{"paper":{"title":"Adversarial Fragility and Language Vulnerability in Clinical AI: A Systematic Audit of Diagnostic Collapse Under Imperceptible Perturbations and Cross-Lingual Drift in Low-Resource Healthcare Settings","license":"http://creativecommons.org/licenses/by/4.0/","headline":"Clinical AI for chest X-rays loses accuracy from 89 percent to 62 percent under tiny invisible image changes and drops further on Nigerian dialects.","cross_cats":["cs.AI","cs.LG"],"primary_cat":"cs.CY","authors_text":"Ahmad Rufai Yusuf, Anthonio Oladimeji Gabriel","submitted_at":"2026-05-16T13:33:47Z","abstract_excerpt":"Current clinical artificial intelligence (AI) systems are evaluated almost exclusively on clean, standardised, English-language inputs, conditions that do not reflect the realities of healthcare delivery in low-resource settings. This study presents the first systematic dual audit of two orthogonal safety vulnerabilities in clinical AI: adversarial image fragility and cross-lingual diagnostic drift. Using DenseNet121, the architecture underlying CheXNet, fine-tuned on the COVID-QU-Ex chest X-ray dataset (85,318 images; COVID-19, Non-COVID Pneumonia, Normal), we demonstrate that diagnostic accu"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"Diagnostic accuracy collapses from 89.3% to 62.0% under a Fast Gradient Method (FGM) perturbation of epsilon=0.021, a magnitude imperceptible to the human eye, and language models drop from 85.0% to 55.0% on Nigerian Pidgin and Yoruba-inflected English cases.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The 20 clinical cases and the specific perturbation magnitude of epsilon=0.021 are representative of real deployment conditions in Nigerian Primary Health Centres without further validation against actual noisy images or spoken dialects.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"The study shows clinical AI accuracy collapsing from 89% to 62% on X-rays under imperceptible adversarial perturbations and from 85% to 55% on clinical cases in Nigerian Pidgin and Yoruba-inflected English.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"Clinical AI for chest X-rays loses accuracy from 89 percent to 62 percent under tiny invisible image changes and drops further on Nigerian dialects.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"a386328fd7960a48a8918cad9b4778fd69f3b0aebb7ae425ac34b31eaaf7a8f4"},"source":{"id":"2605.16993","kind":"arxiv","version":1},"verdict":{"id":"333489ca-8c1e-4306-9664-47b8bdb204cf","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-19T19:14:59.408701Z","strongest_claim":"Diagnostic accuracy collapses from 89.3% to 62.0% under a Fast Gradient Method (FGM) perturbation of epsilon=0.021, a magnitude imperceptible to the human eye, and language models drop from 85.0% to 55.0% on Nigerian Pidgin and Yoruba-inflected English cases.","one_line_summary":"The study shows clinical AI accuracy collapsing from 89% to 62% on X-rays under imperceptible adversarial perturbations and from 85% to 55% on clinical cases in Nigerian Pidgin and Yoruba-inflected English.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The 20 clinical cases and the specific perturbation magnitude of epsilon=0.021 are representative of real deployment conditions in Nigerian Primary Health Centres without further validation against actual noisy images or spoken dialects.","pith_extraction_headline":"Clinical AI for chest X-rays loses accuracy from 89 percent to 62 percent under tiny invisible image changes and drops further on Nigerian dialects."},"integrity":{"clean":false,"summary":{"advisory":0,"critical":1,"by_detector":{"doi_compliance":{"total":1,"advisory":0,"critical":1,"informational":0}},"informational":0},"endpoint":"/pith/2605.16993/integrity.json","findings":[{"note":"Identifier '10.1016/j.media.2025.103375' is syntactically valid but the DOI registry (doi.org) returned 404, and Crossref / OpenAlex / internal corpus also have no record. The cited work could not be located through any authoritative source.","detector":"doi_compliance","severity":"critical","ref_index":3,"audited_at":"2026-05-19T19:20:45.870301Z","detected_doi":"10.1016/j.media.2025.103375","finding_type":"unresolvable_identifier","verdict_class":"cross_source","detected_arxiv_id":null}],"available":true,"detectors_run":[{"name":"citation_quote_validity","ran_at":"2026-05-19T19:50:03.210650Z","status":"completed","version":"0.1.0","findings_count":0},{"name":"doi_title_agreement","ran_at":"2026-05-19T19:31:18.907270Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"cited_work_retraction","ran_at":"2026-05-19T19:23:43.980259Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"doi_compliance","ran_at":"2026-05-19T19:20:45.870301Z","status":"completed","version":"1.0.0","findings_count":1},{"name":"claim_evidence","ran_at":"2026-05-19T18:41:56.205160Z","status":"completed","version":"1.0.0","findings_count":0},{"name":"ai_meta_artifact","ran_at":"2026-05-19T18:33:26.293984Z","status":"skipped","version":"1.0.0","findings_count":0}],"snapshot_sha256":"0a9bb97437194dad8dbe39681e1355bd80dd973fac2bb2d645af179850cfda6f"},"references":{"count":35,"sample":[{"doi":"10.1136/bmjgh-2018-000798","year":2018,"title":"Wahl B. et al. Artificial intelligence (AI) and global health: how can AI contribute to health in resource-poor settings? BMJ Global Health, 3(4), e000798 (2018). https://doi.org/10.1136/bmjgh-2018-00","work_id":"7cfcb28d-e508-4547-8fec-feaa8699d0ca","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2024,"title":"Okafor C. et al. The utilization of artificial intelligence (AI) and machine learning (ML) for health in Nigeria: a rapid review. Journal of Medical Artificial Intelligence (2024). https://jmai.amegro","work_id":"9be5f6b5-4b8d-4d0a-80b5-d71e3e591b19","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1016/j.media.2025.103375","year":2025,"title":"Amgad M. et al. Robust and Interpretable Chest X-ray Classification via Diffusion Purification and Concept-Based Adversarial Detection. Journal of Object Technology in Biomedical Research, 2025. https","work_id":"47c23a9d-1d4d-4b62-ab98-f24bc681fe1d","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"10.1016/j.compbiomed.2021.105002","year":2021,"title":"Tahir A.M. et al. COVID-19 infection localization and severity grading from chest X-ray images. Computers in Biology and Medicine, 139, 105002 (2021). https://doi.org/10.1016/j.compbiomed.2021.105002","work_id":"e07ef64e-f79e-441f-ba69-42557d1d1ba3","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"Adeyemi O. et al. WeCAViT: A Weighted CNN-ViT model for Pneumonia Detection in Chest X-rays. IEEE Access, 2025. https://www.researchgate.net/publication/389527548","work_id":"2ef2710b-8fd2-43e3-9406-78640bdaa6de","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":35,"snapshot_sha256":"32ab889e2735745be62100cdd2e25ba027e363cfb12585b5c7605f68c72a3274","internal_anchors":2},"formal_canon":{"evidence_count":2,"snapshot_sha256":"21b2769af6479c51ba8ab8e1a55f3f5c6222ab0294b0b0a2a5188c68f0e66e21"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}