately explain how differences in susceptibility testing standards were handled.<\/i> \n<i>Were CLSI and EUCAST breakpoints both present? How were differences between laboratory methods and breakpoint systems managed? Was the interpretive standard included as a model feature? Were analyses stratified by breakpoint methodology? Could changes in breakpoint definitions over time have influenced model performance?<\/i> \n<b>Response<\/b> We have revised the Methods to clarify the susceptibility interpretations used in the study. The Pfizer ATLAS programme reports MIC measurements together with standardised susceptibility interpretations derived using Clinical and Laboratory Standards Institute breakpoints. Although the ATLAS platform also permits users to explore susceptibility using European Committee on Antimicrobial Susceptibility Testing breakpoints, the downloaded dataset used for the present analysis contained the standardised CLSI-derived susceptibility categories. All model outcomes in this study were therefore based on those CLSI-derived categories. Because a single interpretive framework was used for the analysed outcomes, CLSI and EUCAST categories were not combined or harmonised, breakpoint methodology was not included as a predictor, and analyses were not stratified by interpretive system. We have also clarified that the study was restricted to isolates collected between 2019 and 2023 partly to improve contemporary epidemiological relevance and reduce the influence of older resistance patterns. We nevertheless acknowledge that breakpoint definitions and laboratory practices may change over time and that residual heterogeneity in multicentre surveillance data remains a possible source of bias. This issue is now recognised in the discussion of dataset limitations and generalisability.  \n<b>Comment 3: Rule-based MIC interpretation tool is insufficiently described<\/b> \n<i>One of the stated objectives is development of a rule-based tool to convert MIC values into susceptibility categories. However, there is very little detail describing the underlying decision rules, breakpoint standards, handling of discrepancies, breakpoint updates, validation, and how the tool was used within the modelling workflow.<\/i> \n<b>Response<\/b> The description of the rule-based MIC interpretation tool has been expanded substantially. The revised Methods now clarify that the tool did not derive new clinical breakpoints. Instead, it operationalised the CLSI-derived susceptibility interpretations contained within the Pfizer ATLAS dataset. For each organism-antibiotic combination, MIC values and their corresponding susceptibility categories were extracted and represented as structured JavaScript Object Notation lookup rules. Within the dashboard, the user selects the organism and antibiotic and enters an MIC value, after which the application returns the corresponding susceptible, intermediate, or resistant category. We have also clarified that the rule-based component and the machine-learning component perform different functions. The rule-based tool translates an observed MIC into a susceptibility category, whereas the machine-learning models predict susceptibility from routinely collected epidemiological and isolate-level metadata when MIC information is not used as a predictor. No formal independent validation of the dashboard-based interpretive tool was conducted. We therefore revised the manuscript to describe it as a research prototype that demonstrates the operationalisation of existing interpretations rather than as a newly validated diagnostic or clinical tool. The manuscript now explicitly states that it should not replace established laboratory AST interpretation procedures.  \n<b>Comment 4: Machine-learning model development lacks transparency<\/b> \n<i>The manuscript should clearly specify which algorithms were developed and evaluated, hyperparameter optimisation procedures, feature selection methods, data partitioning strategies, internal validation approaches, and measures taken to prevent data leakage.<\/i> \n<b>Response<\/b> We agree that the original description of the model-development process was insufficient. The revised Methods now provide a detailed account of the modelling workflow. We clarify that Random Forest and baseline XGBoost approaches were explored during model development. The final models were antibiotic-specific XGBoost classifiers. XGBoost was selected because of its suitability for structured tabular data and its capacity to represent nonlinear relationships and interactions among categorical surveillance variables. For each antibiotic, observations were divided into stratified training and testing sets using an 80:20 split. RandomOverSampler was applied only to the training set to address class imbalance, while the test set retained the original class distribution. This separation was used to prevent synthetic or duplicated training observations from entering the test dataset. Hyperparameter optimisation was performed using GridSearchCV with three-fold stratified cross-validation. The search grid included maximum tree depths of 3, 5, and 7; 50 and 100 boosting trees; learning rates of 0.05 and 0.10; and subsampling fractions of 0.8 and 1.0. The model achieving the highest weighted F1-score across the cross-validation folds was selected and retrained on the balanced training set before evaluation on the independent test set. No separate feature-selection procedure was applied. All prespecified routinely collected surveillance predictors that met data-quality requirements were retained. This has now been stated explicitly.  \n<b>Comment 5: Unclear specification of predictor variables<\/b> \n<i>The manuscript refers to predictor variables and key features but provides limited information regarding the actual metadata used. The authors should provide a complete description of all candidate predictors, variables included in the final model, their definitions and sources, and the rationale for their inclusion.<\/i> \n<b>Response<\/b> The revised Methods now provide a complete list of the candidate and final predictor variables used in the antibiotic-specific models:  bacterial species; bacterial family; country of isolation; patient sex; age group; clinical specialty; and specimen source.  These variables were obtained directly from the Pfizer ATLAS surveillance dataset. They were selected because they are routinely recorded in microbiological surveillance systems and therefore support the study\u2019s aim of evaluating a modelling framework that does not depend on genomic sequencing or detailed electronic health-record data. All predictors were categorical and were converted into numerical representations using one-hot encoding. No numerical scaling or normalisation was required. Laboratory testing methodology and interpretive standard were not included as predictor variables because the model outcomes used the standardised CLSI-derived susceptibility classifications supplied in the analysed dataset. We have removed nonspecific wording such as \u201camong others\u201d and now state the full predictor set consistently throughout the manuscript.  \n<b>Comment 6: Representativeness and generalisability of the ATLAS dataset<\/b> \n<i>The manuscript should provide a more critical discussion of the representativeness of ATLAS data and the extent to which findings are generalisable.<\/i> \n<b>Response<\/b> We agree that the representativeness of the dataset and the limits of generalisability required greater attention. The revised Discussion now explains that ATLAS is a multicountry surveillance programme based on participating centres and should not be interpreted as a population-representative sample of all infections occurring within the included countries. We acknowledge that participating laboratories may disproportionately represent tertiary, urban, or better-resourced facilities and that countries and centres were not evenly represented. South Africa, Nigeria, and Morocco contributed the largest numbers of observations in the analytical dataset. Referral patterns, local testing practices, isolate selection, healthcare access, and site participation may therefore influence the observed susceptibility distributions and model performance. We also clarify that the models were intentionally developed using recent African isolates as a regional proof of concept. Their performance should not be assumed to generalise automatically to other geographic regions, healthcare settings, time periods, or laboratory networks. External validation in independently collected and appropriately harmonised datasets is required before broader application. Claims regarding broad applicability have consequently been moderated throughout the Introduction, Discussion, and Conclusion.  \n<b>Comment 7: Justification of organism and susceptibility-test selection<\/b> \n<i>The manuscript does not sufficiently describe which organisms were included, why those organisms were selected, and why only a relatively small number of susceptibility tests appear to have been used.<\/i> \n<b>Response<\/b> The revised Methods now state that the analysis was restricted to bacterial species included in the WHO Bacterial Priority Pathogens List because these organisms represent major priorities for global AMR surveillance and public-health action. The revised Results describe the principal organisms represented in the analytical dataset, including \n<i>Escherichia coli<\/i>, \n<i>Klebsiella pneumoniae<\/i>, \n<i>Pseudomonas aeruginosa<\/i>, \n<i>Staphylococcus aureus<\/i>, and \n<i>Acinetobacter baumannii<\/i>. Antibiotic-specific models were developed for seven antibiotics with sufficient susceptibility observations to support model training and hold-out evaluation: gentamicin, ceftazidime-avibactam, amikacin, imipenem, piperacillin-tazobactam, ceftolozane-tazobactam, and ceftazidime. Antibiotic selection was therefore driven by data availability and suitability for model development rather than an intention to provide comprehensive coverage of all clinically relevant antimicrobial agents. &nbsp;  \n<b>Comments on the Results<\/b> \n<b>Result Comment 1: Model performance is insufficiently described<\/b> \n<i>Model performance metrics are not described in enough detail.