Risk stratification nomogram for diabetic foot ulcer recurrence after healing in type 2 diabetes

Abstract

BACKGROUND

Diabetic foot ulcer (DFU) is one of the severe complications of diabetes mellitus, with a lifetime prevalence ranging from 15% to 25% in diabetic patients. Most studies that have evaluated the association of wound closure with DFU healing are well designed and show high rates of initial healing, between 75% to 95%, nonetheless the recurrence in some reports can be as high as 40% to 65% within a year after successful wound closure. Recurrent foot ulcers are a high burden on patients due to a high rate of hospitalisation, higher amputation risk and lower quality of life and they result in much higher costs. Currently, risk prediction models incorporate clinical, metabolic and biological factors which are features linked to ulcer recurrence but aggregation of these parameters in a composite measure has not been performed.

AIM

To create a model for predicting DFU recurrence after wound healing based on nomogram-based risk stratification in individuals with type 2 diabetes.

METHODS

This study was a single-center retrospective cohort analysis of patients with type 2 diabetes mellitus who achieved complete healing of their first DFU at Shanghai Municipal Hospital of Traditional Chinese from January 2019 to July 2023. A median of 24 months follow up was performed to assess ulcer recurrence in patients. In detail, comprehensive clinical parameters and laboratory biomarkers along with metabolic indices were assessed during their first ulcers healed. Independent predictors and a risk stratification nomogram were identified using a two-stage statistical approach combining Least Absolute Shrinkage and Selection Operator regression for variable selection and Cox proportional hazards modeling for risk estimation.

RESULTS

Over the follow-up period, DFU recurred in 128 (41.0%) patients. Multivariate analysis revealed independent predictors: Prior ulcer duration > 8 weeks [hazard ratio (HR) = 3.24, 95% confidence interval (CI): 2.15-4.88], peripheral arterial disease (HR = 2.87, 95%CI: 1.92-4.29), peripheral neuropathy severity score ≥ 7 (HR = 2.45, 95%CI: 1.63-3.68), median hemoglobin A1c ≥ 8.0% (HR = 2.18, 95%CI: 1.46-3.25), ankle-brachial-index < 0.9 (HR = 2.56, 95%CI: 1.71-3.83 ), serum albumin < 35 g/L (HR = 1.92; 95%CI: 1.28-2.88) and offloading compliance insufficiently sufficient (HR = 2.73, 95%CI: 1.82-4.09). The C-index of the developed nomogram was 0.826 (95%CI: 0.787-0.865), indicating good discrimination, and also showed satisfactory calibration. A simple risk stratification of patients on the basis of points accorded to each variable as follows: 0-100 = low, 101-180 = intermediate, and > 180 points = high-risk groups translated into 2-year recurrence rates of; low: 12.3%, intermediate: 38.7% and high: 71.2%. Heavy temporal validation (C-index 0.814; 95%CI: 0.764-0.864) on an independent cohort (n = 156) from the same institution confirmed strong model performance.

CONCLUSION

This study developed a risk stratification model for predicting recurrence of DFUs after initial wound healing in type 2 diabetes. This model combines easily obtainable clinical and metabolic characteristics of patients, outperforming established risk scores to serve as an easily accessible tool for derivation of sex-specific personalized individual risk profiles and targeted prevention efforts in the clinic.

Key Words: Diabetic foot ulcer; Type 2 diabetes mellitus; Risk stratification; Peripheral arterial disease; Peripheral neuropathy; Nomogram

Core Tip: This study introduces a new risk stratification model for recurrent diabetic foot ulcer based on clinical parameters, metabolic biomarkers and vascular assessments. The tool, based on a nomogram, achieves a high predictive accuracy (C-index 0.826) and successfully divides patients into three well-differentiated groups with statistically significant dissimilar recurrence values. This clinically relevant model allows for recognition of patients at high risk, and will guide targeted prevention efforts in this patient population and could lead to a substantial reduction in the burden of recurring diabetic foot ulcers.

INTRODUCTION

Diabetes mellitus (DM) has emerged as one of the most significant 21^st century global health challenges with an estimated prevalence of 537 million adults aged 20 years to 79 years worldwide living with diabetes in 2021 according to the International Diabetes Federation organization and will rise upward of about 783 million adults by the year of 2045[1,2]. The complications of diabetes, in general, are complex and one of the extremely devastating diabetic foot ulcer (DFU) that problem with diabetic patients varies frequencies between 15%-25% during their lifetime[3,4]. DFU is a tip of iceberg, a door that opens chain pathophysiological sequelae such as infection, gangrene, lower extremity amputation and premature mortality.

While advances in multidisciplinary care like use of specialist wound management, vascular intervention and early systemic antibiotics to control infection can increase initial healing rates[5-8], ulcer recurrence is still a significant clinical challenge. Recurrence rates can be as high as 40%-65% in patients who achieved complete wound closure within one year, and almost 70% across five years for those with healed ulcers[6]. The high frequency of recurrence transforms DFU from an acute clinical problem to a chronic relapsing disease with morbidity, mortality and health care expenditure comparable to that of malignancy[9-11].

