Abstract
Abstract
Objectives
To identify patterns of co-occurring macrovascular complications among people with diabetes-related foot ulceration (DFU) using latent class analysis (LCA) and describe characteristics associated with identified profiles.
Design
Single-centre retrospective clinical audit with secondary observational analysis of routinely collected data.
Setting
Townsville University Hospital, North Queensland, Australia.
Participants
381 adults with type 1 or type 2 diabetes and at least one documented DFU managed between 2015 and 2020.
Primary and secondary outcome measures
LCA of ischaemic heart disease, stroke, peripheral vascular disease and heart failure. Model selection considered information criteria, class size, classification certainty, stability and clinical interpretability. Secondary analyses compared demographic and clinical characteristics between classes. Mortality was examined descriptively because follow-up duration was inconsistently available.
Results
A two-class model had the lowest Akaike information criterion (1464.6), Bayesian information criterion (BIC) (1500.1) and sample-size-adjusted BIC (1471.5), identifying a higher-burden profile (94/381, 24.7%) and lower-burden profile (287/381, 75.3%). Conditional probabilities in higher-burden versus lower-burden profiles were 80.7% vs 20.9% for ischaemic heart disease, 29.6% vs 5.1% for stroke, 63.1% vs 25.8% for peripheral vascular disease and 37.5% vs 5.6% for heart failure. Compared with the lower-burden profile, the higher-burden profile had older age (68.83 vs 63.89 years; p=0.001), lower Estimated Glomerular Filtration Rate (eGFR) (51.24 vs 65.77 mL/min/1.73 m²; p<0.001), higher creatinine (156.69 vs 120.39 µmol/L; p=0.013), and more hypertension (93.6% vs 78.0%; p=0.002), dyslipidaemia (87.2% vs 70.3%; p=0.002) and chronic kidney disease (54.3% vs 24.4%; p<0.001). Smoking status differed overall (p=0.017), with current smoking more common in the higher-burden profile (38.0% vs 24.3%). Glycated hemoglobin (HbA1c) did not differ (8.30% vs 8.46%; p=0.628). Recorded mortality was higher in the higher-burden profile (29/78, 37.2% vs 28/269, 10.4%; p<0.001).
Conclusions
Among people with DFU, LCA identified higher-macrovascular-burden and lower-macrovascular-burden profiles that largely reflected cumulative macrovascular disease rather than distinct biological phenotypes. These exploratory findings require replication in contemporary multicentre cohorts.
Keywords: Diabetic foot; Diabetes Mellitus, Type 2; DIABETES & ENDOCRINOLOGY
STRENGTHS AND LIMITATIONS OF THIS STUDY.
The study used routinely collected data from a regional Australian multidisciplinary diabetes foot service, reflecting real-world clinical practice.
Macrovascular conditions were analysed jointly rather than individually, allowing co-occurrence patterns to be assessed within this clinically high-risk population.
The robustness of the identified profiles was examined using complementary sensitivity and stability analyses, including complete-case analysis and bootstrap resampling.
The use of latent class analysis introduces assumptions of local independence and probabilistic class assignment, which limit individual-level interpretation of the derived profiles.
The single-centre design, use of retrospective routinely collected data and missingness in some clinical variables limit generalisability, and the findings require confirmation in larger contemporary multicentre cohorts.
Introduction
Diabetes-related foot ulceration (DFU) is a major and debilitating complication of diabetes mellitus, with a lifetime risk of up to 25% in many populations including Australian populations.1 2 In Australia, DFU affects approximately 1.2%–1.5% of people with diabetes at any given time, with up to 300 000 individuals at risk of developing DFU, and it remains a leading cause of hospitalisation, amputation and disability.3 DFU is also strongly linked to macrovascular disease, including ischaemic heart disease (IHD), stroke, peripheral vascular disease (PVD) and heart failure. Peripheral arterial disease contributes to 50%–70% of DFU cases and markedly increases risks of poor healing, infection and lower limb amputation.3 4 Accurate identification of PAD in people with diabetes-related foot disease requires systematic lower-limb vascular assessment, although substantial variation in assessment techniques and adherence to guideline-recommended testing has been reported in podiatry practice.5
In Australia, macrovascular complications are highly prevalent among people with DFU, and DFU itself serves as a potent marker of heightened cardiovascular risk, with studies showing increased mortality from cardiovascular causes in this population.3 Despite this association, the clustering of macrovascular complications within the DFU population and their potential to define clinically meaningful phenotypes remain underexplored. Conventional risk stratification often addresses macrovascular conditions individually, potentially missing synergistic effects and heterogeneous risk profiles that could inform tailored interventions.
