C-reactive protein (CRP), produced by hepatocytes in response to proinflammatory cytokines like interleukin-6 (IL-6) and tumor necrosis factor-α (TNF-α), is a well-established biomarker of systemic inflammation. Its rapid elevation during infection, tissue injury, or inflammation, followed by a decline with effective treatment, makes it a cornerstone of clinical monitoring for disease activity and therapeutic response (1). In the context of wound healing, CRP dynamics reflect the balance between inflammatory clearance and tissue repair—an equilibrium particularly fragile in chronic wounds, where persistent inflammation often impedes progression to the proliferative and remodeling phases (2).

Diabetic wounds, such as foot and lower extremity skin ulcers, are a significant complication of diabetes mellitus, impacting up to 25% of patients during their lifetime (3). These wounds exhibit impaired angiogenesis, neuropathy, and dysregulated inflammation, resulting in extended healing durations, frequent recurrences, and an elevated risk of amputation (4). The clinical burden is substantial: even small-to-moderate diabetic wounds (5–8 cm²) can require weeks to months to heal, with approximately 30%–40% of cases progressing to prolonged healing (>8 weeks) despite standard care (5, 6). This unpredictability not only increases healthcare costs but also diminishes patients’ quality of life, underscoring the need for reliable tools to identify high-risk individuals early.

Existing research on diabetic wound healing has primarily focused on static baseline factors, such as initial wound size, depth, and comorbidities like peripheral artery disease (7). While these factors provide valuable context, they fail to capture the dynamic biological processes that drive healing outcomes. In recent years, studies have begun to explore inflammatory markers as predictors: for example, elevated baseline CRP levels have been associated with prolonged healing (8), and serial measurements of proinflammatory cytokines (e.g., IL-6, TNF-α) have shown promise in assessing treatment response (9). However, these studies often lack standardization in wound size, include heterogeneous patient populations (e.g., mixing minor debridement with major amputations), or rely on single-time-point measurements, limiting their clinical applicability.

The importance of dynamic inflammatory trajectories—rather than isolated values—has gained recognition in other chronic conditions, such as sepsis and rheumatoid arthritis, where patterns of marker decline (e.g., CRP “clearance rate”) strongly predict outcomes (10, 11). In diabetic wounds, this principle is equally relevant: an initial elevation in CRP may indicate a productive inflammatory response to clear pathogens, but a failure to decline after interventions (e.g., antibiotics, debridement) signals persistent infection or impaired healing mechanisms (12). Surprisingly, few studies have systematically analyzed serial CRP measurements to predict prolonged healing, especially in well-characterized cohorts with standardized wound sizes and uniform surgical interventions (e.g., debridement alone, excluding amputations).

Machine learning (ML) has become a valuable asset in wound healing research, surpassing traditional regression models by effectively capturing nonlinear relationships and interactions among multiple variables. Existing ML models for diabetic wounds have integrated features such as wound imagery, microbiome data, and baseline clinical metrics to predict healing time or amputation risk (13–15). For instance, a recent study using random forest algorithms achieved an AUC of 0.82 for predicting 12-week healing in diabetic foot ulcers, leveraging features like wound area and HbA1c (13). However, these models rarely incorporate dynamic inflammatory markers, and none have specifically focused on CRP trajectories across critical treatment time points (e.g., post-antibiotic, pre-operative, and post-operative phases).

Our study seeks to create a machine learning model to predict diabetic wound healing times exceeding 8 weeks, emphasizing the dynamic patterns of CRP. By restricting the cohort to patients with standardized wound sizes (5–8 cm²) who underwent only debridement (excluding amputations), we minimize confounding by wound severity and surgical complexity. We hypothesize that serial CRP measurements—specifically the magnitude and rate of decline after antibiotic treatment and surgery—will outperform static baseline factors in predicting prolonged healing. This model will not only enhance risk stratification but also highlight critical intervention windows, enabling clinicians to adjust anti-inflammatory and antimicrobial strategies proactively. Our objective is to develop a clinically useful tool to enhance outcomes for patients with diabetic wounds. While we utilized a temporal validation cohort to assess model robustness, the single-center design warrants future multi-center replication to confirm generalizability across diverse clinical settings.