The clinical and economic burden associated with recurrent DFU is staggering. Recurrent DFU requires aggressive medical care, often marked by prolonged hospital stays, multiple surgical interventions, and high resource utilization. Even more importantly recurrent ulceration is associated with a significantly higher risk of receiving major adverse limb events such as amputation where - ulcer recurrence has been shown in various studies to have a domino effect on progressing the rate of amputation over time with each episode[7]. Mortality may be fatal, with estimates of almost 70% five-year mortality following a major amputation for causes related to diabetes complication cascade[4,10].

The aetiopathophysiology of recurrent DFU is probably heterogeneous and complex, being predominantly attributed to exacerbation of pre-existing causes or inadequately compensated interplay among persistent peripheral neuropathy, progressive peripheral arterial disease (PAD), chronic hyperglycaemia, gait stress and impaired healing potential. Up to 50% of people with diabetes develop some form of peripheral neuropathy. Classification includes diabetic and distal symmetric polyneuropathy[12]. That results in loss of protective sensation and alters foot biomechanics causing increased plantar pressure at the most vulnerable sites for ulceration. Unfortunately, PAD also complicates perfusion and oxygen delivery to the pedal tissue creating an environment less favorable for erosion of studies into DFU viability[9,10] where vascular insufficiency is associated with 30%-50% of patients regenerating DFU. This phenotype of neuropathy and ischemia is called “neuroischemic foot” which is especially vulnerable to recurrent ulceration.

Most of the risk factors are known and verified, but no sequentially validated tool for systemic risk stratification is available in clinical practice (apart from acute wound closure). Most approaches for risk assessment study neuropathy severity or vascular status in isolation as if none overlapped and could interact with each other to further increase the ultimate recurrences risk together with clinical, metabolic and biochemical features[13-18]. It limits clinicians to correctly tailor information on high-risk individuals who may benefit from intensive preventive interventions, and closer surveillance.

With recent advancements in machine learning and predictive modeling, this vast new potential of risk stratification may finally be realized: Sophisticated tools that can process many variables at once to identify complex interaction patterns[19]. Additionally, the use of clinical parameters that are readily available together with objective biomarkers may lead to an applicable and implementable risk stratification in a common clinical setting. Nonetheless, the majority of prediction models developed to date have methodological shortcomings, such as small sample sizes and short follow-up periods, failure to externally validate or omission of clinically relevant metabolic and biochemical factors[20,21].

To address these gaps, we performed a large single-center retrospective cohort study at Shanghai Municipal Hospital of Traditional Chinese with the following aims: (1) To identify independent clinical, vascular, metabolic and biochemical predictors of DFU recurrence post-complete wound healing; (2) To develop and internally validate a practical risk stratification model for prediction of DFU recurrence; (3) To provide an easy-to-use nomogram tool based on such a model for clinical use; and (4) To conduct temporal validation of the model in an independent patient cohort from same institute. We hypothesize that a comprehensive multi-domain model would better predict risk-than single-domain assessment and generate easily applicable ambient risk stratification for targeted individualized prevention.

MATERIALS AND METHODS

Study design and setting

This is a retrospective cohort study conducted at Shanghai Municipal Hospital of Traditional Chinese Medicine, a tertiary care diabetic foot center. Although the institutional DFU database encompasses records from January 2015 onwards, patients treated between January 2015 and December 2018 were excluded from the formal analysis because the standardized composite baseline assessment protocol - including the Composite Neuropathy Severity Score (CNSS) and structured offloading compliance interview - had not been uniformly implemented across the department until January 2019, resulting in incomplete ascertainment of several key model variables in that earlier period. The study therefore covers the period January 2019 to July 2023. The study protocol was approved by the Institutional Review Board and Ethics Committee of Shanghai Municipal Hospital of Traditional Chinese Medicine, approval No. 2025SHL-KY-67-01 and was performed in accordance with all procedures embodied in the Declaration of Helsinki. The requirement for informed consent was waived by the ethics committee because of the retrospective nature of our study. It should be noted that while the formal ethics approval was issued in 2023, the study analyzed de-identified clinical records from patients treated between 2019 and 2023; the Ethics Committee confirmed that retrospective review of de-identified data collected in the course of routine clinical care does not require prospective approval under applicable institutional and national regulations, and the 2023 approval formally ratified this retrospective analysis. Comprehensive baseline data were collected at the time of confirmed complete wound healing, not at the time of ulcer onset. The development cohort comprised patients who achieved wound healing between January 2019 and December 2021; the temporal validation cohort comprised patients treated between January 2022 and July 2023. These two periods are non-overlapping. Following the baseline assessment at wound healing, patients entered a structured prospective surveillance program (monthly visits for the first 6 months, then quarterly thereafter) or until recurrent ulceration occurred.

