Research shows AI surpasses endocrinologists in diagnosing acromegaly using hand photos — Evidence Review
Published in Journal of Clinical Endocrinology & Metabolism, by researchers from Kobe University, Fukuoka University, Hyogo Medical University, 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, Konan Women's University
Researched by
— the AI search engine for science
Table of Contents
Researchers at Kobe University have developed an AI system that detects acromegaly using only hand photos, achieving higher diagnostic accuracy than experienced endocrinologists and protecting patient privacy. Related studies broadly support the effectiveness of AI for endocrine disease diagnosis and emphasize the value of privacy-aware and image-based approaches (1,3,5). See the original study in the Journal of Clinical Endocrinology & Metabolism for further details.
- Numerous studies demonstrate that AI and machine learning can reach high diagnostic accuracy for various endocrine disorders, not just acromegaly, particularly when standardized data and privacy concerns are addressed (1,3,5).
- Prior research mainly used facial images for AI-driven diagnosis of endocrine diseases, raising privacy issues; the new study’s hand-focused method addresses this gap while maintaining or exceeding prior diagnostic performance (5).
- AI-based tools have been shown to assist clinicians in both routine and specialized settings, improving access and consistency in diagnosis, particularly in underserved areas or where specialist expertise is limited (3,4,13).
Study Overview and Key Findings
The slow, subtle progression of acromegaly often leads to delayed diagnosis, with patients sometimes waiting years before receiving appropriate care. Early detection is crucial, as untreated acromegaly can lead to severe health issues and reduced life expectancy. This study addresses both the diagnostic challenge and patient privacy concerns by leveraging AI to analyze non-facial, hand-based photographs, potentially expanding access to screening in clinical and remote settings.
| Property | Value |
|---|---|
| Organization | Kobe University, Fukuoka University, Hyogo Medical University, 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, Konan Women's University |
| Journal Name | Journal of Clinical Endocrinology & Metabolism |
| Authors | Hidenori Fukuoka, Yuka Ohmachi |
| Population | Patients with suspected acromegaly |
| Sample Size | 725 patients |
| Outcome | Diagnostic accuracy of AI in detecting acromegaly |
| Results | AI outperformed experienced endocrinologists in diagnosis |
Literature Review: Related Studies
To contextualize these findings, we searched the Consensus paper database, which includes over 200 million research papers. The following queries were used to identify relevant literature:
- AI diagnosis endocrine disorders
- hand photo disease detection
- endocrinologist performance comparison AI
Related Studies Table
| Topic | Key Findings |
|---|---|
| How effective is AI in diagnosing endocrine disorders compared to clinicians? | - AI and machine learning achieve high diagnostic accuracy for several endocrine conditions, sometimes matching or surpassing non-expert clinicians, but not always outperforming experts (1,3,12,13). - AI-based diagnostic tools can support but not fully replace clinical expertise, suggesting a complementary role in clinical workflows (3,12,13,14). |
| What are the privacy and practicality considerations for AI-based image diagnosis? | - Most prior studies have focused on AI using facial images, which raises privacy concerns; using hand images offers a privacy-protecting alternative while maintaining diagnostic value (5,9). - Limiting image types (e.g., using only the back of the hand) can facilitate broader participation and reduce risk of re-identification (5,9). |
| Can AI improve healthcare access and reduce disparities in endocrine disease diagnosis? | - AI tools can support early detection and specialist referral, especially in underserved or regional healthcare settings where expert endocrinologists may be scarce (3,4,13,14). - Language and disease-specific AI model performance varies, but improvements could enhance diagnostic equity and access across diverse populations (4,13). |
| What are the limitations and future directions for AI in endocrine diagnostics? | - Despite promising results, AI models require further validation, standardized input criteria, and transparency to ensure reliability and clinical adoption (1,2,12). - Integration with clinical data beyond images (e.g., labs, history) and ongoing evaluation of model generalizability are important next steps (1,3,13). |
How effective is AI in diagnosing endocrine disorders compared to clinicians?
Multiple studies indicate that AI and machine learning systems can achieve high diagnostic accuracy across a range of endocrine diseases, sometimes matching or even outperforming non-expert clinicians. However, while AI may surpass less experienced physicians, it typically does not outperform expert clinicians across all diagnostic tasks (1,3,12,13). The Kobe University study is notable in that its AI tool outperformed experienced endocrinologists using only hand images—a method that may have unique application for acromegaly and similar disorders.
