Baishideng Publishing Group Inc, 7041 Koll Center Parkway, Suite 160, Pleasanton, CA 94566, USA
Share the Article
Yang ZJ, Zhao JY, Hu SC, Chen Y, Kang YF, Li J, Wang LC, Chen HM, Zhang ZY, Yang CZ, Zhu D. Explainable machine learning model for predicting minor amputation in diabetic foot ulcers. World J Diabetes 2026; 17(11): 124921 [DOI: 10.4239/wjd.124921]
Zhong-Jing Yang, Shao-Cong Hu, Yi Chen, Yun-Fei Kang, Juan Li, Liang-Chen Wang, Hong-Mei Chen, Zu-Yuan Zhang, Cai-Zhe Yang, Di Zhu, Department of Endocrinology, Air Force Medical Center, Beijing 100142, China
Jun-Yan Zhao, Department of Endocrinology, Jilin Hospital, Beijing Anzhen Hospital, Capital Medical University, Changchun 130123, Jilin Province, China
Co-first authors: Zhong-Jing Yang and Jun-Yan Zhao.
Co-corresponding authors: Cai-Zhe Yang and Di Zhu.
Author contributions: Yang ZJ and Zhao JY had primary responsibility for writing the original draft and developing the methodology as co-first authors; Hu SC, Li J, Chen HM, and Zhang ZY conducted the formal analysis; Chen Y and Kang YF were involved in reviewing the writing; Wang LC was responsible for supervision and overall oversight of the study; Yang CZ and Zhu D defined the research theme and revised the manuscript critically as co-corresponding authors; all of the authors read and approved the final version of the manuscript to be published.
AI contribution statement: The authors take full responsibility and accountability for all content of this manuscript. No AI tools were used in any stage of manuscript drafting, data analysis, result interpretation, or conclusion formulation. All text, analyses and scientific conclusions were independently completed, reviewed, validated and approved by all authors.
Supported by National Natural Science Foundation of China, No. 82470389; Capital’s Funds for Health Improvement and Research, No. 2024-4-5122; and Young Talent Program of Air Force Medical Center, No. 22BJQN005.
Institutional review board statement: Data collection for this study was approved by the Medical Ethics Committee of Air Force Medical Center (No. 2024-34-PJ01).
Informed consent statement: This retrospective study was approved by the Ethics Committee, and the requirement for written informed consent was waived due to the retrospective nature.
Conflict-of-interest statement: All authors declare no conflict of interest in publishing the manuscript.
STROBE statement: The authors have read the STROBE Statement – checklist of items, and the manuscript was prepared and revised according to the STROBE Statement – checklist of items.
Data sharing statement: The raw data supporting the conclusions of this article will be made available by the authors, without undue reservation.
Corresponding author: Di Zhu, MD, Chief Physician, Department of Endocrinology, Air Force Medical Center, No. 28 Fucheng Road, Beijing 100142, China. judy3481@163.com
Received: June 29, 2026 Revised: August 3, 2026 Accepted: October 20, 2026 Published online: November 15, 2026 Processing time: 132 Days and 20.1 Hours
Abstract
BACKGROUND
Minor amputations severely impact the quality of life in patients with diabetic foot ulcers (DFUs), yet early and transparent risk stratification tools remain limited.
AIM
To identify clinical risk factors and develop an explainable machine learning model for predicting minor amputation within 3 months after admission in patients with DFUs, and to apply SHapley Additive exPlanations (SHAP) analysis to interpret the model’s predictions.
METHODS
A total of 685 patients with DFUs were enrolled and randomly partitioned into training and validation sets based on minor amputation status within 3 months after admission. Predictive factors were identified using least absolute shrinkage and selection operator regression. Six machine learning algorithms were trained and compared, with the optimal model selected based on the area under the receiver operating characteristic curve. SHAP analysis was applied to evaluate model interpretability and feature importance.
RESULTS
Least absolute shrinkage and selection operator regression identified four independent predictors of minor amputation: (1) Fibrinogen (FIB); (2) Pedal vessel calcification (PVC); (3) Albumin (ALB); and (4) The C-reactive protein-ALB-lymphocyte (CALLY) index (all P < 0.05). Subsequent model comparison revealed the random forest model as the best predictive model, achieving area under the receiver operating characteristic curves of 0.844 (95%CI: 0.810-0.878) in the training set and 0.766 (95%CI: 0.700-0.828) in the validation set. SHAP analysis confirmed that FIB made the greatest contribution to the model’s prediction, followed by the CALLY index, ALB, and PVC, with elevated FIB and a reduced CALLY index strongly driving higher predicted risk.
