This retrospective, multicenter study was conducted across three hospitals in China (ChiCTR2300077202). The study flowchart was depicted in Figure 1. Patients eligible for inclusion in this study were diagnosed with AIS through either CT or MRI within 24 h of symptom onset and subsequently underwent EVT. Exclusion criteria comprised prestroke modified Rankin Scale (mRS) scores exceeding 2 and documented BP measurements exceeding more than 20% per hour within the initial 24 h. Finally, enrollment was carried out consecutively at 3 medical institutions in China: Beijing Hospital (277 cases from January 2016 to March 2023), Peking University Shenzhen Hospital (92 cases from January 2020 to March 2024), Hainan general hospital (82 cases from January 2021 to January 2023). The study was approved by Beijing Hospital Ethics Committee (2023BJYYEC-364-01).

A comprehensive collection of baseline clinical data was obtained from the hospital medical records system. The information collected includes basic demographic details, (age and sex), medical histories [diabetes, dyslipidemia, hypertension, atrial fibrillation (AF), coronary atherosclerotic heart disease (CHD)], and present use of antithrombotic drugs, NIHSS score at onset, occlusion site, and intravenous tissue-type plasminogen activator (IV-tPA) therapy. Recanalization status was assessed using a modified cerebral infarction thrombolysis (mTICI) grading system, where an mTICI score of 2b or 3 defined successful recanalization. The primary outcome variable that determined the patient’s 90-day functional status using the modified Rankin Scale (mRS). Good functional outcomes were defined by mRS Scores on a scale of 0 to 2, while poor outcomes were scored on a scale of 3 or higher. The mRS Scores were assessed by experienced neurologists, mainly during scheduled outpatient consultations. In cases where direct assessment was impractical, scores were obtained by telephone with the patient’s relative or caregiver.

The threshold for SBP follows guidelines, the specific target is determined by the physician performing EVT. The BP monitoring was measured hourly in the ward after endovascular thrombectomy (EVT) and recorded for at least the first 24 h after surgery. BP readings were systematically recorded using Philips equipment and sphygmomanometer cuffs and transmitted to the nurse station terminal via a manual/bedside non-invasive BP monitor. BP assessments at all three participating centers used a combination of manual and automated methods to ensure comprehensive and accurate readings. As SBP was the predominant risk factor in the elderly (15–17), the study mainly focused on SBP.

To further investigate the circadian rhythm of BP, nighttime and daytime BP recordings were distinguished. Nighttime BP was recorded between 10 pm and 4 am, while daytime BP was measured between 8 am and 6 pm. These indicators provide a comprehensive understanding of the key dynamic BP fluctuations during the first 24 h after EVT.

To reduce model complexity and the risk of overfitting, as well as optimize training speed, we performed least absolute shrinkage and selection operator (LASSO), support vector machine-recursive feature elimination (SVM-RFE), and Boruta to select potential features. SVM-RFE aims to find an optimal subset of features by iteratively removing the least important features based on their weights or rankings obtained from the SVM model, ensuring a global optimal solution (18). The Boruta algorithm provides an importance score for each feature, which helps to understand the relative importance between features (19). The intersection of the selected features from these methods was utilized as the model variables. We selected 15 features, including age, sex, BMI, hypertension, CHD, diabetes, dyslipidemia, AF, stroke history, smoking history, present antithrombotic therapy, IV-tPA, NIHSS on admission, mTICI grade and occluded site. In this study, cases from Beijing hospital (BJH) cohort were utilized for model development, while the Peking University Shenzhen hospital- Hainan general hospital (PKUSZ-HN) cohorts served as an external validation set. The aim of external validation was further assessed the generalizability and predictive efficacy of the selected models. In order to evaluate the optimal predictive performance and predictive effectiveness of different SBP parameters, we constructed nine learning models, including eXtreme Gradient Boosting (XGboost), Logistic regression (LR), Decision Tree (DT), Adaptive boost (AdaBoost), GaussianNB (GNB), Gradient Boosting Decision Tree (GBDT), Multi-layer Perceptron (MLP), Support Vector Machine (SVM), and K-Nearest Neighbor Machine (KNN). In addition, we report several parameters related to model performance in this study, including area under subject operating characteristic curve (AUC), sensitivity, specificity, and accuracy. After comparing the performance of nine different ML models, the final model was determined based on the highest AUC value.

To assess the normality of the data distribution, a Kolmogorov–Smirnov test was first conducted. Patients with missing SBP data exceeding four time points were excluded, while for other cases with missing SBP data, imputation was performed using the average of the 24-h SBP values. Descriptive statistics were used to summarize baseline characteristics and outcomes. Categorical variables were expressed in numbers and percentages, and chi-square tests were used to identify significant differences between risk factors and clinical outcomes. The normally distributed continuous variables were expressed as mean ± standard deviation (SD), and the comparison was performed using a T-test. All statistical analyses were performed using SPSS v25 (IBM Corporation, NY), and an α-level of 0.05 was adopted as the threshold for significance.

