Exploration of a Multimodal Machine Learning Model Integrating Ultrasound and Clinical Indicators for the Diagnosis of Diabetic Peripheral Neuropathy

Abstract

Objectives

Based on ultrasound technology and clinical indicators, this study intends to develop multiple risk prediction models for diabetic peripheral neuropathy (DPN), conduct comparative analyses of these models, and further evaluate and validate the diagnostic efficacy of the optimal model for DPN as well as its potential in clinical application.

Methods

The study included 235 patients grouped according to criteria for diagnosis of DPN. Ultrasound and clinical data were collected concurrently. The dataset was randomly partitioned into training and testing sets at a 7:3 ratio. Four machine learning models were developed: logistic regression (LR), extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM). These models were evaluated using a 10‐fold cross‐validation, comparing accuracy and areas under the curve (AUC) to determine the most optimal model. Based on SHAP (Shapley additive explanation) value visualization, decision curve analysis (DCA) and clinical impact curve (CIC) were employed to assess clinical utility.

Results

Comparisons of different models indicate that the RF model performed best overall across all models, reaching an AUC of 0.852 in the test set while also producing the highest recall rates and F1 scores. SHAP analysis revealed key risk factors identified in the RF model, in order of importance: body mass index (BMI), 2‐hour C‐peptide (2hC‐P), diabetes duration, triglycerides (TG), CSA‐CPN1, and age. In addition, DCA and CIC demonstrate the model's clinical utility.

Conclusions

The RF model demonstrated optimal performance for DPN prediction and shows significant potential for clinical application in DPN risk assessment.

Abbreviations

2hC‐P

2‐hour postprandial C‐peptide

AIC

Akaike information criterion

AUC

area under the curve

BMI

body mass index

CIC

clinical impact curve

CPN

common peroneal nerve

CSA

cross‐sectional area

DCA

decision curve analysis

DPN

diabetic peripheral neuropathy

HbA1c

glycated hemoglobin

HFUS

high‐frequency ultrasound

IQR

interquartile range

LR

logistic regression

MRI

magnetic resonance imaging

NEE

neuroelectrophysiological examination

RF

random forest

ROC

receiver operating characteristic

SHAP

Shapley additive explanation

SVM

support vector machine

SWV

shear wave velocity

T3

triiodothyronine

TG

triglycerides

TN

tibial nerve

XGBoost

extreme gradient boosting

Diabetic peripheral neuropathy (DPN) refers to symptoms or signs associated with peripheral nerve dysfunction in patients with diabetes mellitus after other causes have been excluded. DPN is one of the most common complications of diabetes mellitus, with the prevalence of diabetic DPN exceeding 50%. ¹ Its pathogenesis is complex and may involve multiple contributing factors. ² Of these, lower limb neuropathy was dominant in DPN patients. ³ Diagnosis of DPN mainly involves clinical evaluation based on characteristic symptoms and signs to rule out other causes of neuropathy. Neuroelectrophysiological examination (NEE) can be used for further confirmation if necessary. However, current diagnostic tools such as NEE, magnetic resonance imaging (MRI), and skin neurobiopsies still have significant limitations. NEE reflects sensory and motor conduction abnormalities in DPN, but mainly targets large nerve fibers and is insensitive to small early neurofibrosis. MRI provides direct visualization of peripheral nerve microstructure, but is expensive and time‐consuming, limiting its use as a routine screening tool. Skin neurobiopsies offer advantages in diagnosing small fiber neuropathy, often involved in early‐stage DPN, but their invasive nature limits routine application for DPN detection. Although the results of these 3 diagnostic methods vary, none of them is suitable for routine screening tools. ⁴ , ⁵ , ⁶ In recent years, numerous studies ⁷ have shown that ultrasound plays an important role in screening for DPN and assessing the severity of neuropathy. High‐frequency ultrasound (HFUS) can visualize structural details, including nerve morphology and dimensions, ⁸ , ⁹ facilitating the detection of DPN‐related morphological changes and assessment of neuropathy severity. SWE can assess nerve stiffness by reflecting nerve stiffness to identify conditions such as nerve rigidity. ¹⁰ Currently, commonly used ultrasound metrics for DPN assessment include width (W), thickness (T), cross‐sectional area (CSA), and shear wave velocity (SWV).

