We adopted a prospective study design with primary data collection. From September 2024 to June 2025, a convenience sampling method was used to recruit and survey patients with AIS admitted to the Comprehensive Treatment Center for Neurological Diseases at a tertiary hospital in China.

This study was conducted according to the Declaration of Helsinki. The study protocol was reviewed and approved by the Institutional Ethics Committee of the participating tertiary hospital (Approval No: 2024SA21). All participants provided informed consent before data collection and could withdraw at any time.

Before data collection, all research team members underwent standardized training. After the training was completed, clinical data collection began. Research participants were rigorously selected based on the study objectives. Patients were informed of the study purpose, procedures, and rights before data collection. All patients provided written informed consent. Researchers collected data using paper questionnaires designed for the study objectives. During patient visits, one-on-one administration of the paper questionnaires with standardized instructions was conducted in the ward, and patient responses were recorded. All completed paper questionnaires were subjected to double-entry data verification by two independent researchers before being electronically input. Each questionnaire was individually reviewed for completeness and accuracy. The data were entered into Microsoft Excel 2019 by two researchers, followed by a final cross-check to ensure data authenticity and reliability.

To ensure that data management was rigorous, a series of quality control measures was implemented during the electronic data entry and storage process. Specific quality control measures included the following: ① Data entry performed by two independent researchers, with discrepancies resolved by consulting original paper questionnaires; ② Ensuring standardized entry to guarantee consistency, such as decimal places, thousand separators, numerical or percentage formats, and even alignment; ③ Access control: Electronic datasets were stored on password-protected hospital servers. Editing privileges were restricted to designated research team members, whereas supervisors had read-only access to prevent data tampering or modification. ④ All data modifications were tracked using the “Track Changes” feature in Excel, recording the editor, revision time, and details to ensure full traceability of alterations. These measures collectively protected the integrity and reliability of the data.

The data were analyzed using SPSS 26.0 statistical software. Continuous variables were first tested for normality using the Shapiro–Wilk test. Variables that followed a normal distribution were reported as the mean ± standard deviation. The differences were determined by conducting the independent samples t-test (between two groups) or one-way analysis of variance (ANOVA) (among multiple groups). Non-normally distributed continuous variables (e.g., NIHSS score and time from symptom onset to admission) were reported as medians and interquartile ranges [M (Q1, Q3)]. The differences were determined by conducting the Mann–Whitney U test (between two groups) or the Kruskal-Wallis H test (among multiple groups). Categorical variables were presented as frequencies and percentages [n (%)], and the data were compared using the χ² test or Fisher’s exact test (for the expected frequency <5). Univariate analysis was performed to initially screen potential predictors of prehospital delay. For continuous predictors (e.g., patient age and the NIHSS score), univariate logistic regression was performed to assess their associations with prehospital delay, and the results were reported as odds ratios (ORs) with 95% confidence intervals (CIs). For categorical predictors (e.g., education, history of cerebral infarction, place of onset, and first measures after onset), univariate logistic regression was also conducted after dummy variable coding, with ORs and 95% CIs reported to quantify the strength of association.

Multivariate binary logistic regression (enter method) was then performed to identify independent predictors for prehospital delay. All variables with p < 0.10 from the univariate analysis were entered into the initial model. The outcome was coded as 1 for prehospital delay (symptom-to-hospital time >4.5 h) and 0 for non-prehospital delay (symptom-to-hospital time ≤4.5 h). The results are presented as adjusted ORs with 95% CIs. A risk prediction model using a nomogram was constructed using the R 4.5.1 software and the loaded rms package.

The specific methods for constructing the nomograms were as follows: first, the unstandardized regression coefficients (β) of all independent predictors in the final logistic model were extracted; second, these β coefficients were scaled to a common range to facilitate graphical representation of a unified point system; third, the scaled coefficients were converted into a user-friendly point scale for each predictor level, with the reference category set at 0 points; and finally, the total point score (sum of individual predictor points) was mapped to the predicted probability of prehospital delay using the inverse logistic function. Then, the nomogram was visualized with axes for each predictor, a total point axis, and a predicted probability axis.

Internal validation of the prediction model was performed using the bootstrap resampling method with 1,000 repetitions. The procedure involved repeatedly drawing random samples with replacement from the original dataset to create multiple bootstrap samples of the same size. The model was rebuilt on each bootstrap sample and then tested on the original sample to estimate its optimism, i.e., overfitting. Model performance was evaluated based on three key metrics: the concordance index (C-index) for discrimination, the calibration curve for calibration, and decision curve analysis (DCA) for clinical utility.

