We retrospectively collected Patients with AIS between January 2022 and December 2024 at the Affiliated Hospital of Xuzhou Medical University. The study was conducted in accordance with the Declaration of Helsinki and approved by the Ethics Committee of the Affiliated Hospital of Xuzhou Medical University. Inclusion criteria:(1) diagnosed with AIS according to World Health Organization criteria with complete baseline and follow-up data; (2) confirmed large vessel occlusion by imaging and treated with MT; and (3) time from onset to start of thrombolysis was <6 h, which could be extended to 16 h or 24 h by meeting the inclusion criteria for the DEFUSE 3, DAWN, or BAOCHE studies (14–16); (4) Age > 18 years. Exclusion criteria: (1) CT of the head indicates intracranial hemorrhage or other conditions (e.g., intracranial aneurysm) associated with high bleeding risk that would contraindicate thrombectomy; (2) severe allergy to contrast media; (3) pregnant and breastfeeding females; (4) patients with serious diseases, such as renal failure, severe hepatic insufficiency, or cancer; (5) patients with autoimmune diseases; (6) patients with immunosuppressive drugs or antibiotic use; (5) 90d postoperative loss of visit.

We collected patients’ general information and medical history (including age, sex, diabetes mellitus, hypertension, and atrial fibrillation), preoperative National Institutes of Health Stroke Scale (NIHSS) scores, modified Rankin Scale (mRS) scores, and 90-day postoperative mRS scores. Stroke characteristics [Trial of Org 10,172 in Acute Stroke Treatment (TOAST) classification and site of onset] and procedural details (onset-to-puncture time, procedure duration, number of passes, thrombectomy method, pre-procedure intravenous thrombolysis, stent use, and the grade of revascularization after the procedure [assessed by the Thrombolysis In Cerebral Infarction (TICI) grading]) were recorded. The TICI grading is used to assess the degree of cerebrovascular recanalization, primarily evaluating post-procedural blood flow restoration following endovascular treatment for AIS. Laboratory indices (platelet, neutrophil, and lymphocyte counts) were collected from peripheral blood samples obtained immediately upon the patient’s admission, prior to any endovascular intervention. The SII was calculated as (platelet count × neutrophil count)/lymphocyte count, and we also calculated the PLR which was calculated as platelet count/lymphocyte count and NLR which was calculated as neutrophil count/lymphocyte count.

Patients or their family members were followed up by telephone or outpatient clinic at 90d postoperatively, and the functional outcome was assessed according to the mRS score. mRS ≤ 2 points were included in the good functional outcome group, and mRS > 2 points were included in the bad functional outcome group, and the data of the two groups were statistically analyzed to explore the effects of NLR, PLR, and SII levels on the patients’ poor functional outcome at 90d postoperatively after MT.

All the measured data were statistically described by mean ± standard deviation to satisfy normality, and the t-test was used for the differences between the two groups; non-normal distribution was statistically described by median (percentile), and the Wilcoxon’s rank sum test was used for the differences between the two groups. The statistical description of the count data was based on the number of cases (%), and the differences between groups were tested by X² or Fisher’s exact probability test. The Wilcoxon rank sum test was used for differences between the two groups for grade information. Based on the results of intergroup variability, in order to study the analysis of risk factors related to the functional outcome of AIS patients after MT, a one-way logistic regression model was used to analyze the statistically significant differences between good and poor functional outcome of AIS patients after MT, and based on these results, we identified key clinical predictors and constructed logistic regression models. The baseline clinical model included onset-to-puncture time, procedure duration, and admission NIHSS. We then built three additional models by adding each inflammatory index to the baseline variables: Model 2 = baseline + SII, Model 3 = baseline + PLR, and Model 4 = baseline + NLR. Model performance was evaluated by ROC curves, and the optimal model was selected using NRI and IDI analyses, as well as calibration curves and DCA. The area under the ROC curve (AUC) was used to predict the performance of the models, and the optimal prediction model was constructed by using the NRI and IDI analyses, and the area under the ROC curve (AUC) was used to evaluate the value of the different prediction models and the performance of the models for predicting the clinical outcomes of patients with AIS after mechanical thrombus extraction. The calibration curves and the Hosmer–Lemeshow test were used to evaluate model fit, while internal validation was performed using bootstrap resampling with 1,000 repetitions to ensure the robustness of the primary model. Model fit was assessed using calibration curves and the Hosmer–Lemeshow test. To evaluate the internal validity and correct for potential overfitting, internal validation was performed using bootstrap resampling with 1,000 repetitions. This generated an optimism-corrected estimate of model performance (e.g., AUC), providing a more realistic assessment of the model’s generalizability within the study cohort. Finally, decision curve analysis (DCA) was employed to assess the clinical value of the models. To further assess the predictive efficacy of SII, PLR and NLR on the functional outcome of patients with AIS after MT, ROC curves were analyzed using unadjusted analysis and multivariable-adjusted ROC analysis to account for potential confounders. p < 0.05 was considered statistically significant. R4.3.2 software was used for basic statistical analysis, and Python3.9 software was used for SHapley Additive exPlanations (SHAP) visualization.

