We performed a retrospective analysis of patients with AIS-LVO who received EVT at Dongguan Hospital of Guangzhou University of Chinese Medicine (September 2018 to March 2025). The inclusion criteria were as follows: (1) confirmed LVO (internal carotid artery, M1/M2 middle cerebral artery, or basilar artery) by computed tomography angiography (CTA), magnetic resonance angiography (MRA), or digital subtraction angiography (DSA); (2) AIS diagnosis according to World Health Organization criteria with neuroimaging confirmation (computed tomography [CT] or magnetic resonance imaging [MRI]); (3) age ≥18 years; Exclusion criteria were: (1) pre-existing pulmonary or systemic infection (confirmed by microbiological or radiographic evidence); (2) severe comorbidities (e.g., end-stage organ failure, hematologic disorders); (3) immunocompromised status (HIV, immunosuppressant use, etc.); (4) modified Thrombolysis in Cerebral Infarction (mTICI) grade ≤ 2a without recanalization after EVT; (5) incomplete clinical data or loss to follow-up. After screening 650 patients, 412 met eligibility criteria and were subsequently allocated into a training cohort (n = 288) and a validation cohort (n = 124) in a 7:3 ratio. This research was approved by the Institutional Review Board of Dongguan Hospital (No. PJ[2025]88), with informed consent waived due to the retrospective use of anonymized data.

The HP was diagnosed within 7 days after EVT by two independent neurologists blinded to the predictors; any diagnostic discrepancies were arbitrated by a senior neurologist experienced in managing stroke complications. The diagnosis required simultaneous fulfillment of the following three criteria (8, 12, 13): (1) Clinical manifestations: ≥2 of cough, fever (>38 °C), purulent sputum, and deteriorated respiratory status, coupled with prolonged bed rest and reduced or absent sputum excretion; (2) Laboratory findings: Peripheral white blood cell (WBC) count >10 × 10⁹/L or <4 × 10⁹/L; (3) Radiographic evidence: Chest X-ray showing irregular, tiny patchy high-density shadows (with blurred edges and uneven density) in the lower lung fields.

Given the lack of unified diagnostic guidelines for HP, this study established the diagnostic framework and observation window with reference to key literature (8, 12, 13): a comprehensive approach combining clinical, radiological, and laboratory assessments was adopted, along with a 7-day observation window. This window reflects how HP develops—specifically, HP is a secondary infection driven by pulmonary secretion accumulation from post-EVT immobilization.

We identified 28 predictor variables for HP in EVT patients through literature retrieval, and these variables were then systematically extracted from electronic medical records. Meanwhile, multicollinearity tests were performed on these variables using SPSS software. The results showed that the variance inflation factor (VIF) of all variables was less than 10 (Supplementary Material 2), indicating no significant multicollinearity among the variables. These variables included: demographics (age, gender), medical history (hypertension, diabetes, atrial fibrillation, prior stroke, smoking, drinking), clinical status [dysphagia, postoperative 48 h fever, admission sbp, admission dbp, admission NIHSS (14), admission GCS, admission mRS (1, 15)], laboratory indicators [postoperative 24 h CRP, postoperative 24 h NLR (16), postoperative 24 h SII, postoperative 24 h SIRI (17), postoperative 24 h PLR, postoperative 24 h TyG, postoperative 24 h SHR], imaging/etiologic data [ASPECTS (18, 19), TOAST classification, fazekas scale, infarction site], and procedure-related parameters (operation time, symptom onset to puncture time).

The SPSS software (version 27.0), R software (version 4.2.3), and Origin 2024 were the primary tools utilized for data analysis and visualization. To reduce selection bias and ensure baseline balance between the training and validation cohorts, propensity score matching (PSM) was first conducted to adjust for confounders. Descriptive statistical analysis was then performed for 412 participants after PSM.

