Distinct urine and plasma metabolic signatures in diabetic foot: early diagnostic biomarkers and predictive modeling

Abstract

Background and Objective

Diabetic foot (DF), a limb-threatening complication associated with high risk of amputation, currently lacks reliable early diagnostic biomarkers. This study aims to identify novel DF-specific metabolic biomarkers and develop predictive models for early diagnosis by integrating serum and urine metabolomic profiling.

Methods

Serum and urine samples were collected from patients with diabetic foot and those with diabetes mellitus without foot complications. Metabolomic and lipoprotein profiles were quantitatively analyzed using multivariate statistical methods to identify metabolic alterations associated with DF. Differential metabolites were used to construct a machine learning-based predictive model for early DF diagnosis.

Results

Distinct metabolic profiles differentiated DF from DM patients. Serum analysis revealed significantly lower hemoglobin, albumin, calcium, and apolipoprotein A1 levels in DF (P < 0.05). Urine metabolomics identified elevated N-isovaleroylglycine (OR = 12.89) and valine (OR = 2.23) as key DF-associated metabolites (P < 0.05). Lipidomics demonstrated increased triglyceride-rich LDL subtypes (L2TG, L4TG) and reduced high-density lipoprotein components (H4CH, H4PL) in DF. A predictive model integrating urinary metabolites (N-isovaleroylglycine, valine) and clinical profiles (albumin, apolipoprotein A1, calcium) achieved robust diagnostic accuracy (AUC = 0.91).

Conclusion

This study reveals distinct metabolic disturbances in DF through integrated metabolomic analysis. The combination of urinary metabolites and clinical biomarkers provides a non-invasive approach for early detection of DF, highlighting the potential utility of metabolomics in improving early diagnosis and management of diabetic foot.

Clinical trial number

Not applicable.

Introduction

Diabetic foot (DF), a severe complication of diabetes mellitus, is characterized by chronic non-healing wounds accompanied by deep tissue destruction, primarily resulting from lower extremity neuropathy and peripheral arterial diseas [1]. Globally, DF affects approximately 18.6 million people and is associated with a 5-year mortality rate of 50%, which rises to over 70% in patients requiring major amputations [2]. Current clinical management often relies on the Wagner-Meggitt (WM) classification system (G1–G6), which guides therapeutic decisions based on ulcer depth and severity [3]. However, this system—like other existing classification tools—depends on the presence of visible wounds, thereby limiting its applicability for early-stage detection and intervention.

Although early diagnostic approaches such as the Michigan Neuropathy Screening Instrument (MNSI), ankle-brachial index (ABI), and transcutaneous oxygen pressure (TcPO₂) do not require visible wounds, their diagnostic sensitivity and specificity for DF remain controversial [4]. Metabolomics, the large-scale study of small molecules, has become an indispensable approach for deciphering metabolic alterations in biological systems and offers unique insights into health, disease [5], and therapeutic interventions [6].

As the downstream product of genomic, transcriptomic, and proteomic activity, the metabolome provides a direct snapshot of the physiological state and phenotypic manifestation of a biological system [7], making metabolomics a powerful tool for discovering biomarkers and understanding disease mechanisms [8]. In DF research, serum and urine metabolite profiling has demonstrated significant potential. Although prior studies have suggested that certain metabolic factors impact the profile of DF patients [9, 10], a holistic analysis of DF-specific metabolites and their dynamic changes during disease progression remains incomplete. Elucidating the relationship between metabolite alterations and DF advancement is critical for developing novel diagnostic biomarkers and targeted therapies.

This study aims to provide a comprehensive characterization of the metabolic signature in individuals at risk of diabetic foot (DF), by analyzing urine metabolic profiles, plasma metabolic profiles, and plasma lipoprotein components. The systematic evaluation of differential metabolite expression across this multi-biosample combination enables a more holistic revelation of the pathophysiological state through information complementarity and cross-validation. Based on this strategy, “consensus metabolic features” that exhibit significant and consistent changes across multiple dimensions can be screened. This approach significantly enhances the sensitivity and specificity of early diagnosis. Our findings hold the potential to establish a foundation for early diagnostic strategies and inform future therapeutic interventions.