This study developed and validated 12 machine learning models to predict prolonged diabetic wound healing, with the GradientBoosting model demonstrating superior performance in both training and validation sets. Through SHAP (SHapley Additive exPlanations) analysis, we identified the core factors driving model predictions: dynamic inflammatory markers—specifically CRP_2nd (C-reactive protein measured after admission antibiotic therapy but before surgical debridement), CRP_3rd (CRP at discharge following debridement), and baseline CRP (CRP at admission)—emerged as the most influential predictors. Among these, CRP_2nd (mean absolute SHAP value 0.460) ranked first. This finding aligns with the well-established role of inflammation as a central regulator of wound repair, where both insufficient and excessive inflammation can disrupt the healing cascade (16). Physiological mechanisms for CRP_2nd to be more predictive than other time points: (1) specificity of time window: CRP_2nd was measured after antibiotic treatment and before surgery (3–7 days after admission). This time point is a key window to evaluate the effect of anti-infective treatment and the ability to clear inflammation - if the antibiotic treatment is effective and the pathogen is controlled, CRP should decrease significantly. If CRP persistently increased (CRP_2nd > 50 mg/L), it indicated that the infection was not controlled or tissue repair was impaired, and it was an important warning signal for subsequent prolonged healing. CRP_2nd directly reflects the balance between inflammation and repair. CRP on admission may be disturbed by acute stress and complications, and CRP_3rd after surgery may be affected by surgical traumatic stress, with low predictive specificity. (3) The time window corresponding to CRP_2nd provides an opportunity for preoperative intervention. Timely adjustment of treatment (such as replacement of antibiotics and albumin supplementation) for patients with elevated CRP_2nd can effectively improve the healing outcome, while CRP measurement at other time points lack such a critical intervention window.

Our results demonstrated that the prolonged healing group displayed significantly elevated CRP levels at all three time points compared to the normal healing group (Supplementary Table S1). The prolonged group exhibited a significantly slower rate of CRP reduction (p < 0.001), emphasizing that “inflammatory resolution” rather than the initial inflammation level influences healing outcomes. This aligns with research indicating that prolonged inflammation in diabetic wounds hinders the shift from the inflammatory to the proliferative phase, marked by decreased fibroblast activity and angiogenesis (17–22). For instance, Eming et al. (23) noted that persistent inflammation in chronic wounds leads to tissue damage by promoting excessive release of reactive oxygen species (ROS) and matrix metalloproteinases (MMPs), which directly hinder collagen synthesis.

Consistent with this, recent studies have indicated that non-healing wounds possess the characteristic of “an unstable spatiotemporal balance between inflammation and protease activity”: Uncontrolled inflammation leads to excessive secretion of proteases, which in turn damages growth factors and matrix proteins, thereby hindering the healing process (24).

The AGE-RAGE axis, a central pathway in diabetic complications, offers a mechanistic framework underlying our findings (25–27). The activation of RAGE by AGEs or EN-RAGEs induces the release of proinflammatory cytokines such as TNF-α and IL-1β, which perpetuates inflammation and protease activity (27–30). While our study did not directly measure AGEs or RAGE, the prominence of CRP—a downstream marker of this axis—in prediction suggests that our model implicitly encapsulates this pathological cascade. This reinforces the clinical relevance of our machine learning approach: by prioritizing serial CRP monitoring, clinicians can indirectly assess RAGE-mediated inflammation and intervene to restore the “self-limited” inflammatory phase necessary for healing.

Beyond CRP dynamics, hematological indices revealed distinct inflammatory phenotypes between groups. The prolonged healing group exhibited significantly higher neutrophil percentage (Neut%) and count (Neut#), alongside lower lymphocyte percentage (Lymph%) and count (Lymph#) (Supplementary Table S1). This pattern reflects a “persistent neutrophilic inflammation” phenotype, which is pathological in the context of chronic wounds.