A total of 451 AIS patients aged above 18 underwent EVT treatment. Forty-four patients were excluded due to missing BP data or loss to follow-up, with 36 excluded from the development group and 8 from the external cohort. Ultimately, 407 patients were included in the analysis (Figure 1), comprising 241 in the development group and 166 in the validation group. Baseline characteristics of the patients were shown in Table 1. The median age of these patients was 70 years, with 253 males (62.2%). The median baseline NIHSS score was 14 (range: 9–19), with 129 (31.7%) receiving IV-tPA treatment before EVT. One hundred ninety-six patients (81.3%) and 132 patients (79.5%) achieved successful recanalization in the two cohorts, respectively. Finally, 182 patients (44.7%) had favorable outcomes. Missing values were imputed for BP readings (n = 27), BMI (n = 43), atrial fibrillation (n = 1), diabetes (n = 1), hypertension (n = 1), dyslipidemia (n = 1), smoking history (n = 1), IV-tPA (n = 1), coronary heart disease (n = 1), antithrombotic therapy (n = 18), and stroke history (n = 1). The average daytime and nighttime SBP values for the population before and after imputation were both 131 mmHg (SD: 14.5) and 127 mmHg (SD: 14.5), respectively. In the development group, the postoperative daytime, nighttime, and 24-h BPs were mm 133, 129, and 131 mmHg, respectively, while in the validation group, they were 128, 125, and 127 mmHg, respectively.

Three machine learning algorithms, LASSO, SVM-RFE, and Boruta, were performed to identify and select potential variables associated with the 3-month functional outcome (Figure 2). The LASSO analysis identified potential prognostics factors as NIHSS, mTICI grade, CHD, diabetes, hypertension, AF, IV-tPA, age and BMI, while SVM-RFE determined gender, IV-tPA, AF, hypertension, dyslipidemia, diabetes, stroke, CHD, antithrombotic therapy, occluded site, mTICI grade, NIHSS, and Boruta established NIHSS, mTICI, CHD, AF, hypertension. By selecting the variables overlapped among the three algorithms, the final variables for modeling were determined as NIHSS, mTICI, CHD, AF, hypertension as potential factors.

After determining the final 5 variables, we applied 9 machine learning algorithms to establish model and evaluated performance (Figure 3). The specific parameters of different machine learning methods were shown in Supplementary Table S1. The AUC of all algorithms increased to varying degrees after the addition of SBP-related parameters. The AUC of LR algorithm in the baseline model was 0.772, and after adding daytime BP, nighttime BP, and 24-h BP, the AUCs were 0.776, 0.795, and 0.785, respectively. The AUC of XGboost algorithm in the baseline model was 0.800, and AUC were 0.841, 0.840 and 0.838 after adding daytime BP, night BP and 24-h BP, respectively. The evaluation indicators of models were shown in Table 2. Among all algorithms, XGBoost with baseline plus nighttime SBP demonstrated the best performance in terms of AUC, with accuracy of 0.751, specificity of 0.863, and sensitivity of 0.751. Other evaluation metrics for the models are presented in Table 2.

To provide a better understanding of the effects of different models, further validation was conducted in the external validation set (Table 3). Similarly, after adding different BP parameters, the performance of all models improved in the validation set. The AUC of XGboost algorithm in baseline model is 0.742, 0.728, 0.752, 0.734, respectively, after increasing daytime BP, night BP and 24-h BP, respectively. For Adaboost and KNN algorithms in the baseline plus nighttime SBP model, the AUCs were 0.745 and 0.704, with accuracy, sensitivity, and specificity of 0.695, 0.500, 0.917, and 0.634, 0.870, 0.444, respectively. In summary, adding SBP levels in the predictive model established by 9 algorithms can improve AUC, with nighttime SBP showing the most significant improvement. XGBoost emerged as the best-performing algorithm.

To assessed the bias and clinical benefit of the different models, we applied the DCA curve and calibration curve to evaluate the models generated by nine algorithms (Figure 4). In the models of baseline and baseline + nighttime SBP, XGboost showed the lowest brier score of 0.203 and 0.198, respectively, demonstrating the robustness compared to the other algorithmic models. The logistic algorithm exhibited the highest robustness in the baseline + daytime SBP and baseline +24-h SBP models, with scores of 0.203 and 0.202, respectively, followed by the XGboost algorithm, with scores of 0.216 and 0.206, respectively.