Materials and methods

Study population

This study prospectively collected clinical data from patients with type 2 diabetes mellitus (T2DM) enrolled at the Department of Vascular Surgery, Fujian Medical University Affiliated First Hospital, during a predefined observational period from March to June 2023. Diabetes diagnosis followed the World Health Organization (WHO) Diagnosis and Management of Type 2 Diabetes [11], The diagnosis of diabetic foot (DF) was established in accordance with the 2023 International Working Group on the Diabetic Foot (IWGDF) guidelines, defined by the absence of amputation history and no evidence of gangrene or severe ulceration, as classified by the Wagner risk stratification system (grades 0-III) [1]. Exclusion criteria included: (1) renal insufficiency or history of dialysis; (2) severe cardiovascular or cerebrovascular diseases; (3) acute infections or antibiotic use within the past month; (4) missing fasting blood glucose or glycated hemoglobin (HbA1c) data; and (5) incomplete blood biochemical or urinalysis results(Fig. 1). This study was approved by the Ethics Committee of the First Affiliated Hospital of Fujian Medical University (approval number: FMU [2023] 508). All participants provided informed consent.

Fig. 1.

The flowchart of patient screening and the case-control study design. DF, Diabetic Foot; DM, Diabetes Mellitus

Medical data collection

Medical data collection involved trained researchers retrieving historical medical records and conducting telephone follow-ups to obtain demographic and clinical information, including age, sex, hypertension, dyslipidemia, cardiovascular diseases, and cerebrovascular diseases. Laboratory analyses focused on routine hematological parameters, urinalysis, and biochemical indicators (e.g., serum albumin, calcium, lipid profiles, and platelet counts) at enrollment.

Metabolite quantification

Carefully thaw the plasma samples, which have been stored at -80 °C, at room temperature. Then, pipette 250 µL of Bruker plasma/serum buffer into Eppendorf container, and add 250 µL of serum/plasma. Shake the mixture gently for 1 min, and transfer 500 µL of well mixed sample into 5 mm SampleJet rack tube. And urine samples are centrifuged and aliquoted. To account for concentration variations, metabolic profiles are normalized to urinary creatinine concentration measured by a standard enzymatic assay. All aliquots are snap-frozen and stored at -80 °C for stability.

Nuclear magnetic resonance (NMR) spectroscopy was performed on BRUKER AVANCE IVDr spectrometer (Bruker BioSpin, GmBH, Rheinstetten, Germany). All 1 H-NMR spectra were obtained by employing a 310 K and 600.13 MHz proton Larmor frequency NMR spectrometer with a 5 mm BBI probe at the Zhejiang NUTRIEASE Science and Technology corporation [12].

Automated quantitative analysis of 41 metabolites was performed using the Bruker IVDr Plasma/Serum Quantification Platform (B.I.Quant-PS™). Lipoprotein subclasses and lipid profiles were analyzed via the Bruker IVDr Lipoprotein Subclass Analysis Platform (B.I.LISA™), quantifying 114 lipid parameters. These included 15 lipoprotein subclasses (5 VLDL [0.950–1.066 kg/L], 6 LDL [1.019–1.063 kg/L], and 4 HDL [1.063–1.210 kg/L]) and their compositional metrics—such as apolipoproteins (apoA1, apoA2, apoB), cholesterol esters, free cholesterol, phospholipids, triglycerides, and particle counts of specific subtypes [13, 14].Additionally, plasma inflammatory markers (GlyA, GlyB, GlyC, SPC, and Glyc/SPC) were quantified using the Bruker PhenoRisk PACS RuO platform [15, 16] (Figure 2).

Fig. 2.

Workflow of ¹H-NMR-based metabolomic analysis​. The schematic illustrates the pipeline for metabolomic profiling: biological samples (blood, urine) undergo preparation, ¹H-NMR spectroscopy data acquisition, and subsequent multivariate statistical analysis (e.g. heatmap, PCA, OPLS-DA) to identify significant metabolic features

Statistical analysis

Categorical variables were presented as numbers and percentages, while continuous variables were expressed using the interquartile range (IQR). Group comparisons for categorical variables were performed using Fisher’s exact test or Pearson’s Chi-squared test. Continuous variables were analyzed using the Wilcoxon rank sum test. To control the false positive rate in multiple hypothesis testing, false discovery rate (FDR) correction via the Benjamini-Hochberg method was applied to the statistical tests. To ensure model robustness and external validity, the dataset was randomly partitioned into a training set (70%) and a validation set (30%).

We first employed heatmap analysis and principal component analysis (PCA) technology to systematically explore differences in metabolic components among different groups, thereby evaluating the predictive value of each metabolic indicator in the early diagnosis of diabetic foot (DF). Based on the above analysis results, key metabolic biomarkers were screened, and significant covariates were selected via univariate logistic regression. Finally, a multivariable logistic regression model was constructed to analyze the diagnostic efficacy of the blood metabolomics model, urine metabolomics model, lipid metabolomics model, and clinical indicator model while controlling for confounding factors.