Risk prediction models are widely applied tools across various diseases. To achieve a more comprehensive evaluation, this study compared 4 distinct machine learning models: logistic regression (LR), extreme gradient boosting (XGBoost), random forest (RF), and support vector machine (SVM). This approach aims to cover a spectrum of algorithms, from interpretable linear models to complex nonlinear models, ensuring the identification of the optimal model for predicting DPN.

Currently, clinical data commonly incorporated into risk prediction models for DPN include: age, diabetes duration, ¹¹ , ¹² body mass index (BMI), ¹³ triglycerides (TG), ¹⁴ glycated hemoglobin (HbA1c), ¹⁵ 2‐hour postprandial C‐peptide (2hC‐p), ¹⁶ Triiodothyronine (T3), ¹⁷ etc. Age and diabetes duration are independent factors contributing to DPN onset. As age and diabetes duration increase, the extent of neural aging and damage progressively worsens, indicating the progression of DPN. ¹⁸ BMI, the ratio of weight to height squared, reflects an individual's degree of obesity. Obesity, characterized by fat accumulation, is closely associated with insulin resistance in diabetic patients, linking to the pathogenesis of diabetes. ¹⁹ Similar to BMI, TG also reflects fat accumulation. Studies indicate elevated TG levels correlate closely with neuropathy, potentially related to loss of myelinated fiber density in nerves. ²⁰ HbA1c is the most commonly used indicator in blood glucose testing and serves as a marker for monitoring glycemic control in diabetes. HbA1c, the primary biomarker for glycemic control, reflects the average blood glucose level over the preceding 2–3 months. Studies have established a positive correlation between elevated HbA1c levels and the incidence of DPN. ²¹ Consequently, integrating HbA1c into risk prediction models enables the assessment of long‐term glycemic control on DPN progression. C‐peptide, a key component in insulin synthesis, demonstrates anti‐inflammatory and proliferative properties by modulating Na+‐K+‐ATPase and endothelial nitric oxide synthase activity. ²² This physiological mechanism may elucidate the role of C‐peptide deficiency in the pathogenesis of diabetic microvascular complications, positioning C‐peptide as a potential biomarker for early DPN diagnosis. Thyroid hormone T3, a critical indicator of thyroid function, exhibits synergistic effects with insulin in regulating glucose and lipid metabolism. Moreover, elevated T3 levels can inhibit insulin secretion. ²³ Previous research further indicates that low T3 syndrome is associated with an increased risk and severity of DPN in patients with type 2 diabetes. ²⁴ A significant limitation of existing DPN risk prediction models is their reliance on HFUS data without integration with clinical indicators. ²⁵ Furthermore, few studies have investigated the inclusion of advanced ultrasound technologies, such as SWE. To address these gaps, this study aims to develop a comprehensive risk prediction model by integrating clinical indicators with multi‐modal ultrasound data. The tibial nerve (TN) and common peroneal nerve (CPN), as major lower extremity nerves, have demonstrated significant predictive value for DPN onset and progression. ²⁶ , ²⁷ Therefore, we will collect ultrasound metrics from various segments of the TN and CPN, alongside relevant clinical data, to construct and compare 4 distinct risk prediction models. The objective is to identify the model with the optimal predictive performance for DPN, with the ultimate aim of determining the independent risk factors through this optimal model. This approach will validate the efficacy of a combined ultrasound‐clinical model and further explore its value in improving early DPN diagnosis.