CONCLUSION
A random forest model including FIB, PVC, serum ALB, and the CALLY index demonstrates significant potential for the early prediction of minor amputation in patients with DFUs, providing clinicians with a transparent and interpretable tool to optimize risk stratification.
Core Tip: This study developed and validated an explainable machine learning model to predict minor amputation within 3 months of admission in patients with diabetic foot ulcers. Among six compared algorithms, the random forest model performed best (validation area under the receiver operating characteristic curve = 0.766), identifying fibrinogen, pedal vessel calcification, albumin, and the C-reactive protein-albumin-lymphocyte index as independent predictors. SHapley Additive exPlanations analysis provided visual, patient-level interpretation of feature contributions. The model relies exclusively on routinely available laboratory and radiographic variables, offering a practical and objective tool for early risk stratification that may help clinicians identify high-risk patients with diabetic foot ulcers and initiate timely targeted interventions.
Citation: Yang ZJ, Zhao JY, Hu SC, Chen Y, Kang YF, Li J, Wang LC, Chen HM, Zhang ZY, Yang CZ, Zhu D. Explainable machine learning model for predicting minor amputation in diabetic foot ulcers. World J Diabetes 2026; 17(11): 124921
Diabetic foot ulcers (DFUs) are a prevalent and severe chronic complication of diabetes mellitus (DM) and are associated with a significantly increased risk of secondary infection and subsequent amputation[1,2]. It is estimated that 537 million individuals worldwide have diabetes, among whom 19%-34% are at risk of developing DFUs[3]. Studies indicate that approximately 25% of patients with DFUs eventually require a lower extremity amputation[4]. In clinical practice, these procedures are categorized into minor amputation (at or below the ankle) and major amputation (above the ankle)[5]. A minor amputation of diabetic foot specifically refers to surgery performed below the ankle joint (such as at the toes and plantar) to resect irreversible necrotic, infected, or ischemic tissues in the foot[6]. By eliminating the infection source and necrotic tissue, this procedure aims to “maximize limb preservation with limited local sacrifice”[7], thus preventing major amputation and preserving the patient’s ambulatory function and independence[8].
However, even after minor amputation, patients often experience considerable loss of mobility, which not only leads to physical dysfunction but can also cause severe psychological distress[9]. These overlapping consequences can have a serious impact on the patient’s overall sense of security, social participation, and quality of life. Currently, extensive research has focused on identifying risk factors and constructing predictive models for major amputations of the diabetic foot. Although several recent studies have examined risk factors for minor amputations, machine learning-based prediction models specifically designed for this outcome remain scarce.
To address this gap, this study aimed to construct a clinically applicable prediction model and validate its performance using machine learning algorithms[10]. These algorithms are well-suited for verifying the predictive value of candidate indicators by capturing complex data patterns, which enhances the reliability of the model. Notably, machine learning models often suffer from an inherent “black box” issue that masks their internal decision-making processes and limits their clinical interpretability[11,12]. To resolve this, we employed the SHapley Additive exPlanations (SHAP) algorithm developed by Lundberg et al[13]. This tool quantifies the contribution of each predictive indicator to the model’s output and clarifies the direction of their influence (via positive or negative SHAP values), effectively unlocking the “black box” of machine learning. Having demonstrated successful application in multiple medical fields, SHAP analysis ensures the model’s interpretability while maintaining predictive accuracy[14,15]. Thus, the integration of machine learning for indicator validation and SHAP for performance explanation provides a robust basis for early identification of minor amputation risk factors in DFUs, thereby facilitating individualized clinical management[16,17].
MATERIALS AND METHODS
Patient selection
This retrospective cohort study screened 819 hospitalized patients with type 2 DM and DFUs from the Department of Endocrinology at Air Force Medical Center between January 2019 and January 2024. The exclusion criteria were as follows: (1) Age below 18 years; (2) Type 1 DM; (3) Ulcers located above the ankle joint; (4) Complete absence of digits one through five on the affected foot; (5) Respiratory or urinary tract infections at the time of admission; (6) History of transplant surgeries; (7) Malignant tumors, autoimmune diseases, hematologic disorders, or acute cardiovascular or cerebrovascular events; (8) History of amputation above the ankle joint; and (9) Medical records with missing data exceeding 20%. Finally, 685 patients were enrolled for the development of the machine learning model. This retrospective cohort study was conducted in compliance with the Declaration of Helsinki and was approved by Medical Ethics Committee of Air Force Medical Center (approval No. 2024-34-PJ01).