The objective of this study was to apply latent class analysis (LCA) to identify latent macrovascular phenotypes among Australian adults with diabetes and DFU, using key indicators (IHD, stroke, PVD and heart failure), and to characterise the demographic and clinical profiles of these phenotypes. By delineating distinct high-burden macrovascular and low-burden macrovascular subgroups, we aim to enhance risk stratification and identify opportunities for targeted cardiovascular prevention and smoking-cessation strategies in this high-risk Australian population.
Methods
Study design
This was a single-centre retrospective clinical audit with a secondary observational analysis of routinely collected data. The audit included adults living with diabetes and DFU who were being treated at Townsville University Hospital between 2015 and 2020.
Setting data sources and study population
Townsville University Hospital is a regional tertiary referral centre in North Queensland, Australia. This study used a retrospective clinical audit of routinely collected service data. There was no prospective recruitment or sequential screening pathway. All available records meeting the predefined eligibility criteria during the audit period were included. Cases were identified from several routine clinical sources, including hospital diagnostic coding, wound clinic databases and multidisciplinary diabetes-related foot service records; the cohort was not assembled using a single prespecified code list or automated algorithm. Eligibility was confirmed from the clinical record. Eligible patients were aged 18 years or older, had clinician-recorded type 1 or type 2 diabetes, and had at least one documented diabetes-related foot ulcer during the study period. A diabetes-related foot ulcer was defined as a break in the skin of the foot in a person with diabetes, in line with International Working Group on the Diabetic Foot guidance. Diabetes-related foot disease is a broader term that includes ulceration, infection, peripheral artery disease, neuropathy, neuro-osteoarthropathy, gangrene and amputation. As entry to this audit required documented ulceration, the term DFU is used when referring specifically to the study cohort. Demographic, clinical, laboratory and outcome data were extracted from the electronic medical record (ieMR) and entered into a secure study database. During database preparation, duplicate records were removed, variable formats and units were standardised, and implausible or internally inconsistent values were checked against the source ieMR where possible. Key diagnoses and outcomes were based on clinician-recorded documentation rather than diagnostic coding alone. The historical audit did not retain a complete list of all diagnostic codes used during initial case finding, and a stage-by-stage screening log was not generated. Records relating to people without diabetes or ulcers not related to diabetes were not included. Patients with partial missingness in the macrovascular indicators were retained in the primary analysis. The final analytic cohort comprised 381 patients; counts at earlier identification and exclusion stages cannot now be reconstructed reliably.
Variables and definitions
The index date was the first qualifying diabetes-related foot-ulcer encounter recorded during the audit period. Diabetes type was taken from the clinician-recorded diagnosis in the medical record. Age was determined at the index date. Sex and Aboriginal and Torres Strait Islander status were obtained from the hospital demographic record. Aboriginal and Torres Strait Islander status was based on recorded self-identification and was not statistically imputed. Smoking status was recorded as current, former or never smoking using the status documented closest to the index date. IHD, stroke, PVD, heart failure, hypertension, dyslipidaemia and chronic kidney disease were coded as present when a clinician-recorded diagnosis was documented in the medical history, problem list, specialist correspondence or hospital diagnostic record on or before the index date. Chronic kidney disease was based on the recorded clinical diagnosis rather than being derived from a single eGFR measurement. Dialysis was defined as receipt of maintenance haemodialysis or peritoneal dialysis at the index date. HbA1c, serum creatinine and eGFR were taken from the measurement closest to the index date within 3 months. HbA1c was recorded as a percentage, creatinine as µmol/L and eGFR as mL/min/1.73 m². The eGFR reported by the treating laboratory was used. Mortality analyses were restricted to patients for whom vital status was available. As follow-up duration was not available consistently, mortality was treated as an unadjusted descriptive outcome rather than a time-to-event endpoint.
Study size
As this was a clinical audit, all available patients meeting the eligibility criteria during the predefined period were included. No formal a priori sample-size calculation was undertaken. The suitability of the available sample for the LCA was examined through class size, classification certainty, repeated random initialisations, bootstrap resampling and complete-case sensitivity analysis.