The discriminatory performance of predictive models was evaluated using receiver operating characteristic (ROC) curves, with the area under the curve (AUC) reflecting diagnostic accuracy. Calibration curves assessed the agreement between predicted probabilities and observed outcomes, while decision curve analysis (DCA) quantified clinical utility across varying decision thresholds. These metrics collectively evaluated the predictive accuracy and generalizability of metabolomic models. All analyses were conducted using R software (version 4.2.1). A two-sided P-value < 0.05 was considered statistically significant.

Results

Baseline characteristics

This study analyzed serum and urine samples from the DM and DF groups. The blood metabolomics cohort included 65 participants (23 DF and 42 DM) with a median age of 73 years (60% male). Significant differences were identified in age, hemoglobin (HB), platelet count (PLT), albumin (ALB), high-density lipoprotein (HDL), low-density lipoprotein (LDL), apolipoprotein A1 (APOA1), calcium (Cal), and sodium (Na) (all P < 0.05). Notably, all indicators except PLT were significantly lower in the DF group compared to DM (Table 1).

Table 1.

Patient characteristics of blood metabolomics cohort

Characteristic	Overall, N = 651	DF, N = 231	DM, N = 421	p-adj2
Female, n (%)	26 (40%)	10 (43%)	16 (38%)	0.672
Age	73 (65, 77)	67 (59, 74)	76 (67, 79)	0.004
HT	42 (65%)	14 (61%)	28 (67%)	0.640
DisLIP	4 (6.2%)	1 (4.3%)	3 (7.1%)	> 0.999
Celebro	2 (3.1%)	0 (0%)	2 (4.8%)	0.536
Cardio	3 (4.6%)	1 (4.3%)	2 (4.8%)	> 0.999
WBC	7.08 (5.98, 9.17)	7.16 (6.02, 11.32)	7.03 (5.70, 9.06)	0.414
Neu%	69 (61, 75)	72 (60, 75)	68 (63, 74)	0.588
Lym	1.51 (1.12, 1.96)	1.61 (1.37, 1.96)	1.44 (1.10, 1.95)	0.287
Mono	0.51 (0.40, 0.63)	0.51 (0.36, 0.75)	0.51 (0.43, 0.63)	0.832
HB	118 (102, 130)	102 (91, 117)	122 (110, 132)	< 0.001
PLT	263 (214, 345)	326 (254, 401)	245 (195, 306)	0.003
TBI	5.90 (4.80, 7.90)	5.60 (4.50, 6.50)	6.35 (4.85, 7.98)	0.138
ALB	38.6 (34.4, 41.0)	35.3 (30.1, 38.6)	39.0 (37.5, 42.4)	0.001
ALT	14 (10, 23)	14 (11, 19)	14 (9, 24)	0.863
AST	17 (14, 21)	17 (14, 19)	17 (14, 22)	0.492
CREA	74 (62, 95)	75 (62, 92)	74 (62, 100)	0.934
UA	301 (251, 353)	300 (241, 336)	303 (259, 373)	0.281
GLU	6.5 (5.3, 8.5)	7.1 (5.4, 10.2)	6.4 (5.1, 7.6)	0.172
TG	1.22 (0.90, 1.59)	1.22 (0.98, 1.35)	1.22 (0.79, 1.73)	0.721
HDL	0.91 (0.82, 1.13)	0.89 (0.79, 0.99)	0.96 (0.83, 1.16)	0.074
LDL	2.22 (1.59, 3.08)	2.00 (1.36, 2.47)	2.24 (1.89, 3.20)	0.073
VLDL	0.56 (0.41, 0.72)	0.56 (0.45, 0.61)	0.56 (0.37, 0.78)	0.907
APOA1	1.11 (0.96, 1.30)	1.03 (0.94, 1.11)	1.17 (1.02, 1.36)	0.003
APOB	0.83 (0.64, 1.03)	0.83 (0.62, 1.03)	0.83 (0.66, 1.07)	0.462
Cal	2.20 (2.13, 2.25)	2.15 (2.08, 2.22)	2.21 (2.18, 2.25)	0.017
Mg	0.91 (0.87, 0.97)	0.91 (0.83, 0.98)	0.91 (0.88, 0.97)	0.885
K	4.42 (4.09, 4.67)	4.46 (4.16, 4.67)	4.41 (4.08, 4.67)	0.550
Na	139.9 (137.4, 141.7)	138.5 (137.0, 140.2)	140.4 (138.5, 142.2)	0.064
Cl	102.8 (100.0, 104.3)	102.8 (98.2, 103.4)	102.8 (101.6, 104.8)	0.091
EPI	78 (62, 92)	78 (70, 92)	78 (59, 92)	0.660

¹n (%); Median (IQR)

²Pearson’s Chi-squared test; Wilcoxon rank sum test; Fisher’s exact test; FDR adjustment

