Mortality prediction following non-traumatic amputation of the lower extremity

Abstract

Background:

Patients who undergo lower extremity amputation secondary to the complications of diabetes or peripheral artery disease have poor long-term survival. Providing patients and surgeons with individual-patient, rather than population, survival estimates provides them with important information to make individualized treatment decisions.

Methods:

Patients with peripheral artery disease and/or diabetes undergoing their first unilateral transmetatarsal, transtibial or transfemoral amputation were identified in the Veterans Affairs Surgical Quality Improvement Program (VASQIP) database. Stepdown logistic regression was used to develop a 1-year mortality risk prediction model from a list of 33 candidate predictors using data from three of five Department of Veterans Affairs national geographical regions. External geographical validation was performed using data from the remaining two regions. Calibration and discrimination were assessed in the development and validation samples.

Results:

The development sample included 5028 patients and the validation sample 2140. The final mortality prediction model (AMPREDICT-Mortality) included amputation level, age, BMI, race, functional status, congestive heart failure, dialysis, blood urea nitrogen level, and white blood cell and platelet counts. The model fit in the validation sample was good. The area under the receiver operating characteristic (ROC) curve for the validation sample was 0·76 and Cox calibration regression indicated excellent calibration (slope 0·96, 95 per cent c.i. 0·85 to 1·06;intercept 0·02, 95 per cent c.i. −0·12 to 0·17). Given the external validation characteristics, the development and validation samples were combined, giving a total sample of 7168.

Conclusion:

The AMPREDICT-Mortality prediction model is a validated parsimonious model that can be used to inform the 1-year mortality risk following non-traumatic lower extremity amputation of patients with peripheral artery disease or diabetes.

Introduction

Shared decision-making at the time of amputation surgery may enhance postamputation adjustment and recovery¹. Survival², along with other key outcomes such as quality of life¹, mobility^3,4, and the risk of failure of amputation healing and need for additional surgery⁵, have been shown to be important outcomes that should be considered.

Operative and longer-term survival for patients undergoing lower extremity amputation because of critical limb ischaemia are poor^2,6,7 and mortality risks exceed those of the majority of cancer diagnoses^8,9. Communicating future mortality risk for individual patients is therefore a critical component of the amputation decision-making process².

Many studies have identified preoperative and perioperative factors associated with operative mortality¹⁰, and the average mortality risk associated with a given amputation level¹¹. Average risks, although important in understanding the outcomes of populations, are of limited value in shared amputation decision-making because they do not provide information about individual-patient risk. In contrast, prediction models can play an important role in the provision of individualized care, and shared decision-making^12,13. Previous mortality prediction models associated with amputation owing to the complications of diabetes or peripheral artery disease (PAD) have all been developed to predict operative risk^14–16. None provides an estimate of longer-term survival that might better assist patients in making value-concordant treatment decisions.

The purpose of this research was to present the development and validation of a 1-year mortality risk prediction model. The model was developed in patients who require an incident unilateral lower extremity amputation at the transmetatarsal (TM), transtibial (TT) or transfemoral (TF) amputation level secondary to diabetes and/or PAD.

Methods

Administrative, quality improvement and clinical data from two primary sources were used in accordance with procedures approved by the participating institution’s human subjects review board.

Veterans Affairs Surgical Quality Improvement Program

The Veterans Affairs Surgical Quality Improvement Program (VASQIP) database was used to define the inception cohort as well as several preamputation risk predictors. It includes information on 30-day surgical outcomes, and preoperative, perioperative and postoperative co-variables from 110 Veterans Affairs Medical Center inpatient surgical programmes. VASQIP is a surgical quality improvement data set developed to monitor the quality of surgical care in the Veterans Affairs (VA) Health Care System. Data are collected on approximately 70 per cent of all major operations and about 25 per cent of all operations in the VA Health Care System^17,18. The calendar is subdivided into consecutive 8-day cycles, each starting on a different day of the week. In each 8-day cycle, the first 36 consecutive eligible surgical patients in that cycle are entered into VASQIP. Eligible non-cardiac procedures include those performed by a physician that require general, spinal or epidural anaesthesia.