Latent class analysis
LCA was used to examine patterns of co-occurrence across four binary macrovascular indicators: IHD, stroke, PVD and heart failure. Models containing one to six classes were fitted using the poLCA package in R. Each model was fitted using 500 random initialisations and a maximum of 5000 iterations. Model fit was assessed using log-likelihood, Akaike information criterion (AIC), Bayesian information criterion (BIC), sample-size-adjusted BIC, likelihood-ratio/deviance statistic (G²) and Pearson χ². Full fit statistics, class sizes and the frequency with which the maximum log-likelihood was reproduced are reported for all candidate models. The number of classes was selected by considering AIC and BIC together with class size, clinical interpretability, model stability and classification certainty. Formal likelihood-ratio tests were not used to compare models with different numbers of classes because the standard χ² assumptions do not hold for finite-mixture models. Class-specific conditional probabilities were used to describe the contribution of each macrovascular condition. Patients were assigned to the class with the highest posterior probability for descriptive comparisons. Normalised entropy, the mean maximum posterior probability and the distribution of maximum posterior probabilities were calculated to describe classification certainty. The local-independence assumption was examined using pairwise bivariate residuals (BVRs) for the four manifest indicators. Residual associations were considered when interpreting the model and the distinction between the identified classes.
Comparison of clinical characteristics
Continuous variables are presented as mean and SD, and categorical variables as number and percentage. Between-class differences in continuous variables were assessed using linear regression. Binary categorical variables were assessed using logistic regression. Smoking status was treated as a three-category variable. Its association with class membership was assessed using a global test rather than separate category-specific tests. χ2 statistics were combined across the imputed datasets using the D2 method. Aboriginal and Torres Strait Islander status was not imputed and was analysed using observed records only. Because of small cell counts, the overall comparison was made using Fisher’s exact test with class-specific observed denominators. Standardised mean differences (SMDs) were calculated to describe the magnitude of between-class differences independently of p values. SMDs were calculated within each imputed dataset and summarised across imputations. For smoking status, a single overall SMD and global p value were reported. The four macrovascular indicators used to construct the classes were presented as class-specific conditional probabilities and were not treated as independent between-class hypothesis tests. All tests were two-sided, with p<0.05 taken to indicate statistical significance. Mortality proportions were compared using Pearson’s χ2 test. No adjusted logistic or survival analysis was undertaken because follow-up time was not recorded consistently and deaths outside the health service may not have been fully captured.
Handling of missing data
The amount of missing data was calculated for each variable, and the pattern of missingness was examined before imputation. Missingness among the four LCA indicators was handled within the latent class model using the available indicator responses. Missing clinical variables used to describe the classes were handled using multiple imputation by chained equations with the mice package. 50 imputed datasets were generated, with 20 iterations for each dataset. Continuous variables were imputed using predictive mean matching, binary categorical variables using logistic regression and smoking status using polytomous regression. Latent class membership was not imputed. It was included as a predictor in the imputation model to preserve the observed relationship between class membership and the clinical variables being imputed. Aboriginal and Torres Strait Islander status was not imputed. Analyses were undertaken separately within each imputed dataset. Regression estimates, SEs, CIs and p values were combined using Rubin’s rules. Descriptive means, SD, counts and percentages were summarised across all 50 completed datasets rather than taken from one selected dataset. The missing-at-random assumption was considered plausible after conditioning on the observed variables included in the imputation model, but this assumption could not be tested directly. The possibility of data being missing not at random was considered when interpreting the results. Variable-specific available-case analyses using observed values were performed as sensitivity analyses.
Internal stability and sensitivity analyses
The selected two-class model was assessed using several within-sample checks. Repeated random initialisations were used to examine whether the maximum log-likelihood was reproduced and to reduce the risk of retaining a local solution. A complete-case LCA was conducted among patients with complete records for all four macrovascular indicators. Models containing one to six classes were refitted and compared using AIC and BIC. Sampling stability was examined using 500 non-parametric bootstrap resamples. The two-class model was refitted in each bootstrap sample. To account for arbitrary label switching, the class with the higher mean conditional probability across the four indicators was labelled the higher-burden class. Bootstrap distributions of class prevalence and class-specific conditional probabilities were summarised using medians and 95% percentile intervals. A simple macrovascular burden score was also calculated by summing the number of recorded conditions across IHD, stroke, PVD and heart failure, producing a score from 0 to 4. This analysis was restricted to patients with complete data for all four indicators. Scores were compared between the latent classes, and agreement between a threshold of two or more complications and the higher-burden LCA class was assessed using percentage agreement, sensitivity, specificity, accuracy and Cohen’s κ. This was treated as a concordance analysis because both approaches used the same four conditions.