DF: Diabetic foot, DM: diabetes mellitus, Neu%:​​ Neutrophil percentage (%), Lym:​​ Lymphocyte count (10⁹/L), Mono:​​ Monocyte count (10⁹/L), HB:​​ Hemoglobin, PLT:​​ Platelet count, TBI:​​ Total bilirubin, ALB:​​ Albumin, ALT:​​ Alanine aminotransferase, AST:​​ Aspartate aminotransferase, CREA:​​ Creatinine, UA:​​ Uric acid, GLU:​​ Glucose, TG:​​ Triglycerides, HDL:​​ High-density lipoprotein cholesterol, LDL:​​ Low-density lipoprotein cholesterol, VLDL:​​ Very low-density lipoprotein cholesterol, APOA1:​​ Apolipoprotein A1, APOB:​​ Apolipoprotein B, Cal:​​ Calcium, Mg:​​ Magnesium K:​​ Potassium, Na:​​ Sodium, Cl:​​ Chloride

In the urine metabolomics cohort (39 participants: 16 DF vs. 23 DM), consistent differences were observed for HB, PLT, ALB, HDL, LDL, APOA1, APOB, Cal, and Na (P < 0.05), with DF values predominantly lower than DM, aligning with blood metabolomics results (Table 2).

Table 2.

Patient characteristics of urine metabolomics cohort

Characteristic	Overall, N = 391	DF, N = 161	DM, N = 231	p-adj2
Female, n (%)	18 (46%)	8 (50%)	10 (43%)	0.688
Age	71 (65, 76)	69 (63, 73)	72 (67, 78)	0.109
HT	22 (56%)	8 (50%)	14 (61%)	0.501
HB	113 (102, 127)	103 (87, 114)	122 (108, 132)	0.002
PLT	270 (222, 353)	313 (264, 413)	244 (198, 310)	0.023
TBI	5.80 (4.75, 7.45)	5.55 (4.60, 5.83)	6.30 (4.90, 7.75)	0.162
ALB	38.6 (34.0, 40.8)	36.8 (31.6, 38.7)	39.1 (37.0, 42.6)	0.021
ALT	18 (12, 25)	18 (14, 23)	18 (11, 28)	0.689
AST	19 (15, 23)	18 (16, 20)	20 (15, 24)	0.607
CREA	75 (57, 94)	75 (64, 89)	75 (56, 104)	0.700
UA	306 (227, 371)	303 (205, 335)	312 (235, 415)	0.247
GLU	6.50 (5.44, 7.79)	7.42 (6.04, 10.18)	6.32 (5.15, 6.80)	0.056
TG	1.27 (0.97, 1.66)	1.25 (1.00, 1.29)	1.31 (0.86, 1.80)	0.637
HDL	0.90 (0.83, 1.13)	0.85 (0.74, 0.90)	1.05 (0.87, 1.18)	0.011
LDL	2.01 (1.61, 3.02)	1.73 (1.27, 2.03)	2.51 (1.96, 3.20)	0.005
VLDL	0.58 (0.46, 0.75)	0.57 (0.46, 0.59)	0.62 (0.45, 0.82)	0.407
APOA1	1.10 (0.96, 1.24)	1.04 (0.93, 1.10)	1.15 (1.10, 1.39)	0.002
APOB	0.83 (0.68, 1.01)	0.70 (0.60, 0.83)	0.85 (0.74, 1.05)	0.039
Cal	2.19 (2.12, 2.23)	2.15 (2.08, 2.19)	2.21 (2.17, 2.24)	0.024
Mg	0.89 (0.82, 0.97)	0.89 (0.82, 0.92)	0.89 (0.85, 0.97)	0.474
K	4.19 (4.05, 4.54)	4.33 (4.12, 4.58)	4.19 (4.05, 4.48)	0.529
Na	140.3 (137.2, 142.1)	138.2 (136.9, 140.3)	140.4 (139.2, 143.0)	0.023
Cl	102.8 (99.7, 103.8)	101.7 (97.7, 103.1)	102.8 (101.8, 104.4)	0.100
EPI	75 (58, 94)	77 (72, 89)	70 (57, 94)	0.440

¹n (%); Median (IQR)

²Pearson’s Chi-squared test; Wilcoxon rank sum test; Fisher’s exact test; FDR adjustment

DF: Diabetic foot, DM: diabetes mellitus, HT: Hypertension, HB: Hemoglobin, PLT: Platelet count, TBI: Total Bilirubin, ALB: Albumin, ALT: Alanine Aminotransferase, AST: Aspartate Aminotransferase, CREA: Creatinine, UA: Uric Acid, GLU: Glucose, TG: Triglycerides, HDL: High-Density Lipoprotein cholesterol, LDL: Low-Density Lipoprotein cholesterol, VLDL: Very Low-Density Lipoprotein cholesterol, APOA1: Apolipoprotein A1, APOB: Apolipoprotein B, Cal: Calcium, Mg: Magnesium, K: Potassium, Na: Sodium, Cl: Chloride, EPI:​​ Estimated glomerular filtration rate (eGFR) calculated using the Chronic Kidney Disease Epidemiology Collaboration (CKD-EPI) formula

