Applying mobility prediction models to real-world patients with major amputations.

Background: Outcome prediction models (PMs) have become commonplace and are promoted to aid in counseling patients. The aim of this study is to evaluate the performance of existing mobility PMs for post-major amputation (MA) patients in a real-world, socioeconomically disadvantaged population.

Methods: A retrospective review of patients with MA secondary to peripheral arterial disease from 2016 to 2022 was performed. Patients who were nonambulatory before MA or with contralateral MA were excluded. Three published PMs were investigated: (1) AmpPredict (predicts 1-year mobility), (2) Amputee Single Item Mobility Measure, (predicts degree of mobility with prosthesis at 1 year), both derived from Veteran's Affairs data, and (3) a Vascular Quality Initiative (VQI) data-derived model (predicts 1-year mobility). Predicted mobility rates vs actual mobility rates were compared.

Results: The study cohort consisted of 126 patients, 71% male, 60% non-White race, with a mean state Area Deprivation Index of 9 of 10. Baseline characteristics were significantly different between the study and derivation cohorts. Actual mobility at 1 year was 43%. Of the 38 patients with an AmpPredict 1-year mobility of ≥70%, 45% actually achieved mobility. Of 101 patients with a high predicted probability from the VQI score (≥71%), 48% achieved mobility. The mean difference between AmpPredict and VQI for a given patient was 36% (range, 1%-81%). The Amputee Single Item Mobility Measure predicted 87% of patients would be community (vs home) ambulators at 1 year and 32% of patients actually achieved community ambulation (sensitivity of 91%, specificity of 14%, positive predictive value of 33%, and negative predictive value of 79%).

Conclusions: Published models dramatically overestimated the likelihood of mobility in our patient cohort. This result may be related to the demographics and comorbidities of our cohort being significantly different from the derivation cohorts. We recommend caution when applying PMs to a population with significantly different characteristics from the population used to derive the model.