Patient and public involvement
Patients and the public were not involved in the design, conduct, reporting or dissemination plans of this study.
Results
Study cohort and missing data
The demographic and clinical characteristics of the cohort are summarised in table 1. The study consisted of 381 adults living with DFU. The mean age was 65.1 years, 130 participants (34.1%) were female, and First Nations status was recorded for 310 participants, of whom 71 (22.9%) identified as Aboriginal and/or Torres Strait Islander. Hypertension was present in 312 participants (81.9%), dyslipidaemia in 284 (74.5%), chronic kidney disease in 121 (31.8%) and dialysis treatment in 24 (6.3%). The extent of missing data for individual variables is reported in online supplemental table S1.
Table 1. Demographic and clinical characteristics of the study cohort.
| Characteristic | Overall cohort (n=381) |
|---|---|
| Age, years | 65.11 (13.12) |
| HbA1c, % | 8.42 (2.09) |
| eGFR, mL/min/1.73 m² | 62.19 (25.25) |
| Creatinine, µmol/L | 129.35 (97.91) |
| Female sex, n (%) | 130 (34.1) |
| First Nations status, n/N (%) | 71/310 (22.9) |
| Current smoker, n (%) | 106 (27.8) |
| Former smoker, n (%) | 104 (27.3) |
| Never smoker, n (%) | 171 (44.9) |
| Hypertension, n (%) | 312 (81.9) |
| Dyslipidaemia, n (%) | 284 (74.5) |
| Chronic kidney disease, n (%) | 121 (31.8) |
| Dialysis, n (%) | 24 (6.3) |
| Ischaemic heart disease, n/N (%) | 138/381 (36.2) |
| Stroke, n/N (%) | 44/380 (11.6) |
| Peripheral vascular disease, n/N (%) | 135/379 (35.6) |
| Heart failure, n/N (%) | 69/314 (22.0) |
| Recorded mortality, n/N (%) | 57/347 (16.4) |
Values are pooled mean (SD) or pooled n (%) across 50 imputed datasets where applicable. First Nations status, the four macrovascular indicators and mortality are shown using observed-data denominators.
eGFR, Estimated Glomerular Filtration Rate; HbA1c, Glycated hemoglobin.
Latent class model selection and classification
We examined models containing one to six classes. In this analysis the two-class model had the lowest AIC (1464.6), BIC (1500.1) and sample-size-adjusted BIC (1471.5). The two-class model separated the cohort into a higher-macrovascular-burden profile (n=94, 24.7%) and a lower-burden profile (n=287, 75.3%). All 500 random initialisations of the two-class model converged to the same maximum log-likelihood (−723.3065). The three-class model had higher information criteria, included a class of only 12 participants, and reproduced its best log-likelihood in 44 of 500 starts. Models with four to six classes produced small classes and negative residual df. Full model-fit results are provided in online supplemental table S2 and shown graphically in figure 1. Classification certainty was moderate. Normalised entropy was 0.533 and the mean maximum posterior probability was 0.866. Detailed characteristics of the selected two-class model are provided in online supplemental table S3.
Figure 1. Model fit across competing latent class solutions. Information criteria for one-class to six-class latent class models. Lower values indicate better fit after penalising model complexity. AIC, Akaike information criterion; BIC, Bayesian information criterion.
These findings supported use of the two-class solution as a parsimonious description of higher and lower macrovascular burden, while indicating that individual class assignment was not error-free.
Local independence
Pairwise bivariate-residual analysis was based on the 314 participants with complete responses for all four indicators. Evidence of residual association remained between IHD and PVD (BVR 21.45), stroke and PVD (BVR 20.40) and PVD and heart failure (BVR 19.67); all three associations remained significant after Benjamini-Hochberg and Bonferroni adjustment (all adjusted p<0.001). The remaining three indicator pairs did not show significant residual association (online supplemental table S4). Thus, the local-independence assumption was not fully satisfied.