Clinical data collection
Demographic and clinical data were extracted from electronic medical records, including sex, age, body mass index, duration of DM, duration of DFUs, history of smoking and alcohol use, ulcer site, wound size, Wagner grade, minor amputation status, and comorbidities, such as hypertension, coronary artery disease, cerebral infarction, diabetic peripheral vascular disease, peripheral arterial disease (PAD), diabetic kidney disease (DKD), diabetic retinopathy, and cerebrovascular disease. Diabetic peripheral vascular disease was defined as a clinical diagnosis based on suggestive symptoms (intermittent claudication, rest pain) and abnormal physical examination findings (diminished or absent pedal pulses). PAD was defined as atherosclerotic stenosis or occlusion of the arteries of the lower extremities, confirmed by duplex ultrasound.
All 685 enrolled patients completed a standardized 3-month follow-up after admission. Follow-up data were collected from inpatient medical records, outpatient follow-up records, and telephone interviews. Minor amputation was defined as surgical resection of tissues below the ankle joint performed within 3 months after admission due to uncontrollable infection, tissue necrosis, or ischemia refractory to conservative treatment. This outcome was adjudicated by two independent researchers through a cross-review of surgical notes, discharge summaries, and follow-up records. All amputation records, including operation time and operative indications, were double-verified from operation notes and follow-up files.
All laboratory indicators were measured at baseline (on the second day after admission) before the initiation of definitive treatment for DFUs. Laboratory tests included the following: (1) White blood cell count; (2) Red blood cell count; (3) Hemoglobin (Hb); (4) Platelet count (PLT); (5) Neutrophil count (NEU); (6) Lymphocyte count (LYM); (7) Monocyte count (MONO); (8) Neutrophil percentage; (9) Lymphocyte percentage; (10) C-reactive protein (CRP); (11) Fasting plasma glucose; (12) HbA1c; (13) Serum creatinine; (14) Blood urea nitrogen; (15) Total cholesterol; (16) Triglycerides; (17) Low-density lipoprotein cholesterol; (18) High-density lipoprotein cholesterol (HDL-C); (19) Albumin (ALB); (20) Blood uric acid (BUA); (21) Serum calcium; and (22) Fibrinogen (FIB). The ankle-brachial index was assessed at baseline for all included patients. An ankle-brachial index between 0.9 and 1.3 was considered normal; values outside this range were considered abnormal. The presence or absence of pedal vessel calcification (PVC) was evaluated on baseline foot radiographs by two experienced clinicians working independently[18]. PVC was defined as radiographically visible, radiopaque linear densities along the course of the foot arteries, indicating pathologic mineral deposition within the vessel walls. The clinicians noted the presence of any pedal vessel that showed calcification. Calcification in one or more pedal vessels was recorded as “yes” for PVC; otherwise, it was recorded as “no”. If the two clinicians disagreed regarding the PVC status, a senior vascular radiologist reviewed the X-rays with them to reach a final consensus.
Local wound characteristics were systematically documented. Ulcer location was classified as forefoot, midfoot, hindfoot/lower leg, or multiple sites. Wound area was calculated as length × width for superficial ulcers and as depth × length for sinus-tract ulcers, then categorized as small (< 5 cm2), medium (5-15 cm2), or large (> 15 cm2). Wagner grade was dichotomized into ≤ 3 (no gangrene) and > 3 (gangrene, grades 4-5).
Inflammatory markers were derived from the laboratory test results, including the neutrophil-lymphocyte ratio (NLR), neutrophil-platelet ratio (NPR), platelet-lymphocyte ratio (PLR), and additional indices such as the monocyte-HDL-C ratio (MHR), neutrophil-HDL-C ratio (NHR), systemic immune-inflammation index (SII), systemic inflammation response index (SIRI), and the CRP-ALB-lymphocyte (CALLY) index. The formulas for these indices are presented as follows: (1) NLR = NEU/LYM; (2) NPR = NEU/PLT; (3) PLR = PLT/LYM; (4) MHR = MONO/HDL-C level; (5) NHR = NEU/HDL-C level; (6) SII = PLT × NEU/LYM; (7) SIRI = NEU × MONO/LYM; and (8) CALLY index = ALB × LYM/(CRP × 10).
Statistical analysis
Data analysis and machine learning model development were carried out using SPSS version 29.0 and R version 4.4.0 statistical software. All subsequent preprocessing and feature selection steps were performed exclusively within the training set. Missing data were handled using multiple imputation by chained equations (MICE, m = 5) via the mice package in R. Imputation was performed within the training set only, and the imputation model was then applied to the validation set to prevent data leakage.
