Interpretable machine learning for predicting major amputation risk in hospitalized diabetic foot ulcer patients: a single-center study with temporal external validation.
A random forest model using routine admission variables achieved an area under the receiver operating characteristic curve of 0.977 in internal testing and 0.984 in temporal validation for predicting major lower-extremity amputation in hospitalized diabetic foot ulcer patients.
Key Findings
Results
The random forest model demonstrated the best overall discrimination among the four compared models for predicting major amputation in hospitalized diabetic foot ulcer patients.
Area under the receiver operating characteristic curve (AUC) of 0.977 in internal testing
AUC of 0.984 in temporal external validation using a 2024 cohort
Models compared included logistic regression, elastic net, random forest, and extreme gradient boosting
The model also showed acceptable calibration in both cohorts
Methods
The study used a temporal external validation design, developing models in a 2019-2020 cohort and validating them in a later 2024 cohort.
Retrospective review of consecutive admissions for diabetic foot ulcers at a single center
Development cohort was from 2019-2020 and temporal validation cohort was from 2024
The outcome was defined as in-hospital major lower-extremity amputation above the ankle
This single-center study design limits generalizability but allows temporal validation of model stability
Results
The most influential predictors of major amputation reflected limb perfusion and infection severity.
Top predictors included perfusion grade, ankle-brachial index, maintenance dialysis, white blood cell count, C-reactive protein, and prior minor amputation
All candidate predictors were routinely available admission variables collectible within 24 hours
Predictors included comorbidities, bedside limb/ulcer assessment, and standard laboratory tests
Shapley additive explanations (SHAP) were used to provide patient-level interpretability of predictor contributions
Methods
Shapley additive explanations (SHAP) were applied to provide interpretability at the individual patient level.
SHAP values were used to explain the contributions of individual predictors for each patient
This approach addresses the 'black box' nature of ensemble machine learning models such as random forest
Patient-level interpretability was identified as a key feature supporting clinical use for risk stratification
Conclusions
The authors concluded that an explainable admission-data model can support early inpatient risk stratification for major amputation in diabetic foot ulcer patients.
The model relies solely on variables available within 24 hours of admission
The authors suggest the model 'may help prioritize timely multidisciplinary care'
Early identification of high-risk patients was described as 'challenging' prior to this work
Diabetic foot ulcers were described as 'a leading cause of non-traumatic lower-limb amputation'
What This Means
This research suggests that a type of machine learning algorithm called a random forest, trained on routine information collected within the first 24 hours of a hospital admission, can very accurately predict which patients admitted for diabetic foot ulcers are at high risk of needing a major amputation (removal of a limb above the ankle). The model was developed using patient data from 2019-2020 and then tested on a separate group of patients from 2024, and it performed very well in both groups, correctly distinguishing high-risk from low-risk patients over 97% of the time. The most important factors the model used were related to blood flow in the limb (such as a measurement called the ankle-brachial index and a perfusion grade) and signs of infection (such as white blood cell count and C-reactive protein), as well as whether the patient was on dialysis or had a previous minor amputation.
A key feature of this study is that the researchers used a technique called SHAP (Shapley additive explanations) to make the model's predictions understandable for individual patients, showing which specific factors were driving the risk estimate for each person. This matters because many powerful machine learning models are difficult to interpret, which can make clinicians reluctant to trust or use them. By making the model transparent, the researchers aimed to make it more practical for real clinical settings.
This research suggests that such a tool, applied at hospital admission, could help clinical teams quickly identify patients who need urgent attention and prioritize them for multidisciplinary care — including vascular surgeons, infectious disease specialists, and wound care teams — potentially improving outcomes. However, the study was conducted at a single center, so further validation across different hospitals and patient populations would be needed before the model could be broadly applied.
Zou M, Ju S. (2026). Interpretable machine learning for predicting major amputation risk in hospitalized diabetic foot ulcer patients: a single-center study with temporal external validation.. Frontiers in endocrinology. https://doi.org/10.3389/fendo.2026.1821550