Macrovascular class profiles
The selected model and its class-specific conditional probabilities are reported in online supplemental table S3. The higher-burden profile had conditional probabilities of 80.7% for IHD, 29.6% for stroke, 63.1% for PVD and 37.5% for heart failure. Corresponding probabilities in the lower-burden profile were 20.9%, 5.1%, 25.8% and 5.6%, respectively (figure 2). The greatest separation between profiles was seen for IHD and PVD.
Figure 2. Class-specific conditional probabilities conditional probability of each macrovascular condition within the higher-macrovascular-burden and lower-macrovascular-burden profiles. IHD, ischaemic heart disease; PVD, peripheral vascular disease.
Demographic and clinical characteristics of the classes
Table 2 shows demographics and clinical characteristics of the two risk profiles. Participants in the higher-burden profile were older than those in the lower-burden profile (68.83 vs 63.89 years; mean difference 4.94 years, 95% CI 1.91 to 7.97; p=0.001; SMD 0.396). Renal function was poorer in the higher-burden profile: mean eGFR was 51.24 vs 65.77 mL/min/1.73 m² (p<0.001; SMD 0.583), and mean creatinine was 156.69 vs 120.39 µmol/L (p=0.013; SMD 0.356). HbA1c did not differ between profiles (8.30% vs 8.46%; p=0.628; SMD 0.115).
Table 2. Clinical and demographic characteristics by macrovascular-burden profile.
| Characteristic | Higher-burden profile (n=94) | Lower-burden profile (n=287) | P value | SMD |
|---|---|---|---|---|
| Age, years | 68.83 (11.47) | 63.89 (13.41) | 0.001 | 0.396 |
| HbA1c, % | 8.30 (2.05) | 8.46 (2.10) | 0.628 | 0.115 |
| eGFR, mL/min/1.73 m² | 51.24 (25.86) | 65.77 (24.03) | <0.001 | 0.583 |
| Creatinine, µmol/L | 156.69 (110.91) | 120.39 (91.71) | 0.013 | 0.356 |
| Female sex, n (%) | 27 (28.9) | 103 (35.9) | 0.216 | 0.15 |
| First Nations status, n/N (%) | 5/53 (9.4) | 66/257 (25.7) | 0.011 | 0.437 |
| Current smoker, n (%) | 36 (38.0) | 70 (24.3) | 0.017 | 0.403 |
| Former smoker, n (%) | 15 (16.4) | 89 (31.1) | ||
| Never smoker, n (%) | 43 (45.6) | 128 (44.5) | ||
| Hypertension, n (%) | 88 (93.6) | 224 (78.0) | 0.002 | 0.458 |
| Dyslipidaemia, n (%) | 82 (87.2) | 202 (70.3) | 0.002 | 0.423 |
| Chronic kidney disease, n (%) | 51 (54.3) | 70 (24.4) | <0.001 | 0.642 |
| Dialysis, n (%) | 10 (11.0) | 14 (5.0) | 0.054 | 0.225 |
Values are pooled mean (SD) or averaged n (%) across 50 multiply imputed datasets unless otherwise indicated. P values for continuous and binary variables were obtained from linear and logistic regression, respectively, and pooled using Rubin’s rules. Smoking status has one global D2 pooled p value and one overall SMD. First Nations status was not imputed and is shown using observed class-specific denominators; its p value was obtained using Fisher’s exact test.
SMD, standardised mean difference.
Hypertension (93.6% vs 78.0%; p=0.002; SMD 0.458), dyslipidaemia (87.2% vs 70.3%; p=0.002; SMD 0.423), and chronic kidney disease (54.3% vs 24.4%; p<0.001; SMD 0.642) were more common in the higher-burden profile. Smoking status differed overall between profiles (D2=4.08; p=0.017; SMD 0.403): current smoking was more common in the higher-burden profile, whereas former smoking was more common in the lower-burden profile. Dialysis was numerically more frequent in the higher-burden profile, but the pooled difference did not reach conventional statistical significance (11.0% vs 5.0%; p=0.054; SMD 0.225). Sex distribution did not differ significantly (p=0.216). Among participants with observed First Nations status, the proportion identifying as Aboriginal and/or Torres Strait Islander was lower in the higher-burden profile (5/53, 9.4%) than in the lower-burden profile (66/257, 25.7%; Fisher exact p=0.011; SMD 0.437). Pooled effect estimates and corresponding available-case analyses are presented in online supplemental table S5.