Neutrophils are critical for clearing pathogens in acute wounds, but their persistent infiltration in diabetic wounds leads to excessive degranulation and NETosis (neutrophil extracellular trap formation), inducing damage to surrounding healthy tissue and impairing re-epithelialization (31). Our findings align with those of Rui et al. (32), who reported that neutrophil hyperactivity in diabetic ulcers correlates with elevated MMP-8 levels, which degrade growth factors and extracellular matrix. Conversely, reduced lymphocytes in the prolonged group may indicate impaired adaptive immunity—lymphocytes are essential for transitioning to the proliferative phase by secreting cytokines like TGF-β1, which promotes fibroblast recruitment (33). This imbalance between neutrophils and lymphocytes underscores the need for targeted anti-inflammatory strategies (e.g., neutrophil apoptosis promoters or lymphocyte modulators) in high-risk patients.

Higher levels of albumin (ALB) and positive therapeutic responses (therapeutic_response_all) were identified as protective factors, both associated with a decreased risk of prolonged healing. Our results further demonstrated that inflammation does not operate in isolation but interacts with nutritional and metabolic status to influence healing. The prolonged group exhibited notably lower albumin (ALB) levels (p < 0.001), with SHAP analysis highlighting ALB as a crucial predictive feature (mean absolute SHAP value = 0.020). Albumin, a major plasma protein, modulates inflammation by binding proinflammatory mediators (e.g., LPS) and supporting immune cell function; hypoalbuminemia thus exacerbates inflammation and impairs tissue repair (34). A study supports this by demonstrating that adding albumin to the diet of malnourished diabetic patients speeds up wound healing by lowering CRP levels and boosting neutrophil phagocytosis.

The prolonged group exhibited increased glucose levels (p = 0.021), aligning with findings that hyperglycemia enhances oxidative stress and advanced glycation end product accumulation, thereby intensifying inflammation through NF-κB pathway activation (35). Notably, our data showed that short-term glucose fluctuations (Glu) exerted a more pronounced effect than long-term glycemic control (HbA1c, p = 0.092), emphasizing the need for tight perioperative glucose management to mitigate inflammation.

The cumulative evidence from our analyses—highlighting dynamic inflammation (particularly CRP_2nd) as a central driver, alongside its interplay with neutrophilic phenotypes, nutritional status (ALB), and metabolic factors (perioperative glucose)—translates into actionable clinical insights for optimizing diabetic wound management:

First, in terms of early risk stratification, the prominence of CRP_2nd—measured after initial antibiotic therapy but before debridement—marks a critical window for identifying high-risk patients. This aligns with our key finding that inflammatory resolution, rather than initial inflammation magnitude, dictates outcomes. By monitoring CRP_2nd alongside early therapeutic response (therapeutic_response_1), clinicians can stratify patients pre-operatively, enabling timely adjustments to antibiotic regimens or nutritional support (e.g., albumin supplementation) to mitigate persistent inflammation before it entrenches.

Second, the distinct neutrophilic inflammation phenotype (elevated Neut#/Neut% with reduced Lymph#/Lymph%) in prolonged healers points to the need for targeted anti-inflammatory strategies. This phenotype, linked to excessive NETosis and MMP-mediated tissue damage in our mechanistic framework, supports interventions aimed at dampening neutrophil hyperactivity—such as topical anti-TNF-α therapy to curb proinflammatory cytokine release or neutrophil elastase inhibitors to limit matrix degradation—thereby restoring the inflammatory balance required for progression to proliferation.

Finally, these findings underscore the value of integrated management strategies that address the interplay between inflammation, nutrition, and metabolism. Given our observations that hypoalbuminemia exacerbates inflammation and perioperative hyperglycemia amplifies NF-κB-mediated inflammatory responses, concurrent efforts to optimize glycemic control, supplement albumin in malnourished patients, and modulate inflammation through targeted therapies could synergistically improve the wound microenvironment, ultimately accelerating healing (36, 37).

