Deep learning model uses hand images to improve acromegaly detection

Objective:

To evaluate a deep learning model's ability to identify acromegaly from hand photographs while maintaining patient privacy, which is crucial for ethical medical practices.

Key Findings:

• The model achieved 89% sensitivity and 91% specificity, with an F1 score of 0.89 and AUROC of 0.96, showing consistent performance across age, remission status, and sex.
• It outperformed 10 board-certified endocrinologists, whose F1 scores ranged from 0.43 to 0.63, with subgroup analyses indicating sensitivity and specificity of 89% and 90% in younger patients and 86% and 95% in older patients.

Interpretation:

The deep learning model demonstrates significant potential for accurate acromegaly detection based on external physical traits, aiding healthcare providers without specialized training.

Limitations:

• Incomplete biochemical confirmation in some control participants may affect the reliability of results.
• Higher prevalence of acromegaly than in general practice limits generalizability.
• Study limited to Japanese patients from specialist centers may not reflect broader populations.

Conclusion:

The study highlights the promise of AI in enhancing diagnostic accuracy for acromegaly while preserving patient anonymity, potentially improving healthcare delivery.