Materials and Methods

Subjects

A retrospective analysis was performed on the inpatient records of patients clinically diagnosed with type 2 diabetes mellitus (T2DM) who were consecutively admitted to our hospital between November 2021 and May 2024. Based on the predefined inclusion and exclusion criteria, a total of 235 patients were enrolled in this study, comprising 434 TNs and 435 CPNs. All patient diagnoses adhered to the T2DM and DPN criteria specified in the Chinese Diabetes Prevention and Treatment Guidelines (2024 Edition). ²⁸ Participants were stratified into 2 groups: a DPN group and a non‐DPN (NDPN) group, based on the confirmed presence or absence of DPN. The workflow of this study is shown in Figure 1. All patients provided written informed consent prior to participation.

Figure 1.

Flowchart of participant recruitment and study design.

Inclusion Criteria

1. General Criteria: ① Age ≥ 18 years; ② Mental competence and full capacity for civil conduct; ③ Provision of informed consent and voluntary participation in the study.

2. Fulfillment of the T2DM diagnostic criteria as defined by the Chinese Diabetes Prevention and Treatment Guidelines (2024 Edition). ²⁸

Exclusion Criteria

• ① Peripheral neuropathy attributable to other etiologies;

• ② Diagnosis with other types of diabetes;

• ③ Comorbidities that could interfere with the examination;

• ④ Acute diabetic complications;

• ⑤ Incomplete clinical data or inability to cooperate with the required examinations. ²⁸

Instruments and Methods

Tibial Nerve Examination

The transducer was initially positioned 3–5 cm superior to the popliteal fossa, corresponding to the origin of the TN (Figure 2A). It was then advanced along the nerve's anatomical course, tracing its path to its distal termination near the sole of the foot. The nerve's W, T, and CSA were measured and recorded at 3 specific locations: the TN origin (TN1), the popliteal fossa (TN2), and the posterior medial malleolar groove (TN3).

Figure 2.

Ultrasonic parameters of common tibial and peroneal nerves. A, Two‐dimensional ultrasound measurement of the tibial nerve; B, Elastography measurement of the tibial nerve; C, Two‐dimensional ultrasound measurement of the common peroneal nerve; D, Elastography measurement of the common peroneal nerve.

Common Peroneal Nerve Examination

The transducer was placed 3–5 cm above the popliteal fossa at the CPN origin (Figure 2C). Using the fibular head as a key anatomical landmark, the nerve was traced along the fibula toward the foot, following its course to the dorsum. The nerve's course, perineural envelope contour, intraneural echogenicity, blood flow signals, and spatial relationship with adjacent tissues were observed at the CPN origin (CPN1) and the fibular head (CPN2). Measurements of W, T, and CSA were taken at these 2 sites.

Shear Wave Elastography Protocol

Following the transverse and longitudinal B‐mode scans at all specified TN and CPN sites, the transducer was rotated to obtain longitudinal views of each nerve. The SWE mode was subsequently activated, and the SWV of the nerve was measured at each location. A total of 5 SWV measurements were taken and recorded per site (Figure 2B and D).

All ultrasound examinations were performed by experienced sonographers following a standardized protocol.

Data Collection of Clinical Variables

Based on clinical experience and a review of relevant literature, potential risk factors associated with DPN onset were collected. These included general demographic characteristics, such as age, gender, diabetes duration, and BMI, as well as laboratory parameters, including HbA1c, 2hC‐P, TG, and T3.

Statistical Methods

All statistical analyses were performed using R software (version 4.4.2). Continuous variables were tested for normality using the Kolmogorov–Smirnov test. Normally distributed data are presented as mean ± standard deviation and were compared between groups using the Student's t‐test. Non‐normally distributed data are presented as median and interquartile range (IQR) (M [P25, P75]), and the Mann–Whitney U test (rank‐sum test) was used for intergroup comparisons. Categorical data are expressed as counts and percentages (n [%]) and were analyzed using the chi‐square test. Prior to model development, continuous variables were examined for outliers using the IQR method.

Model Construction and Evaluation

The presence or absence of DPN was designated as the dependent variable. Independent variables comprised ultrasonographic parameters—such as the CSA and SWV from various tibial and CPN sites—and the collected clinical indicators.

