Privacy-focused AI improves screening for rare endocrine diseases

AI can accurately diagnose rare endocrine disorders simply by analyzing a photo of the back of a hand or clenched fist. Kobe University’s privacy-friendly results are expected to establish a more efficient referral system and reduce medical disparities between communities.

Acromegaly is a rare, incurable disease that usually develops in middle age or later, causing enlarged limbs, changes in the appearance of the face, and affects the growth of bones and organs throughout the body. This condition, caused by excessive secretion of growth hormone, develops slowly over decades, but if left untreated can lead to life-threatening complications that can reduce life expectancy by about 10 years. “Because the symptoms progress very slowly and it is a rare disease, it is not uncommon for it to take up to 10 years before it is diagnosed,” says Hidenori Fukuoka, an endocrinologist at Kobe University. He further explains, “With advances in AI tools, there are attempts to use photographs for early detection, but this has not been implemented in clinical settings.”

The group examined current AI research challenges and found that much of it relies on facial photographs, which can cause privacy concerns. “In an attempt to address this concern, we decided to focus on the hand, which is a body part that is routinely examined for diagnostic purposes in clinical settings, along with the face, especially since changes in the hands often appear in acromegaly,” said Yuka Omachi, a graduate student at Kobe University. However, I decided to double down on privacy by only using images of the back of the hand and clenched fist, and avoiding the more discrete palm line pattern. This resulted in the support of 725 patients from 15 medical facilities across Japan, who donated over 11,000 images to train and validate the AI ​​model.

in Journal of Clinical Endocrinology and Metabolisma team from Kobe University have now announced that their model recognizes this condition with very high sensitivity and specificity. In fact, their models outperform even experienced endocrinologists who are asked to evaluate the same photos. “I was honestly surprised that we were able to achieve such a high level of diagnostic accuracy using only photos of the back of the hand and clenched fist. What I felt was particularly important was that we achieved this level of performance without using facial features, making this approach very practical for disease screening,” says Omachi.

As a next step, the group is considering extending the model to other conditions that can be identified from photos, such as rheumatoid arthritis, anemia, and finger clubbing. Omachi says, “This result may be the gateway to expanding the possibilities of medical AI.”

In medical practice, doctors rely on a wide range of factors and data to make diagnoses, rather than just using hand images. The Kobe University research team therefore sees their newly developed model as an opportunity to “complement clinical expertise, reduce missed diagnoses, and enable early intervention,” as they write in their paper. Mr. Fukuoka, the principal investigator, said, “Further development of this technology may lead to the construction of a medical infrastructure that connects patients suspected of having hand-related diseases during medical checkups with specialists.We also believe that it can support non-specialist doctors at local medical sites and contribute to reducing medical disparities.”

This research was supported by the Hyogo Science and Technology Foundation. This study was conducted in collaboration with researchers from Fukuoka University, Hyogo College of Medicine, Nagoya University, Hiroshima University, Toranomon Hospital, Nippon Medical School, Kagoshima University, Tottori University, Yamagata University, Okayama University, Hyogo Prefectural Kakogawa Medical Center, Hokkaido University, International University of Health and Welfare, Moriyama Memorial Hospital, and Konan Women’s University.