Corporate Data Warehouse

The VA Corporate Data Warehouse (CDW) includes inpatient and outpatient data as well as demographic information. Data from the CDW were used to assess whether individuals had previously undergone an amputation procedure (and were thus ineligible for the present study) and to acquire additional predictor variables (such as weight and laboratory values) that were not available through VASQIP. The CDW’s Vital Status File was used to determine date of death of those who died¹⁹.

Study sample

The target population was patients undergoing their first unilateral TM, TT or TF amputation assumed to be secondary to diabetes and/or PAD (based on ICD-9-CM codes; diabetes: 249.7, 250.7, 785.4, 443.81, 785.4, 249.8, 250.8, 707.1, 707.11, 707.12, 707.13, 707.14, 707.15, 707.19; PAD: 440.22, 440.23, 440.24, 440.4, 442.3, 444.22). VASQIP subjects were included in the study cohort if they were aged 40 years or older (to ensure that the study population was restricted to those undergoing amputations related to diabetes and/or PAD¹⁵) and had undergone an incident major unilateral lower extremity amputation, defined as TM (Current Procedures Terminology (CPT): 28800, 28805; ICD-9: 84.12), TT (CPT: 27880, 27881, 27882, 27888, 27889; ICD-9: 84.14, 84.15) or TF (CPT: 27590, 27591, 27598; ICD-9: 84.16, 84.17, 84.18) amputation between 1 October 2004 and 31 December 2014. Those who had a record of a previous major lower extremity amputation in the CDW occurring between 2 days and 5 years before the VASQIP operation were excluded. Subjects were also excluded if they had a preoperative diagnosis of coma, paraplegia, quadriplegia, disseminated cancer, a tumour of the central nervous system, or were ventilator-dependent, because these conditions result in atypical risks of death which may more commonly lead to a higher level of amputation. It was presumed that patients undergoing a guillotine procedure at the TT (CPT: 27881; ICD-9: 84.13) and TF (CPT: 27592) levels would have a closure procedure within 3weeks of the guillotine procedure; therefore, research staff searched forward 3 weeks for the next procedure code to classify the incident amputation level. If the next subsequent procedure occurred more than 3 weeks after the initial procedure, it was presumed that the initial guillotine code was an error, and the initial guillotine procedure was accepted as a definitive level of amputation and any subsequent procedure as a reamputation. For those coded as having guillotine procedures, without a subsequent closure procedure from VASQIP or CDW, VA Informatics and Computing Infrastructure chart annotation services were used to identify the definitive closure level and laterality.

Outcome

The primary outcome of this study was death within 1 year of index amputation.

Candidate predictor variables

The databases were used to retrieve 50 potential predictor variables, identified a priori from evidence in the literature and expert clinical opinion, based on an informal process that included four epidemiologists (1 an expert in VASQIP data, another in CDW data and 2 with amputee research experience), a physical medicine physician with national recognition in amputation care and research, a vascular surgeon, an internal medicine physician with clinical and research expertise in diabetic foot ulceration and amputation risk, and a rehabilitation psychologist. Thirty-three variables were ultimately considered as candidates in developing a predictive model (Table 1), after dropping or combining 17 variables because of low frequency of occurrence, difficulty in clinical measurement, or because they were combined with other variables (Table S1, supporting information). When VASQIP laboratory values were missing, the nearest CDW value within 3 months before the date of surgery was used; missing values for other VASQIP predictors that were also recorded in the CDW were dealt with similarly. To calculate BMI, the median CDW value for height and the VASQIP recorded value for weight were used, as these are documented just before surgery in the hospital setting. Patients were excluded if their height, weight or BMI was considered implausible (less than 1–2 or more than 2·1 m; below 34 or above 318 pounds²⁰; and less than 15 or over 52 kg/m²) because these represent patients who may be severely malnourished or obese and may be more likely to be a candidate for a TF amputation, and/or there has been a data entry error.