The calibration of a model serves as a key indicator for evaluating the accuracy of a disease risk model while predicting the probability of an individual experiencing an outcome event in the future. It reflects the degree of consistency between the risk predicted by the model and the actual occurrence of risk. The Hosmer–Lemeshow test was conducted to validate the model’s goodness of fit.

The discriminatory ability of a model is assessed by plotting a receiver operating characteristic (ROC) curve and evaluating the AUC. A larger AUC indicates better predictive discrimination capability. Generally, an AUC < 0.6 indicates poor discriminatory ability, 0.6–0.75 indicates moderate discriminatory ability, and >0.75 indicates good discriminatory ability.

Decision curve analysis (DCA) was conducted to evaluate the clinical utility of the predictive model. By comparing the model against the “All” (complete treatment) and “None” (no treatment) reference strategies, the DCA visually illustrates the net benefit of the model across different threshold probabilities. If the model curve consistently demonstrates a net benefit superior to both reference strategies within a specific threshold probability range, it indicates that using the model to guide clinical decisions within that range yields greater net benefit and holds practical clinical value. In contrast, if the model curve approaches the reference strategy curves within a certain threshold range (showing no significant net benefit advantage), it suggests limited clinical utility for the model in that range. This outcome provides a direct reference for clinicians in the formulation of individualized intervention strategies. Model predictive performance was determined to confirm sensitivity, specificity, and accuracy, with p < 0.05 indicating statistically significant differences.

A total of 348 inpatients were included in this study, including 231 males (66.4%) and 117 females (33.6%), with an age of 68.1 ± 12.2 years. Among all participants, 231 participants (66.4%) received primarily school-level education or below. There were 240 urban residents (69.0%) and 94 rural residents (27.0%). There were 137 participants (39.4%) with a history of smoking and 112 participants (32.2%) with a history of alcohol consumption. There were 254 participants (73.0%) with a history of hypertension. A total of 258 patients (74.1%) themselves discovered the initial symptoms of the disease. There were 277 cases (79.6%) in which the individuals were not aware that a stroke had occurred after the onset of symptoms. A total of 115 patients (33.0%) had never undergone a physical examination before, whereas 149 patients (42.8%) first chose to wait and observe their symptoms after the onset of illness.

Among the 348 hospitalized patients, 281 had prehospital delay, leading to an incidence rate of prehospital delay of 80.8%. The duration from patients’ self-perceived onset of illness to obtaining effective treatment at a medical institution was 33.7 ± 3.0 h, and the median duration was 24 h. Among these patients, 10 patients (2.9%) arrived at the hospital within 4.5–6 h, 134 patients (38.5%) arrived within 6–24 h, and 136 patients (39.1%) arrived after more than 24 h.

Univariate analysis of the prompt care group and delayed care group revealed that educational attainment, place of residence, monthly income, history of cerebral infarction, history of atrial fibrillation, Place of onset, initial medical facility, awareness of stroke onset after symptom onset, Subsequent measures, transport, distance from initial hospital, NIHSS score, PSSS score, and PBHSD-C score were factors influencing prehospital delay seeking among patients with acute ischemic stroke. Specific details regarding univariate variance information for patients with AIS experiencing prehospital delay are presented in Table 1.

Statistically significant factors from univariate analysis were used as independent variables, and binary logistic regression was performed, where prehospital delay (0 = no delay, 1 = delay) served as the dependent variable. The coding details are presented in the Supplementary material. The results of the binary logistic regression are presented in Table 2. Education level (specifically, high school/vocational school level), a history of cerebral infarction, subsequent measures, and the NIHSS score were identified as independent predictors of prehospital delay (all p < 0.05). The variable “place of onset” was retained in the final model because of its a priori clinical relevance for prehospital delay, although it did not reach statistical significance (p = 0.100), as detailed in Table 2.

Based on the results of binary logistic regression analysis, a predictive model for the risk of prehospital delay in AIS patients was established using educational attainment, history of cerebral infarction, location of onset, initial measures taken after onset, and the NIHSS score as predictors. The prediction formula was as follows: Logit(P) = 0.97–0.31 × Junior high school−1.46 × High school/vocational school−1.19 × Bachelor’s degree or higher−0.91 × History of Cerebral Infarction−0.90 × Workplace−0.73 × Outdoors+1.34 × Immediately contacted family/friends to go to hospital+3.52 × Waited and observed symptom changes-0.12 × NIHSS score. A scorer plot is illustrated in Figure 1. Each predictor in the figure corresponds to a score on the upper scale. By summing all the factors, the total score was derived. The vertical line extending downward from the total score axis reflected the predicted probability of delayed presentation in AIS patients. Internal validation through 100 bootstrap resamples revealed that the optimal cutoff value was 0.625, with a sensitivity of 0.79 and a specificity of 0.84, suggesting that the model was well-fitted. The area under the ROC curve was 0.87 (95% CI, 0.82–0.91), which indicated that the constructed predictive model for prehospital delay in AIS patients possessed clinical diagnostic significance, as illustrated in Figure 2.