The clinical transformation path of the model is as follows: (1) Integration into the electronic medical record (EMR) system: The model can be embedded into the hospital’s existing EMR system via an API interface to automatically extract patients’CRP test results, biochemical indicators, and other relevant data, generate real-time prolonged healing risk scores (0–100 points), and assist clinicians in formulating personalized treatment plans; (2) Outpatient risk stratification tool: For outpatients with diabetic wounds, rapid risk assessment can be performed using pre-debridement CRP_2nd test results, with high-risk patients (score ≥70 points or CRP_2nd > 80 mg/L) prioritized for intensive interventions such as albumin supplementation (target ≥35 g/L) and targeted anti-inflammatory therapy, intermediate-risk patients (score 40–69 or 30 mg/L ≤ CRP_2nd ≤ 80 mg/L) receiving intensified wound care and close CRP monitoring, and low-risk patients (score < 40 or CRP_2nd < 30 mg/L) continuing standard care with discharge CRP_3rd measurement; (3) Remote follow-up monitoring: Combined with dynamic CRP monitoring data, the model updates risk assessments in real time to guide home care practices and optimize follow-up visit timing (weekly for high-risk, biweekly for intermediate-risk patients), with the risk threshold (score ≥70, predicted probability ≥0.75) derived from the validation set ROC curve’s Youden index (J = 0.856) to balance sensitivity and specificity.

Compared with the existing studies, the core innovation of this study is as follows: (1) For the first time, CRP_2nd was identified as the optimal predictor by focusing on the “dynamic CRP trajectory” instead of the static baseline value; (2) Standardized wound size (5–8 cm²) and unified intervention method (debridement only) were used to reduce heterogeneous interference; (3) The generalization of the model was verified by an external independent cohort (140 cases in 2025), while most of the previous studies lacked external validation or had high cohort heterogeneity. (4) Combining SHAP analysis to reveal the predictive mechanism and improve the clinical interpretability of the model.

This study has several limitations: (1) Single-center retrospective design: Training and temporal validation cohorts were from the same institution, limiting generalizability to diverse populations and care protocols. Future multi-center validation (planned n ≥ 500) is needed to confirm broad applicability. (2) Follow-up bias: Healing status was confirmed via in-person (68%) or telephone (32%) follow-up. Telephone follow-up may introduce misclassification bias (e.g., overreporting of healing), mitigated by standardized follow-up questions and cross-referencing with outpatient records. (3) Limited biomarker sampling: CRP was measured at only three time points; more frequent sampling (e.g., postoperative days 1–3) could capture finer inflammatory fluctuations. (4) Missing variables: Microbiological and imaging data were excluded due to high missing rates (>30%), which may improve model performance if integrated in future studies.

Future research directions: (1) expand the multi-center cohort validation and include patients from different regions and hospitals of different levels to improve the representation of the model population; (2) integrating imaging indicators (such as wound ultrasound and infrared imaging) and microbiome data to further optimize the performance of the model; (3) develop mobile applications for convenience of primary medical institutions and patients; (4) Prospective clinical trials should be conducted to verify whether model-guided treatment strategies can improve the prognosis of patients.

Taken together, our findings underscore the central role of inflammatory dynamics—particularly the post-antibiotic, pre-operative CRP trajectory—in governing diabetic wound healing outcomes. By unraveling the interplay between inflammation, nutrition, and metabolism through explainable machine learning, this work not only enhances our understanding of diabetic wound pathophysiology but also provides a foundation for developing targeted, inflammation-modulating strategies to improve clinical outcomes.

This study developed a GradientBoosting model to predict prolonged healing of diabetic wounds, with SHAP analysis revealing that dynamic inflammatory markers—especially CRP_2nd (post-antibiotic, pre-debridement)—are the most critical predictors. Inflammatory resolution, rather than initial inflammation magnitude, may determine healing outcomes, with persistent neutrophilic inflammation, hypoalbuminemia, and perioperative hyperglycemia potentially exacerbating delays. This work highlights the post-antibiotic, pre-operative phase as a critical window for risk stratification and provides a rationale for targeted interventions (e.g., anti-inflammatory therapies, albumin supplementation) to restore inflammatory balance. Future multi-center studies with more granular inflammatory monitoring may help translate these insights into clinical practice.