Normally distributed continuous variables were expressed as the mean ± SD, and differences were compared using the independent t-test. Non-normally distributed continuous variables were expressed as the median (interquartile range), and the Mann-Whitney U test was employed for comparisons. Categorical data were presented as n (%) and compared using the χ2 test. A two-tailed P < 0.05 was considered statistically significant. Variables exhibiting significant intergroup differences in the training set were further screened using least absolute shrinkage and selection operator (LASSO) regression, with the optimal penalty parameter λ selected via 5-fold cross-validation (1-SE criterion) within the training set. Logistic regression (LR) analysis was subsequently employed to identify the key variables for inclusion in the model. Results were presented as odds ratio with 95%CI.
Six machine learning models were developed and compared using R 4.4.0 with the tidymodels ecosystem to predict the risk of minor amputation. To ensure full reproducibility, the random seed was set to 123 at the beginning of the R environment. This seed controlled three procedures: (1) Data splitting; (2) Five-fold cross-validation fold assignment; and (3) Hyperparameter grid search. All preprocessing steps were performed exclusively within the training set. A recipe framework was used for preprocessing.
Categorical variables were transformed into factor format and then dummy-coded via step_dummy(). Continuous variables were Z-score standardized via step_normalize(). The preprocessing parameters were fitted using prep() on the training set. These fitted parameters were then applied to both the training and validation sets using bake(). This approach ensured complete data isolation between the two sets.
Several core R packages and their functions were applied. The tidymodels ecosystem was used for model building and tuning. Key functions included recipe(), prep(), bake(), rand_forest(), workflow(), vfold_cv(), and tune_grid(). Model performance was evaluated using yardstick and pROC. Decision curve analysis (DCA) was performed with the dcurves package. Data manipulation and visualization were conducted using dplyr and ggplot2. Data reading and exploration were handled by readr and skimr, respectively. The dataset was randomly partitioned into a training set (70%) and a validation set (30%) using stratified sampling based on the distribution of minor amputation events. The class distribution was relatively balanced (minor amputation: 46.0%; non-amputation: 54.0%); therefore, no class imbalance correction techniques (e.g., class weighting, oversampling, or undersampling) were applied. The predictive model was constructed using the training set, with model hyperparameters optimized through 10-fold cross-validation combined with grid search. The area under the receiver operating characteristic curve (AUC) was utilized as the primary evaluation metric to select the machine learning algorithm with the best discriminatory performance for the final model.
To assess the generalizability of the model, internal validation was conducted using the validation set data. Evaluation metrics included the receiver operating characteristic curve, DCA, calibration curve, and Brier score to assess the model’s calibration performance. Additionally, SHAP analysis was performed to determine the feature contributions in the best-performing model. A flowchart illustrating the study is presented in Figure 1.
Figure 1 Machine learning model development and validation workflow.
DFUs: Diabetic foot ulcers; LASSO: Least absolute shrinkage and selection operator; SHAP: SHapley Additive exPlanations.
RESULTS
Demographic and baseline data
A cohort of 819 patients with type 2 DFUs was screened for the study. Among these, 49 patients had missing medical records exceeding 20%; 33 patients had respiratory or urinary tract infections at admission; 11 patients had malignant tumors; 3 patients had hematologic disorders; 3 patients had acute cardiovascular or cerebrovascular events; 2 patients had autoimmune diseases; 2 patients had a history of transplant surgery; and 31 patients experienced an amputation above the ankle joint. Ultimately, 685 patients were included in the final analysis.
The median duration of diabetes was 18 (10, 22) years, and the median duration of DFUs was 1.33 (0.7, 3) months. The median age of the patients was 63 (55, 70) years, with 502 males and 183 females included. The minor amputation group included 315 patients, while the non-amputation group included 370 patients. A comparison of the general characteristics and relevant indicators between the two groups revealed statistically significant differences in the following variables: (1) Gender; (2) Neutrophil percentage; (3) Lymphocyte percentage; (4) PVC; (5) White blood cell count; (6) Red blood cell count; (7) Hb; (8) PLT; (9) NEUT; (10) MONO; (11) LYM; (12) CRP; (13) HbA1c; (14) HDL-C; (15) Serum calcium; (16) BUA; (17) ALB; (18) FIB; (19) NLR; (20) NPR; (21) PLR; (22) MHR; (23) NHR; (24) SII; (25) SIRI; (26) CALLY index; (27) PAD; (28) DKD; and (29) Wagner grade > 3 (P < 0.05; Table 1).
Table 1 Comparison of basic information between the two groups, n (%)/mean ± SD/median (interquartile range).