We initially conducted hierarchical analysis on blood, urine, and lipid metabolites to identify differentially expressed metabolic profiles. In blood metabolomics (Fig. 3A), the heatmap revealed notable increase of 2-oxoglutaric acid, glutamine, and citric acid in the diabetes mellitus (DM) group (P < 0.01), while the methionine-K-EDTA conjugate was markedly decreased (p < 0.05). For urine metabolomics (Fig. 3B), seven metabolites exhibited significant differential expression: leucine, N-isovaleroylglycine, valine, 4-hydroxyhippuric acid (two isoforms), pantothenic acid, and cytosine (P < 0.05), all of which were increased in the DM group. Lipidomics analysis (Fig. 3C) demonstrated distinct alterations in lipid subtypes in the DM group (P < 0.01), characterized by significant accumulation of triglyceride subtypes (TPA1, LDTG, H4TG, L2TG, L4TG, L5TG, L6TG) and depletion of phosphatidylcholine (H4PL), cholesteryl esters (H3CH, H4CH), and fatty acid-binding proteins (H4A1, H4A2). Additionally, differential increase was observed for TPA2, HDA1, HDA2, and H4FC (P < 0.05). These findings highlight systemic metabolic dysregulation across biological matrices, with urine metabolites showing the most pronounced discriminative power between DM and diabetic foot groups.

Fig. 3.

Preliminary screening of discriminant metabolites between diabetic foot (DF) and diabetes mellitus (DM)​​: (A-C) Differential metabolite heatmaps show distinct expression patterns between diabetic foot (DF) and diabetes mellitus (DM) groups across serum (A), urine (B), and lipid (C) profiles. Red indicates high abundance, blue indicates low abundance. (D-F) Principal component analysis reveals separation trends in urine metabolome (D), serum (E) and lipid (F) profiles. Ellipses represent 95% confidence intervals. PCA score plots exhibited substantial spatial overlap between the DM and DF groups across all three metabolic profiles, (G) OPLS-DA demonstrates enhanced group separation in urinary metabolites, confirming discriminative capacity. (H) VIP analysis identifies N-isovalerylglycine (VIP = 2.38) as the top discriminant metabolite, with 6 additional metabolites exceeding VIP > 1.0 threshold

To further investigate intergroup disparities, we conducted principal component analysis (PCA) on blood, urine, and lipid metabolomic datasets (Fig. 3D-F). PCA score plots exhibited substantial spatial overlap between the DM and DF groups across all three metabolic profiles, demonstrating no statistically significant separation in global metabolic variation. This phenomenon may stem from the high heterogeneity of individual basal metabolic backgrounds, where disease-specific metabolic changes are masked by background metabolites with high abundance or substantial inter-individual variability.​.

Orthogonal partial least squares-discriminant analysis (OPLS-DA) was subsequently applied for in-depth metabolomic discrimination (Fig. 3G). The OPLS-DA results revealed that the urine metabolome achieved significant group separation, N-Isovaleroylglycine emerged as the primary discriminatory metabolite through variable importance in projection (VIP = 2.38, P < 0.001, Fig. 3H)), with its contribution significantly surpassing other metabolites (second-highest VIP = 1.74). ​This suggests that the metabolic signature of diabetic foot is more likely to reflect fine-tuned regulation of specific pathways rather than global metabolic disturbance.​.

In order to evaluate the clinical potential of the screened metabolites for early diagnosis of diabetic foot, we constructed multivariate regression models based on differential metabolites, including a clinical indicator model, a plasma metabolomics model, a urinary metabolomics model, and a lipid metabolomics model. In the clinical indicator model ( Figure 4 A ), all five indicators showed significant associations with DF occurrence: platelet count (PLT, OR = 2.23, P = 0.017) was identified as an independent risk factor for DF, while hemoglobin (OR = 0.29, P = 0.002), albumin (OR = 0.22, P = 0.004), serum calcium (OR = 0.46, P = 0.021), and APOA1 (OR = 0.40, P = 0.013) exhibited significant protective effects. In the urinary metabolomics model ( Figure 4B ), N-isovaleroylglycine (OR = 12.89, P = 0.026) and valine (OR = 2.23, 95% CI: 1.07–5.42, P = 0.048) were significantly associated with increased DF risk. The plasma metabolomics model ( Figure 4 C ) revealed that only methionine (OR = 0.57, P = 0.041) was negatively correlated with DF risk, suggesting its potential protective role, whereas other metabolites showed no statistically significant associations. The lipid metabolomics model ( Figure 4D ) demonstrated that low-density lipoprotein components such as L2TG (OR = 2.71, P = 0.003), L4TG (OR = 2.92, P = 0.002), and L5TG (OR = 2.00, P = 0.021) were significant risk factors, while high-density lipoprotein components including H4CH (OR = 0.31, P = 0.001), H4PL (OR = 0.33, P = 0.001), and H4A1 (OR = 0.32, P = 0.001) exerted strong protective effects. These findings indicate that the accumulation of low-density lipoprotein subclasses might be positively associated with DF risk, whereas the maintenance of high-density lipoprotein subclasses might confer protection against its pathological progression.