Table 1.

Candidate predictors included in the AMPREDICT 1-year mortality full prediction model, stratified by development and validation samples

		% of patients*
		Development sample (n = 5028)	Validation sample (n = 2140)	Total (n = 7168)
Amputation level
Transmetatarsal	20·6	21·9	21·0
Transtibial	45·1	46·4	45·5
Transfemoral	34·3	31·8	33·5
Demographics
Age (years)†	68(11)	68(10)	68(11)
Male sex	98·9	99·4	99·1
Married	40·3	41·8	40·8
Race
Caucasian	57·2	63·4	59·1
Black	34·1	25·3	31·5
Hispanic	8·3	9·3	8·6
Other	0·4	1·7	0·8
Co-morbidities
Coronary atherosclerosis	50·7	47·3	49·7
Hypertension requiring medications	85·8	85·4	85·7
Previous cardiac surgery	21·0	20·7	20·1
Previous PTCA or PCI	13·5	12·1	13·1
Any past diagnosis of CHF	35·6	34·9	35·4
Diabetes with oral agents or insulin	62·5	62·9	62·6
Dialysis in 2 weeks before operation	13·0	12·3	12·8
History of COPD	18·6	18·2	18·5
Dyspnoea (minimal or at rest)	18·5	19·2	18·7
Stroke with or without neurological deficit	21·4	22·2	21·7
Health factors
Smoker within 1 year of operation	39·5	39·6	39·5
> 2 alcoholic drinks per day in 2 weeks before admission	8·1	7·6	8·0
Illicit drug use within 1 year of operation	12·0	10·0	11·2
Nutritional status
> 10% weight loss in 6 months before operation	9·9	9·2	9·7
BMI (kg/m2)†	26(6)	27(6)	26(6)
Mental health
Any mental health diagnosis‡	42·1	40·5	41·6
Physical function
Independent	51·9	51·4	51·8
Partially dependent	35·8	39·4	36·9
Totally dependent	12·3	8·7	11·3
Medication
Use of beta-blockers	42·4	43·4	42·7
Outpatient anticoagulation using warfarin	13·6	11·1	12·9
Outpatient antiplatelet medication	17·2	17·2	17·2
Preoperative laboratory values
Blood urea nitrogen (mg/dl)†	24(17)	24(17)	24(17)
WBC count (per μl)
< 11 000	58·2	58·1	58·2
≥ 11 000	41·8	41·9	41·9
eGFR (ml per min per 1·73 m2)
< 15	10·4	10·3	10·4
15–29	7·8	8·3	8·0
30–59	23·6	25·6	24·2
60–89	28·8	30·9	29·5
≥ 90	29·3	24·9	28·0
Platelet count (× 106/ml)†	334(139)	333(145)	334(140)
Potassium (mEq/l)†	4(<1)	4(< 1)	4(< 1)
Haematocrit (%)†	32(5)	32(5)	32(5)
Vascular/limb status
Rest pain/gangrene in 30 days before operation	75·1	71·0	73·9
Open wound/wound infection	78·9	82·8	80·1

^*Unless indicated otherwise;

^†values are mean(s.d.).

^‡Any diagnosis of depression, anxiety, post-traumatic stress disorder, bipolar disorder or schizophrenia.

PTCA, percutaneous transluminal coronary angioplasty; PCI, percutaneous coronary intervention; CHF, congestive heart failure; COPD, chronic obstructive pulmonary disease; WBC, white blood cell; eGFR, estimated glomerular filtration rate.

Model development and validation samples

Geographical validation was used to validate the prediction model externally. The prediction model was developed using three of the five regions within VA (East, South and Midwest), and validated in the Mountain plus Texas and West regions. The selection of regions was based both on geography (dividing the USA nearly in half) and on sample size, allowing a larger sample size in the development (training) cohort.