Study population

The study included adult patients (≥ 18 years) with type 2 DM who completely healed their first DFU during the study at Shanghai Municipal Hospital of Traditional Chinese Medicine. The development cohort included patients with confirmed wound healing between January 2019 and December 2021 (n = 312); the temporal validation cohort included patients treated between January 2022 and July 2023 (n = 156). These cohorts are non-overlapping and represent consecutive patient groups from the same institution.

Inclusion criteria: (1) A documented diagnosis of type 2 DM according to American Diabetes Association criteria; (2) Complete healing of first DFU defined as full epithelialization without drainage for ≥ 2 consecutive weeks; (3) Availability of relevant clinical, laboratory and follow-up data; and (4) Minimum duration of follow-up was set at 12 months or occurrence of recurrent ulcer within 12 months.

Exclusion criteria: (1) Type 1 DM; (2) Previous history of DFU or lower extremity amputation; (3) Active malignancy requiring chemotherapy or radiation therapy; (4) End-stage renal disease requiring dialysis; (5) Severe cognitive impairment preventing accurate follow-up, per physician assessment; (6) A life expectancy less than 12 months according to the treating physician’s judgment; and (7) Incomplete baseline data collection or loss-to-follow up during the first 12-months study period.

Data collection and outcome assessment

At complete wound healing, 2 stringently collected comprehensive “zero” baseline data were available for recurrence risk assessment. Recurrent foot ulceration, which was defined as the formation of a new full-thickness skin break below the ankle on either foot at least 4 weeks after complete wound closure from an incident ulcer, was the main outcome. Patients were followed either through regularly scheduled clinical visits (monthly for the first 6 months, then quarterly thereafter) or until recurrent ulceration developed.

Clinical parameters

Clinical variables age, sex, diabetes duration and body mass index were prospectively collected in all participating patients as well as smoking status (current/former/never), alcohol consumption and diabetes-related comorbidities (diabetic retinopathy, nephropathy cardiovascular disease). The characteristics of index ulcers were: Anatomical location (forefoot, midfoot, hindfoot), Wagner classification grade, time from the occurrence of the wound to healing and whether or not there was an infection during its treatment. Offloading compliance was operationally defined as patient use of the prescribed offloading device for more than 80% of daily ambulatory time (adequate compliance) vs 80% or less (inadequate compliance). This was ascertained through structured patient interviews conducted at each follow-up visit and cross-referenced against medical record documentation. The > 80% threshold was adopted in accordance with prior validated offloading intervention trials. It should be noted that this measure is susceptible to recall and social desirability bias, as objective electronic monitoring was not available in this retrospective cohort; this is acknowledged as a limitation.

Peripheral neuropathy assessment

Peripheral neuropathy was defined as ≥ 1 of: 10-g Semmes-Weinstein monofilament test (10 plantar sites/foot); 128-Hz tuning fork vibration perception threshold or signs of motor neuropathy (muscle wasting, foot deformities, Charcot neuroarthropathy). These were summed into a CNSS (0-10 scale) adapted from the validated Neuropathy Disability Score. Severe neuropathy was defined using a pre-specified cut-off of CNSS ≥ 7.

PAD evaluation

Vascular status was assessed non-invasively. Standardized Doppler ultrasound was used to measure ankle brachial index (ABI); an ABI < 0.9 confirms PAD while and ABI > 1.3 necessitates toe-brachial index (TBI) measurement in vivo, where a TBI < 0.7 indicates significant disease[15]. Pedal pulses and TcPO₂ (critical limb ischaemia threshold: < 30 mmHg) were also assessed. PAD was defined according to current International Working Group on the Diabetic Foot/The European Society for Vascular Surgery/The Society for Vascular Surgery guidelines and required one criterion to be met (ABI < 0.9, TBI < 0.7, absent pedal pulses or TcPO₂ < 30 mmHg).

Laboratory and metabolic parameters

Blood samples were obtained fasting at baseline. Methods hemoglobin A1c (HbA1c) was assessed by high-performance liquid chromatography (≥ 8.0% classified as poor glycaemic control). Other assessments included a standard lipid panel, renal function (serum creatinine and estimated glomerular filtration rate as determined by the chronic kidney disease-epidemiology equation), nutritional markers (serum albumin, total protein), inflammatory markers (white blood cell, C-reactive protein, erythrocyte sedimentation rate), and 25-hydroxyvitamin D.

Statistical analysis

Analyses were performed in R 4.2.0. Continuous variables were summarised as mean ± SD or median (inter-quartile range); categorical variables as frequency and percentage. For continuous variables, a between-group comparison used the Student’s t-test or Mann-Whitney U test; for categorical variables, χ² or Fisher’s exact test. The development cohort was divided into a training set and validation set (70:30, stratified by recurrence). Candidate predictors that had P < 0.10 on univariable Cox regression were subjected to Least Absolute Shrinkage and Selection Operator (LASSO) regression using 10-fold cross-validation (λ selected by minimum cross-validation error). A final multivariable Cox model estimated adjusted hazard ratios (HRs) and 95% confidence intervals (CIs) including variables with non-zero LASSO coefficients. We validated the proportional hazards assumption using Schoenfeld residuals. Model discrimination was evaluated using Harrell’s C-index (> 0.75 = good) and calibration determined by comparing predicted to observed recurrence probabilities at 1-year and 2-year.

