A multicenter retrospective cohort included 1,020 adults (≥18 years) undergoing elective craniotomy at four tertiary centers (2005–2025). Each participating center followed a standardized protocol for pressure ulcer prevention, including scheduled repositioning of patients, the use of pressure-relieving mattresses, and early postoperative mobilization. Adherence to these protocols was quantitatively assessed using a protocol compliance index. Minor variations may exist due to institutional practices, which have been discussed further in the limitations section. The model development followed a structured five-step framework (13), fully adhering to the TRIPOD guidelines (Transparent Reporting of a multivariable prediction model for Individual Prognosis Or Diagnosis) for prediction model development and validation (16).

Patients were excluded if they met any of the following criteria: 1. pre-existing PU or skin breakdown at baseline; 2. emergency craniotomy due to life-threatening conditions (e.g., intracranial hemorrhage, severe traumatic brain injury); 3. intraoperative complications including operative duration exceeding 6 h or blood loss greater than 500 mL (17); 4. postoperative complications such as sepsis (defined as fever >38.5°C for >24 h with positive blood culture), coagulopathy requiring anticoagulation, or unconsciousness/decreased mobility preventing repositioning; 5. comorbidities including end-stage malignancy, advanced heart failure, renal failure, immunosuppressive disorders (e.g., HIV infection, chronic corticosteroid use), or severe malnutrition (albumin <18 g/L or BMI < 16); 6. inability to complete the 7-day postoperative follow-up period (e.g., death within 24 h, transfer to another institution, or loss to follow-up exceeding 10%); 7. concurrent dermatological conditions (e.g., eczema, psoriasis) that could interfere with ulcer assessment; 8. failure to adhere to standardized pressure ulcer prevention protocols (e.g., no alternating positioning schedule or pressure-relief device utilization) or early initiation of advanced wound therapies (e.g., negative pressure wound therapy within 48 h postoperatively); or 9. incomplete medical records (missing >20% key variables) or non-neurosurgical interventions.

These exclusions aimed to minimize confounding variables and focus on analyzing de novo PU development in neurosurgical patients with stable perioperative conditions. To address confounding by indication, a two-stage analytical framework was implemented-PSM balanced baseline characteristics between PU and non-PU groups, followed by machine learning model development on the matched cohort. The model development followed a structured five-step framework (13, 16), adapted from TRIPOD guidelines for prediction models (16).

A directed acyclic graph identified 20 confounders. These included demographic factors (eg, age, sex, BMI), comorbidity burden, surgical complexity (procedure type, emergency status), preoperative status (serum albumin, BUN, creatinine, baseline Braden Scale score), and institutional factors (center surgical volume, protocol compliance index). Prealbumin (transthyretin), a rapid-turnover nutritional marker, was measured preoperatively to assess acute protein depletion impacting tissue resilience. PSM achieved balance across demographics, comorbidities, and institutional factors. Protocol compliance scores were integrated as stabilizing weights. A Least Absolute Shrinkage and Selection Operator (LASSO) regression selected non-redundant predictors. The optimal λ (λ = 0.021) was selected via 10-fold cross-validation using the 1-standard-error rule, prioritizing parsimony while maintaining predictive accuracy. LASSO identified variables (e.g., diabetes duration) with non-zero coefficients.

A multivariable logistic regression model was trained on the LASSO-selected predictors. To capture non-linear relationships, an XGBoost model (learning rate = 0.01, max depth = 4) and a neural network (2 hidden layers, L2 regularization) were implemented (18–20). Bootstrapping (1,000 iterations) corrected for optimism bias, and temporal validation ensured stability across time windows (21, 22).

PSM was performed using the nearest neighbor algorithm with a caliper width of 0.2 standard deviations of the logit propensity score. Covariate balance was assessed through SMD, with an absolute SMD < 0.1 considered indicative of adequate balance. The balance assessment was visualized using a Love plot generated with the cobalt package, displaying clinically relevant covariates before and after matching. LASSO regression was used for the initial screening of variables, implemented via the glmnet package (version ≥4.1) in R (version 4.4.3). Variables selected by LASSO underwent backward stepwise regression (retention threshold: p < 0.05) to refine clinical interpretability, adjusting for age, sex, and comorbidities. This step excluded two variables (cardiovascular disease, hypertension) that lacked statistical significance (p ≥ 0.05) without compromising model performance (ΔAUC<0.01 in sensitivity analysis). Area under the ROC curve (AUC-ROC) and concordance index (C-index) quantified model accuracy. Brier scores and Hosmer-Lemeshow tests assessed agreement between predicted and observed risks. Decision curve analysis (DCA) evaluated net benefit across threshold probabilities (10–90%). Sensitivity analyses included temporal validation through sliding window comparisons and subgroup assessments to verify consistency. All analyses were implemented in R version 4.4.3.