The patient cohort was randomly split into a training set (70% of patients) for model development and a testing set (30%) for validation. Four distinct machine learning models were constructed on the training set: LR, RF, XGBoost, and SVM. Consistent outlier handling procedures were applied across all models.

Feature selection strategies were tailored to each model's strengths. For the SVM model, features were selected based on univariate analysis. The LR model employed a combination of univariate analysis and Least Absolute Shrinkage and Lasso regression. This pre‐screening was applied because SVM and LR both benefit from pre‐screening to reduce dimensionality and enhance stability. The RF and XGBoost models leveraged their built‐in, impurity‐based feature importance measures. This approach was feasible because tree‐based ensemble methods can efficiently evaluate a large number of features internally through impurity‐based measures. Finally, to ensure robustness and mitigate the instability of a single data split, all models underwent a 10‐fold cross‐validation procedure during the training phase.

Model performance was evaluated and compared on the testing set using several metrics: the area under the receiver operating characteristic (ROC) curve (AUC), accuracy, F1‐score, precision, recall, and specificity. The SHAP (Shapley additive explanation) framework was employed to interpret the models and analyze feature importance. Finally, the clinical utility of the models was assessed using decision curve analysis (DCA) and clinical impact curves (CIC).

Hyperparameter Tuning and Validation Strategy

All machine learning models were optimized using a nested 10‐fold cross‐validation strategy conducted exclusively on the training set.

LR was tuned via a 2‐stage procedure: first, fast backward stepwise elimination based on the Akaike information criterion (AIC) was used for feature selection; second, LASSO regularization was applied, with the optimal penalty parameter (λ) determined through 10‐fold cross‐validation.

RF hyperparameters were optimized using a grid search over mtry and min.node.size, while the number of trees was fixed at 10. The configuration yielding the highest mean AUC in 10‐fold cross‐validation was selected.

XGBoost underwent hyperparameter optimization via a random search with 10‐fold cross‐validation, exploring key parameters including tree depth, learning rate, regularization terms, and sampling rates. Early stopping was employed to mitigate overfitting.

SVM parameters were tuned through a grid search that optimized the kernel width (sigma) and regularization constant (C) using 10‐fold cross‐validation.

Upon completion of the nested tuning, final models were retrained on the full training set using the optimal hyperparameters and subsequently evaluated once on the held‐out test set.

Results

Comparison of Baseline Data between Training Set and Test Set

The study cohort was randomly divided into a training set (n = 305 observations) and a test set (n = 129 observations). Baseline characteristics were well‐balanced between the 2 sets, with no statistically significant differences observed for the vast majority of variables (Table 1). Minor differences in the distribution of 2 parameters—SWV‐TN1 (p = .0381) and SWV‐CPN1 (p = .0200)—likely attributable to random sampling variability, did not compromise the overall reliability of the model development. Furthermore, SHAP analysis indicated that these variables were not among the most important predictors in the final RF model (Figure 5). Thus, the observed slight intergroup differences in these elastographic parameters are expected to have minimal impact on the overall evaluation of model performance.

Table 1.