Fig. 4.

Multivariate analysis of clinical and metabolic predictors for diabetic foot risk​. This composite figure presents the association between various parameters and diabetic foot risk using forest plots. Each panel displays odds ratios (OR) with 95% confidence intervals (parentheses) and P-values for different variable categories. The vertical dashed line at OR = 1 indicates the null value. (A) Forest plot of clinical indicators:​​ Hemoglobinand albumin (OR = 0.22, P < 0.001) demonstrate significant protective effects, while platelet count shows elevated risk. (B) Forest plot of Serum metabolomics markers:​​ Glutamine methionine is risk factors. (C) Forest plot of Urine metabolomics markers:​​ Valineand N-isovalerylglycine show significant risk associations. (D) Forest plot of Lipoprotein subclass parameters:​​ HDL subfractions shows protective effects, while L2TG, L4TG and L5TG indicate elevated risk

Establishment and validation of prediction models

We developed a clinical prediction model integrating selected clinical indicators and urinary metabolomics to stratify diabetic foot risk. During training, the initial model combining full clinical features and metabolomics achieved an AUC of 1.00 (Fig. 5A), but dropped to 0.71 in the validation set, indicating overfitting (Fig. 5B). After excluding overfitting contributors (hemoglobin and platelet count), the refined integrated model demonstrated strong performance with an AUC of 0.90 in both cohorts, confirming its generalizability (AUC = 0.90). A nomogram was constructed for clinical translation (Fig. 5E), incorporating stable metabolomic and clinical predictors. Calibration curves showed excellent agreement between predicted and observed outcomes, with a consistent mean absolute error (MAE) of 0.077 in both sets (Fig. 5C-D). Furthermore, Decision curve analysis (DCA) confirmed superior net clinical benefit over default strategies across cost-to-benefit thresholds from 1:4 to 4:1, peaking at 2:3 (Fig. 5F). In summary, the finalized model achieves high predictive accuracy, robustness, and clinical applicability, enabling personalized risk assessment and intervention planning without overtreatment.

Fig. 5.

Comprehensive validation of the diabetic foot prediction model:​ Multidimensional performance evaluation using training and independent validation cohorts. (A) Receiver operating characteristic (ROC) curves of the training set. (B) ROC curves of the validation set. (C) Training set calibration curve. (D) Validation set calibration curve. (E). Clinical implementation nomogram: The nomogram integrates five predictors: Valine; N-Isovalerylglycine; ALB: albumin; APOA1:apolipoprotein A1 and Cal: serum calcium. part clinical: ALB; APOA1and Cal.(​F). Decision curve analysis (DCA): Net clinical benefit across threshold probabilities (x-axis). Red curve: Nomogram model outperforms “treat-all” and “treat-none” strategies, demonstrating clinical utility for risk thresholds 10%-35%.AUC: Area under the ROC curve; urine Met: Urine Metabolomics; plasma Met: Serum metabolomics; plasma LP: Lipoprotein subclasses; Clinical: clinical indicators; part Clinical: clinical indicators excluded HB and PLT. Apparent: Apparent calibration, uncorrected for overfitting; Bias-correct: Bias-corrected calibration; and Ideal: Ideal reference line, y = x

Dissussion

In this study, we evaluated the comprehensive metabolic profiles of diabetic foot patients and compared them with non-complicated DM individuals to identify distinct metabolic signatures and develop diagnostic models. A predictive model incorporating two metabolites (N-isovaleroylglycine and valine) and three clinical indicators (albumin [ALB], apolipoprotein A1 [APOA1], and calcium [Cal] demonstrated robust performance in predicting DF.

