AMPREDICT MoRe: Predicting Mortality and Re-amputation Risk after Dysvascular Amputation

Objective

This study aimed to create a novel prediction model (AMPREDICT MoRe) that predicts death and re-amputation after dysvascular amputation, which overcomes prior implementation barriers by using only predictors that are readily available in the electronic health record (EHR).

Methods

This was a retrospective cohort study of 9 221 patients with incident unilateral transmetatarsal, transtibial, or transfemoral amputation secondary to diabetes and or peripheral arterial disease identified in the Veterans Affairs Corporate Data Warehouse between 1 October 2015 and 30 September 2021. The prediction model evaluated factors falling into several key domains: prior revascularisation; amputation level; demographics; comorbidities; mental health; health behaviours; laboratory values; and medications. The primary outcome included four categories: (i) no death and no re-amputation (ND/NR); (ii) no death and re-amputation (ND/R); (iii) death and no re-amputation (D/NR); and (iv) death and re-amputation (D/R). Multinomial logistic regression was used to fit one year post-incident amputation risk prediction models. Variable selection was performed using LASSO (least absolute shrinkage and selection operator), a machine learning methodology. Model development was performed using a randomly selected 80% of the data, and the final model was externally validated using the remaining 20% of subjects.

Results

The final prediction model included 23 predictors. The following outcome distribution was observed in the development sample: ND/NR, n = 4 254 (57.7%); ND/R, n = 1 690 (22.9%); D/NR, n = 1 056 (14.3%); and D/R, n = 376 (5.1%). The overall discrimination of the model was moderately strong (M index 0.70), but a deeper look at the c indices indicated that the model had better ability to predict death than re-amputation (ND/NR vs. ND/R, 0.64; ND/NR vs. D/NR, 0.78; grouped ND vs. D, 0.79 and NR vs. R, 0.67). The model was best at distinguishing individuals with no negative outcomes vs. both negative outcomes (ND/NR vs. D/R, 0.82).

Conclusion

The AMPREDICT MoRe model has been successfully developed and validated, and can be applied at the time of amputation level decision making. Since all predictors are available in the EHR, a future decision support tool will not require patient interview.