Comparison of Baseline Data Between Training Set and Test Set

Variable (Unit)	Training set (n = 305)	Test set (n = 129)	p
Sex (Male/Female)	238/67	101/28	≈1
Age (years)	54.00 [45.00, 58.00]	52.00 [42.00, 57.00]	.0955
Diabetes duration (years)	3.00 [0.00, 10.00]	3.00 [0.00, 9.00]	.7187
BMI (kg/m2)	23.67 ± 3.24	23.58 ± 3.32	.7813
HbA1C (%)	11.20 [9.08, 12.98]	10.80 [8.95, 12.40]	.1907
TG	1.43 [0.94, 2.15]	1.43 [1.00, 2.25]	.3366
2hC‐P (ng/mL)	1.55 [0.69, 2.66]	1.77 [0.84, 3.12]	.1792
T3	4.77 [4.35, 5.35]	4.88 [4.43, 5.30]	.6939
W‐TN1 (mm)	9.00 [8.40, 9.89]	9.10 [8.30, 9.70]	.5051
W‐TN2 (mm)	8.10 [7.50, 8.70]	8.20 [7.60, 9.00]	.4304
W‐TN3 (mm)	7.02 ± 0.89	7.18 ± 1.01	.1158
T‐TN1 (mm)	7.10 [6.70, 7.60]	7.10 [6.50, 7.60]	.6397
T‐TN2 (mm)	7.00 [6.50, 7.50]	7.00 [6.40, 7.60]	.9132
T‐TN3 (mm)	4.93 ± 0.67	4.91 ± 0.67	.7630
CSA‐TN1 (mm2)	44.00 [40.00, 48.00]	43.00 [39.00, 49.00]	.8721
CSA‐TN2 (mm2)	39.00 [36.00, 44.00]	40.00 [35.00, 45.00]	.8005
CSA‐TN3 (mm2)	25.00 [21.00, 27.00]	25.00 [22.00, 28.00]	.4633
SWV‐TN1 (m/s)	2.95 [2.47, 3.38]	3.09 [2.60, 3.66]	.0381
SWV‐TN2 (m/s)	3.42 [2.92, 3.97]	3.55 [3.03, 4.05]	.1408
SWV‐TN3 (m/s)	4.15 [3.67, 4.67]	4.11 [3.57, 4.63]	.3848
W‐CPN1 (mm)	6.10 [5.80, 6.50]	6.20 [5.80, 6.60]	.7570
W‐CPN2 (mm)	6.30 [5.80, 6.90]	6.40 [6.00, 6.90]	.1263
T‐CPN1 (mm)	5.10 [4.80, 5.40]	5.10 [4.80, 5.40]	.7305
T‐CPN2 (mm)	4.50 [4.20, 4.90]	4.40 [4.00, 4.80]	.1928
CSA‐CPN1 (mm2)	24.00 [22.00, 25.00]	24.00 [21.00, 26.00]	.7018
CSA‐CPN2 (mm2)	21.00 [19.00, 23.00]	21.00 [19.00, 23.00]	.5904
SWV‐CPN1 (m/s)	3.04 [2.58, 3.62]	3.29 [2.68, 3.98]	.0200
SWV‐CPN2 (m/s)	4.50 [4.03, 4.98]	4.59 [4.09, 5.12]	.2993

BMI, body mass index; HbA1c, glycated hemoglobin; 2hC‐P, 2‐hour postprandial C‐peptide; TG, triglycerides; T3, triiodothyronine; W, width; T, thickness; CSA, cross‐sectional area; SWV, shear wave velocity; TN, tibial nerve; CPN, common peroneal nerve. [Correction added on 11 April 2026, after first online publication: Chinese text in Table 1 has been translated to English in this version.]

Figure 5.

SHAP summary plot.

Comparative Analysis of DPN Diagnostic Performance across Different Machine Learning Models

The performance comparisons of the 4 models are presented in Tables 2 and 3, while 10‐fold cross‐validation results are shown in Table 4.

Table 2.

Effectiveness Comparison of Different Models in Training Set

Model	AUC	Accuracy	Recall	Specificity	Precision	F1‐score	Brier Score
LR	0.798	0.759	0.951	0.462	0.732	0.827	0.175
RF	0.927	0.846	0.918	0.745	0.837	0.875	0.119
XGBoost	0.994	0.954	0.978	0.920	0.946	0.962	0.088
SVM	0.905	0.830	0.683	0.933	0.878	0.768	0.137

LR, logistic regression; RF, random forest; XGBoost, extreme gradient boosting; SVM, support vector machine; AUC, area under the curve.

Table 3.

Performance Comparison of Different Models in Test Set

Model	AUC	Accuracy	Recall	Specificity	Precision	F1‐score	Brier Score
LR	0.744	0.702	0.916	0.450	0.663	0.769	0.207
RF	0.852	0.767	0.855	0.642	0.774	0.813	0.155
XGBoost	0.843	0.762	0.816	0.685	0.785	0.800	0.168
SVM	0.789	0.744	0.585	0.855	0.738	0.653	0.182

LR, logistic regression; RF, random forest; XGBoost, extreme gradient boosting; SVM, support vector machine; AUC, area under the curve.

