Diabetes, Lower-Extremity Amputation, and Death

Ole Hoffstad; Nandita Mitra; Jonathan Walsh; David J. Margolis


Diabetes Care. 2015;38(10):1852-1857. 

In This Article

Research Design and Methods

This is a cohort study of patients from The Health Improvement Network (THIN). THIN data source contains anonymized data collected from the electronic medical records used by General Practice (GP) physicians in the U.K. This data source was chosen because diabetes is a chronic illness. THIN is rich in prospectively collected primary care medical information and has been used in the past to study diabetes and chronic wounds.[21–24] Ethics approval for this study was from the Scientific Review Committee U.K. and from the institutional review board from the University of Pennsylvania.

Inclusion Criteria

All individuals were cared for by a GP whose practice was up to THIN standards for data collection and contributed information to THIN during the years 2003–2012. All subjects had diabetes and at least 1 year of follow up and were at least 25 years of age at the time of diagnosis with diabetes. The diabetes diagnosis was based on previously used coding algorithm(s) and has a positive predictive value of 98%.[24–26] In addition, for ensuring that all subjects had diabetes, they had to have had at least two visits for diabetes at least 2 months apart. Our primary exposure of interest was any LEA. An LEA was thought to be incident if it was not preceded by a code for LEA in the previous 6 months. Because we were interested in death associated with LEA, any individual who died within the first 2 weeks of the procedure were excluded from study, as their death was thought to be due to preexisting sepsis or due to the surgical event itself. Because it can be difficult to ascertain whether repeated codes for LEA are reporting a new LEA or previous LEA, our primary evaluation of the time to death was from the patient's first LEA. Because coding for LEA can be very broad or quite specific, our primary analysis was based on any code for LEA. However, we also conducted additional analyses using codes for "major amputation" (defined as transtibial or higher). Our coding algorithms have previously been described and implemented.[27]

Additional Study Variables

The outcome of interest was all-cause death. Our primary clinical risk factors of interest were selected a priori because they were known to be associated with the most frequent causes of death in those with diabetes. Many of these risk factors are also known to be associated with an increased risk of LEA. Our primary "risk factor variables" included a history of cardiovascular disease such as a history of MI, cerebrovascular accident (CVA), congestive heart failure (CHF), and peripheral vascular disease/arterial insufficiency (PVD); the Charlson comorbidity index (an index that is used to predict mortality); and a history of chronic kidney disease (CKD) as determined by estimate glomerular filtration rate. Coding algorithms for these risk factors have previously been reported.[21,24,25,28] We also evaluated a history of malignancy, history of cigarette smoking, sex, HbA1c, and age over 65 years. In addition, we accounted for practice-level variability.


The goal of this study was to determine whether complications of diabetes known to be associated with death in those with diabetes, such as cardiovascular disease and renal failure, fully explain the higher rate of death in those who have undergone an LEA. All variables were first evaluated descriptively in the full diabetes cohort, among those with LEA, and among those who died. We then created explanatory models to estimate the effect of LEA on death using Cox proportional hazards models, allowing the effect of LEA to be time varying. In all models, death was our primary outcome and LEA was the primary exposure. Additional covariates and potential confounders were added to the model to determine whether they diminished the effect of LEA on death.[29]

We hypothesized that, since LEA is a treatment on the causal pathway to death, the pathway was initiated by another medical risk factor. We therefore expected that when these "risk factor" variables were added to the proportional hazard models, the association of LEA with death would be diminished (i.e., the hazard ratio [HR] would move toward 1). We hypothesized that this statistical process would show that LEA was a surrogate marker for these other factors that were already known to be the most frequent causes of death in those with diabetes. This analysis was modeled after a conceptual framework developed by Prentice.[29] In his classic description of a surrogate marker, Prentice proposed that a true surrogate must yield a valid test of the null hypothesis of no association (i.e., HR = 1) between the treatment and the true response.[29] In our setting, the treatment is LEA and the outcome is death. To simplify, we hypothesized that known risk factors for death in those with diabetes, like a history of MI (i.e., the leading cause of death in those with diabetes), should be able to fully explain the association of LEA with death. Since many of these risk factors are also predictors of LEA, many of these associations may also represent confounding, which is often thought to be statistically important if it changes the primary variable's effect estimate by >10–15%.[7]

We report HRs for each variable with respect to all-cause death (i.e., unadjusted models), for LEA by itself, and individually with each covariate (i.e., LEA and an additional variable) and for all covariates together (i.e., a fully adjusted model). The proportionality assumption was examined using diagnostic complimentary log-log plots. The fully adjusted Cox models were also evaluated for their ability to predict death using the area under the receiver operator curve (AUC), a statistic often used to describe the ability of a model to discriminate. Finally, sensitivity analyses were performed to determine the magnitude of an unknown risk factor that would be needed to render our findings no longer significant.[30] Analyses were conducted using STATA 13.1 and R 3.3.0.