ORIGINAL RESEARCH article
Front. Sociol.
Sec. Medical Sociology
Multimodal Antibiogram Completion from MALDI-TOF Spectra and Partial Antibiograms: Distribution-Free Error Control for Antibiotic Resistance Prediction Across Hospitals
Sadia Mehjabeen Kanta 1
- MH
MD Arman Hossen 2
- NE
Nabeeha Ejaz 3
- AO
Adel T. Osman 4
- MM
Md Samir Murad 5
- MA
Mohammad A. A. Al-Najjar 6
- AF
Adiba Binte Farid 7
- OK
Oun Deli Khudhair 8
- KM
Khaldoon Mustansir 9
1. Department of Public Health, Canadian University of Bangladesh, Dhaka, Bangladesh
2. College of Software Engineering, Sichuan University, Chengdu, China
3. Ayaan Institute of Medical Sciences, Kaloji Narayana Rao University of Health Sciences, Hyderabad, Telangana, India
4. Department of Pharmacology and Toxicology, Faculty of Pharmacy, Al-Azhar University, Assiut Branch, Assiut, Egypt
5. Data Science Program, College of Engineering and Applied Sciences, Stony Brook University, Stony Brook, NY, USA
6. Department of Pharmaceutical Sciences and Pharmaceutics, Faculty of Pharmacy, Applied Science Private University, Amman, Jordan
7. Department of Biochemistry and Molecular Biology, Jahangirnagar University, Dhaka, Bangladesh
8. Pusat Kanser Tun Abdullah Ahmad Badawi, Universiti Sains Malaysia, 13200 Kepala Batas, Penang, Malaysia
9. Department of Biomedical Engineering, University of Houston, Houston, TX, USA
Select one of your emails
You have multiple emails registered with Frontiers:
Notify me on publication
Please enter your email address:
If you already have an account, please login
You don't have a Frontiers account ? You can register here
Abstract
Background: MALDI-TOF mass spectrometry identifies a bacterial isolate within minutes of colony growth, but the full antimicrobial susceptibility panel takes another one to two days and is released in stages. Models that predict resistance from the spectrum and from results already released could close this gap, but only if their errors are controlled where they are used.
Methods: We used two public datasets: DRIAMS (four Swiss hospitals; 53,854 isolates) and MS-UMG (University Medical Center Göttingen, Germany, 2020–2021; 73,736 isolates), with 1.45 million susceptibility results for 35 antibiotics. PRISM is a transformer that reads the 200 most intense spectral peaks as tokens, together with one token per antibiotic carrying its class, mechanism and state (susceptible, non-susceptible or untested). Its encoder was pre-trained without labels by masked-peak reconstruction. A Mondrian conformal rule per antibiotic × organism group reports an untested result or sends it to the laboratory, targeting very major error (VME) ≤1.5% and major error (ME) ≤3%. Models were trained on DRIAMS-A 2015–2017, tested on 2018, and evaluated on five external cohorts.
Results: At identification time, PRISM outperformed LightGBM with spectral features (macro-AUROC 0.845 vs 0.826; within-species AUROC 0.641 vs 0.600; differences +0.019 [95% CI 0.015–0.023] and +0.041 [0.023–0.059]). Each released result added information: within-species AUROC rose from 0.62 with none to 0.81 with two and 0.86 with four. Pre-training helped mainly at identification time and at external sites. Thresholds transported from the development hospital broke the ME target at all five external cohorts (up to 5.9%) and the VME target at four (up to 3.4%). Recalibration on each site's own isolates restored both targets on average at every cohort; none of 490 random calibration draws exceeded the VME target. With all local isolates, 10–16% of untested results were released at identification time and 18–25% once the first-line panel was known.
Conclusion: Combining the spectrum with partial susceptibility results completes the antibiogram more accurately than tabular baselines. Its error guarantees, however, do not travel between hospitals and must be re-established with local calibration data before use.
Summary
Keywords
antimicrobial resistance, MALDI-TOF mass spectrometry, antimicrobial susceptibility testing, transformer, self-supervised learning, conformal prediction, very major error, external validation
Received
20 October 2026
Accepted
01 November 2026
Copyright
© 2026 Kanta, Hossen, Ejaz, Osman, Murad, Al-Najjar, Farid, Khudhair and Mustansir. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) or licensor are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Sadia Mehjabeen Kanta
Disclaimer
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article or claim that may be made by its manufacturer is not guaranteed or endorsed by the publisher.