Nomogram construction and risk stratification

Predicting risk using a nomogram: To visualize the data obtained from the multivariable Cox model, a nomogram was constructed to provide an easily accessible and user-friendly graphical tool for individualized prediction of clinical outcomes. The effect of each predictor was also reflected in points calculated according to its value of regression coefficient, while total points were transformed into estimated recurrence rates at 1-year follow up as well as 2-year follow up. To aid clinical decision-making, patients were categorized into three risk groups [low (0-100), intermediate (101-180) and high (> 180)] for recurrence rates, with cut-points identified using recursive partitioning analysis to optimize between-group separation in recurrence rates based on total nomogram points.

Model validation

The developed model was subjected to stringent internal validation using the 30% hold-out validation subset within the development cohort. We also validated in a second cohort of 156 patients from Shanghai Municipal Hospital of Traditional Chinese Medicine who fulfilled the same inclusion/exclusion criteria but were treated at a later time (temporal validation). C-index and calibration plots were used to assess the performance of models in validation cohorts. Clinical usefulness was assessed through decision curve analysis, which quantifies the net benefit at different threshold probabilities to obtain the intervention. Two-sided statistical tests were used, and P values < 0.05 was considered statistically significant. Because predictive selection is exploratory in nature, we do not adjust for multiple comparisons during univariable screening steps and used shrinkage methods to account for multiple testing in final multivariable model inference.

RESULTS

Patient characteristics and follow-up

Overall, 312 patients were included (mean age 64.3 ± 10.8 years; 63.5% male). The median diabetes duration was 14.0 years and the median follow-up was 24.0 months; 128 patients (41.0%) had a recurrent ulcer (median time to ulceration, 10.5 months). In comparison with the non-recurrence group, patients who developed recurrence had longer diabetes duration (16.0 years vs 12.0 years, P < 0.001), poorer glycemic control (HbA1c 8.5% ± 1.3% vs 7.3% ± 1.2%, P < 0.001) and roughly twice as common a prevalence of HbA(1c) ≥ 8.0% (62.5% vs 33.7%, P < 0.001). In addition, nutritional status was worse as reflected by lower mean serum albumin (34.7 ± 4.1 g/L vs 38.2 ± 3.7 g/L, P < 0.001) and increased prevalence of hypoalbuminemia (51.6% vs 28.3%, P < 0.01, Table 1).

Table 1 Baseline characteristics of study population, n (%)/mean ± SD.

Variable	All patients (n = 312)	No recurrence (n = 184)	Recurrence (n = 128)	Test statistic	P value
Age, years	64.3 ± 10.8	63.8 ± 11.2	65.1 ± 10.2	t = 0.98	0.328
Male sex	198 (63.5)	114 (62.0)	84 (65.6)	χ2 = 0.42	0.517
Diabetes duration, years	14.0 (10.0-19.0)	12.0 (9.0-17.0)	16.0 (12.0-21.0)	Z = 4.12	< 0.001
Body mass index, kg/m2	27.8 ± 4.3	27.5 ± 4.1	28.2 ± 4.6	t = 1.35	0.178
Current smoking	87 (27.9)	42 (22.8)	45 (35.2)	χ2 = 5.64	0.018
Hemoglobin A1c (%)	7.8 ± 1.4	7.3 ± 1.2	8.5 ± 1.3	t = 7.82	< 0.001
HbA1c ≥ 8.0%	142 (45.5)	62 (33.7)	80 (62.5)	χ2 = 25.3	< 0.001
Serum albumin, g/L	36.8 ± 4.2	38.2 ± 3.7	34.7 ± 4.1	t = 7.45	< 0.001
Serum albumin < 35 g/L	118 (37.8)	52 (28.3)	66 (51.6)	χ2 = 17.2	< 0.001
Peripheral neuropathy score	6.2 ± 2.1	5.4 ± 1.9	7.3 ± 1.8	t = 8.12	< 0.001
Neuropathy score ≥ 7	156 (50.0)	68 (37.0)	88 (68.8)	χ2 = 31.5	< 0.001
Peripheral arterial disease	164 (52.6)	74 (40.2)	90 (70.3)	χ2 = 27.9	< 0.001
ABI	0.94 ± 0.21	1.02 ± 0.18	0.82 ± 0.19	t = 8.93	< 0.001
ABI < 0.9	147 (47.1)	58 (31.5)	89 (69.5)	χ2 = 44.2	< 0.001
Previous ulcer duration > 8 weeks	136 (43.6)	52 (28.3)	84 (65.6)	χ2 = 43.1	< 0.001
Inadequate offloading compliance	143 (45.8)	56 (30.4)	87 (68.0)	χ2 = 43.7	< 0.001

HbA1c: Hemoglobin A1c; ABI: Ankle brachial index.

