Retrospective machine learning models predict amputation risk in diabetic
A separate model was built to account for patients who died before having an amputation. Both models were combined in a public online calculator. The study looked back at patient records from four hospitals in China. It covered January 2018 to September 2024, with follow-up ending in September 2025. Overall, 14.0% of patients had an amputation within 12 months. The findings suggest that machine-learning predictions may be useful alongside traditional methods for estimating risk. However, the calculator is still an additional research tool and has not been tested in a real-world clinical setting through a study that follows patients forward in time.
The patients had both lower extremity arterial disease and diabetes. Lower extremity arterial disease occurs when arteries supplying blood to the legs become narrowed or blocked. The patients had an average age of 72.7 years, and 65.7% were men. The researchers divided the patients into three groups: 967 for training the models, 414 for internal validation and 538 for external validation.
During the standard 12-month follow-up period, 269 patients had a major or minor amputation, 118 died before an amputation without losing a limb, and 1,532 were censored, meaning their full outcomes were not available for the analysis. Amputation rates were 14.6% in the training group, 13.3% in the internal-validation group and 13.6% in the external-validation group. Death before amputation matters because a standard model may simply count these patients as people who did not have an amputation. This can lead to an inaccurate estimate of the chance that a patient will have an amputation while still alive and at risk.
To avoid using information from the validation groups when choosing predictors, the researchers screened possible predictors using only the training group. A review of published research first identified 75 possible factors linked to amputation. The researchers then reduced this list to 48 variables based on whether the information was available in clinical practice and whether the records were complete. They used three methods to help select the most useful factors: LASSO logistic regression, a random forest classifier and the Boruta algorithm.
A factor was kept if at least two of the three methods supported its use. This produced 14 predictors: Rutherford classification, financial burden, age, Barthel Index, diabetic foot, statin use, fibrinogen, neutrophil count, residence, prealbumin, C-reactive protein, activated partial thromboplastin time, body-mass-index classification and the presence of a wound. Both the machine-learning model and the competing-risk model used the same 14 predictors. This allowed the researchers to compare their performance using the same patient population and 12-month prediction period. They trained 10 algorithms, adjusted their settings to improve performance and assessed how well they separated higher-risk from lower-risk patients, how closely their predictions matched actual outcomes and other measures of accuracy.
The 10 algorithms were random forest, logistic regression, decision tree, naive Bayes, artificial neural network, k-nearest neighbors, XGBoost, gradient boosting, CatBoost and light gradient boosting machine. The GraBoost model, a type of gradient-boosting algorithm, performed best in internal validation. Its AUC was 0.860, accuracy was 0.787, F1 score was 0.782 and Brier score was 0.156. In the separate external-validation group, its AUC was 0.828, accuracy was 0.760, F1 score was 0.695 and Brier score was 0.178.
Eight algorithms had external AUC values above 0.70, while decision tree and k-nearest neighbors gave less consistent results. DeLong tests showed statistically significant differences between GraBoost and most other models in internal validation, although some differences in external validation were not statistically significant. These results show how well the models distinguished between risk levels and how well their predictions matched outcomes in the study groups. They do not prove that using the model will improve medical decisions or patient outcomes.
SHAP analysis, a method that shows how much each input contributes to a prediction and in which direction, found that Rutherford classification had the greatest influence on the model's predictions. Financial burden and age were the next most influential factors. Barthel Index, diabetic foot, statin use, fibrinogen, neutrophil count, residence and prealbumin had moderate effects. C-reactive protein, activated partial thromboplastin time, body-mass-index classification and wounds had smaller effects. Higher Rutherford grades, lower Barthel Index scores, younger age, wounds and diabetic foot were linked to higher predicted amputation risk.
Higher fibrinogen, activated partial thromboplastin time, C-reactive protein and neutrophil levels also pushed predictions toward higher risk. Higher prealbumin levels were linked to lower predicted risk. Living in an urban area, having a lower financial burden, having a normal body-mass-index classification and using statins were linked to lower predicted risk. These are patterns found in an observational study. They do not prove that changing any one of these factors would cause amputation risk to fall.
The Fine-Gray competing-risk model used the same 14 predictors. It treated amputation as the main event and death from any cause before amputation as a competing event. Its time-dependent AUC at 12 months was 0.762 in the training group, 0.711 in internal validation and 0.811 in external validation. The corresponding Brier scores were 0.092, 0.106 and 0.085.
Calibration curves for outcomes at 3, 6 and 12 months showed that the predicted and observed cumulative chances of amputation were generally in agreement across different risk groups. Gray's test found no statistically significant difference in the cumulative rate of amputation at 12 months among the three groups, with p = 0.722. The internal-validation group was relatively small, which led to wide confidence intervals for its time-dependent AUC and made the model's ability to distinguish between risk levels less certain. The researchers also created a nomogram, a chart that assigns points to each of the 14 predictors and converts the total into estimated cumulative amputation risk at 3, 6 and 12 months while accounting for competing risks.
