JMIR Publications

ABSTRACT

Background:

Major amputation proximal to the ankle is a devastating outcome of diabetes-related foot disease. Decisions about source control, revascularization, limb salvage, and reconstruction begin at admission, when prognosis remains uncertain. The Wagner classification is routinely available but places many hospitalized patients in the same grade, and few prediction models for major amputation have undergone temporal validation.

Objective:

We aimed to develop and temporally validate an admission-time model for in-hospital major amputation using routinely collected data. Secondary objectives were to compare flexible machine learning algorithms with penalized logistic regression and to assess whether segmental limb pressures add prognostic information.

Methods:

We conducted a retrospective cohort study at a tertiary center in Taiwan from January 2015 through December 2025. Adults admitted with diabetes-related foot disease were divided by admission date into a development cohort (2015-2021) and a temporal test cohort (2022-2025). Twenty admission variables were modeled. Five algorithms were compared by nested cross-validation in the development cohort; the algorithm with the lowest out-of-fold log loss was designated the primary model, and its operating thresholds were fixed before one-time evaluation in the temporal test cohort.

Results:

The development cohort included 1400 patients with 126 major amputations (9.0%), and the temporal test cohort included 771 patients with 51 major amputations (6.6%). Ridge logistic regression was selected as the primary model. In the test cohort, its area under the receiver operating characteristic curve (AUROC) was 0.882 (95% CI 0.841-0.917), compared with 0.804 (95% CI 0.737-0.867) for Wagner grade alone, which performed identically as a categorical factor or a linear term (paired difference +0.078, 95% CI 0.012-0.142). The calibration slope was 1.33, and the observed-to-expected ratio was 0.69, indicating moderate overprediction. At the prespecified screening threshold, sensitivity was 0.980, specificity was 0.593, and negative predictive value was 0.998. No amputation occurred among 396 patients (51.4%) in the lowest risk band (0%, 95% CI 0%-1.0%), whereas 22 of 54 patients (40.7%) in the highest risk band underwent amputation. Random forest and gradient boosting achieved higher test AUROCs (0.915 and 0.912; paired differences +0.033, 95% CI 0.007-0.060, and +0.030, 95% CI 0.005-0.059), but these were secondary test-cohort comparisons. In complete cases, segmental limb pressures did not clearly improve discrimination (change in AUROC -0.004, 95% CI -0.044 to 0.035).

Conclusions:

A transparent model based on routinely collected admission data stratified the risk of in-hospital major amputation beyond Wagner grade alone, but moderately overpredicted risk in the later cohort. The model may support early multidisciplinary assessment and prognostic communication; it must not be used to indicate amputation or to withhold vascular or limb-salvage evaluation. External validation, local recalibration, and prospective evaluation are required before clinical use. Clinical Trial: Not applicable; this study was not registered.