Internal stability and sensitivity analyses
The two-class solution remained preferred in the complete-case LCA of 314 participants. The two-class model had the lowest AIC (1156.1) and BIC (1189.9); BIC values for the one-class, three-class, four-class, five-class and six-class models were 1193.4, 1212.8, 1238.3, 1266.0 and 1294.8, respectively.
Across 500 non-parametric bootstrap resamples, the median prevalence of the higher-burden class was 26.1% (95% percentile interval 7.6%–56.4%), compared with 24.7% in the primary analysis. Median bootstrap conditional probabilities in the higher-burden profile were 81.4% for IHD, 29.2% for stroke, 63.6% for PVD and 36.4% for heart failure. Corresponding values in the lower-burden profile were 21.0%, 4.4%, 25.4% and 5.7%. The percentile intervals for the two profiles did not overlap for any indicator, although the interval for higher-burden class prevalence was wide (online supplemental figure S1).
Available-case analyses produced the same direction of association as the multiply imputed analysis for all clinical variables. Age, eGFR, hypertension, dyslipidaemia and chronic kidney disease remained statistically significant; HbA1c and sex remained non-significant. The creatinine association was attenuated in the available-case analysis (p=0.078 compared with p=0.013 after multiple imputation), and dialysis remained non-significant (online supplemental table S5).
In the comparison with a simple 0–4 macrovascular complication count, the higher-burden LCA class had a median score of 2 (IQR 2–3), compared with 0 (IQR 0–1) in the lower-burden class (p<0.001). A threshold of two or more complications agreed with LCA membership in 98.4% of complete cases, with sensitivity 100%, specificity 98.1% and Cohen’s κ=0.945. This high concordance indicated that the two-class solution predominantly represented cumulative macrovascular burden (online supplemental table S6).
Mortality data were available for 347 participants (91.1%): 78 of 94 participants in the higher-burden profile and 269 of 287 in the lower-burden profile. There were 57 recorded deaths. The unadjusted proportion with recorded mortality was higher in the higher-burden profile (29/78, 37.2%) than in the lower-burden profile (28/269, 10.4%; p<0.001). Because follow-up time was not recorded consistently and deaths outside the health service may not have been fully ascertained, this finding was treated as descriptive rather than as formal outcome validation.
Discussion
In this retrospective study of 381 Australian people with DFU we showed presence of two distinct macrovascular burden profiles in people with DFU. The high macrovascular burden phenotype, comprising 24.7% of participants, was marked by elevated probabilities of IHD, PVD, stroke and heart failure and associated with older age, poorer renal function, higher rates of hypertension, dyslipidaemia, chronic kidney disease and current smoking. This subgroup exhibited a strikingly higher recorded mortality (37.2%) compared with the low-burden phenotype (10.4%).
The application of LCA to identify cardiovascular phenotypes in diabetes has emerged as an important tool for uncovering latent subgroups. For example, in the CANVAS Programme and CREDENCE trial, LCA identified four clinically distinct phenotypes in type 2 diabetes, with varying responses to canagliflozin and implications for cardiovascular outcomes. Similarly, LCA has also been used to identify cardio-metabolic subtypes of prediabetes and clinical risk phenotypes associated with adverse events in diabetes.6 Our study extends LCA specifically to the DFU population, to explore the clustering of macrovascular complications of diabetes. We have identified two distinct subgroups with one having higher prevalence of macrovascular complications and mortality. This high-risk subgroup is characterised by higher prevalence of hypertension, dyslipidaemia, chronic kidney disease and smoking.