Compared to the DM group, DF patients exhibited significantly lower serum levels of hemoglobin (HB), ALB, Cal, and APOA1. These findings align with prior studies suggesting that reduced APOA1 may serve as an early marker of cardiovascular complications in diabetic foot ulcer (DFU) patients [9, 17]. Declines in HB, Cal, and ALB likely reflect systemic nutritional impairment due to diabetes progression, particularly in DFU patients with limited protein intake [17]. Xie et al. further demonstrated that the Geriatric Nutritional Risk Index (GNRI) independently predicts amputation and mortality in DFU after adjusting for confounding risk factors [17, 18]. Notably, low hemoglobin levels exacerbate limb ischemia due to diminished oxygen-carrying capacity, while anemia-induced hyperdynamic circulatory state may promote thrombogenesis, accelerating DFU progression [19, 20].

In our blood metabolomics analysis, methionine (Met) was identified as a protective factor against diabetic foot (DF) (OR = 0.57, 95% CI: 0.32–0.96, P = 0.041). This finding appears somewhat paradoxical in light of existing evidence, which implicates homocysteine (Hcy)—a key metabolite in methionine metabolism—as a well-established inflammatory marker. Hyperhomocysteinemia (HHcy) has been linked to an elevated risk of cardiovascular diseases and diabetic complications [21]. Nevertheless, emerging evidence suggests that methionine itself may confer protective effects under certain conditions. For instance, methionine and its derivative S-adenosylmethionine (SAM) have been shown to enhance mitochondrial DNA density in skeletal muscle [22, 23], improve insulin sensitivity, and prevent weight gain [24]. These observations indicate a potential dual role of methionine in diabetic metabolism: while its metabolite Hcy may promote inflammatory pathways and complications, methionine might also activate beneficial mechanisms that protect against metabolic dysfunction. Furthermore, future research should also consider the metabolic characteristics and lifestyle differences among different populations to better understand the complex relationship between methionine and diabetic foot.

Urine metabolomics analysis identified elevated N-isovalerylglycine and valine as significant and novel biomarkers associated with increased risk of diabetic foot (DF). Critically, these markers are detectable in routine urine samples, offering a potential non-invasive approach for early risk stratification in diabetic patients. N-isovalerylglycine, a conjugate derived from branched-chain amino acid metabolism, reflects underlying mitochondrial stress and compensatory detoxification processes [25, 26]. As an intermediate in branched-chain amino acid metabolism (BCAA), N-isovalerylglycine facilitates the removal of toxic isovaleryl-CoA from mitochondrial pathways, thereby mitigating metabolic toxicity and supporting mitochondrial energy function [25]. Its elevation signifies a state of metabolic overload that occurs prominently in advanced diabetes, particularly when complicated by DFT. Similarly, elevated valine aligns with well-established evidence linking branched-chain amino acids to insulin resistance and diabetes progression [27–29]. Within the context of DF—a condition marked by severe neuropathy, microvascular compromise, and chronic infection—N-isovalerylglycine functions to maintain energy metabolism and internal environmental balance. Neuropathy-induced sensory impairment and foot deformities exacerbate localized mechanical stress, while the energy metabolism overload during tissue repair accelerates the catabolism of branched-chain amino acids. Concurrently, tissue ischemia and hypoxia resulting from microangiopathy, combined with the systemic hypercatabolic state triggered by infection, promote protein breakdown (particularly in muscle tissue). This leads to increased release of BCAAs (especially valine and leucine) [30], consequently generating more N-isovalerylglycine to maintain energy metabolism and internal homeostasis. Thus, elevated urinary N-isovalerylglycine serves as an integrative biomarker reflecting enhanced BCAA catabolism, metabolic dysregulation, and excretory adaptation in DF, highlighting its potential utility in predicting DF onset. Additionally, urine-based metabolomic screening, leveraging its non-invasive nature and ease of repeated sampling, can be effectively integrated into existing longitudinal monitoring systems for diabetic patients [31]. This approach serves as a valuable complement to blood tests and is particularly suitable for the regular screening of high-risk populations and dynamic monitoring of disease progression. Compared to Liquid Chromatography-Tandem Mass Spectrometry (LC-MS/MS) technology, Nuclear Magnetic Resonance (NMR) spectroscopy, despite its relatively lower sensitivity, offers advantages such as minimal sample preparation, high reproducibility, and reliable quantitative capabilities. These features make NMR more feasible for high-throughput screening in clinical laboratory settings. By synergistically combining these two technologies, a tiered testing strategy can be established to effectively address diverse clinical needs [32].