Univariable Cox regression analysis

We performed univariable Cox regression to identify candidates from 11 variables which could have an independent effect on DFU recurrence (Table 2). The strongest multivariable-adjusted recurrence predictors (highest HRs) were previous ulcer duration > 8 weeks (HR = 3.68, P < 0.001), inadequate offloading compliance (HR = 3.42, P < 0.001), PAD (HR = 3.21, P < 0.001), ABI < 0.9 (HR = 3.15; P < 0.001) and peripheral neuropathy score ≥ 7 (HR = 2.94; P < 0.001). Similarly, moderate and strong associations were found with metabolic factors of HbA1c ≥ 8.0% (HR = 2.53, P < 0.001) and serum albumin < 35 g/L (HR = 2.18, P < 0.001) respectively. Diabetes duration (HR = 1.24 per 5 years; P < 0.001) and current smoking (HR = 1.68; P = 0.007) were moderate associations. Neither sex or age was associated with the incidence of recurrence legitimacy. All variables with P < 0.10 underwent multivariable modelling.

Table 2 Univariable Cox proportional hazards regression analysis.

Variable	Hazard ratio (95%CI)	P value
Age (per 10-year increase)	1.08 (0.92-1.27)	0.342
Male sex	1.16 (0.79-1.70)	0.452
Diabetes duration (per 5-year increase)	1.24 (1.11-1.39)	< 0.001
Current smoking	1.68 (1.15-2.45)	0.007
HbA1c ≥ 8.0%	2.53 (1.74-3.68)	< 0.001
Serum albumin < 35 g/L	2.18 (1.51-3.14)	< 0.001
Peripheral neuropathy score ≥ 7	2.94 (2.02-4.28)	< 0.001
Peripheral arterial disease	3.21 (2.19-4.70)	< 0.001
ABI < 0.9	3.15 (2.16-4.59)	< 0.001
Previous ulcer duration > 8 weeks	3.68 (2.52-5.37)	< 0.001
Inadequate offloading compliance	3.42 (2.34-5.00)	< 0.001

HbA1c: Hemoglobin A1c; ABI: Ankle brachial index.

Independent predictors and multivariable Cox regression

Seven independent predictors of recurrence were determined using both LASSO variable selection and multivariable Cox regression, and all obtained results met the proportional hazards assumption. The strongest predictor was: Previous ulcer duration > 8 weeks (HR = 3.24, 95%CI: 2.15-4.88, P < 0.001), followed by; PAD (HR = 2.87, P < 0.001), inadequate offloading compliance (HR = 2.73, P < 0.001), ABI < 0.9 (HR = 2.56, P < 0.001), severe peripheral neuropathy (HR = 2.45, 95%CI: 1.63-3.68, P < 0.001), HbA1c ≥ 8.0% (HR = 2.18, P < 0.001), and hypoalbuminemia (HR = 1.92, P = 0.002, Table 3).

Table 3 Multivariable Cox proportional hazards regression analysis.

Variable	Hazard ratio (95%CI)	P value
Previous ulcer duration > 8 weeks	3.24 (2.15-4.88)	< 0.001
Peripheral arterial disease	2.87 (1.92-4.29)	< 0.001
Inadequate offloading compliance	2.73 (1.82-4.09)	< 0.001
ABI < 0.9	2.56 (1.71-3.83)	< 0.001
Peripheral neuropathy score ≥ 7	2.45 (1.63-3.68)	< 0.001
HbA1c ≥ 8.0%	2.18 (1.46-3.25)	< 0.001
Serum albumin < 35 g/L	1.92 (1.28-2.88)	0.002

HbA1c: Hemoglobin A1c; ABI: Ankle brachial index.

Nomogram development and point assignment system

Nomogram construction and risk stratification: Based on the multivariable Cox regression coefficients, we established a nomogram point assignment system (Figure 1). Previous ulcer > 8 weeks were given the highest allocated points (40) followed by PAD (35), poor adherence to offloading (33), ABI < 0.9 (32), neuropathy score ≥ 7 (30), HbA1c ≥ 8.0% (27) and serum albumin < 35 g/L (23). The maximum score is 220 points, with each present risk factor contributing a certain amount to the total score (0 for none). This risk score can therefore provide a rapid assessment of individual patient risk at the bedside by using total point scores, which correspond directly with predicted recurrence probabilities for use in counseling and planning management.

Figure 1 Nomogram for predicting 1-year and 2-year diabetic foot ulcer recurrence probability. The nomogram consists of point scales for each of the seven predictors, a total points scale, and prediction scales for 1-year and 2-year recurrence probabilities. ABI: Ankle brachial index; HbA1c: Hemoglobin A1c.