Table 4.

Performance Comparison of Different Models Using 10‐Fold Cross‐Validation

Model	AUC	Accuracy	Recall	Specificity	Precision	F1‐score	Brier Score
LR	0.772 ± 0.098	0.781 ± 0.063	0.859 ± 0.133	0.664 ± 0.225	0.817 ± 0.104	0.824 ± 0.060	0.190 ± 0.035
RF	0.810 ± 0.071	0.754 ± 0.060	0.756 ± 0.073	0.757 ± 0.062	0.596 ± 0.133	0.661 ± 0.103	0.181 ± 0.019
XGBoost	0.850 ± 0.093	0.763 ± 0.095	0.833 ± 0.114	0.645 ± 0.144	0.784 ± 0.090	0.804 ± 0.085	0.156 ± 0.048
SVM	0.804 ± 0.072	0.764 ± 0.007	0.617 ± 0.098	0.867 ± 0.084	0.774 ± 0.136	0.682 ± 0.097	0.178 ± 0.032

LR, logistic regression; RF, random forest; XGBoost, extreme gradient boosting; SVM, support vector machine; AUC, area under the curve.

The predictive performance of the 4 machine learning models on the training set is summarized in Table 2. The RF, XGBoost, and SVM models all demonstrated high discriminative ability, with AUC values exceeding 0.9. In contrast, the LR model exhibited a comparatively lower AUC of 0.798. Other metrics, including accuracy, recall, specificity, precision, F1 score, and Brier score, showed considerable variation across the models.

On the independent test set (Table 3), the LR model demonstrated moderate performance (AUC = 0.744) and exhibited the highest prediction error (Brier score = 0.207). Although the SVM model achieved the highest specificity (0.855), its low recall (0.585) suggests a substantial risk of missing positive cases, thereby limiting its utility for early DPN screening. The RF model achieved a marginally higher AUC than the XGBoost model (0.852 versus 0.843). Furthermore, RF outperformed XGBoost in both recall (0.855 versus 0.816) and the F1 score (0.813 versus 0.800), indicating a superior capacity to identify true positive cases and a better overall balance between precision and recall. Notably, the RF model also achieved the lowest Brier score (0.155) among all models, signifying that its predicted probabilities were the most calibrated and closest to the true observed outcomes.

The 10‐fold cross‐validation results (Table 4) revealed that the RF model attained a slightly lower mean AUC than XGBoost (0.810 ± 0.071 versus 0.850 ± 0.093). However, the RF model's markedly smaller standard deviation indicates superior generalization capability and stability—a critical attribute for reliable clinical prediction models. Additionally, the RF model offers greater clinical interpretability through feature importance analysis than the more opaque SVM and XGBoost models, aligning more closely with practical clinical requirements. Therefore, based on its robust test‐set performance, superior stability, and clinical interpretability, the RF model was selected as the optimal model for subsequent analysis in this study.

Performance Evaluation of the Random Forest Model

The tuned RF model demonstrated strong predictive performance, outperforming traditional models. The ROC curve for the test set is presented in Figure 3, with an AUC of 0.852, indicating excellent discriminative ability. Furthermore, the probability calibration curve closely aligned with the ideal diagonal line (Figure 4), confirming the high reliability of the model's predicted probabilities. These results collectively affirm the robustness of the final RF model.

Figure 3.

ROC curve of RF model.

Figure 4.

RF model probability calibration curve.

Analysis of Feature Importance

The global importance and directional influence of the top predictive features are visualized in the SHAP summary plot (Figure 5). In this analysis, the SHAP value‐based feature importance is considered more robust than the model's inherent Gini importance. The results identified BMI as the most influential predictor, followed by 2hC‐P, diabetes duration, TG, the CSA of CPN‐CSA1, and age.