In the Australian context, DFU is known to disproportionately affect Aboriginal and Torres Strait Islander peoples. A systematic review of eleven studies including 157 892 participants found that Aboriginal and Torres Strait Islander Australians had a threefold to sixfold higher likelihood of a diabetes-related foot complication than non-Indigenous Australians.7 Population-based data from regional Australia also show marked over-representation of Aboriginal and Torres Strait Islander peoples among those with diabetic foot ulceration, with geographic remoteness associated with greater risk and severity.8 Studies from Central and Northern Australia have likewise identified First Nations status as an important predictor of lower-extremity amputation after adjustment for clinical factors.9 In the present cohort, First Nations individuals were represented in both latent classes rather than being confined to the higher-burden profile. This is an unadjusted finding. We did not fit a multivariable model to estimate an independent association between First Nations status and macrovascular burden, so the class distribution should not be interpreted as evidence for or against such an association. The observed pattern may reflect the combined influence of clinical, social, geographic and healthcare-access factors within this cohort. The broader excess burden of diabetes-related foot disease and vascular complications among Aboriginal and Torres Strait Islander peoples is multifactorial; delayed access to specialist care, social and geographic disadvantage and structural barriers to preventive care may all contribute. These findings support the need for earlier intervention, culturally safe models of care and integrated vascular and diabetes services for First Nations communities.
In this study smoking was overrepresented in class with high risk of macrovascular complication in those with DFU. Contemporary literature shows that smoking impairs all stages of DFU onset and healing, via mechanisms like peripheral neuropathy, vascular alterations and wound infection.10 In active disease, smoking correlates with lower systolic toe pressures, indicating impaired perfusion.11 The high burden of smoking, dyslipidaemia and CKD in high-risk group for macrovascular complications suggests prioritising multifaceted cardiovascular prevention, including aggressive lipid and blood pressure control, renal protection and smoking cessation programmes. The two-class solution provided by this study primarily distinguished people with DFU according to cumulative macrovascular burden rather than identifying clearly separate biological phenotypes. This interpretation is supported by the high concordance between latent class membership and the simple macrovascular complication count. The higher-burden profile was also associated with older age, poorer renal function, hypertension, dyslipidaemia and current smoking. However, the classes should not be interpreted as clinical risk categories or used to determine treatment intensity. Patients in the lower-burden profile still had substantial cardiovascular disease, including IHD and PVD and management should continue to be based on individual clinical conditions and established cardiovascular risk.
Our study has several limitations. The study period (2015–2020) may limit generalisability to contemporary practice, given subsequent advances in diabetes treatment, cardiovascular risk reduction and multidisciplinary diabetic foot care. The retrospective, single-centre design and reliance on routinely collected electronic records may also limit generalisability and may have led to underascertainment of some complications. The original audit did not generate a stage-by-stage screening log, and the historical source systems did not retain reproducible counts at each case-ascertainment and exclusion step. The final eligible cohort can therefore be reported, but a conventional participant flow cannot be reconstructed reliably. Mortality data were incomplete and cause-specific mortality was unavailable. The LCA classification is exploratory and should not be used to determine treatment intensity. Missingness was substantial for several clinical variables, particularly HbA1c and measures of renal function. Although multiple imputation was used and estimates were pooled across 50 imputed datasets, this approach assumes that data were missing at random conditional on the variables included in the imputation model, and residual bias due to unmeasured determinants of missingness cannot be excluded.
Latent class solutions are also data-dependent and may not reproduce identically in other populations. Although repeated random initialisations, complete-case analysis of the manifest indicators and bootstrap resampling supported the internal robustness of the overall two-class structure, the moderate entropy and variability in bootstrap estimates of class prevalence indicate uncertainty in individual class assignment and the precise relative size of the classes. Furthermore, these analyses assess internal stability rather than external validity. The identified classes should therefore be considered exploratory, data-derived patterns requiring replication in larger, independent and multicentre cohorts before their use for individual-level clinical risk stratification or phenotype-directed care can be recommended. Finally, the high concordance between latent class membership and a simple macrovascular complication count indicates that the identified classes predominantly reflect cumulative macrovascular burden. Although LCA provides class-specific probabilities and allows conditions to contribute unequally, its incremental clinical value over a simple count appears limited in this cohort. The classes are therefore interpreted as exploratory higher-burden and lower-burden profiles rather than distinct biological phenotype.
Future research should prospectively validate these phenotypes in multicentre cohorts, incorporating DFU-specific outcomes like healing rates and amputations. Interventional trials testing phenotype-targeted therapies, such as enhanced smoking cessation or poly-vascular risk reduction, are warranted.
In conclusion, in this regional Australian cohort of people with DFU, LCA identified higher-macrovascular-burden and lower-macrovascular-burden profiles. The two-class solution showed very high concordance with a simple count of macrovascular complications, indicating that it predominantly represented cumulative disease burden rather than distinct biological phenotypes. These findings are exploratory and require replication in independent cohorts to determine whether LCA provides clinically useful information beyond conventional assessment of individual macrovascular conditions or a simple complication count.