Model discrimination and calibration performance

Table 4 shows the model performance metrics in both training and internal validation cohorts, which indicated excellent discrimination and calibration for each respective model. The C-index values across both datasets were > 0.82, confirming better predictive accuracy. In both training and internal validation cohorts, the model exhibited strong discrimination [C-index = 0.832 (95%CI: 0.789-0.875) in the internal training cohort; C-index = 0.826 (95%CI: 0.787-0.865) in the internal validation cohort] which means that if two patients are randomly selected, then approximately 83% of the time their true recurrence risk order is correctly rank ordered by our model (i.e., area under the curve). C-indexes were well above the 0.75 threshold commonly regarded as a good predictor, and outperformed most previously reported DFU prediction models. The small distance of 0.006 between training and validation C-index values indicates a satisfactory level of overfitting, demonstrating the model’s good generalizability.

Table 4 Model performance metrics.

Metric	Training cohort (n = 218)	Validation cohort (n = 94)
C-index (95%CI)	0.832 (0.789-0.875)	0.826 (0.787-0.865)
Sensitivity at 1 year	82.3%	80.6%
Specificity at 1 year	78.9%	77.4%
Positive predictive value	74.2%	73.1%
Negative predictive value	85.7%	84.3%
Calibration slope	0.98	0.96
Calibration intercept	0.02	0.04

CI: Confidence interval.

Risk stratification and clinical categories

Patients were stratified into three groups, with outcomes clearly differing on the basis of nomogram total points. Patients in the low-risk category (0-100 points, n = 89) had 1- and 2-year recurrence rates of merely 8.1% and 12.3%, thereby supporting standard surveillance protocols. As a result, the rates of 28.4% and 38.7% among intermediate-risk patients (101-180 points, n = 135) were much higher and should be supported by more intensive modification of risk factors and follow-up. Patients with high risk (> 180 points, n = 88) demonstrated high rates of recurrence at 1 year and 2 years (58.9% and 71.2%, respectively), necessitating intensive multifactorial prevention involving continued follow-up appointments, individualized in-depth treatment footwear/orthosis provision, as well as proactive management of hyperglycemia and malnutrition. The model was able to discriminate between the low- and high-risk groups with a 2-year recurrence difference of almost six-fold (12.3% vs 71.2%), which indicates the strong discriminatory power and clinical utility of this model (Table 5).

Table 5 Risk stratification groups and recurrence rates.

Risk group	Point range	1-year recurrence rate	2-year recurrence rate
Low risk	0-100	8.1% (4.2%-12.0%)	12.3% (7.5%-17.1%)
Intermediate risk	101-180	28.4% (22.1%-34.7%)	38.7% (31.8%-45.6%)
High risk	> 180	58.9% (51.2%-66.6%)	71.2% (63.8%-78.6%)

Kaplan-Meier survival analysis

Kaplan-Meier curves stratified by risk group are displayed in Figure 2, which graphically shows the impressive separation of recurrence-free survival among the three groups. The curves diverge quickly within the first 6 months after initial wound healing, with clear separation up to at least the 24-month follow-up period. Log-rank test showed highly significant differences in recurrence-free survival distributions by risk group (P < 0.001). Median recurrence-free survival was not reached for the low-risk group (indicating that > 50% remained ulcer free throughout follow-up), 18.5 months (95%CI: 16.2-20.8 months) for the intermediate-risk group, and only 9.8 months (95%CI: 8.3-11.3 months) for those with high-risk features high-risk vs low-risk patients had an 8.7 (95%CI: 5.4-13.9, P < 0.001). HR for recurrence at any given time point, conferring instantaneous risk of recurrence approximately nine times that of the low-risk cohort. These curves visually solidify the ability of the model to classify patients into significantly different prognostic groups and may represent a convenient method to communicate risk to providers or patients (Figure 2).

Figure 2 Kaplan-Meier curves showing recurrence-free survival stratified by risk group (low, intermediate, high). Patients with healed diabetic foot ulcers were stratified by total nomogram points into low-risk (0-100 points; n = 89), intermediate-risk (101-180 points; n = 135), and high-risk (> 180 points; n = 88) groups. Recurrence-free survival showed clear stepwise separation across risk categories, with progressively worse outcomes in higher-risk patients. The curves diverged within the first 6 months after wound healing and remained separated through follow-up. Differences among groups were significant by log-rank test (P < 0.001). Median recurrence-free survival was not reached in the low-risk group, compared with 18.5 months in the intermediate-risk group and 9.8 months in the high-risk group.

Temporal validation

An independent cohort of 156 patients from the same institution (Shanghai Municipal Hospital of Traditional Chinese Medicine, treated during the later period January 2022-July 2023) was used for temporal validation. The overall model showed good discrimination (C-index: 0.814, 95%CI: 0.764-0.864) with only a modest performance decrement compared to internal validation, confirming temporal stability of model performance. The three-tier risk stratification also showed consistent performance with 2-year recurrence rates for the validation cohort being similar to those in the development cohort across all risk groups (low-risk 14.1% vs 12.3%, intermediate-risk 41.2% vs 38.7%, high-risk 68.7% vs 71.2%) and within less than a median of ± 3 percentage points. No systematic miscalibration across the full risk range, with calibration plot showing close agreement of predicted and observed probabilities: Slope 0.94 (95%CI: 0.87-1.01); intercept 0.06 (95%CI: -0.02 to 0.14). Nomogram was also confirmed clinically useful by decision curve analysis, where its net benefit was greater than either the “treat all” or the “treat none” strategy from threshold probabilities of 10% to 80%, especially when between 30% and 40% which are corresponding to intermediate-risk patients. Taken together, these results validate the robustness of the nomogram across multiple performance metrics and support its clinical application in personalized preventive management (Figures 3, 4, and 5).