To maximize predictive performance and reduce the risk of overfitting, the LASSO binary LR model was employed for feature selection using the training set data, and continuous variables were standardized prior to inclusion in the LASSO regression. All variables from the prior group comparison analysis, with a P value < 0.05, were incorporated into the LASSO regression model. Based on the binomial deviance criterion with the 1-standard error (1-SE) rule, 11 non-zero coefficient variables were selected: (1) Lymphocyte percentage; (2) CALLY index; (3) PVC; (4) PLT; (5) MONO; (6) SII; (7) CRP; (8) HDL-C; (9) ALB; (10) FIB; and (11) BUA (Figure 2).
Figure 2 Least absolute shrinkage and selection operator coefficient path plot and cross-validation curve for feature selection.
A: Coefficient path plot showing variable trajectories against the log(λ) sequence; each colored line represents one variable, with variable names labeled at the right margin; B: Cross-validation curve showing binomial deviance vs log(λ); the left vertical dashed line indicates the minimum deviance (λmin), and the right dashed line indicates the 1-standard-error criterion (λ1se). LASSO: Least absolute shrinkage and selection operator; SII: Systemic immune-inflammation index; FIB: Fibrinogen; PLT: Platelet count; Hb: Hemoglobin; NPR: Neutrophil-platelet ratio; HbA1c: Hemoglobin A1c; CRP: C-reactive protein; NEUT: Neutrophil count; LYM: Lymphocyte count; Ca: Serum calcium; BUA: Blood uric acid; NLR: Neutrophil-lymphocyte ratio; CALLY: C-reactive protein-albumin-lymphocyte; HDL-C: High-density lipoprotein cholesterol; RBC: Red blood cell count; WBC: White blood cell count; MHR: Monocyte-high-density lipoprotein cholesterol ratio.
The selected features were subsequently subjected to multivariable LR analysis on the training set to identify predictive variables for subsequent model development. The multivariable LR analysis indicated that PVC, ALB, CALLY index, and FIB were independent risk factors for minor amputation in patients with DFUs (P < 0.05; Table 2).
Of note, DKD and Wagner grade > 3, despite significant univariate associations, were eliminated during LASSO shrinkage. This elimination is biologically plausible: DKD closely tracks with chronic inflammation and malnutrition, which are pathways already captured by FIB, ALB, and the CALLY index. Similarly, Wagner grade > 3 (gangrene) was highly prevalent across both groups (approximately 80%), which limited its discriminative utility. Furthermore, although PAD exhibited a strong univariate association (P < 0.001), it was eliminated during LASSO shrinkage because its predictive information was largely subsumed by PVC, which directly visualizes pedal arterial calcification and more specifically reflects the vascular pathology driving ischemic tissue loss in the foot. FIB independently contributed to the model by capturing the inflammatory and hypercoagulable state associated with disease severity.
Machine learning model
The dataset was randomly divided into a training set (70%) and a validation set (30%). No significant differences in baseline characteristics and other variables were observed between the training and validation sets (P > 0.05). In the training set, the tidymodels package was employed in combination with grid search and 10-fold cross-validation to optimize hyperparameters and construct six predictive models (Table 3). The bootstrap method was applied with 5000 resamples to compute the CIs for the AUC, sensitivity, specificity, accuracy, and F1 score.
Table 3 Hyperparameter optimization for predictive models.
The AUC was used as the primary evaluation metric to compare model performance. As shown in Table 4 and Figure 3, the mean AUC and its 95%CI for each model in the training and validation sets were as follows: (1) Random forest (RF): 0.844 (0.810-0.878) and 0.766 (0.700-0.828); (2) K-Nearest Neighbors: 0.785 (0.744-0.824) and 0.734 (0.665-0.800); (3) Extreme gradient boosting: 0.757 (0.714-0.799) and 0.749 (0.682-0.814); (4) LR: 0.754 (0.710-0.795) and 0.755 (0.689-0.817); (5) Multilayer perceptron: 0.754 (0.711-0.797) and 0.756 (0.685-0.816); and (6) Support vector machine: 0.753 (0.708-0.795) and 0.753 (0.684-0.815).
Figure 3 Receiver operating characteristic analysis of six machine learning models in training and validation sets.
A: Training set (n = 480); B: Validation set (n = 205). The X-axis represents 1-specificity, and the Y-axis represents sensitivity. ROC: Receiver operating characteristic; KNN: K-nearest neighbors; LR: Logistic regression; MLP: Multilayer perceptron; RF: Random forest; SVM: Support vector machine; XGBoost: Extreme gradient boosting.
Table 4 Predictive performance comparison of six machine learning models.
The RF model demonstrated the highest AUC values in both the training and validation sets, indicating the best discriminatory ability and predictive performance. Therefore, the RF algorithm was selected for the final model construction.
