This bioRxiv preprint is titled “Evaluating performance bias in face-to-BMI vision transformer models across diverse human populations.” The work is indexed on bioRxiv with DOI 10.64898/2026.09.02.748815. The provided source text includes the full title, author list, institutional affiliations, the DOI, a note that the manuscript is a preprint and has not been peer reviewed, and the opening sentence of the abstract. The available excerpt does not include the full abstract, methods, results, figures, tables, or conclusions.
The central technical focus conveyed by the title and opening sentence is the use of vision transformer models to estimate body mass index (BMI) from facial images — a task described by the authors as face-to-BMI estimation. The title explicitly frames the study as an evaluation of performance bias across diverse human populations, indicating that comparative performance or fairness across demographic groups is a principal concern.
The supplied excerpt contains the beginning of the abstract. That opening frames facial-image-based BMI estimation as a non-invasive, cost-efficient alternative to direct measurement. The authors list potential use cases for such tools, including applications in telemedicine, emergency care settings where scales or measuring tools are unavailable, automated self-monitoring, and large-scale population assessment.
No additional sentences from the abstract were present in the supplied source text. As a result, specific claims, study objectives, hypotheses, or summaries of findings from the abstract cannot be reported here because they are not present in the excerpt.
The excerpt provides the full author list and institutional affiliations as reported on the bioRxiv record. Lead and contributing authors are affiliated with multiple institutions, including the University of Utah, University of California Santa Barbara, Chapman University, Arizona State University, Max Planck Institute for Evolutionary Anthropology, University of New Mexico, University of Toronto, Vanderbilt University, University of Calgary, Universiti Malaya and others. For correspondence the contact listed on the bioRxiv page is thomas.kraft@utah.edu.
The author list and affiliations indicate a multi-disciplinary, multi-institution collaboration. Specific author contributions, funding declarations, competing interests, and acknowledgements were not included in the provided excerpt.
This manuscript is posted on bioRxiv as a preprint and is explicitly noted as not peer reviewed on the bioRxiv page. The DOI assigned to the preprint is 10.64898/2026.09.02.748815. The record includes links to the abstract, article info/history, metrics, and a preview PDF on the bioRxiv site. The supplied content also shows the standard bioRxiv navigation and metadata blocks (channels, institutional groups, and author links).
The material supplied for this rewrite is incomplete. Only the front matter (title, authors, affiliations, DOI), the preprint notice, and the opening clause of the abstract were present. Critical sections that would be required to evaluate the study’s methods and findings — including full abstract, introduction, dataset descriptions, cohort demographics, model architecture and training details, validation approach, bias metrics, quantitative performance results by subgroup, statistical analyses, limitations discussed by the authors, and conclusions — were not present in the source excerpt and therefore cannot be summarized.
Because key content is missing, this summary does not attempt to infer results, numerical outcomes, or the authors’ conclusions about bias or model performance. Any assessment of whether the vision transformer models demonstrated bias, how bias was measured, which populations were included, and the magnitude or direction of performance differences must rely on the complete preprint text.
To evaluate the study’s evidence and conclusions, readers should consult the full bioRxiv preprint (DOI 10.64898/2026.09.02.748815) for the complete abstract, full methods, results, figures, and discussion. The provided excerpt is sufficient to identify the research topic — algorithmic face-to-BMI estimation using vision transformer architectures and an expressed focus on performance bias across populations — but not to assess empirical findings or their clinical or public-health implications.
If you’d like, I can retrieve and summarize the full preprint content (abstract, methods, results, tables and figures) if you provide the complete text or allow me to access the full bioRxiv record. As presented here, details beyond the opening abstract sentence were not reported in the source excerpt and therefore are not available for synthesis.