Statistical analysis

The 33 candidate predictors were evaluated rigorously through exploratory data analysis followed by bivariable analysis, analysing the association between each predictor and 1-year mortality. There were few missing data for these candidate variables and so a complete-case analysis was used as opposed to other approaches, such as multiple imputation of missing values. Interactions were not considered among the candidate predictors because biological plausibility could not be justified, nor, after a thorough review of the literature, could any evidence for such interactions be identified. For all continuous measures, non-linearity of the association with (logit) risk of 1-year mortality was explored graphically using non-parametric smoothing and by comparison of linear with fractional polynomial models²¹.

A full logistic regression model with all 33 candidate predictors was fitted to provide a reference for comparison with more parsimonious models. Variable selection using both backwards stepwise and stepdown²² logistic regression methods was considered. For the backwards stepwise variable selection approach, a P value cut-off of 0·157 was chosen, which has been shown to approximate the best subset of predictors using the Akaike information criterion²³. For the stepdown variable selection approach, models that explained 99 and 95 per cent of the variability in the risk predictions from the full model were considered. Calibration and discrimination of the fitted models were assessed separately in the development (training) and validation samples. Calibration was assessed by means of the Hosmer–Lemeshow (H–L) goodness-of-fit test²⁴ and Cox calibration regression^24–26. A plot of observed fraction of 1-year deaths versus the average of predicted risks of 1-year death for each decile of predicted risk was assessed visually. Discrimination was assessed quantitatively by calculating the area under the receiver operating characteristic (ROC) curve (AUC), the discrimination slope (the difference in mean predicted 1-year mortality risk for those who did and those who did not die within the first year) and the difference in mean estimated mortality risk in the highest and lowest deciles of predicted risk.

Results

Of 7495 eligible people, two with no CDW data were excluded, leaving 7493 (East, 1896; South, 1849; Midwest, 1459; Mountain plus Texas, 1139; West, 1150). Some 325 individuals were missing at least one observation and therefore excluded. The data set analysed included 7168 people with non-traumatic amputation owing to diabetes and/or PAD: 1504 TM amputees (21·0 per cent), 3261 TT amputees (45·5 per cent) and 2403 TF amputees (33·5 per cent) (Fig. 1, Table 1). This represented 95·6 per cent of those eligible.

Fig. 1.

STROBE diagram showing total numbers acquired by the Veterans Affairs Surgical Quality Improvement Program and total number excluded to achieve final inception cohort

Outcome

Some 1901 patients (26·5 per cent) died within the first year after incident amputation. The mortality risk increased by amputation level (TM, 17·7 per cent; TT, 24·7 per cent; TF, 34·5 per cent).

Risk prediction model development

The development sample consisted of 5028 subjects from the East, South and Midwest VA regions. Platelet count and serum potassium concentration were modelled in a non-linear manner using fractional polynomials. Estimated glomerular filtration rate and white blood cell counts were divided into clinically relevant categories. Other continuous predictors were modelled linearly and categorical factors as binary variables (yes/no) unless specified otherwise (Table 2).

Table 2.

Candidate predictors for the AMPREDICT 1-year mortality prediction model, stratified by survival status (7168 patients)

