A foot-amputation AI scored 0.83 in another hospital—why isn’t it clinic-ready?
A four-hospital Chinese study predicted 12-month amputation among 1,919 people with diabetes and lower-extremity arterial disease. External validation is useful, but retrospective design, pooled imputation and table inconsistencies block clinical adoption.
The short version
- What happened
- The retrospective study included 1,919 patients at four Chinese hospitals: 967 for training, 414 for internal validation and 538 at another hospital for external validation.
- Who it affects
- People with diabetes and arterial disease need rapid vascular and multidisciplinary foot assessment; a risk score must not delay urgent referral, revascularisation or treatment of infection.
- What remains unproven
- Retrospective records cannot show that model-guided care improves amputation, survival, healing or quality of life.
Keep in mind
Retrospective records cannot show that model-guided care improves amputation, survival, healing or quality of life.
Multiple imputation on the pooled dataset before cohort splitting may leak information across development and validation data.
The cohorts differed at baseline, all came from four Chinese hospitals and 98.1% of participants were recorded as Han.
The published baseline table contains apparent internal count and percentage inconsistencies that should be corrected before independent reuse.
The outcome combined major and minor amputation; performance for clinically different amputation levels was not established in the reported headline metrics.
Source analysis · 1 linked source. Primary analysed source date: 20 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The model was tested at a geographically separate hospital
Researchers reviewed 1,919 patients with diabetes and lower-extremity arterial disease treated at four hospitals in Sichuan between 2018 and 2024. The outcome was any major or minor amputation within 12 months. Of the full cohort, 269 people—14.0%—had an amputation, 118 died before amputation and 1,532 were censored without either recorded event.
Three hospitals supplied 967 training and 414 internal-validation records; a fourth supplied 538 external-validation records, including 73 amputations. After comparing ten algorithms, the paper selected a gradient-boosting model using 14 variables. Its area under the receiver-operating curve was 0.860 internally and 0.828 externally. A separate Fine–Gray model used the same predictors to account for death before amputation.[1]
Good discrimination is not proof of safer care
The model emphasised Rutherford stage, reported financial burden, age, functional status, wounds, diabetic-foot status, statin use and inflammatory or coagulation markers. Those inputs could help organise risk review, but the study did not compare model-guided care with ordinary multidisciplinary assessment, and it measured no change in revascularisation, healing, amputation, survival or quality of life.
The authors acknowledge that missing values were imputed on the pooled dataset before the train–validation split, creating potential information leakage. Cohorts also differed at baseline, the internal set was relatively small and all participants were Chinese; 98.1% were recorded as Han. The published baseline table contains apparent count and percentage inconsistencies. Prospective external validation, corrected tables, locked code and calibration in other health systems are needed before clinical use.[1]
The distinction that matters
What the study tested
The evidence supports risk-model validation, not a claim that using the calculator improves care.
Development
967 retrospective records
Fit 14-predictor models
Internal check
414 records from the same hospital group
AUC 0.860
External check
538 records from another hospital
AUC 0.828; 73 amputations