Ten board-certified endocrinologists looked at photographs of 138 people's hands and tried to say which ones had acromegaly. Their best F1 score, a combined measure of precision and recall, was 0.63. An AI model reading the same images scored 0.89.
The study, from a Kobe University team with graduate student Yuka Ohmachi as first author and endocrinologist Hidenori Fukuoka as senior author, runs in The Journal of Clinical Endocrinology & Metabolism, in the August 2026 print issue after publishing online in February. It drew on 716 participants and 11,480 images from 15 Japanese pituitary centers.
The design choice that makes it interesting is what the researchers refused to photograph.
They Deliberately Avoided Faces and Palms
Acromegaly is caused by excess growth hormone, usually from a pituitary tumor. It enlarges hands, feet, and facial features slowly enough that patients and their families often fail to notice.
Previous AI attempts used facial photographs and reached strong accuracy, but none reached clinical practice. The Kobe team attributes that partly to privacy: facial images and fingerprints warrant heightened safeguards under data protection principles.
So they restricted inputs to two views, the back of the extended hand and the clenched fist with the thumb outside, which they term the "dorsal hand sign" and the "fist sign." Palms were excluded because they carry fingerprints. Hand and foot enlargement occurs in 90% of acromegaly patients, making the hand a reasonable substitute.
The fist sign itself comes from earlier clinical reports and is defined as the inability to cover the fingernails with the center of the palm when making a fist. The paper is the first to formally test it as a diagnostic feature. The authors also note that a previous hand-image model, trained on a dataset where more than half the images came from acromegaly patients, included palm views and did not address identifiability.
The Numbers, and What Sits Behind Them
The final dataset included 317 people with acromegaly and 399 controls, split by facility so that training and testing drew on entirely different hospitals. A ResNet-50 model was trained with five-fold cross-validation, and each patient's prediction averaged four images.
On the held-out test set, the model reached a sensitivity of 0.89, specificity of 0.91, positive predictive value of 0.88, negative predictive value of 0.93, an F1-score of 0.89, and an area under the ROC curve of 0.96. The ten comparison endocrinologists, averaging 9.8 years of experience, produced F1 scores between 0.43 and 0.63.
Heatmap analysis suggested the model was attending to finger joints, the fingernails of a clenched fist, and the region near the base of the thumb.
The heatmaps also served a housekeeping purpose, confirming the model was reading hand shape rather than backgrounds or sleeves, a common failure mode in medical imaging AI.
Error analysis was revealing. All six false negatives were women with a history of pituitary surgery, and in five of those cases the endocrinologists correctly classified the image 20% of the time or less. The model missed the same hard cases humans miss. The seven false positives ran the other way: those participants had a median age of 48, five of the seven were men, and in six of the seven cases the endocrinologists got the answer right at least 80% of the time.
The Real Problem Is a Decade of Waiting
Diagnostic performance only matters because acromegaly is so often caught late.
Prevalence runs 8 to 24 per 100,000. Untreated, mortality is two to three times that of the general population and life expectancy drops by roughly a decade. Patients typically present first to orthopedics, dental and maxillofacial surgery, gynecology, gastroenterology or cardiology rather than endocrinology, and about 24% experience diagnostic delays exceeding 10 years.
"Because the condition progresses so slowly, and because it is a rare disease, it is not uncommon to take up to a decade for it to be diagnosed," Fukuoka said in Kobe University's announcement.
Where the Model Falls Down
The authors are unusually candid about limits, and readers should not treat this as a self-check tool.
Participants were recruited exclusively at Japanese specialist pituitary centers, so performance in multiethnic or primary-care populations is unknown. More than half the acromegaly patients had achieved biochemical remission when photographed, meaning their features may have partially normalized. Controls were selected by expert judgment, and 191 of them never had IGF-I measured, so undiagnosed cases could theoretically sit in that group.
Most importantly, acromegaly was far more common in this dataset than in the real world. That inflates positive predictive value. The subgroup analysis makes the point concretely: at one test facility PPV was 0.93, at another it was 0.44. In a genuine screening population, most positives would be false ones, each carrying costs and anxiety.
The team plans to extend the approach to rheumatoid arthritis, anemia and finger clubbing. External validation in diverse, non-specialist cohorts comes first. Anyone concerned about changes in their hands, feet, jaw, ring size or shoe size should raise it with a physician rather than compare their fingers to a research paper.
Key Questions Answered
What is acromegaly?
A disorder caused by excess growth hormone, usually from a pituitary tumor, producing gradual enlargement of the hands, feet, and facial features along with effects on bone and organs.
How well did the AI perform?
Sensitivity 0.89, specificity 0.91, and an ROC-AUC of 0.96 on a test set from three hospitals not used in training, with a higher F1-score than any of the ten endocrinologists who reviewed the same photographs.
Why hands instead of faces?
Facial images and fingerprints carry heavier legal and ethical requirements. The back of the hand and clenched fist show the disease's characteristic changes with far lower re-identification risk.
Can I photograph my own hand and get an answer?
No. This is an experimental research model, not an approved or publicly available diagnostic tool, and it was never validated for self-assessment.
Would it work outside Japan?
Unknown. The authors state that generalizability to other ethnic groups and community settings remains unverified pending external validation.
What symptoms should prompt a doctor visit?
Gradual enlargement of hands or feet, changes in ring or shoe size, altered facial appearance, joint pain, headaches, or a deepening voice are worth discussing with a clinician.
