📱 Smartphone thermography AI vs nurse foot risk scoring

In a cross-sectional study of 100 adults with Diabetes mellitus (DM) undergoing routine foot screening, a smartphone-based thermography AI flagged all nurse-defined moderate-to-high risk cases with 100% sensitivity and 96.8% specificity (PPV 66.7%, NPV 100%). The AI risk score (03) strongly correlated with the diabetic foot nurses reference score (=0.973).

Why It Matters To Your Practice

▪ Diabetic foot complications are time-sensitive; missing moderate-to-high risk patients can mean preventable ulcers and amputations.

▪ Smartphone thermal imaging paired with AI could extend screening capacity beyond specialist clinics into primary care and remote settings.

▪ A very high NPV (100%) suggests potential value as a “rule-out” tool to prioritize who needs urgent in-person assessment.

Clinical Implications

▪ If deployed, expect AI outputs as a 03 risk score; plan workflows for how scores trigger education, offloading, podiatry referral, or expedited follow-up.

▪ Given the modest PPV (66.7%), anticipate false positives and ensure confirmatory clinical evaluation pathways are in place.

▪ Thermal asymmetry increased with higher risk levels for both AI and nurse scoringsupporting thermography as a physiologic signal you can track over time.

Insights

▪ This was a single-visit (cross-sectional) comparison using nurse scoring as the reference standardnot longitudinal outcomes like ulceration or amputation.

▪ Performance metrics hinge on prevalence and the nurse-defined threshold for “moderate-to-high risk,” which may vary by clinic and training.

▪ The tools appeal is standardization: it converts thermal images into reproducible scores, potentially reducing inter-rater variability in thermography interpretation.

The Bottom Line

▪ In 100 adults with DM, smartphone thermography + AI detected all nurse-identified moderate-to-high risk feet (100% sensitivity) with high specificity (96.8%).

▪ Clinically, it looks promising as a scalable adjunctespecially to rule out risk and triage follow-upbut needs longitudinal validation before it can be treated as outcome-predictive.