RF model performance and interpretation
The model’s discriminatory ability (its capacity to correctly identify positive samples) was evaluated. In this study, the AUC was used for evaluation. Typically, an AUC value greater than 0.75 is indicative of moderate discriminatory performance. As shown in Table 5 and Figure 4A and B, the RF model achieved a mean AUC (95%CI) of 0.844 (0.810-0.878) in the training set and 0.766 (0.700-0.828) in the validation set. The mean sensitivity was 0.747 in the training set and 0.628 in the validation set, whereas the mean specificity was 0.761 and 0.703, respectively. The mean F1 score was 0.737 and 0.633 for the training and validation sets, respectively. All performance metrics were calculated using a classification threshold of 0.5, which is the default probability cutoff for binary classification. For the validation set (n = 205), the confusion matrix yielded true positive = 59, false negative = 35, true negative = 78, and false positive = 33, with a positive predictive value of 0.641 and a negative predictive value of 0.690. These results indicated that the RF-based clinical prediction model for minor amputation within 3 months after admission among patients with DFUs had moderate discriminatory power.
Figure 4 Validation and evaluation of the random forest model.
A and B: Receiver operating characteristic curves for the training set (n = 480; area under the receiver operating characteristic curve =0.844) and validation set (n = 205; area under the receiver operating characteristic curve = 0.766), respectively; C and D: Calibration curves for the training set (Brier score = 0.161) and validation set (Brier score = 0.198), respectively; the dashed diagonal line represents ideal calibration; E and F: Decision curve analysis for the training set and validation set, respectively. The Y-axis denotes mean net benefit, and the X-axis denotes threshold probability (%). AUC: Area under the receiver operating characteristic curve; ROC: Receiver operating characteristic.
Table 5 Predictive performance evaluation of the random forest model.
To assess the accuracy and consistency of the RF model’s predicted probability of positive events with actual occurrences, the Brier score was used for calibration. As shown in Figure 4C and D, the calibration curves for the RF model in both the training and validation sets closely approximated the ideal curve, with Brier scores of 0.161 and 0.198, respectively. The calibration intercepts were 0.036 and 0.075, and the calibration slopes were 0.915 and 0.738 for the training and validation sets, respectively. These values indicated acceptable calibration of the RF model.
Additionally, DCA was employed to further evaluate the clinical utility of the RF model. The DCA curve, which assesses the net benefit of clinical decisions across various threshold probabilities, quantitatively integrates model performance with real-world applicability. This provides decision support for choosing an optimal treatment strategy. The results indicated that, in the training set, when the threshold probability ranged from 7% to 95%, the RF model demonstrated higher clinical net benefit. In the validation set, this range was between 15% and 80% (Figure 4E and F). These findings further confirmed that the RF predictive model can provide significant practical value in clinical settings.
The SHAP feature importance for the RF model is presented in Figure 5A. The evaluated features are ranked from highest to lowest based on their mean absolute SHAP values, reflecting the impact of features on the prediction. The results indicated that, among all the included clinical features, FIB made the greatest contribution to the model’s prediction, followed by CALLY, ALB, and PVC. The SHAP summary plot for the RF model illustrates the impact of features on the prediction model (Figure 5B). In this prediction model, a higher SHAP value for a feature indicates that the feature contributes positively to the predicted risk of minor amputation. For example, elevated FIB was associated with higher predicted risk of minor amputation in this RF model. Additionally, lower CALLY index values correspond to an increased predicted probability of minor amputation.
Figure 5 SHapley Additive exPlanations-based interpretation of the random forest model.
A: Mean SHapley Additive exPlanations (SHAP) value bar plot ranking the four key features by importance (fibrinogen > C-reactive protein-albumin-lymphocyte > albumin > pedal vessel calcification); B: SHAP summary (beeswarm) plot, where each dot represents one patient, color indicates feature value (yellow = high, blue = low), and the X-axis shows SHAP value; C-F: SHAP dependence plots for fibrinogen, pedal vessel calcification, C-reactive protein-albumin-lymphocyte, and albumin; blue dots indicate non-amputation and orange dots indicate amputation, and the horizontal dashed line marks SHAP value = 0. ALB: Albumin; CALLY: C-reactive protein-albumin-lymphocyte; FIB: Fibrinogen; PVC: Pedal vessel calcification; SHAP: SHapley Additive exPlanations.