	% of patients*
	Survived (n = 5267)	Died (n = 1901)
Amputation level
Transmetatarsal	23·5	14·0
Transtibial	46·6	42·3
Transfemoral	29·9	43·7
Demographics
Age (years)†	66(10)	73(10)
Male sex	99·0	99·1
Married	40·2	42·5
Race
Caucasian	59·2	58·9
Black	32·0	29·8
Hispanic	7·5	10·9
Other	1·0	0·4
Co-morbidities
Coronary atherosclerosis	46·5	58·4
Hypertension requiring medications	85·6	85·8
Previous cardiac surgery	19·3	25·4
Previous PTCA or PCI	12·4	14·9
Any past diagnosis of CHF	30·1	49·8
Diabetes with oral agents or insulin	63·1	61·5
Dialysis in 2 weeks before operation	9·1	23·1
History of COPD	16·5	24·0
Dyspnoea (minimal or at rest)	16·7	24·1
Stroke with or without neurological deficit	19·4	27·8
Health factors
Smoker within 1 year of operation	42·9	30·0
>2 alcoholic drinks per day in 2weeks before admission	8·8	5·5
Illicit drug use within 1 year of operation	11·8	7·0
Nutritional status
>10% weight loss in 6 months before operation	8·7	12·6
BMI (kg/m2)†	26(6)	25(6)
Mental health
Any mental health diagnosis‡	42·6	38·8
Physical function
Independent	57·3	36·6
Partially dependent	35·0	42·1
Totally dependent	7·7	21·4
Medication
Use of beta-blockers	42·5	43·2
Outpatient anticoagulation using warfarin	12·6	13·5
Outpatient antiplatelet medication	17·7	15·7
Preoperative laboratory values
Blood urea nitrogen (mg/dl)†	22(15)	29(21)
WBC count (per μl)
< 11 000	60·0	53·0
≥11 000	40·0	47·0
eGFR (ml per min per 1·73 m2)
< 15	7·5	18·5
15–29	6·5	11·9
30–59	23·8	25·4
60–89	30·4	26·8
≥ 90	31·8	17·4
Platelet count (× 106/ml)†	343(142)	308(133)
Potassium (mEq/l)†	4(<1)	4(<1)
Haematocrit (%)†	32(5)	32(5)
Vascular/limb status
Rest pain/gangrene in 30 days before operation	72·6	78·0
Open wound/wound infection	79·8	80·7

^*Unless indicated otherwise;

^†values are mean(s.d.).

^‡Any diagnosis of depression, anxiety, post-traumatic stress disorder, bipolar disorder or schizophrenia.

PTCA, percutaneous transluminal coronary angioplasty; PCI, percutaneous coronary intervention; CHF, congestive heart failure; COPD, chronic obstructive pulmonary disease; WBC, white blood cell; eGFR, estimated glomerular filtration rate.

The backwards stepwise and 99 per cent stepdown variable selection resulted in the same model with 17 predictor variables, whereas the 95 per cent stepdown variable selection resulted in a model with ten predictor variables (a subset of the 17 predictors). The 17- and ten-predictor models were compared with the full model as the criterion standard. Both models performed similarly to the full model with respect to qualitative (graphical) and quantitative (AUC and discrimination slope) assessments (Table S2 and Fig. S1, supporting information). The ten-predictor model was chosen, as it was more parsimonious and would result in a lower clinical burden for the provider to obtain the information. The mean of the 1-year mortality risks predicted for each individual in the development sample using the ten-predictor risk prediction model was 26·5 (range 0·7–94·9) per cent. The estimated AUC was 0·77 and the H–L goodness-of-fit test for this model indicated good calibration (P = 0·523). The discrimination slope was 18·9 per cent.

Risk prediction model validation

The external validation sample consisted of 2140 subjects from the Mountain plus Texas and West regions (Table 1). The overall risk of death in the validation sample was 26·7 per cent. The mean predicted risk of 1-year mortality using the ten-predictor risk prediction model was 25·7 (range 0·6–95·1) per cent. Table 3 and Fig. 2 summarize observed versus predicted risks of 1-year mortality by deciles of predicted risk. The fit of the model to the validation sample was good, and this was further supported by the H–L goodness-of-fit test result (P = 0·283). The estimated AUC for the validation sample was 0·76 and Cox calibration regression yielded an estimated slope of 0·96 (95 per cent c.i. 0·85 to 1·06) and intercept of 0·02 (−0·12 to 0·17), again indicating good calibration (perfect calibration is represented by a slope of 1·00 and intercept of 0). The discrimination slope was 18·2 per cent. The difference in mean estimated risk of death in the highest versus lowest deciles of predicted risk was 62·3 per cent.

Table 3.