To elucidate the nature of these relationships, SHAP dependency plots were generated (Figure 6). These plots illustrate the monotonic or non‐linear relationship between individual feature values and their impact on the SHAP value, directly showing how a change in a feature influences the predicted risk of DPN.

Figure 6.

Shap dependency graph.

For local interpretability, individual force plots were analyzed (Figure 7). These plots deconstruct the model's prediction for a single patient into the contributions of each feature. Features pushing the prediction above the base value (positive SHAP values) represent risk factors for DPN, while those pulling it below (negative SHAP values) act as protective factors. A comparative analysis of these plots across multiple patients in the DPN and non‐DPN groups allows for the intuitive identification of consistent risk and protective patterns.

Figure 7.

SHAP summary plot for sample‐level explanations.

Assessment of Clinical Utility

The clinical utility of the RF model was evaluated on the test set using DCA and a CIC, as shown in Figure 8.

Figure 8.

Decision curve analysis (DCA) and clinical impact curve (CIC).

The DCA evaluates the net benefit of using the model across a range of probability thresholds for clinical intervention. The model demonstrates clinical value when its curve lies above the reference lines for the “treat‐all” and “treat‐none” strategies. The results indicate that for risk thresholds between 10 and 100%, the model's curve is superior, meaning the model provides a superior net benefit compared to the default strategies across this range.

The CIC further illustrates the model's clinical impact by plotting the number of individuals classified as high‐risk against the actual number of true positives at each threshold. The gap between these 2 curves represents the number of false positives. Our analysis shows that the false positive rate decreases substantially when the decision threshold exceeds 0.4, indicating a more specific identification of true DPN cases at higher risk levels. Together, these analyses confirm the model's strong potential for practical clinical application.

Discussion

DPN represents one of the most prevalent and serious complications of diabetes mellitus, characterized by complex and multifactorial pathophysiology. Ultrasound offers distinct advantages over other diagnostic modalities, being non‐invasive, cost‐effective, readily operable, and providing excellent reproducibility, thus presenting an ideal tool for routine screening applications. The development of risk prediction models has emerged as a major focus in DPN research due to their demonstrated efficacy across various disease contexts. In this study, we systematically developed and compared 4 machine learning models for DPN risk prediction by integrating multiparametric ultrasonographic data from both tibial and CPNs with key clinical indicators. Our results establish the feasibility of this integrated approach and identify the RF model as optimal, thereby providing novel insights for early DPN identification and risk assessment.

Model Performance and Selection

The RF model demonstrated superior predictive performance in our comprehensive evaluation. On the test set, it achieved an AUC of 0.852, significantly outperforming both LR (0.744) and SVM (0.789) models, while showing marginal superiority over XGBoost (0.843). The RF model further excelled in recall (0.855) and F1‐score (0.813), indicating enhanced capability to identify true DPN cases while maintaining optimal precision‐recall balance—particularly crucial for reducing missed diagnoses in early screening. Its superior calibration was evidenced by the lowest Brier score (0.155) among all models, and its robustness was confirmed through 10‐fold cross‐validation showing minimal performance variation. The RF algorithm's inherent advantages in capturing complex non‐linear relationships and feature interactions, combined with its resistance to overfitting, make it particularly suitable for moderate‐sized datasets, explaining its optimal performance in this study.

Feature Importance and Clinical Interpretation

SHAP analysis identified 6 predominant features: BMI, 2hC‐P, diabetes duration, TG, CSA of CSA‐CPN1, and age.

Notably, BMI emerged as the most influential predictor, with SHAP dependency plots revealing an unexpected inverse relationship where lower BMI values correlated with increased DPN risk. While previous studies have reported conflicting associations between BMI and DPN—including positive, null, and U‐shaped relationships ²⁹ , ³⁰ , ³¹ —our findings align with Wang et al ²¹ in identifying low BMI as a significant risk factor. This association may therefore reflect underlying metabolic disturbances, nutritional deficiencies, or sarcopenia‐related mechanisms contributing to neural vulnerability.

