Integrating Technical and Clinical Perspectives to Address Key Challenges in AI-Driven Health Care

To the Editor,

We read with great interest the recent article “Artificial Intelligence and Machine Learning in Diabetic Foot Ulcer Care: Advances in Diagnosis, Treatment, and Prognosis.” ¹ The authors provide a timely and comprehensive review of AI/ML applications across the DFU care continuum, from diagnosis and individualized treatment to prognostic modeling. We especially appreciate the emphasis on multimodal integration—combining RGB images, thermography, biomechanical data, wearable sensors, and molecular profiling—which reflects a holistic vision beyond single-modality approaches. The discussion of remote monitoring, cost-effective wearables, and emerging generative AI further underscores the article’s forward-looking perspective.

While this work makes a valuable contribution, we would like to highlight several limitations not explicitly discussed and offer constructive suggestions for future research. First, the review follows a narrative approach without a PRISMA-based flow diagram or formal bias assessment (e.g., QUADAS-2, PROBAST).^2,3 This increases the risk of publication and language bias, as searches were limited to a few databases and primarily English sources. Future reviews could adopt systematic methods, expand database coverage (EMBASE, Scopus, Cochrane), and include non-English literature to ensure a more comprehensive and globally representative evidence base.

Second, although standardization issues are mentioned, the problem of label noise and inter-rater variability in DFU datasets is overlooked. Many AI models rely on subjective clinician labels without reporting agreement metrics, raising concerns about training on inconsistent annotations. We recommend standardized multi-expert annotation protocols, reporting of inter-rater statistics, and adoption of uncertainty-aware training methods such as probabilistic labeling or noise-robust loss functions to improve reliability.^4,5

Third, model generalizability and lifecycle management. The review notes standardization needs but does not address domain shift from variations in devices, acquisition settings, or patient populations, nor the risk of performance degradation over time (data drift). Future work should evaluate cross-device and cross-site performance, apply domain-adversarial or meta-learning strategies, and implement lifecycle protocols for ongoing monitoring, drift detection, and scheduled retraining to maintain safety and performance in real-world use.

In conclusion, this review effectively consolidates current advances in AI/ML for DFU care and underscores their clinical promise. By strengthening evidence synthesis, improving annotation quality, and ensuring model robustness across settings and over time, future research can accelerate the safe translation of AI/ML tools into practice—ultimately reducing amputations, improving healing, and enhancing quality of life for patients. We commend the authors for their excellent work and hope these suggestions will further enrich this important field.