<\/i> \n<b>Response<\/b> The Results have been expanded to distinguish training from hold-out test performance across the seven antibiotic-specific models and to describe variation in predictive performance between antibiotics. Training accuracy ranged from 73% to 86%, while test accuracy ranged from 57% to 76%. The Methods now specify accuracy, precision, recall, F1-score, and confusion matrices as the evaluation measures used in the modelling workflow. The revised Results also interpret model performance in the context of the substantial class imbalance observed across the antibiotic-specific datasets. Because class-specific precision, recall, and F1-score values were not retained as complete numerical outputs for all seven definitive models, these metrics are not tabulated in the main manuscript. &nbsp;  \n<b>Result Comment 2: Confidence intervals and uncertainty estimates<\/b> \n<i>Confidence intervals and uncertainty estimates should be provided.<\/i> \n<b>Response<\/b> We agree that uncertainty estimates would strengthen the interpretation of model performance. Confidence intervals were not generated during the original model-development workflow, and we have not retrospectively introduced estimates that were not produced by the analysis. We have therefore avoided overstating differences between models and now present the reported performance measures descriptively. The absence of confidence intervals and resampling-based uncertainty estimates is acknowledged as a limitation. Future evaluations should use approaches such as bootstrap resampling or repeated cross-validation to quantify uncertainty around performance metrics. Because this request was not fully addressed analytically, we have stated this transparently rather than implying that confidence intervals were calculated. &nbsp;  \n<b>Result Comment 3: Comparative performance across organisms, antimicrobials, and regions<\/b> \n<i>Comparative performance across organisms, antimicrobials, and geographic regions would strengthen interpretation.<\/i> \n<b>Response<\/b> We thank the reviewer for this suggestion. The revised Results provide clearer comparisons across antibiotic-specific models, including variation in training and testing accuracy and the influence of class imbalance and sample size. However, separate models stratified by organism and geographic region were not developed as part of the original analytical plan. Conducting such subgroup analyses with the present data could also produce unstable estimates where organism-country-antibiotic combinations contain limited observations. We have therefore not added post hoc subgroup performance claims that were not supported by the completed analysis. Instead, the revised Discussion identifies organism-specific and geographic validation as priorities for future work and cautions against assuming uniform performance across countries, organisms, or clinical settings.  \n<b>Result Comment 4: Susceptibility methodology as a predictor<\/b> \n<i>It remains unclear whether susceptibility testing methodology was included as a predictor variable and whether its inclusion affected performance.<\/i> \n<b>Response<\/b> We have clarified that susceptibility-testing methodology and breakpoint standard were not included as model predictors. The analysed outcomes were the standardised CLSI-derived susceptibility categories provided in the downloaded ATLAS dataset. Consequently, no comparison of model performance with and without breakpoint methodology was conducted. This decision and its rationale are now stated explicitly in the Methods.  \n<b>Result Comment 5: Feature importance and explainability<\/b> \n<i>Additional information on feature importance and explainability outputs is required.