Lower 2hC‐P levels constituted another major risk factor, supporting the proposed neuroprotective role of C‐peptide through its neurotrophic, anti‐inflammatory, and microcirculatory effects. ³² This underscores the importance of evaluating pancreatic β‐cell function in DPN risk stratification. Diabetes duration confirmed its established role as a fundamental risk determinant, ³³ while elevated TG levels demonstrated a significant association with DPN risk, consistent with known pathways involving inflammatory cascades, oxidative stress, and microvascular compromise. ³⁴ Interestingly, SHAP analysis indicated a threshold effect beyond which further TG elevation did not substantially increase risk.

Among ultrasonographic parameters, CSA‐CPN1 emerged as the most significant predictor, with nerve enlargement likely reflecting underlying pathological processes including edema, inflammation, and structural remodeling. ³⁵ This finding highlights the particular sensitivity of specific nerve site evaluation in detecting DPN‐related morphological alterations. Advanced age completed the top predictors, reaffirming its well‐established role in DPN pathogenesis, ³⁶ likely through accumulated neural vulnerability and degenerative processes.

Innovation and Clinical Implications

The principal innovation of this study lies in successfully integrating quantitative ultrasonographic measurements with clinical parameters within a machine learning framework. This integrated approach yielded substantially improved predictive performance compared to single‐modality models. For instance, this study's RF model (AUC = 0.852) demonstrated marked superiority over the clinical data‐only LR model developed by Liu et al (AUC = 0.704), ³² establishing the complementary value of ultrasound in DPN risk assessment. The prominent position of CSA‐CPN1 among core predictors strongly supports the incorporation of site‐specific nerve ultrasound in DPN evaluation protocols.

Clinical application analysis through DCA confirmed significant net benefit across the 10–100% threshold probability range. The CIC further demonstrated that threshold selection above 0.4 substantially reduces false positive rates while maintaining sensitivity, thus optimizing the balance between intervention benefits and potential harms.

Limitations and Future Directions

Several limitations warrant consideration. First, as a single‐center retrospective design, this study is inherently susceptible to certain biases, despite consecutive enrollment and stringent criteria. Therefore, external validation through future multi‐center, prospective studies is paramount to establish the generalizability of our findings. Second, to capture potential asymmetrical manifestations of DPN, our analysis was conducted at the nerve level. The inherent correlation within subjects remains a methodological consideration. Utilizing aggregated patient‐level data in future studies could further enhance the generalizability of the findings. Third, the ultrasonographic assessment, while comprehensive for lower limb nerves, did not include upper extremity evaluation, which might provide additional diagnostic information. Future investigations could benefit from incorporating advanced ultrasonographic techniques such as microvascular flow imaging. Fourth, the reproducibility of ultrasound measurements was not formally assessed in this study. Although measurements were conducted by trained operators using a standardized protocol, future prospective studies would benefit from incorporating such reliability analyses to further strengthen the methodological rigor. Additionally, while we included established clinical predictors, expansion to include other potentially relevant clinical and serological markers might further enhance model performance. Future studies should also explore the integration of additional modalities and more extensive clinical profiling to refine predictive accuracy and clinical applicability. Finally, despite employing methods such as nested cross‐validation to mitigate overfitting, a performance gap persists between the training and test sets, particularly for more complex models like XGBoost. This is an expected characteristic when applying flexible algorithms to moderate‐sized, heterogeneous clinical datasets. Future prospective studies with larger, multi‐center cohorts will be valuable for external validation and to further enhance the robustness of these models.

Conclusion

The results of this study demonstrate that the RF model achieved superior predictive performance for DPN compared to LR, XGBoost, and SVM models constructed from ultrasound and clinical data. SHAP‐based analysis identified BMI, 2hC‐P, diabetes duration, TG, CPN‐CSA1, and age as key risk factors, with BMI being the most significant contributor. DCA and CIC confirmed the model's substantial net clinical benefit, supporting its potential for clinical implementation.

Data Availability Statement

The data that support the findings of this study are available on request from the corresponding author. The data are not publicly available due to privacy or ethical restrictions.