A new manuscript in the Journal of Diabetes Science and Technology from our global team explores how machine learning is being used to detect, diagnose, and even predict complications of diabetes. The paper, authored by Agatha Scheideman, Mandy Shao, Henry Zelada, Jorge Cuadros, Joshua Foreman, Pinaki Sarder, Cindy Ho, Niels Ejskjaer, Jesper Fleischer, Simon Lebech Cichosz, David G. Armstrong, Nestoras Mathioudakis, Tao Wang, Yih Chung Tham, and David C. Klonoff, highlights the promise—as well as the current limitations—of applying machine learning in this space .
Why this matters
Diabetes remains one of the world’s leading causes of disability-adjusted life years. Despite major advances in therapy, complications such as retinopathy, nephropathy, neuropathy, autonomic dysfunction, and foot ulcers still account for significant morbidity and mortality. Early diagnosis is key, and machine learning may help identify risk patterns invisible to human reviewers.
Key themes from the review
- Retinopathy: Deep learning models using fundus images are already FDA-cleared and in national screening programs. Future systems are moving toward multimodal risk prediction and smartphone-based deployment.
- Nephropathy: Both biopsy-based and non-invasive machine learning models show diagnostic promise, with transformer-based multimodal systems on the horizon.
- Peripheral and autonomic neuropathy: Early detection remains challenging; machine learning applied to wearable sensors, ECGs, and corneal imaging may help bridge this gap.
- Diabetic foot ulcers: Thermal imaging combined with convolutional neural networks can identify risk earlier than human examiners.
- Other systemic complications: External eye imaging is emerging as a surprisingly powerful marker for systemic disease signals.
- Hospital outcomes: Machine learning has been applied to predict hypoglycemia, adverse events, and mortality in hospitalized patients, with integration into EHRs expected to expand.
The road ahead
The review underscores that while many models demonstrate impressive accuracy, widespread clinical use is still constrained by limited dataset diversity, labeling inconsistencies, and lack of external validation. The authors argue for multicenter, multiethnic datasets, explainable AI approaches, and user-friendly integration into workflows.
This work provides a roadmap for how data, imaging, and clinical insight may converge to make diabetes care more proactive and precise.
📖 Full citation:
Scheideman A, Shao M, Zelada H, Cuadros J, Foreman J, Sarder P, Ho C, Ejskjaer N, Fleischer J, Cichosz SL, Armstrong DG, Mathioudakis N, Wang T, Tham YC, Klonoff DC, “Machine Learning to Diagnose Complications of Diabetes,” Journal of Diabetes Science and Technology, 2025;1–21. doi:10.1177/19322968251365245 .
Leave a Reply