The y-axis represents the SHAP values of features, and the values of certain features are displayed on the X-axis. Each dot represents the SHAP value of a feature for an individual patient; the color coding is as shown in Figure 5C-F. As shown in Figure 5C-F, the dependence plot demonstrated that higher FIB and PVC values significantly increased the risk of minor amputation, whereas lower CALLY and ALB levels were also associated with an elevated risk of amputation. This visual representation aids clinicians in better understanding the specific contributions of each variable to the risk of minor amputation in patients with DFUs, thereby enabling more effective prediction of disease progression and optimization of treatment strategies.
DISCUSSION
DFUs are one of the most serious and costly complications of diabetes. In China, the overall amputation rate among DFUs patients is approximately 19.03%, with major and minor amputation rates accounting for 2.14% and 16.88%, respectively[19]. While a minor amputation can help control infection, alleviate pain, promote healing, and prevent disease progression by removing limited affected tissues, it still compromises foot function and increases mortality risk[20-22]. The aim of this study was to develop and validate an interpretable machine learning model to predict the risk of minor amputation occurring within 3 months after admission among patients with DFUs. Four key indicators – FIB, the CALLY index, ALB, and PVC – were identified as core predictive drivers through multi-stage variable screening, and six machine learning models were constructed based on these indicators[23,24]. Among the six algorithms compared, the RF model showed the best overall performance (AUC = 0.844), demonstrating superior discriminative ability, calibration, and interpretability.
Our findings strongly support the central role of chronic inflammation in the progression of DFUs and subsequent minor amputation[8,25]. FIB levels not only reflect the systemic inflammatory state but may also directly lead to microcirculatory thrombosis through enhanced blood coagulability, thereby exacerbating local ischemia and accelerating tissue gangrene, ultimately leading to amputation[26]. This finding is consistent with those of Ardelean et al[27], and together these findings support FIB as a key predictor of lower-limb amputation risk that may help clinicians identify high-risk patients and intervene early[28].
In this study, the CALLY index emerged as a novel predictor that integrates inflammatory status, nutritional status, and immune function[29,30]. Although previous studies have highlighted its prognostic value in critical conditions such as oncology and sepsis, few studies have applied the CALLY index specifically to DFUs. Our findings revealed a significant negative correlation with the risk of minor amputation. The CALLY index has been correlated significantly with markers of infection, inflammation, and ulcer severity[31]. A lower CALLY index implies a superimposed triple burden of increased inflammation, malnutrition, and immune dysfunction, a state that severely impairs wound healing capacity. This triple burden collectively impairs angiogenesis, delays re-epithelialization, and compromises local immune defense, ultimately leading to poor wound healing and disease progression. Given that its components are easily accessible in clinical practice, the CALLY index provides a simple and efficient tool for routine risk assessment[32,33].
ALB is an important indicator of the nutritional status of patients with DFUs and serves as a crucial reference for assessing the severity and prognosis of their condition[34]. This study showed that reduced ALB levels were significantly correlated with an increased risk of minor amputation in patients with diabetic foot[35]. Reduced ALB levels reflect a common malnutritional state in patients with DFUs, which involves excessive protein loss and increased metabolic demands due to wound exudation[8]. Hypoalbuminemia increases the risk of amputation by weakening immune function and tissue repair, leading to a vicious cycle of “malnutrition-uncontrolled infection-poor wound healing”[36].
Furthermore, PVC was identified as a significant anatomical risk factor. Its pathogenesis involves disorders of glucose and lipid metabolism coupled with chronic inflammation, which lead to arterial stiffening and reduced elasticity, ultimately resulting in vascular stenosis or occlusion[37,38]. This significantly compromises blood perfusion and elevates the risk of tissue ischemic necrosis[39,40]. Our findings corroborate existing evidence linking PVC to adverse outcomes in diabetic foot. The detection of vascular calcification on plain foot radiographs provides a simple, cost-effective, and widely accessible method for identifying patients at high risk, thereby offering significant value for preoperative planning and prognostic assessment[41].
DKD and Wagner grade > 3 were associated with minor amputation in univariate analyses but were not retained after LASSO selection. This should not be interpreted as evidence against their clinical relevance. Rather, these variables likely lost independent significance because FIB, ALB, the CALLY index, and PVC capture the same pathophysiological axes – chronic inflammation, malnutrition, and vascular calcification – that drive DKD progression and gangrene. Importantly, the final model relies exclusively on routine laboratory and radiographic variables, enabling objective, reproducible risk stratification upon admission.
This study presents a prediction model for minor amputation risk in DFUs patients that integrates the CALLY index with SHAP-based interpretability. While the CALLY index has been correlated with major amputation[31] and SHAP has been applied in models for lower extremity amputation[42,43], evidence on their combined use specifically for minor amputation in DFUs remains limited. This model enables the early identification of high-risk patients, thereby creating a critical window for targeted interventions. This stratified approach not only enhances the efficiency of prevention and treatment but also mitigates the risk of overtreatment in low-risk individuals[44].
