Common Questions

Face2Gene RESEARCH uses holdout cross-validation splits to compare cohorts of images. In each split, 50% of the images are randomly selected to train a classifier (train set) and 50% for validation (validation set). This process is repeated 10 times, each time for different random splits. The classification results are computed on the validation set and reported both numerically and graphically.

Mean results: an average of the area under the curve (AUC) of each of the 10 results is computed and the standard deviation is reported, visible in the top table.

Aggregated results: a score distribution curve and a receiver-operating-characteristic (ROC) curve for aggregated results for each photo used in the validation set. This is visible in the bottom results where the graphs are displayed, and allows calculation of p-value (see explanation below).

Face2Gene RESEARCH uses two different methods to calculate classification performance.

See “How are the results of the Binary Comparisons calculated?” above for more detail.


To measure the statistical significance of the binary comparisons, we use a permutation test by measuring the distribution of the validation-set accuracy statistic under the null hypothesis. We randomly permutate and train models 1000 times, then test the models on the validation set to get new AUCs. From the distribution of AUCs, we then calculate the one-sided p-value for the original AUC value.


For three or more cohorts, a multi-class experiment is conducted using all cohorts and a confusion matrix is computed. Similar to the binary comparisons, a holdout cross-validation splits method is used where in each split 50% are randomly selected for training and 50% for validation. The values presented are the mean accuracy per cohort over all splits.

The results shown in the matrix are normalized by dividing the results of each row by the sum of that row.



The highlighted (green) diagonal represents the true positive values of the classification for each cohort. The other values (white) are the false positives for each cohort. To get a sense of how different or similar the cohorts are, you can compare the true positive values in the matrix to the random chance for comparison below the matrix.


An example:

Three cohorts are compared here: Dr. Kerry’s Cases, Unaffected Controls, and Other Syndrome Controls. 71% of the time, Dr. Kerry’s Cases are classified by the system properly, compared to the random chance of 42.11%. 23% of the time, Dr. Kerry’s Cases are misclassified as Unaffected Controls. 6% of the time, Dr. Kerry’s Cases are misclassified as Other Syndrome Controls. The standard deviation is 10.76%.


There are different possibilities for control groups:

  • If you are interested in unaffected subjects, FDNA might be able to provide these for your use. Please contact us with the demographic description of your test cohort.
  • You can share cohorts (groups of patient photos) with a colleague without sharing personal health information. See details in the question below titled “A colleague will provide the control groups I need. How can this be done?”

If you would like to test which syndromes are possible confounders prior to selecting your control cohort, FDNA can provide you with a list of top-ranking syndromes from within the syndrome matches appearing in Face2Gene CLINIC for your test cohort.

Please contact us at


To share data with a colleague for a research project, every investigator contributing data will need to create a separate project in Face2Gene RESEARCH and upload their cohorts. Once this is done, please inform us at We will connect these data sources together in a de-identified way to comply with HIPAA and GDPR requirements.


Yes, you have full authorship rights. Research collaborations with Face2Gene are governed by the Research Terms & Conditions. We ask that you mention in the methodology of your paper  that the results were obtained using the Face2Gene RESEARCH application (FDNA Inc., Boston, MA, USA.)

There are several such publications. Please click here to view an updated list.

There have been a few publications that have studied this from different perspectives:

Pantel J.T., Zhao M., Mensah M.A., Hajjir N., Hsieh T.H., Hanani Y., Fleischer N., Kamphans T., Mundlos S., Gurovich Y.,  Krawitz P.M. (2018). Advances in computer-assisted syndrome recognition by the example of inborn errors of metabolism J Inherit Metab Dis.


Vorravanpreecha N., Lertboonnum T., Rodjanadit R., Sriplienchan P., Rojnueangnit K., (2018). Studying Down syndrome recognition probabilities in Thai children with de‐identified computer‐aided facial analysis. Am J Med Genet. Part A. 2018;1–6.


Amudhavalli S.M., Hanson R., Angle B., Bontempo K., Gripp K.W. (2018). Further delineation of Aymé‐Gripp syndrome and use of automated facial analysis tool. Am J Med Genet. Part A. 2018;176A:1648–1656.