- Systematic reviews show AI diagnostic accuracy for endocrine diseases, such as PCOS, ranges from 89% to 100%, highlighting AI’s potential for early detection (1).
- AI-assisted interpretation notably improves the performance of junior endocrinologists in image-based tasks, such as bone age assessment, suggesting its utility for clinician training and decision support (13).
- Meta-analyses find that, on average, generative AI matches non-expert clinician performance but still lags behind expert physicians, underscoring the need for human oversight (12).
- Across the literature, AI is consistently presented as a complement to, rather than a replacement for, clinical expertise (3,13,14).
What are the privacy and practicality considerations for AI-based image diagnosis?
While facial recognition-based AI models have demonstrated strong performance in diagnosing endocrine and metabolic conditions, they introduce significant privacy challenges (5). The Kobe University study addresses this by focusing solely on photographs of the back of the hand and a clenched fist. This approach not only preserves privacy but also facilitates broader patient participation and could be more easily adopted in clinical and community settings.
- AI facial recognition models reach high diagnostic accuracy for multiple endocrine syndromes, but privacy concerns and the uniqueness of facial features are major barriers to clinical adoption (5).
- Using less-identifiable features such as hand images, or even nail photos, can still yield useful diagnostic information while reducing privacy risks (9).
- Excluding highly individual features (like palm lines) further mitigates identification concerns and increases participant willingness (5).
- Practicality and privacy are important factors for scaling AI-based screening tools to real-world healthcare environments (5,9).
Can AI improve healthcare access and reduce disparities in endocrine disease diagnosis?
AI-driven tools hold promise for improving access to specialty care, particularly in regions with limited endocrinology expertise. Studies suggest these systems can enable earlier detection, support non-specialist clinicians, and reduce diagnostic variability, which may help address healthcare disparities (3,4,13,14). The Kobe University study’s design—focusing on a simple, privacy-conscious screening tool—could be particularly impactful for underserved populations.
- AI models have been shown to improve diagnostic accuracy for both clinicians and trainees, even in resource-limited settings (13).
- Large-language model tools like ChatGPT can support clinical decision-making, although language and disease context affect performance and reliability (4).
- AI-based screening could expedite specialist referral for rare or complex endocrine disorders, potentially reducing time to diagnosis and improving outcomes (13,14).
- Efforts to optimize AI tools for diverse populations and healthcare settings are ongoing, with the goal of enhancing diagnostic equity (4,14).
What are the limitations and future directions for AI in endocrine diagnostics?
Despite encouraging results, several challenges remain before AI can be routinely implemented in endocrine diagnostics. These include the need for standardization, transparency, and external validation of AI systems (1,2,12). Most current models rely on specific data types or settings, and integration with broader clinical information is required for robust, generalizable performance.
- There is variability in the diagnostic criteria and data sources used in AI studies, which affects clinical applicability and reliability (1,2).
- Transparency and interpretability of AI decisions are critical for clinician trust and regulatory approval (2).
- Combining image-based AI with other clinical data (labs, history) could further enhance diagnostic accuracy and utility (1,3).
- Ongoing research is needed to evaluate generalizability, monitor real-world performance, and address biases in training data (12,14).
Future Research Questions
While the Kobe University study demonstrates strong diagnostic performance and privacy awareness, several important questions remain. Future research should address the generalizability, integration, and ethical implementation of such AI tools in diverse healthcare settings.
| Research Question | Relevance |
|---|---|
| How generalizable is the AI hand image model across different ethnicities and populations? | The current study is based on a Japanese patient cohort; evaluating performance in diverse populations is needed to ensure diagnostic equity and avoid bias (1,5,12). |
| Can AI hand image analysis be used to detect other systemic or rare diseases? | The approach may extend to other conditions with hand manifestations (e.g., arthritis, anemia); research could determine the broader utility and performance of similar AI models (3,13,14). |
| What is the impact of integrating AI hand image analysis with other clinical data on diagnostic accuracy? | Combining image analysis with laboratory, history, and physical exam data could further improve diagnostic accuracy, but requires methodical evaluation (1,3,13). |
| How do patients and clinicians perceive the privacy and acceptability of hand-based AI screening? | Understanding user attitudes toward privacy-preserving AI screening tools is crucial for successful adoption and may differ from perceptions of facial recognition models (5,9). |
| What are the long-term clinical outcomes when AI screening is used for early detection of acromegaly? | Assessing whether earlier detection via AI leads to improved management, reduced complications, and better prognosis is necessary to quantify real-world benefits (12,14). |