ECEESPE2025 Poster Presentations Pituitary, Neuroendocrinology and Puberty (162 abstracts)
Acromegaly detection from facial images using machine learning
Konstantina Vouzouneraki 1 , Erik Ylipää 2 , Tommy Olsson 1 , Katarina Berinder 3,4 , Charlotte Höybye 3,4 , Sophie Bensing 3,4 , Maria Petersson 3,4 , Anna-Karin Ækerman 5 , Henrik Borg 6 , Bertil Ekman 7 , Jonas Robért 7 , Britt Eden Engstrom 8 , Oskar Ragnarsson 9,10,11 , Pia Burman 6 & Per Dahlqvist 1
1Umeå University, Department of Public Health and Clinical Medicine, Umeå, Sweden; 2Linköping University, Department of Science and Technology, AIDA Data Hub, Linköping, Sweden; 3Karolinska institute, Department of Molecular Medicine and Surgery, Stockholm, Sweden; 4Karolinska University Hospital, Department of Endocrinology, Stockholm, Sweden; 5Örebro University Hospital, Department of Internal Medicine, Örebro, Sweden; 6Lund University, Skåne University Hospital, Department of Endocrinology, Malmö, Sweden; 7Linköping University, Departments of Endocrinology in Linköping and Norrköping and Department of Health, Medicine and Caring Sciences, Linköping, Sweden; 8Uppsala University, Uppsala University Hospital, Department of Medical Sciences, Endocrinology and Mineral Metabolism, Uppsala, Sweden; 9Sahlgrenska University Hospital, Department of Endocrinology, Gothenburg, Sweden; 10Institute of Medicine at Sahlgrenska Academy, Department of Internal Medicine and Clinical Nutrition, Gothenburg, Sweden; 11Institute of Medicine, University of Gothenburg, Wallenberg Centre for Molecular and Translational Medicine, Gothenburg, Sweden
JOINT1165
Background: There is a substantial delay in the diagnosis of acromegaly contributing to increased morbidity and mortality. Machine learning-based analysis of facial images has shown promising potential in identifying acromegaly. However, there is a need for further validation in large acromegaly cohorts, incorporating new methodological insights in machine learning development and using human expert assessment for comparison.
Methods: Video recordings from 155 patients with acromegaly and 153 matched controls at all seven Swedish University hospitals were collected using cellphones. Facial pictures from different angles were extracted. Clinical data regarding disease course and status were obtained. Four different deep neural network architectures were trained to distinguish photographs of patients with acromegaly from controls. These neural networks were pretrained on large image datasets including three networks trained on various image categories (ImageNet networks) and one on human faces [Facial Representation Learning (FaRL)]. The performance of these models was compared to that of 12 experienced endocrinologists.
Results: The FaRL based machine learning model presented higher sensitivity compared to the compound majority assessment (ensemble) of experienced endocrinologists (0.82 vs 0.66), with only marginally lower specificity (0.87 vs 0.93). The overall diagnostic performance was comparable for FaRL and the expert ensemble with ROC AUC 0.89 for both. However, the balanced accuracy of each individual endocrinogist was lower (range 0.72-0.79) than the FaRL model (0.80). Accuracy of the other neural networks were inferior to FaRL. The classification agreement between the top-performing models and human experts was high for true negatives (76%) but lower for true positives (55%). Both the machine learning models and human experts showed greater sensitivity in accurately classifying males compared to females, but showed no significant difference in precision metrics between patients with active (n = 33) vs controlled acromegaly (n = 122).
Conclusions: A FaRL-based machine learning model shows comparable accuracy to expert endocrinologists in acromegaly identification by face photographs, but with the advantage of higher sensitivity. This supports that digital face analysis can be useful in acromegaly detection. Further research is required to validate its performance and explore its applicability in clinical practice.
Volume 110
Article tools
My recent searches
My recently viewed abstracts
Authors
Endocrine Abstracts
ISSN 1470-3947 (print) | ISSN 1479-6848 (online)
© Bioscientifica 2025 |
Privacy policy |
Cookie settings
BiosciAbstracts
Bioscientifica Abstracts is the gateway to a series of products that provide a permanent, citable record of abstracts for biomedical and life science conferences.
Find out more