When using Face2Gene, you upload patient photos into your private and personal case list. These photos are not shared with anyone and only you have access to them.

Therefore, many institutions are comfortable using the same consent used for taking photos for their own clinical records. Please click the following links to read more about our privacy policy and download a sample patient consent form, which includes the option for research. The research collaborations with Face2Gene are governed by our Research Terms & Conditions

If your institution requires an IRB, we will be glad to provide assistance–contact

Face2Gene User Community Includes Users From:

  • Using Face2Gene to reference all my department’s cases, share information with my colleagues and quickly look up relevant information in the London Medical Databases Online saves me hours of work every week and allows me to focus on my patients.

    Dr. Ibrahim Akalin

    Assoc. Prof. Ibrahim Akalin, MD, Medical Geneticist from the Istanbul Medeniyet University, Istanbul, Turkey

  • FDNA’s game-changing technology introduces an objective computer-aided dimension to the “art of dysmorphology”, transforming the analysis into an evidence-based science.

    Dr. Michael R. Hayden

    Chairman of FDNA’s Scientific Advisory Board & Steering Committee and Editor in Chief of Clinical Genetics

  • FDNA is developing technology that has the potential to help so many physicians and families by bringing them closer to a diagnosis- there are literally millions of individuals with unusual features around the world that lack a diagnosis and therefore lack information on natural history, recurrence risk and prevention of known complications.

    Dr. Judith G. Hall

    Professor Emerita of Pediatrics & Medical Genetics UBC & Children's and Women's Health Centre of BC

  • FDNA has been “right on the money”, providing me with relevant, accurate and insightful information for differential diagnoses.

    Dr. Cynthia J.R. Curry

    Professor of Pediatrics UCSF, Adjunct Professor of Pediatrics Stanford

  • I am excited to be a part of the FDNA community, promoting broad information sharing with my peers to amplify the scientific and clinical value of our community’s accumulated knowledge for the purpose of efficiently diagnosing individuals with rare genetic disorders.

    Dr. Karen W. Gripp

    Chief, Division of Medical Genetics A.I. duPont Hospital for Children

  • FDNA's idea of incorporating several dysmorphology resources (OMIM, GeneReviews), supported by their visual analytic technology, will be able to improve researching of genetic syndromes - all within a single mobile app.

    Dr. Chad Haldeman-Englert

    Assistant Professor Pediatrics at Mission Fullerton Genetics

  • Given the advancement of visual analytical technology, it’s about time Dysmorphology is supported with computational capabilities and moving this to mobile support, is simply the next logical step.

    Dr. Chanika Phornphutkul

    Associate Professor of Pediatrics Director, Division of Human Genetics Department of Pediatrics Warren Alpert Medical School of Brown University

  • Having an archive of cases easily accessible from my mobile device anytime and anywhere is a long-time unmet need.

    Dr. Lynne Bird

    Rady Children's Specialists of San Diego

  • FDNA's solution is a huge leap forward for dysmorphology. It saves me significant time when I’m evaluating patients in my clinic and provides me with insightful tools that help me generate a differential diagnosis.

    Dr. David A. Chitayat

    Head of the Prenatal Diagnosis and Medical Genetics Program at Mount Sinai Hospital, Toronto

  • Shortly after learning about Face2Gene, I’ve started to incorporate this amazing tool into my workflow. Soon enough, Face2Gene’s analysis flushed out references that I would not have considered for several of my patients, which turned out to be their correct diagnosis

    Dr. Zvi U. Borochowitz

    Chairman (Retired) of The Simon Winter Institute for Human Genetics at Bnai-Zion Medical Center, Technion-Rappaport Faculty of Medicine

  • The Unknown Forum from Face2Gene is a great community platform for exchanging opinions regarding undiagnosed cases. It is straightforward to use and safe for exchange of medical data, thanks to the efforts of its developers and to the involvement of geneticists worldwide.

    Dr. Oana Moldovan

    Clinical Geneticist at the Hospital Santa Maria, CHLN, Lisbon, Portugal