Some of the results challenge common assumptions about risk. Younger age was linked to higher predicted amputation risk, even though arterial disease generally becomes more common as people get older. The researchers suggested possible explanations. For example, doctors might be more likely to choose surgery for younger patients with severe limb problems, or some younger patients might have a more aggressive form of the disease. However, they stressed that these explanations are only possibilities and have not been proven.
The researchers also said they could not rule out the effects of other factors that were not measured or fully accounted for in the study. The link between living in a rural area or town, having greater financial burden and having an amputation might reflect differences in access to healthcare, delays in treatment or difficulties following medical advice. However, the study's retrospective design cannot establish whether these factors explain the observed patterns. The model also showed a U-shaped pattern: both underweight patients and those in higher body-mass-index categories had higher predicted risk.
Being underweight may reflect poor nutrition, loss of muscle or a greater burden of illness. A higher body mass index may be linked to metabolic problems and inflammation. However, the study did not test whether these processes cause amputation. Wounds and diabetic foot were also linked to higher predicted risk, although their influence on the model was smaller than that of Rutherford classification, financial burden and age. These conditions can be associated with poor blood flow to tissues, slow healing and infection, but the model cannot explain why a particular patient developed a wound or diabetic foot.
Statin use was linked to lower predicted risk, while higher fibrinogen levels and a longer activated partial thromboplastin time were linked to higher risk. The researchers discussed these laboratory findings in relation to inflammation, blood clotting and tissue healing. They also highlighted the importance of the Barthel Index, which measures a person's ability to carry out everyday activities. A lower score may reflect reduced mobility or less ability to manage one's own care. This could delay the detection or treatment of foot injuries, although the study did not directly test this explanation.
These biological and social explanations are possible reasons for the patterns seen in the model, not proven causes. The researchers made both models available through an online calculator built with Shiny. Users can enter patient information and receive matching 12-month risk estimates from the binary GraBoost model and the Fine-Gray competing-risk model. The calculator uses dropdown menus for categories and sliders for continuous measurements, allowing users to enter clinical details and calculate both estimates.
The authors described the 14 predictors as measurements that can usually be collected in hospitals at different levels, suggesting that gathering the required information should be practical. Showing both estimates is intended to provide two views of risk: one from the binary machine-learning model and another from a model that directly accounts for death before amputation. However, the online calculator has not been tested prospectively in routine medical care.
The calculator cannot replace a full clinical assessment. The study did not test whether using it changes patient monitoring, procedures to restore blood flow, medication prescribing, limb outcomes or survival. Several limitations affect how the findings should be understood. Because the study looked back at existing records and was observational, its results may be affected by selection bias, missing information and other factors that were not measured.
The three groups differed in their starting characteristics, including Rutherford classification, Barthel Index, residence, statin use and other variables. All participants were Chinese patients treated at hospitals in China, so the results may not apply to people from other ethnic backgrounds or to different healthcare systems. The researchers also used multiple imputation, a method for estimating missing values, on the combined dataset before dividing it into groups. This may have allowed some information to leak between groups, even though the researchers restricted predictor selection to the training group.
Patients who were lost to follow-up also differed significantly from the overall group in most starting characteristics, suggesting that people who dropped out were not a random selection of participants. The models were designed to predict amputation risk over 12 months, so they cannot automatically be expected to predict longer-term outcomes. The observed amputation rate of 14.0% was also within the 13%–48% range reported for similar patient groups. Differences between studies may be due to differences in who was included, how severe their disease was, the healthcare systems involved and how follow-up was defined.
Overall, the study shows promising results when the models were tested in a separate group of patients with a specific set of medical conditions. However, the tools need to be tested prospectively across multiple hospitals before they can be used routinely in clinical care.
Subject of Research: Development and external validation of 12-month amputation risk models for patients with lower extremity arterial disease and diabetes
Article Title: Development and validation of a machine learning-based model for predicting amputation risk in patients with lower extremity arterial disease and diabetes
Article References: Tang, L., Wang, L., Li, J., Liu, X., Chen, Q., Song, P., Quan, X., Li, Q., Liu, G., Liu, Y., Cui, C., & Shi, M. (2026). Development and validation of a machine learning-based model for predicting amputation risk in patients with lower extremity arterial disease and diabetes. Frontiers in Cardiovascular Medicine, 13, Article 1893875. [https://doi.org/10.3389/fcvm.2026.1893875](https://doi.org/10.3389/fcvm.2026.1893875)
Image Credits: AI Generated
DOI: 10.3389/fcvm.2026.1893875
Keywords: amputation risk, lower extremity arterial disease, diabetes, machine learning, gradient boosting, competing risks, external validation, clinical prediction, SHAP interpretability, web calculator, Development, validation