Goodness-of-fit assessment in external validation sample (2140 patients)

Decile of predicted risk	Range of predicted probability (%)*	n	Observed deaths	Expected deaths
1	0·6–6·5	214	13	9·6
2	6·6–9·5	214	20	17·1
3	9·6–12·3	214	19	23·1
4	12·4–15·7	214	32	30·0
5	15·8–19·8	214	51	38·0
6	19·9–25·1	214	44	48·1
7	25·2–32·0	214	60	61·0
8	32·1–41·9	214	83	78·9
9	42·0–55·2	214	108	102·0
10	55·3–95·1	214	141	143·0

^*From the ten-variable prediction model.

Fig. 2.

Observed versus predicted risk of 1-year death by decile of risk in the external validation sample

Combined risk prediction model

Given the strong external validation characteristics, the development and validation samples were combined (Table 4). Amputation level was associated with 1-year mortality, increasing amputation levels being associated with a greater risk of death. Other predictive factors associated with increased risk of death were older age, partially and totally dependent functional status, ever being diagnosed with congestive heart failure, being currently on dialysis, increasing blood urea nitrogen levels, and white blood cell counts of at least 11 000/μl. Individuals with greater BMI, and those who were black or in the ‘other’ category for race were at a decreased risk of death. There was a non-linear relationship between platelet levels and risk of death, with a rapid decrease in risk with increased platelet levels, until approximately 400 × 10⁶/ml, and then stabilizing thereafter. The modelled mortality risk equation in Table 4 can be employed easily for individual mortality risk calculations.

Table 4.

Mortality risk score coefficients for individual predictors in the ten-variable mortality prediction model (7168 combined development and validation samples)

	Coefficient
Baseline for transmetatarsal	−2·192
Baseline for transtibial	−1·897
Baseline for transfemoral	−1·717
Age (years)	+0·047 × (Age – 60)
BMI (kg/m2)	−0·050 × (BMI – 25)
Black race	−0·257
Race other than black/white/Hispanic	−0·985
Partially dependent functional status	+0·303
Totally dependent functional status	+0·955
Ever diagnosed with CHF	+0·545
Currently on dialysis	+0·900
BUN (mg/dl)	+0·016 × (BUN – 25)
WBC count ≥ 11 000/μl	+0·342
Platelet count (× 106/ml)	+1·720 × [(P/100)−0·5 − 0·577]

The mortality risk score (S) for an individual is the sum of the coefficients for all the components that apply to that individual. Predicted 1-year mortality risk = e^S/(1 + e^S). For example, the mortality risk score for a 70-year-old white transfemoral amputee, with a BMI of 20 kg/m², partially dependent functional status, no congestive heart failure (CHF) diagnosis or dialysis, with a blood urea nitrogen (BUN) level of 30 mg/dl, white blood cell (WBC) count less than 11 000/μl and platelet count (P) 300 × 10⁶/ml, is calculated as: S = −1·717 + 0·047 × (70 – 60) − 0·050 × (20 – 25) + 0·303 + 0·016 × (30 – 25) + 1·720 × [3^−0·5 – 0·577] = −0·6134. Predicted 1-year mortality risk = e^−0·6134/(1 + e^−0·6134) = 0·351; this represents a 35·1 per cent predicted risk of death in 1 year.

Discussion

A 1-year mortality risk prediction model, AMPREDICT-Mortality, was developed in a population of patients who experienced an incident lower extremity amputation at the TM, TT or TF levels secondary to diabetes and/or PAD. The model has good calibration characteristics and has been validated externally using VA regional data. The developed model can be used to inform surgeons and their patients about individual-patient mortality risk.

