I
Reflections on new work from the International Flow and Toe Research Team (iFORT)
One of the most persistent challenges in limb preservation is not the first ulcer — it is the next one. Recurrence remains a defining feature of the disease, with up to 60% of people developing a new ulcer within three years of “healing.” This cycle drives avoidable hospitalizations, infections, and amputations, and it underscores the need for tools that help clinicians identify biological risk between episodes of active disease.
A new study from our combined iFORT team, published this month in BMJ Open Diabetes Research & Care, takes an important step toward meeting that need. The group developed an interpretable machine-learning model capable of predicting 3-year recurrence using readily available clinical variables. Their work highlights both the promise of machine learning and the continued importance of clinically grounded, explainable systems.
What the Team Did
Investigators analyzed 494 individuals with diabetic foot ulcers (DFU) drawn from three major centers in Southwest China. They used four complementary feature-selection methods to isolate a robust set of predictors, and then evaluated seven machine-learning algorithms — from logistic regression to XGBoost.
The XGBoost model performed best, achieving:
- AUROC: 0.924
- Brier score: 0.096 after calibration
- High interpretability via SHAP analysis
These are exceptionally strong values for a clinical prediction problem as complex as ulcer recurrence.
What the Model Learned
The model identified ten strongly associated, clinically intuitive contributors to recurrence risk:
- Smoking burden
- Age
- Hemoglobin (anemia)
- Ischemia
- HbA1c
- BMI
- Length of hospital stay
- LDL-cholesterol
- Creatinine
- White blood cell count
Together, these features paint a picture many clinicians will recognize: a biologic loop involving hypoxia, inflammation, and metabolic stress. The SHAP interpretability layer helps visualize each patient’s unique risk profile — an essential step toward trust and real-world adoption.
Why This Matters
This work provides:
- A framework for building and validating DFU recurrence models using multimethod feature selection.
- A clinically interpretable tool that supports personalized recurrence-risk discussions and targeted prevention strategies.
- A direction for future multicenter work, especially in diverse populations and primary-care settings where DFU identification often begins.
As the community continues shifting focus from wounds to remission — the period between episodes of breakdown — tools like this can help clinicians match the right surveillance, offloading, or metabolic optimization strategy to the right patient at the right time.
Full Citation
Mou W, Shan W, Yu S, et al. (Armstrong, Deng) “Interpretable machine learning model for predicting recurrence in patients with diabetic foot ulcers.” BMJ Open Diabetes Research & Care. 2025;13:e005242. DOI:10.1136/bmjdrc-2025-005242.
Leave a Reply