Although the original goal was to enroll a larger number of patients, strict inclusion criteria such as 7-day postoperative follow-up and the exclusion of emergency craniotomy cases limited our sample size. Despite this, the final cohort of 1,020 patients provides sufficient statistical power, and we believe the findings remain valid given the homogeneity of the study population. Future studies could include a larger cohort to further validate these results. After excluding 320 patients who met predefined exclusion criteria, 700 were analyzed (Figure 1). PSM yielded balanced training (n = 340) and validation (n = 360) cohorts. Post-matching SMDs confirmed covariate balance (all <0.1).

Baseline characteristics before and after PSM are summarized in Table 1. Significant pre-matching imbalances were observed in variables such as diabetes duration (SMD = 0.17), BMI (SMD = 0.26), and albumin (SMD = 0.20). Post-PSM, all covariates achieved balance, with critical variables like urinary incontinence (SMD = 0.05) and age (SMD = 0.06) demonstrating equitable distribution. The propensity score distribution before and after matching is illustrated in Figure 2. Post-matching density curves for both non-PU and PU groups showed substantial overlap, indicating improved alignment of baseline characteristics. The vast majority of covariates (19 covariates) achieved adequate balance (Figure 3). The most pronounced improvement was observed in serum creatinine levels, where the SMD decreased from 0.35 (unmatched) to 0.06 (matched). All post-matching SMD values fell below the 0.1 threshold, confirming the robustness of the matching process in reducing confounding bias. The PSM data is available in the Supplementary Table S1.

We employed LASSO regression analysis on a dataset comprising 20 variables, utilizing a 10-fold cross-validation method to fine-tune the regularization parameter λ. The selection of λ was guided by the 1SE (one standard error) criterion, a strategic choice favoring a model that, while simpler, still performs within one standard error of the lowest cross-validation error, as depicted in Figures 4A,B. Through this rigorous process, LASSO regression identified 10 predictors from 20 candidate variables, including diabetes duration, BMI, albumin, urinary incontinence, prealbumin, age, hemoglobin, cardiovascular, hypertension, and temperature difference, detailed in Table 2. Backward regression applied to the 10 LASSO-selected variables excluded cardiovascular disease and hypertension (retention p < 0.05), yielding 8 predictors for the final nomogram (Figure 5). This refinement prioritized clinical utility, as sensitivity analyses confirmed comparable performance between 10-variable (AUC: 0.824) and 8-variable models (AUC: 0.825). Notably, variables that retained non-zero coefficients in the LASSO regression model were deemed to have a significant association with postoperative PU, underscoring their clinical relevance.

In this comprehensive research endeavor, we employed backward regression as a methodological framework to discern key factors associated with the occurrence of PU. Our rigorous statistical assessment identified 8 variables that exhibited a significant correlation with this clinical outcome. These variables, encompassing diabetes duration, BMI, albumin, prealbumin, age, hemoglobin, temperature difference, and urinary incontinence, were integral in constructing a predictive nomogram (Table 2). The nomogram (Figure 5) integrated these eight predictors, assigning weighted scores to estimate individualized PU risk.

The model demonstrated excellent discrimination in the training set (AUC-ROC = 0.825, 95% CI: 0.797–0.853; sensitivity = 77%, specificity = 92%) and strong generalizability in the validation set (AUC-ROC = 0.800, 95% CI: 0.753–0.847; sensitivity = 76%, specificity = 92%) (Figure 6). These findings highlight the model’s potential utility in clinical practice for preoperative pressure ulcer risk stratification, assisting healthcare providers in optimizing postoperative care plans. Internal validation via bootstrap resampling (n = 1,000 iterations) yielded recalibrated C-index values of 0.833 (training) and 0.826 (validation) (Figures 7A,B), confirming model stability. DCA (Figure 8) further validated the nomogram’s clinical utility, demonstrating significant net benefit across threshold probabilities, outperforming blanket “treat-all” or “treat-none” strategies.