Initially, 14,369 records were retrieved through the system. After removing duplicated studies, 9181 articles remained. Upon reviewing titles and abstracts, 9132 articles unrelated to the research topic were excluded. Furthermore, we identified four studies through citation searches of relevant systematic reviews and conducted full-text readings for 39 articles. Among these, we excluded five studies targeting populations with diabetes or peripheral artery disease (PAD), five studies predicting outcomes such as re-amputation or considering indicators like death as adverse events, three studies with fewer than two predictors, two studies with duplicate sample data, and five studies with missing important data or lacking model performance evaluations. Ultimately, 20 articles were included in this review. Figure 1 illustrated a flowchart depicting the Preferred Reporting Items for Systematic Reviews and Meta-Analyses (PRISMA) 2020 guidelines, outlining the comprehensive search process and its outcomes.

PRISMA 2020 flow diagram for systematic review

Among the included literatures, 17 [17, 18, 22, 23, 28,29,30,31,32,33,34,35,36,37,38,39,40] were in English and three [24, 41, 42] were in Chinese, with 12 publications [22,23,24, 28,29,30,31,32,33,34,35, 41] within the last 5 years. Covering studies conducted in nine countries: 10 [22,23,24, 28, 29, 31, 32, 34, 41, 42] from China, two each from the United States [30, 39] and India [36, 37], and one each from Germany [35], the Philippines [33], South Korea [17], Portugal [17], the Netherlands [38], and Spain [40]. Seven studies [18, 33,34,35,36, 38, 42] were prospective, while the remainder were retrospective. Three papers [31, 40, 41] focused on predictive models for DF amputation in T2D patients. Five studies [28,29,30, 38, 39] involved multiple centers, while the remaining were conducted at single centers. Two studies [30, 33] concentrated on major amputation as the primary outcome, one [29] on minor amputation, and four studies [ 18, 23, 35, 38] simultaneously predicted two different outcome indicators (e.g., amputation and minor amputation, amputation and major amputation, minor amputation and major amputation). The remaining studies focused on any form of amputation, with one study [37] reporting simple and complex amputation prediction models. Sample sizes ranged from 23 to 32,685 participants across the studies, with follow-up periods ranging from 3 months to 3.5 years. An overview of the essential study characteristics can be found in Table 1.

A total of 54,265 DF patients were included in these studies. Amputation occurred in 3779 patients, with a prevalence of 5.9–58.5%. Three of the included papers utilized external data to validate the predictive performance of existing amputation scoring systems [17, 18, 33], while the remaining studies focused on developing new models. The range of candidate factors considered in each study varied from seven to 44, with the events per variable (EPV) spanning from 0.194 to 53.556. Among the 17 studies dedicated to model development, five opted for traditional logistic regression (LR) [31, 36, 38,39,40], four employed a single machine learning (ML) method [23, 28, 35, 37], and eight utilized multiple ML techniques [22, 24, 29, 30, 32, 34, 41, 42]. The most prevalent ML method was random forest (RF), used in five studies, followed by extreme gradient boosting (XGBoost) and support vector machine (SVM), each appearing in four studies. Across the 20 papers, a total of 77 prediction models were constructed, with 20 optimal models highlighted across 17 model development studies. Notably, in all five studies, data with incomplete clinical information were excluded, yet details regarding the number of missing values and the handling methods were not explicitly mentioned [22, 28,29,30, 37]. Regarding model evaluation, only 11 studies appropriately assessed differentiation and calibration. Thirteen model development studies employed internal validation techniques, with seven using Bootstrap or K-fold cross-validation methods [23, 24, 29, 32, 41], while four lacked a formal model validation process [34, 35, 37, 39]. In terms of model presentation, Stefanopoulos et al. [30] and Wang et al. [29] developed web-based risk calculators for clinical application and dissemination, Li [24] created amputation risk assessment and prediction software, and Peng et al. [