Supplementary material
Footnotes
Funding: The authors have not declared a specific grant for this research from any funding agency in the public, commercial or not-for-profit sectors.
Prepublication history and additional supplemental material for this paper are available online. To view these files, please visit the journal online (https://doi.org/10.1136/bmjopen-2026-119751).
Provenance and peer review: Not commissioned; externally peer reviewed.
Patient consent for publication: Not applicable.
Data availability free text: The individual-level data used in this study are not publicly available because they contain potentially identifiable clinical information and are subject to institutional and ethical restrictions. Deidentified data may be available from the corresponding author on reasonable request and subject to approval by the relevant data custodian and ethics requirements.
Patient and public involvement: Patients and/or the public were not involved in the design, or conduct, or reporting, or dissemination plans of this research.
Ethics approval: The audit was approved by the Townsville Hospital and Health Service Human Research Ethics Committee (HREC/2020/QTHS/62575). The requirement for individual informed consent was waived because the analysis used routinely collected retrospective data. Data were analysed in de-identified form. All analyses were performed using R V.4.3.0.
Data availability statement
Data may be obtained from a third party and are not publicly available.
References
- 1.Armstrong DG, Swerdlow MA, Armstrong AA, et al. Five year mortality and direct costs of care for people with diabetic foot complications are comparable to cancer. J Foot Ankle Res. 2020;13:16. doi: 10.1186/s13047-020-00383-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.McDermott K, Fang M, Boulton AJM, et al. Etiology, Epidemiology, and Disparities in the Burden of Diabetic Foot Ulcers. Diabetes Care. 2023;46:209–21. doi: 10.2337/dci22-0043. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Zhang Y, van Netten JJ, Baba M, et al. Diabetes-related foot disease in Australia: a systematic review of the prevalence and incidence of risk factors, disease and amputation in Australian populations. J Foot Ankle Res. 2021;14:8. doi: 10.1186/s13047-021-00447-x. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Fitridge R, Chuter V, Mills J, et al. The intersocietal IWGDF, ESVS, SVS guidelines on peripheral artery disease in people with diabetes and a foot ulcer. Diabetes Metab Res Rev. 2024;40:e3686. doi: 10.1002/dmrr.3686. [DOI] [PubMed] [Google Scholar]
- 5.Tehan PE, Fox M, Stewart S, et al. Lower limb vascular assessment techniques of podiatrists in the United Kingdom: a national survey. J Foot Ankle Res. 2019;12:31. doi: 10.1186/s13047-019-0341-2. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Mantovani M, Kwiendacz H, Irlik K, et al. Clinical risk phenotypes in diabetes and their associations with adverse cardiovascular events: A report from the Silesia Diabetes-Heart Project. Diabet Med. 2026;43:e70136. doi: 10.1111/dme.70136. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.West M, Chuter V, Munteanu S, et al. Defining the gap: a systematic review of the difference in rates of diabetes-related foot complications in Aboriginal and Torres Strait Islander Australians and non-Indigenous Australians. J Foot Ankle Res. 2017;10:48. doi: 10.1186/s13047-017-0230-5. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 8.Soonarane YK, Kirk M, Khandaker G, et al. Epidemiology and healthcare access inequities in diabetic foot disease: a retrospective study in Central Queensland, Australia. BMJ Open. 2025;15:e098999. doi: 10.1136/bmjopen-2025-098999. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Ashok A, Bonnefin R, Yerramilli A, et al. Predictive factors for lower extremity amputations in patients with diabetic foot disease in Central Australia. Intern Med J. 2025;55:1907–16. doi: 10.1111/imj.70222. [DOI] [PubMed] [Google Scholar]
- 10.Xia N, Morteza A, Yang F, et al. Review of the role of cigarette smoking in diabetic foot. J Diabetes Investig. 2019;10:202–15. doi: 10.1111/jdi.12952. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Bechara N, Hng TM, Gunton JE. The association between tobacco smoking and systolic toe pressures in active foot ulceration. Sci Rep. 2024;14:8550. doi: 10.1038/s41598-024-59158-5. [DOI] [PMC free article] [PubMed] [Google Scholar]