Figure 3 Temporal validation calibration plot. The calibration plot shows close agreement between predicted and observed 2-year recurrence probabilities in the temporal validation cohort. Observed rates (blue circles) align well with the line of perfect calibration (dashed line), with calibration slope of 0.94 and intercept of 0.06, confirming excellent model calibration across the risk spectrum. CI: Confidence interval.

Figure 4 Risk stratification performance: Development vs temporal validation. The three-tier risk stratification system showed consistent 2-year recurrence rates between development and validation cohorts across low-risk (12.3% vs 14.1%), intermediate-risk (38.7% vs 41.2%), and high-risk (71.2% vs 68.7%) groups, with differences under 3 percentage points confirming model transportability.

Figure 5 Decision curve analysis - clinical utility of the nomogram. Decision curve analysis shows the nomogram (blue) provides substantial net benefit compared to uniform strategies (treat all/none) across 10%-80% threshold probabilities, with peak benefit at 30%-40% (light-blue shaded), supporting risk-stratified clinical implementation.

DISCUSSION

This single-center retrospective cohort study aimed to develop and temporally validate a new risk stratification model for recurrence of DFU post-healing. Based on individual cardiovascular risk factors, our model included 7 clinical, vascular, metabolic and behavioral parameters which provided C-index of 0.826 at one year follow-up with well calibrated models. The nomogram-based tool categorizes patients into three subgroups with very distinct 2-year recurrence rates (12.3%, 38.7% and 71.2% for the low-, intermediate- and high-risk subgroups, respectively) - enabling targeted use of preventive resources and personalized follow-up strategies.

The model detected seven independent predictors of DFU recurrence; previous ulcer duration, PAD, offloading compliance and ABI for the systemic component; peripheral neuropathy severity for the local component; and glycemic control, and serum albumin as components of overall host function to absorb a major wound load. The strongest predictor compared to all other factors was the duration of prior ulcers (HR = 3.24), probably just a reflection, in simple form, of tissue quality, healing potential and mechanical stress at the wound-related site over time. The chronic wound characteristic may imply more extensive tissue loss, decreased vascularity or the presence of biofilm all factor associated with increased chances of recurrent breakdown[22-24].

Low ABI and degeneration in patients with peripheral arterial disease in that our data show that arterial insufficiency of a limb not only impairs early wound healing, but also long-term survival as well as tolerance to mechanical forces. Is PAD a good marker of high vascular morbimortality? Those with PAD had almost 3-fold higher odds for recurrent events thus justifying the aggressive modification of risk factors (smoking cessation, lipid reduction, antiplatelet) and vascular intervention in selected patients. Guidelines have since focused on revascularization to reduce the risk of recurrent ulceration in patients with PAD visible[17,18].

An adjusted HR of 2.45 was identified to predict recurrence independently with a cutoff for severe peripheral neuropathy (score ≥ 7). Protected/corrected sensation should not be performed, as such practice dulls sensory feedback and the normal reflex of promptly redistributing weight away from areas being subjected to excessive pressure is lost; thus allowing continued erosive trauma at sites which may traditionally be considered propitious. Motor neuropathy will create foot deformities (claw toes, prominent metatarsal heads) with an emphasis on areas of concentrated plantar pressure at focal points and increase the risk of ulceration[25-27]. Given the strong association of neuropathy severity with recurrent foot ulceration, patient education regarding appropriate footwear, routine foot inspection and preventive measures (custom orthotics, therapeutic footwear) in at-risk populations is a critical approach[5,28].

Recent advances in the effort to understand how our patients respond to different interventions are being made using real-world data[20,29], and offloading compliance became an important modifiable behavioral predictor with high impact (HR = 2.73). Although it has been shown repeatedly that adequate offloading needs to reduce plantar pressure to prevent recurrent ulceration, the adherence in clinical practice is still very poor[30,31]. Our results are suggestive that poor adherence with offloading increases the risk of recurrent ulceration more than twofold, highlighting the need for novel approaches to enforcing compliance such as patient education, telehealth monitoring and development of less visible offloading technologies maintaining an appropriate balance between efficacy, wearability and cosmesis[32].

The independent predictors for recurrence included poor glycemic control (HbA1c ≥ 8.0%) that is consistent with the extensive literature showing the deleterious effects of hyperglycemia on impaired wound healing through several mechanisms including impaired neutrophil production, decreased angiogenesis, increased oxidative stress[33,34] and advance glycosylated end products accumulation[35]. Both glycemic control is a modifiable exogenous condition subject to treatment. The result of our study suggests that a vigorous control of blood glucose is an important element in broad strategies for prevention of DFU, especially act aggressively to maintain the blood glucose level when the ulcer healed because this period with higher rate of recurrence[36].