<\/i> \n<b>Response<\/b> We thank the reviewer for this valuable suggestion. The revised manuscript now provides additional methodological and interpretive information on model explainability. The Methods specify that predictor importance was quantified using the gain-based feature-importance metric implemented in XGBoost. Gain reflects the improvement in the model objective associated with splits involving a predictor and was used to identify variables contributing most strongly to susceptibility classification. A dedicated Results subsection now reports that bacterial species, bacterial family, and country of isolation were among the most influential predictors, while specimen source and clinical specialty also contributed to prediction. Patient age group and sex generally showed lower relative importance. The Discussion has been expanded to interpret these findings biologically and epidemiologically. Differences between bacterial taxa are plausibly related to intrinsic susceptibility and organism-specific resistance mechanisms, while geographic importance may reflect variation in antimicrobial consumption, healthcare systems, infection-prevention practices, diagnostic capacity, and circulating resistance determinants. We have also clarified the limitations of this approach. Gain-based importance provides a global assessment of predictor contribution but does not explain individual predictions and may be affected by correlations among encoded variables. Future studies should incorporate approaches such as SHAP analysis to provide both global and local explanations of model behaviour.  \n<b>Conclusion<\/b> \n<i>The manuscript lacks the methodological transparency and reporting detail required for proper scientific evaluation. Greater clarity regarding data handling, breakpoint interpretation, predictor variables, model development, validation procedures, and dataset representativeness is essential.<\/i> \n<b>Response<\/b> We thank the reviewer for this candid and constructive assessment. The manuscript has undergone extensive revision in direct response to these concerns. The Methods now provide a substantially more detailed account of data selection, preprocessing, missing-data handling, predictor encoding, model development, class balancing, hyperparameter optimisation, internal validation, hold-out testing, evaluation metrics, and feature importance. We have clarified the use of CLSI-derived susceptibility categories, expanded the description of the rule-based MIC interpretation component, provided a complete list of predictor variables, and strengthened discussion of ATLAS representativeness and geographic generalisability. The dashboard is now framed solely as a research prototype, and statements regarding clinical utility have been moderated. The revised manuscript consistently presents the framework as a proof of concept requiring external validation, calibration assessment, uncertainty estimation, prospective evaluation, and implementation research before clinical use. We believe these revisions materially improve the transparency and scientific interpretability of the study. We also acknowledge that some limitations remain, particularly the absence of external validation, calibration analysis, confidence intervals, and complete local explainability outputs. These limitations are now stated explicitly rather than being obscured by claims of clinical readiness or broad generalisability. We sincerely thank the reviewer again for the detailed critique. The comments prompted major improvements in the reporting, positioning, and interpretation of the study. We hope that the revised manuscript now provides sufficient information to permit a more complete assessment of its methods, findings, and limitations.","URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2#referee-comment-6711","order":2,"name":"referee-comment-6711","label":"Referee Comment","group":{"name":"article-reports","label":"Article Reports"}},{"value":"10.21956\/wellcomeopenres.29157.r159850, Olga Perovic, National Institute for Communicable Diseases, Johannesburg, 20 Jul 2026, version 1, 1 approved, 1 approved with reservations, 1 not approved","URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v1#referee-response-159850","order":3,"name":"referee-response-159850","label":"Referee Report","group":{"name":"article-reports","label":"Article Reports"}},{"value":"<b>Raphael Mutua<\/b>; \n<i>Posted: 30 Aug 2026<\/i>; We sincerely thank the reviewer for the careful assessment of our manuscript and for recognising the relevance of the study to antimicrobial resistance surveillance, particularly in resource-limited settings. We appreciate the positive evaluation of the study design, analytical approach, and overall scientific contribution. We have also carefully considered the limitations and recommendations highlighted in the report and revised the manuscript accordingly. \n<b>Comment 1: Data representativeness and potential surveillance bias<\/b> \n<i>Surveillance data may overrepresent tertiary or urban centres.<\/i> \n<b>Response:<\/b> We agree that the Pfizer ATLAS dataset may not be fully representative of routine clinical populations in the included countries. Participating centres may disproportionately reflect tertiary, urban, or better-resourced facilities, and country-level contributions were uneven. The Discussion has therefore been expanded to acknowledge potential site-selection, referral, testing, and sampling biases. We also clarify that the findings should not be interpreted as population-representative estimates and that external validation is required before broader generalisation. \n<b>Comment 2: Class imbalance<\/b> \n<i>More susceptible than resistant isolates may affect model performance.