Continuous variables were presented as mean ± SD (normal distribution, independent samples t-test) or median (IQR) (non-normal distribution, Mann–Whitney U test). Categorical variables were reported as frequencies (%) and compared via chi-square test (expected frequency ≥5) or Fisher’s exact test (expected frequency <5). Two-sided p < 0.05 was statistically significant.

Missing values were present only for the stress hyperglycemia ratio (SHR) (21/412 cases, 5.10%), resulting in an overall missing rate <10%. Following confirmation of missing at random (MAR), multiple imputation (SPSS, 5 iterations) was used for handling. Sensitivity analysis confirmed model robustness, with consistent AUC (0.827 vs. 0.824) and no substantial differences in key metrics. Detailed comparisons are available in Supplementary Material 1.

Sample size was estimated using the Events Per Variable (EPV) principle for binary logistic regression-based predictive models. To construct a nomogram predicting HP in AIS-LVO patients after EVT (HP as the outcome), key parameters were set as follows: 4–6 expected independent variables, a minimum EPV of 10, an anticipated HP incidence of 35%, and a 70% training set proportion. This yielded a minimum required training set size of ~171.4 cases, translating to a total sample size of ~245 cases (before accounting for invalid samples). Considering a 10% reserve for invalid data, the final minimum total sample size was 273 cases. A total of 412 eligible patients were enrolled, meeting the sample size requirement.

Multicollinearity among all candidate predictor variables was diagnosed using the VIF. LASSO regression was applied for feature selection, with the optimal penalty parameter (lambda, λ₁se = 0.053) determined as the value that corresponds to the cross-validation error within one standard deviation of the minimum error. The variables selected by LASSO were incorporated into a multivariable logistic regression model, and multicollinearity was re-evaluated to ensure the stability of parameter estimates. For model performance assessment, discrimination was evaluated using the Receiver Operating Characteristic (ROC) curve and the Area Under the Curve (AUC); goodness-of-fit and calibration were comprehensively assessed via the Hosmer-Lemeshow test, calibration curves, and Brier score; additionally, Decision Curve Analysis (DCA) and Clinical Impact Curve (CIC) were utilized to evaluate the clinical net benefit across different decision thresholds, thereby verifying the model’s clinical utility.

Figure 1 presents the participant selection flowchart. From an initial cohort of 650 patients with AIS-LVO who underwent EVT, 238 were excluded based on predefined criteria (detailed in Methods). The final analytical cohort comprised 412 patients, who were allocated to either the training cohort (n = 288) or validation cohort (n = 124) in a 7:3 ratio.

To eliminate potential selection bias and ensure the comparability of baseline data between the training and validation cohorts, we first applied propensity score matching (PSM) to adjust for confounding factors. A total of 412 patients were recruited for this study, including 257 patients without HP (62.38%) and 155 patients with HP (37.62%). After PSM, the 412 patients were allocated into a training cohort (n = 288) and a validation cohort (n = 124) at a 7:3 ratio to ensure the reliability of subsequent model development and validation.

Table 1 presents the baseline characteristics of the two cohorts, including demographic data, clinical scores, laboratory indices, procedure-related variables, medical history, and clinical status. The “Statistic” and “P” columns show the results of differential analysis between the two cohorts. No statistically significant differences were observed in any variable (all p > 0.05), indicating that the training and validation cohorts were balanced and comparable, which guaranteed the validity of cross-validation for the prediction model.

LASSO regression was employed to screen 28 candidate variables, identifying predictive factors with non-zero coefficients (Figures 2A,B). Ten-fold cross-validation was performed to select the optimal λ value, which balanced model fit and parsimony. Ultimately, λ₁se (λ = 0.053) was determined as the optimal threshold, leading to the identification of 6 predictive variables: postoperative 24 h fever, dysphagia, admission mRS, admission GCS, ASPECTS, and postoperative 24 h NLR.

