RESUMEN
Objective: One of the most frequent consequences of diabetes mellitus has been identified as diabetic peripheral neuropathy (DPN), and numerous inflammatory disorders, including diabetes, have been documented to be reflected by the neutrophil-to-lymphocyte ratio (NLR). This study aimed to explore the correlation between peripheral blood NLR and DPN, and to evaluate whether NLR could be utilized as a novel marker for early diagnosis of DPN among those with type 2 Diabetes Mellitus (T2DM). Methods: We reviewed the medical records of 1154 diabetic patients treated at Tongji Hospital Affiliated to Tongji University from January 2022 to March 2023. These patients did not have evidence of acute infections, chronic inflammatory status within the past three months. The information included the clinical, laboratory, and demographic characteristics of the patient. Finally, a total of 442 T2DM individuals with reliable, complete, and accessible medical records were recruited, including 216 T2DM patients without complications (DM group) and 226 T2DM patients with complications of DPN (DPN group). One-way ANOVA and multivariate logistic regression were applied to analyze data from the two groups, including peripheral blood NLR values and other biomedical indices. The cohort was divided in a 7 : 3 ratio into training and internal validation datasets following feature selection and data balancing. Based on machine learning, training was conducted using extreme gradient boosting (XGBoost) and support vector machine (SVM) methods. K-fold cross-validation was applied for model assessment, and accuracy, precision, recall, F1-score, and the area under the receiver operating characteristic curve (AUC) were used to validate the models' discrimination and clinical applicability. Using Shapley Additive Explanations (SHAP), the top-performing model was interpreted. Results: The values of 24-hour urine volume (24H UV), lower limb arterial plaque thickness (LLAB thickness), carotid plaque thickness (CP thickness), D-dimer and onset time were significantly higher in the DPN group compared to the DM group, whereas the values of urine creatinine (UCr), total cholesterol (TC), low-density lipoprotein (LDL), alpha-fetoprotein (AFP), fasting c-peptide (FCP), and nerve conduction velocity and wave magnitude of motor and sensory nerve shown in electromyogram (EMG) were considerably lower than those in the DM group (P < 0.05, respectively). NLR values were significantly higher in the DPN group compared to the DM group (2.60 ± 4.82 versus 1.85 ± 0.98, P < 0.05). Multivariate logistic regression analysis revealed that NLR (P = 0.008, C = 0.003) was a risk factor for DPN. The multivariate logistic regression model scores were 0.6241 for accuracy, 0.6111 for precision, 0.6667 for recall, 0.6377 for F1, and 0.6379 for AUC. Machine learning methods, XGBoost and SVM, built prediction models, showing that NLR can predict the onset of DPN. XGBoost achieved an accuracy of 0.6541, a precision of 0.6316, a recall of 0.7273, a F1 value of 0.6761, and an AUC value of 0.690. SVM scored an accuracy of 0.5789, a precision of 0.5610, a recall of 0.6970, an F1 value of 0.6216, and an AUC value of 0.6170. Conclusions: Our findings demonstrated that NLR is highly correlated with DPN and is an independent risk factor for DPN. NLR might be a novel indicator for the early diagnosis of DPN. XGBoost and SVM models have great predictive performance and could be reliable tools for the early prediction of DPN in T2DM patients. This trial is registered with ChiCTR2400087019.
RESUMEN
BACKGROUND: Diabetic peripheral neuropathy (DPN) and lower extremity arterial disease (LEAD) are significant contributors to diabetic foot ulcers (DFUs), which severely affect patients' quality of life. This study aimed to develop machine learning (ML) predictive models for DPN and LEAD and to identify both shared and distinct risk factors. METHODS: This retrospective study included 479 diabetic inpatients, of whom 215 were diagnosed with DPN and 69 with LEAD. Clinical data and laboratory results were collected for each patient. Feature selection was performed using three methods: mutual information (MI), random forest recursive feature elimination (RF-RFE), and the Boruta algorithm to identify the most important features. Predictive models were developed using logistic regression (LR), random forest (RF), and eXtreme Gradient Boosting (XGBoost), with particle swarm optimization (PSO) used to optimize their hyperparameters. The SHapley Additive exPlanation (SHAP) method was applied to determine the importance of risk factors in the top-performing models. RESULTS: For diagnosing DPN, the XGBoost model was most effective, achieving a recall of 83.7%, specificity of 86.8%, accuracy of 85.4%, and an F1 score of 83.7%. On the other hand, the RF model excelled in diagnosing LEAD, with a recall of 85.7%, specificity of 92.9%, accuracy of 91.9%, and an F1 score of 82.8%. SHAP analysis revealed top five critical risk factors shared by DPN and LEAD, including increased urinary albumin-to-creatinine ratio (UACR), glycosylated hemoglobin (HbA1c), serum creatinine (Scr), older age, and carotid stenosis. Additionally, distinct risk factors were pinpointed: decreased serum albumin and lower lymphocyte count were linked to DPN, while elevated neutrophil-to-lymphocyte ratio (NLR) and higher D-dimer levels were associated with LEAD. CONCLUSIONS: This study demonstrated the effectiveness of ML models in predicting DPN and LEAD in diabetic patients and identified significant risk factors. Focusing on shared risk factors may greatly reduce the prevalence of both conditions, thereby mitigating the risk of developing DFUs.