Based on the inclusion and exclusion criteria, 387 cases were extracted from hospital information system (HIS) database of AIS in our hospital, out of which 236 cases occurred with poor functional outcome and the incidence of poor functional outcome was 60.98%. Table 1 compares baseline characteristics of the two outcome groups. Compared to patients with good functional outcome, those with poor functional outcome had significantly longer onset-to-puncture and procedure durations, higher NIHSS scores at reperfusion, and higher SII, PLR, and NLR values (all p < 0.05).

Since the present study found statistically significant differences in the distribution of demographic and clinical indexes such as onset-to-puncture time (min), operating time (min), and NIHSS before MT between the two groups, in order to better assess the effectiveness of SII, PLR, and NLR in predicting clinical outcomes, the present study plotted the ROC curves of multivariable-adjusted ROC analysis and uncorrected covariates to illustrate in depth the predictive value of each comprehensive inflammation index. As shown in Figure 1, SIl had the greatest predictive efficacy, with an AUC = 0.834 (95% CI = 0.790–0.874) for uncorrected confounders and 0.841 (95% CI = 0.841–0.844) for multivariable-adjusted ROC analysis NLR was the next most effective predictor with an AUC = 0.820 (95% CI = 0.774–0.859) for uncorrected covariates; and NLR, with an AUC = 0.820 (95% CI = 0.774–0.859) for uncorrected covariates, and after correcting for other influences its AUC = 0.818 (95% CI = 0.818–0.821); and finally PLR, with an AUC = 0.800 (95% CI = 0.754–0.843) uncorrected for confounders, and after correcting for other influences its AUC = 0.810 (95% C1 = 0.809–0.812) (see Figure 1). The area under the ROC curve based on the above indicators was greater than 0.7, proving that the indicators SII, PLR, and NLR all have some predictive value.

We performed multivariate logistic regression with poor functional outcome as the dependent variable. The baseline model (Model 1) included the key clinical predictors (onset-to-puncture time, procedure time, and NIHSS at reperfusion). We then built three additional models by adding each inflammatory marker: Model 2 = baseline + SII, Model 3 = baseline + PLR, and Model 4 = baseline + NLR. As shown in Tables 2–4 and Figure 2, the model including SII (Model 2) showed the best predictive performance. Although our primary model is based on logistic regression—a standard and highly interpretable method for clinical risk prediction—its interpretability was further enhanced by employing SHAP analysis. This approach allowed us to move beyond traditional regression coefficients, providing a detailed visualization of how each predictor contributes to individual patient risk estimates.

After the Delong test, it was found that the difference in the area of the ROC curve of Model 2 was statistically significant (p < 0.05) compared with that of Model 1 and Model 3, and the difference was not statistically significant (p > 0.05) compared with that of Model 4, and the difference in the area of the ROC curve of Model 2 was statistically significant (p > 0.05). However, after NRI and IDI analyses, it was found that Model 4, when compared with Model 2, had an NRI (categorical) of −0.1568 [(95% CI: −0.2383 to −0.0753); p < 0.001]; an NRI (continuous) of −0.2662 [(95% CI: −0.468 to 0.0643); p = 0.010]; an IDI of −0.0247 [(95% CI: −0.0484 to −0.001); p = 0.041; and an IDI of −0.041. 0.0643; p = 0.010]; IDI was −0.0247 [(95% CI: −0.0484 to −0.001); p = 0.041], which shows that Model 2 has better comprehensive performance than Model 4 and is the optimal prediction model. The result is the formula for the optimal prediction equation: prognostic bad = −3.576 + 0.001 * OnsettoPunctureTime + 0.007 * procedure time + 0.031 * nihss0 + 0.002 * SII.

And, OnsettoPunctureTime, procedure time, nihss0 and SII were prognostic risk factors (OR all >1, p < 0.05). For details, see Table 5.

the optimal prediction model in this study was a multifactorial logistic regression model constructed based on the SII index, with an area under the ROC curve AUC = 0.954 (95% CI: 0.925–0.983), a sensitivity of 0.839 when the model cutoff = 0.497, a specificity of 0.755, PPV of 0.843, NPV of 0.750, and accuracy of 0.806 (95% CI: 0.763–0.844), as detailed in Figure 3. The model was analyzed for goodness-of-fit, and it was found that the model had a good ability to match between the predicted risk and the actual results (Hosmer–Lemeshow test, X² = 4.899, p = 0.265 > 0.05, Figure 4).

The relative importance and impact of variables in the model in predicting postoperative functional outcome were analyzed by SHAP values, and it can be seen in Figure 5 that different variable characteristics have different impacts on prognostic prediction. In the present study, we found that SII, operative duration (min), nihss before MT, and onset-to-puncture time (min) were associated with an increased risk of poor functional outcome, and the higher the value, the higher the probability of poor functional outcome. In addition, we can find that the characteristics of different variables affect the model differently, with SII having the highest influence on the predictive model, accounting for 74.2% of the total model, followed by the length of the procedure (min) and the nihss prior to the removal of the embolus, accounting for 9.6 and 9.3% of the total model, and lastly, the time from onset of the procedure to the puncture (min), at 6.9%.

Figure 6 shows the decision curve analysis. For threshold probabilities (Pt) ≥ 15%, the predictive model yields a higher net benefit than the ‘treat-all’ or ‘treat-none’ strategies. For example, at Pt = 15%, the model would correctly identify approximately 15 additional poor-outcome patients per 100 screened without increasing the false-positive rate. Within the Pt range of 0.15–0.90, the model provides a consistently high net benefit for predicting poor functional outcome.