Deep Learning Model Accurately Detects Acromegaly from Hand Images

Overview

A deep learning model using hand photographs achieved high accuracy in identifying acromegaly while preserving patient privacy by excluding facial and fingerprint features. The model outperformed board-certified endocrinologists and demonstrated consistent performance across age, sex, and remission status.

Background

Acromegaly is a disorder characterized by excessive growth hormone, often diagnosed through clinical features and biochemical tests. Early and accurate detection is crucial for effective management. Traditional diagnosis requires specialized training, which may limit access in some settings. Artificial intelligence offers potential to assist healthcare providers by analyzing external physical traits without compromising patient anonymity.

Data Highlights

Key Findings

• The deep learning model analyzed dorsal hand morphology and the "fist sign" to identify acromegaly.
• At optimal threshold, the model achieved 89% sensitivity and 91% specificity with an AUROC of 0.96.
• The model outperformed 10 board-certified endocrinologists on the same test set.
• Performance was consistent across age groups, remission status, and sex.
• False positives and negatives were often cases difficult for physicians to classify, indicating inherent diagnostic challenges.
• The model preserved patient privacy by excluding facial and fingerprint features from images.

Clinical Implications

This deep learning approach can assist healthcare providers, including those without specialized endocrinology training, in accurately detecting acromegaly using non-invasive hand images. Its high accuracy and privacy-preserving design support equitable healthcare delivery and may facilitate earlier diagnosis and treatment. However, clinicians should consider the model's limitations and validate findings in diverse populations.

Conclusion

The study demonstrates that a deep learning model analyzing hand photographs can accurately identify acromegaly while maintaining patient anonymity, outperforming expert endocrinologists and offering a promising tool for improved diagnostic accuracy.

References

1. Fukuoka et al. 2024 -- Deep learning model uses hand images to improve acromegaly detection