<\/i> \n<b>Response:<\/b> We thank the reviewer for highlighting this important limitation. The revised Methods now provide a more detailed explanation of the class-imbalance strategy. RandomOverSampler was applied exclusively to the training dataset, while the independent test dataset retained the original susceptibility distribution to prevent data leakage. The Results and Discussion now also emphasise that performance was generally stronger for susceptible isolates and weaker for intermediate and resistant classes. We acknowledge that class imbalance remained an important limitation despite the use of oversampling. \n<b>Comment 3: Limited patient-level clinical data<\/b> \n<i>The lack of detailed clinical information limits individualised prediction.<\/i> \n<b>Response:<\/b> We agree. The models were intentionally developed using routinely collected surveillance variables to evaluate feasibility in settings where comprehensive electronic health records and genomic data may be unavailable. However, this design limits patient-specific prediction and does not capture important clinical factors such as prior antimicrobial exposure, comorbidities, recent hospitalisation, treatment history, or previous microbiological results. This limitation is now more explicitly discussed, and the inclusion of richer clinical variables is identified as a priority for future model development. \n<b>Comment 4: Temporal changes in resistance<\/b> \n<i>Resistance patterns change over time and models require regular updating.<\/i> \n<b>Response:<\/b> We agree that temporal drift is an important consideration. The analysis was restricted to African isolates collected between 2019 and 2023 to improve contemporary relevance and reduce the influence of older resistance patterns. Nevertheless, resistance epidemiology, antimicrobial use, laboratory practices, and clinical breakpoints may change over time. The revised Discussion now states that the models would require periodic reassessment, recalibration, and retraining before any prospective or routine application. \n<b>Comment 5: Computational reproducibility<\/b> \n<i>More detailed sharing of scripts, models, and processed data is required.<\/i> \n<b>Response:<\/b> We appreciate this observation. The Methods have been expanded to provide a clearer description of data preprocessing, predictor encoding, class balancing, data partitioning, hyperparameter optimisation, model evaluation, and feature-importance analysis. We have also moderated claims regarding reproducibility and open-source availability where the supporting computational resources are incomplete. We acknowledge that provision of the complete preprocessing pipeline, trained models, breakpoint mappings, feature encodings, and source code would further improve reproducibility. \n<b>Comment 6: Clinical applicability and dashboard evaluation<\/b> \n<i>Further validation is required before clinical adoption, and the dashboard requires formal evaluation.<\/i> \n<b>Response:<\/b> We agree. The dashboard is now consistently described as a research prototype intended to illustrate the visualisation and exploration of surveillance-derived outputs. The manuscript explicitly states that no usability testing, implementation study, or clinical validation was conducted and that the dashboard should not be interpreted as a validated clinical decision-support system. Claims regarding empirical prescribing, antimicrobial stewardship, and clinical utility have also been moderated throughout the manuscript. The framework is now presented as a proof of concept requiring external validation, calibration assessment, prospective evaluation, and implementation research before clinical use. We sincerely thank the reviewer for the positive assessment and constructive recommendations. The comments helped us strengthen the manuscript\u2019s treatment of dataset representativeness, class imbalance, temporal changes, reproducibility, and the limits of clinical applicability. We believe the revised manuscript now presents the findings and limitations more clearly and cautiously.","URL":"https:\/\/wellcomeopenresearch.org\/articles\/11-393\/v2#referee-comment-6712","order":4,"name":"referee-comment-6712","label":"Referee Comment","group":{"name":"article-reports","label":"Article Reports"}},{"value":"This work was supported by Wellcome, grant number [227486]. The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript.","order":5,"name":"grant-information","label":"Grant Information"},{"value":"This is an open access article distributed under the terms of the Creative Commons Attribution Licence, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.","order":0,"name":"copyright-info","label":"Copyright"}]}}