Patients facing amputation benefit from, and want more information about, the anticipated outcome of their surgery, and want to participate in treatment decisions¹. Mortality risk is a key outcome that should be considered at the time of amputation-level decision-making. Individuals who have undergone lower extremity amputation owing to diabetes and/or PAD have 1- and 5-year mortality risks that exceed those of most cancers; they can be as high as 44 and 77 per cent respectively²⁷. The limited lifespan of these patients creates an additional imperative to ensure that treatment decisions allow them best to achieve their personal outcome priorities and optimal quality of life, for their remaining life. A knowledge of individual mortality risk is therefore important in tailoring medical and surgical decisions²⁸. Knowing which patients have a high or low risk of death in the first year after amputation will help surgeons better assist patients in making end-of-life decisions and also making more informed treatment decisions that incorporate their values and goals²⁹.

Mortality risk is certainly not the sole factor to consider in the treatment decision. It is, however, an outcome that acts as a foundation for patients and surgeons to balance the risks of other key outcomes such as mobility³⁰, further ipsilateral amputation³¹, quality of life³, and the body image consequences of each amputation level³². For example, a patient with an 80 per cent risk of 1-year mortality may balance the risk of additional surgery owing to failure of healing of a TM amputation differently than if they had a 20 per cent risk of death.

There are a number of limitations of the study. With respect to the generalizability of the findings, because the model was developed and validated in a US population that comprised predominantly older Caucasian men, some caution is advisable when applying the model outside of the USA, and to women, Hispanic people and patients who fall into the ‘other’ race category such as Asians, Pacific Islanders and Native Americans. Patients aged less than 40 years were excluded to avoid including patients who may have been misclassified. The VASQIP data set does not include amputations that were performed with locoregional anaesthesia, and patients with such amputations were not therefore included in the present study cohort. However, the proportion of TM amputations in this cohort closely resembles that reported by the VA over the same 15-year interval; therefore, it is unlikely that TM amputations were under-represented. It is also unlikely that differences in patient characteristics that may be associated with this type of anaesthetic will confound the results. Therefore, the present mortality risk prediction model should not be used for patients undergoing revisional amputation surgery, contralateral amputations and for patients who have a high risk of death from causes other than those typically seen in diabetes and PAD, such as people with metastatic cancer. Further study is needed to determine the applicability of the mortality risk prediction model in other populations. Potential predictors that were rare (present in less than 5 per cent of the study cohort) were not included in the modelling (Table S1, supporting information). This included amputations such as knee and ankle disarticulations and Syme’s procedures, as well as rare co-morbidities such as acute renal injury and recent myocardial infarction. Predictors with a large number of missing values such as albumin levels, ejection fraction (EF) and HbA1c levels were also dropped from consideration. Interestingly, neither albumin levels, EF nor HbA1c levels were associated with mortality in bivariable assessments of subjects for whom measurements were available. Although the predictive characteristics of this model are good, there may be some unmeasured characteristics that the study was unable to account for, and which might have further improved the performance of the model.

Although previous studies have estimated 1-year mortality, the modelling approach used here differs from that of other prediction models. The present AMPREDICT-Mortality model permits estimation of mortality risk for each amputation level in a single parsimonious model, in a relatively homogeneous population. Furthermore, this study limited the target population to those with no previous amputations, and excluded co-morbidities strongly associated with mortality and a clinical indication for a TF amputation. Previous prediction models, with the exception of those reported by Nelson and colleagues¹⁶, who developed separate models for TT and TF amputations, predicted mortality risk of all patients undergoing a ‘major amputation’, and the effects of amputation level on risk were not modelled. These models also included patients who may have undergone multiple ipsilateral or contralateral amputations previously. The mortality risks of patients included in the sample used to develop those models may therefore differ significantly from those of patients undergoing their first amputation²⁷. Another important difference in the present prediction model is that the TM amputation level was included, acknowledging the increasing rate at which these procedures are being performed to salvage most of the foot. Furthermore, the time interval of analysis of the present model was extended from the perioperative period (30 days after surgery) to 1 year. A 1-year time frame was chosen because of its greater relevance to survival rather than exclusively focusing on operative mortality^27,29. Finally, to optimize the potential clinical value of the model, this study aimed for parsimony and selection of readily available model predictors, while maintaining good discrimination characteristics.