Emerging evidence have proclaimed the protective property of lipoprotiens in diabetic individuals. In the current study, certain low-density lipoprotein (LDL) subfractions, such as LDTG, L2TG, L4TG, and L5TG, have all been found to be associated with the risk of Diabetic Foot (DF) occurrence. A systematic review and meta-analysis showed that elevated LDL levels were significantly associated with an increased risk of Diabetic Foot Ulcer (DFU) development. High LDL levels may promote atherosclerosis, reduce blood circulation in the feet, and thereby increase the risk of ulcer development [17]. Furthermore, a cross-sectional study including 255 patients with type 1 diabetes found that the LDL/HDL ratio was significantly elevated in Diabetic Foot patients, which may serve as an important indicator for assessing the severity of Diabetic Foot [33].

On the other hand, high-density lipoprotein (HDL) subfractions, such as H3CH, H4CH, H4FC, H4PL, H4A1, and H4A2, are all protective factors against the development of Diabetic Foot. In the Framingham Offspring Study, every 1 mg/dL increase in HDL-C level was associated with an approximately 4% lower risk of Type 2 Diabetes (T2D) over a 7-year follow-up period [34]. HDL also plays a significant role in the occurrence and development of Diabetic Foot Ulcer (DFU). HDL possesses anti-inflammatory and antioxidant functions, enabling it to neutralize oxidative stress products in the body and reduce inflammatory responses, thereby lowering the risk of Diabetic Foot occurrence [35, 36]. Additionally, certain subtypes of HDL may further reduce the risk of Diabetic Foot by regulating lipid metabolism and improving insulin sensitivity [37].

The results were derived from a single research center with a relatively small sample size, which may increase the risk of random variability and compromise the reliability of conclusions. Despite efforts to control for potential confounding factors, the predictive model lacks external validation in an independent cohort. While internal validation showed good discrimination, the absence of multi-center data raises concerns about overfitting and generalizability. Before clinical translation, external validation in large, diverse prospective cohorts is essential. the absence of multi-center data raises concerns about overfitting and generalizability. Subsequent studies should use more rigorous designs to validate these findings in larger, multi-center cohorts, improving robustness. Additionally, further exploration is needed to elucidate the mechanisms underlying these metabolic changes and their association with DF progression, as well as to evaluate the impact of lifestyle factors and population-specific metabolic characteristics on DF development.

Conclusion

This study comprehensively analyzed the metabolic profiles of diabetic foot (DF) patients and non-complicated DM individuals, revealing significant differences in metabolic characteristics between the two groups. Through systematic comparison and analysis of blood metabolomics, urine metabolomics, and lipidomics, the integration of urinary metabolomic data with clinical indicators demonstrated exceptional accuracy and stability in predicting DF, providing a foundation for early clinical diagnosis of DF.

From a clinical perspective, the integration of these metabolic biomarkers into existing risk-assessment models—such as those incorporating neuropathy, vascular status, and glycemic control—could significantly improve early detection of high-risk patients. This is especially relevant for personalized patient management, where urinary metabolite profiling may serve as an economical and repeatable monitoring tool.

Abbreviations

DF

Diabetic foot

DM

Diabetes mellitus

HT

Hypertension

HB

Hemoglobin

PLT

Platelet count

TBI

Total Bilirubin

ALB

Albumin

ALT

Alanine Aminotransferase

AST

Aspartate Aminotransferase

CREA

Creatinine

UA

Uric Acid

GLU

Glucose

TG

Triglycerides

HDL

High-Density Lipoprotein cholesterol

LDL

Low-Density Lipoprotein cholesterol

VLDL

Very Low-Density Lipoprotein cholesterol

APOA1

Apolipoprotein A1

APOB

Apolipoprotein B

Cal

Calcium

Mg

Magnesium

K

Potassium

Na

Sodium

Cl

Chloride

EPI

Estimated glomerular filtration rate

Author contributions

CWY and JPG designed the study. FLL and JPG drafted the manuscript. JXH and YXL performed the data analysis. XQZ, XLL, and WLL were involved in data collection and figure preparation. XHH, FLL, and FGC contributed to data interpretation. CWY was responsible for writing review and editing. All authors read and approved the final manuscript.

Funding

This work was supported by grants from the National Natural Science Foundation of China (82204023), Natural Science Foundation of Fujian Province (No. 2025J01734, No. 2025J01758) and Research Start-up Fund of Fujian Medical University (No.2023QH1065).

Data availability

The raw metabolomic data generated in this study have been deposited in the National Genomics Data Center (NGDC) under accession code OMIX012300.

Declarations

Ethics approval and consent to participate

All human subjects provided informed consent at recruitment. This study was conducted in accordance with the ethical principles of the Declaration of Helsinki (https://www.wma.net/policies-post/wma-declaration-of-helsinki/). Ethical approval was obtained from the Ethics Committee of the First Affiliated Hospital of Fujian Medical University (approval number: FMU [2023] 508).

Consent for publication

Not applicable.

Competing interests

The authors declare no competing interests.