Hypoalbuminemia (< 35 g/L) was an indicator of both nutritional status and systemic inflammation, causing global impairment in healing ability[37,38]. Since albumin is a surrogate marker of protein calorie malnutrition and functions to affect collagen synthesis, immune function and treatment of the wound strength. Furthermore, low albumin may also be the reflection of chronic inflammatory and/or other comorbid conditions (e.g., chronic kidney disease, heart failure) that determine a suboptimal overall host status and healing ability. In our paper, we highlight the importance of nutritional assessment and optimization in DFU patients with particular attention paid to protein intake and correction of micronutrient deficiencies.

Strengths of our model one-stop shop for integrating all risk domains into one capable tool; vascular, neurologic metabolic, nutritional and behavioral. Previous prediction model, that evaluated isolated risk factors or only targeted a small number of domains with respect to recurrence risk, are unlikely to explain such interactions necessary to fully account for the pathophysiology of latent-residual disease and recurrence[20,21,29,39]. Conclusion: Our clinical prediction model, using only simple clinical features measured at routine scheduled follow-up visits, outperformed most clinically-implemented models (C-index 0.826).

The clinically actionable three-tier risk stratification system. Standard surveillance protocols (quarterly clinic visits and annual comprehensive foot assessment) may be suitable for low-risk patients (12.3% 2-year recurrence rate) with simple ulceration/predisposing condition. Moderate-risk patients (38.7% recurrence rate) need more intensive preventive measures, including monthly visits in which risk factors are closely monitored and controlled; patient education should also be intensified by emphasizing that the foot should be inspected daily for problems, which must then be treated appropriately. Patients identified as high-risk (e.g. 71.2% the highest risk for recurrence) should be offered comprehensive and individualized prevention programs that include frequent clinic visits every few weeks (e.g. once every 2-4 weeks), customized therapeutic footwear with regular monitoring of shoe fit and wear pattern changes, components required by patients to correct or mitigate specific biomechanical abnormalities, possible prophylactic surgery correction of foot deformities[30], and potentially novel interventions such as telehealth-enabled monitoring devices that detect pre-ulcerative inflammation at least a week before an ulcer occurs[31].

The output C-index on an independent cohort from the same institution was similar (C-index 0.814) and valuable to confirm both the model's reproducibility and its temporal performance stability. The decision curve analysis demonstrated significant clinical utility since the nomogram revealed net benefit over a wide range of decision thresholds, suggesting that risk-based management strategies generated by our model would enhance outcomes over uniform approaches. Therefore, this nomogram can guide the choice of individualised treatment in clinical practice.

There are some limitations of the current study that need to be addressed. First, this was a single-centre study at a tertiary diabetic foot centre in Shanghai; due to differences in ethnic background, pre-existing genetic risk factors, dietary patterns and local treatment protocols between the populations we do not assume that the model’s performance will generalise to other populations or healthcare systems and external validation is required in multicentre, internationally diverse cohorts prior to broad clinical adoption. Second, all predictor variables were assessed at a single time-point; clinically important time-varying factors such as change in glycaemic control, progression of PAD and new-onset renal insufficiency are not recorded and it would be interesting to explore future dynamic models that incorporate such covariates. Third, offloading compliance was determined via patient interview and review of the medical record, which are subject to recall and social desirability bias; objective electronic monitoring was not available, and this should be investigated in future prospective studies. Fourth, the model was created solely as an approach to patients with first-time healed DFU; this is not intended for use in those with recurrent DFU or lower extremity amputation and separate modelling to that population is needed. Fifth, the retrospective design has a risk of selection bias and unmeasured confounding and we could not perform an analysis on the effect of specific interventions (e.g., revascularisation, type of offloading device) on recurrence risk. Lastly, novel prognostic factors (such as inflammatory cytokines, matrix metalloproteinases and growth factors) and new imaging technologies (e.g., magnetic resonance imaging, fluorescence angiography) could contribute more prognostic information and be integrated into future model modifications. The generalizability of the findings in non-specialist care settings and clinical efficacy and cost-effectiveness of this risk stratification tool need to be tested in pragmatic implementation studies.

CONCLUSION

This model represents a successfully developed and validated holistic risk stratification model for DFU recurrence following initial wound healing. A nomogram was formulated using seven simple clinical and metabolic variables (duration of previous ulcer, history of PAD, compliance with offloading, ABI, severity of peripheral neuropathy, glycemic control and serum albumin) allowing individualized risk prediction showing excellent discrimination (C-index 0.826) and calibration as well. The three-tier risk stratification system accurately identifies patients at low, intermediate and high risk with significantly different 2-year recurrence rates (12.3%, 38.7% and 71.2%, respectively) thus allowing personalized prevention strategies and a sustainable use of healthcare resources.