Subsequently, these 6 LASSO-selected variables were subjected to multivariate logistic regression analysis, and four meaningful independent predictors were retained for the final model: postoperative 48 h fever, admission GCS, ASPECTS, and postoperative 24 h NLR. The model was constructed with hypostatic pneumonia (HP) occurrence as the dependent variable, targeting AIS-LVO patients who underwent EVT. Results showed that postoperative 48 h fever (OR = 2.77, p < 0.001) and postoperative 24 h NLR (OR = 1.15, p < 0.001) were independent risk factors for HP, while admission GCS (OR = 0.77, p < 0.001) and ASPECTS (OR = 0.74, p < 0.001) served as independent protective factors. Specifically, patients with postoperative 48 h fever or elevated postoperative 24 h NLR exhibited a higher risk of developing HP, whereas those with higher admission GCS scores or higher ASPECTS scores had a lower risk, as illustrated in Figure 3.

Based on the results of multivariate logistic regression analysis, we constructed a nomogram for patients with HP after EVT (Figure 4A). Each specific value of the included variables corresponds to a respective score. After summing the scores of all variables to obtain a total score, a vertical line is drawn from this total score to estimate the probability of HP occurrence. To facilitate clinical use by physicians and patients, we have developed a web-based tool for this nomogram (Figure 4B), which is publicly accessible at: https://StrokePredMod.shinyapps.io/dynnomapp/.

To comprehensively assess the model’s discriminatory and diagnostic performance, we report key metrics including the area under the receiver operating characteristic curve (AUC), sensitivity (SEN), specificity (SPE), positive predictive value (PPV), and negative predictive value (NPV) at prespecified thresholds (0.219 for the training cohort and 0.494 for the validation cohort), each with 95% confidence intervals (CIs). As summarized in Table 2, the model achieved an AUC of 0.829 (training set) and 0.817 (validation set), indicating favorable discriminatory ability. At their respective thresholds, the training set showed SEN = 0.796, SPE = 0.706, PPV = 0.639, and NPV = 0.838; the validation set exhibited SEN = 0.638, SPE = 0.922, PPV = 0.802, and NPV = 0.802, all with corresponding 95% CIs. These results collectively demonstrate the model’s consistent and robust diagnostic performance across both cohorts (Figure 5).

To address model overfitting, this study employed the bootstrap optimism correction method and evaluated model performance via 10-fold cross-validation: First, a model was built on the entire original dataset and its apparent performance was recorded. Then, 1,000 bootstrap samples were generated through random sampling with replacement from the original dataset. For each sample, a logistic model was constructed, and its performance was evaluated both on the bootstrap sample and the original dataset. The difference between these two performances was calculated as the “optimism.” After 1,000 iterations, the average optimism was obtained and subtracted from the apparent performance of the original model to derive the optimism-corrected performance. For 10-fold cross-validation, the dataset was randomly divided into 10 equally sized subsets. In each iteration, one subset was used as the test set, while the remaining 9 subsets were combined to form the training set for model construction and evaluation. This process was repeated 10 times to ensure each subset served as the test set exactly once. Finally, the performance metrics were averaged across the 10 folds to robustly estimate the model’s generalization performance and mitigate the risk of overfitting (Table 3).

To comprehensively validate the HP nomogram developed in this study, we systematically evaluated its performance across three critical dimensions: statistical calibration, clinical decision-making utility, and real-world clinical impact. Specifically, this section elaborates on: (1) the model’s ability to achieve agreement between predicted probabilities and actual HP incidence, as demonstrated by calibration curves (Figures 6A,B) and the Hosmer–Lemeshow test; (2) its net clinical benefit across a range of risk thresholds, illustrated via decision curve analysis (DCA, Figures 7A,B); and (3) its efficacy in stratifying and identifying high-risk patients, reflected by clinical impact curves (CIC, Figures 8A,B). Collectively, these findings confirm that the nomogram possesses reliable clinical utility in predicting postoperative HP following EVT in patients with AIS-LVO.