The main causes of amputation in patients with DFUs are uncontrollable infection at the site of ulceration due to a persistent high-glycemic environment, microcirculatory disorders, immune dysfunction, and impairment of tissue repair, which ultimately lead to tissue necrosis[19,45,46]. The CALLY Index is particularly suitable for assessing the course of DFUs, a disease characterized by chronic inflammation and malnutrition, as it effectively evaluates the patient’s overall physiological reserve and immune-nutritional balance[47,48]. Through SHAP explainability analysis, we confirmed a significant model-level association between CALLY index reduction and minor amputation risk. Beyond validating the index’s strong tie to DFUs prognosis, this discovery provides a fresh perspective on the systemic disease mechanisms at play[49].
The establishment of this prediction model enables clinicians to achieve early and accurate risk stratification of DFUs patients so that targeted interventions such as intensive anti-infection, nutritional support, and vascular reconstruction can be initiated in a timely manner for the high-risk group, which improves the prognosis and reduces the amputation rate[50,51]. High-risk patients may receive more intensive wound care, including frequent dressing changes and advanced wound dressings. Vascular assessment can be prioritized to identify patients who are eligible for revascularization. Outpatient surveillance can be intensified to detect early signs of deterioration. These strategies may optimize resource allocation and improve clinical outcomes.
While this study presents novel findings, several limitations should be acknowledged. First, its retrospective design carries an inherent risk of selection bias and incomplete data acquisition, which may compromise data integrity. In particular, diabetic peripheral neuropathy was not routinely recorded and could not be included as a candidate variable, which may represent residual confounding. Second, the single-center origin of our cohort limits its diversity and representativeness, potentially affecting the model’s generalizability and external validity. To advance the clinical translation of this prediction model, future efforts should prioritize large-scale, multi-center prospective studies to validate its robustness and universal applicability. Additionally, incorporating the dynamic changes of key indicators into the assessment framework could enhance the timeliness and accuracy of risk prediction.
CONCLUSION
In this study, six machine learning algorithms were employed to construct predictive models for the risk of minor amputation within 3 months after admission in patients with DFUs. The results indicated that the RF model demonstrated the best predictive performance. Additionally, the SHAP algorithm was employed to explain the contribution of variables to the model. The analysis revealed that higher FIB levels, lower CALLY index and ALB levels, and the presence of PVC were significantly associated with an increased risk of minor amputation in DFUs patients. These findings provide clinical evidence for early risk stratification and targeted intervention, which may lower the risk of minor amputation within 3 months after admission in DFUs patients.
Huang J, Chen H, Deng J, Liu X, Shu T, Yin C, Duan M, Fu L, Wang K, Zeng S. Interpretable machine learning for predicting 28-day all-cause in-hospital mortality for hypertensive ischemic or hemorrhagic stroke patients in the ICU: a multi-center retrospective cohort study with internal and external cross-validation.Front Neurol. 2023;14:1185447.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in RCA: 23][Reference Citation Analysis (0)]
Yang M, Lin SQ, Liu XY, Tang M, Hu CL, Wang ZW, Zhang Q, Zhang X, Song MM, Ruan GT, Zhang XW, Liu T, Xie HL, Zhang HY, Liu CA, Zhang KP, Li QQ, Li XR, Ge YZ, Liu YY, Chen Y, Zheng X, Shi HP. Association between C-reactive protein-albumin-lymphocyte (CALLY) index and overall survival in patients with colorectal cancer: From the investigation on nutrition status and clinical outcome of common cancers study.Front Immunol. 2023;14:1131496.
[RCA] [PubMed] [DOI] [Full Text] [Full Text (PDF)][Cited by in RCA: 169][Reference Citation Analysis (1)]
Li P, Li Q, Tang M, Hu X, Tian J, Zhang J, Yang C, Zhu B. Associations of phosphorus concentrations with medial arterial calcification in lower-extremity arteries and diabetic foot in people with diabetes: a retrospective cross-sectional study.Cardiovasc Diabetol. 2024;23:275.
[RCA] [PubMed] [DOI] [Full Text][Cited by in RCA: 4][Reference Citation Analysis (0)]
P-Reviewer: Chen SP, Affiliate Associate Professor, MD, China; Fu L, Associate Professor, PhD, China; Khurram MF, Full Professor, Head, PhD, Professor, India S-Editor: Luo ML L-Editor: A P-Editor: Yang YQ