Calibration curve: This curve primarily evaluates the calibration performance of the predictive model, measuring the consistency between the model’s “predicted probability of prehospital delay” and the “actual probability of delayed medical care occurrence.” The X-axis represents the probability predicted by the regression model for prehospital delay among AIS patients. In contrast, the Y-axis represents the actual probability of prehospital delay among AIS patients. The solid line (Apparent) closely aligns with the ideal 45° dashed line across the entire range of predicted probabilities, indicating strong agreement between the predicted and observed outcomes. The Hosmer–Lemeshow test yielded a χ² value of 9.835 (p = 0.277), further confirming good model fit, as shown in Figure 3.

The clinical utility of risk prediction models needs to be evaluated by conducting DCA to determine whether the model can deliver practical net benefits to clinical practice. DCA compares the constructed model curve with two extreme reference strategy curves (Figure 4): “All” (assuming that all AIS patients are at high risk of prehospital delay and implementing intervention) and “None” (assuming that no patients are at high risk and that no intervention is performed). The X-axis of the curve represents the risk threshold probability (the minimum probability of delayed medical care that clinicians consider worthy of intervention). In contrast, the Y-axis represents the net clinical benefit (balancing the benefits of correctly identifying high-risk patients against the potential harms of unnecessary interventions for low-risk patients).

In this study, the optimal cutoff value derived from the ROC curve was integrated to maximize the overall net clinical benefit. When the risk threshold was set at 0.625, 63 out of every 100 AIS patients could avoid prehospital delay via the predictive model. According to this decision curve, when the model threshold ranged from 0 to 75%, the curve consistently and persistently was above and to the right of both the “All” and “None” reference curves. These findings indicate that the constructed risk prediction model yields sustained positive net clinical benefits and has significant clinical utility.

We enrolled 348 patients with AIS, of whom 281 had prehospital delay, resulting in a delay rate of 80.8%. This finding is comparable to the results of a multicenter study conducted in China, which reported a prehospital delay rate of 82.3% for AIS patients (30); this prehospital delay rate was higher than that (73.3%) reported in AIS patients in South India (31). This might be associated with disparities in regional healthcare resources, cultural characteristics of the population, study design, definitions of prehospital delay, and inclusion and exclusion criteria. A cross-sectional study conducted in Thailand reported a 70.0% delay rate within the first 3 h for patients with AIS (32). A multicenter study conducted in South Korea revealed that only 36.8% of patients with AIS reached the hospital within 4.5 h after symptom onset to undergo intravenous thrombolysis, leading to a delay in seeking medical care for 63.2% of patients (33). In the US, about 50% of AIS patients reach the hospital within 4.5 h, whereas in Spain, the prehospital delay rate for cases within 3 h is 48.7% (6). Although the rates of delayed medical care vary across different countries and regions, they remain consistently high overall. This situation highlights the urgent global need to prevent and control delayed care for AIS patients, with nations still facing significant challenges in this area. Despite an increase in public awareness of cerebrovascular diseases as societies develop, a lack of relevant health consciousness persists among the general population. As key agents in health education and emergency care, nursing staff play an irreplaceable role in this context. The findings of this study indicate a high rate of prehospital delay for AIS patients, suggesting that nurses need further training to strengthen targeted health education for patients with AIS. Therefore, healthcare authorities in all countries should gain a thorough understanding of the situation in their respective regions, collaborate with nursing staff, and develop prevention and control strategies tailored to regional characteristics to increase the level of care for AIS patients.

The AIS delayed prehospital delay prediction model developed in this study demonstrated good predictive performance, featuring excellent discrimination and calibration. Clinicians can use this model as an objective and convenient risk stratification tool that can accurately classify patients with AIS. Generally, an AUC value between 0.7 and 0.9 indicates good discrimination. In this study, the area under the ROC curve was 0.87 (95% CI: 0.82–0.91), accompanied by a sensitivity of 0.79 and a specificity of 0.84, demonstrating robust predictive ability. The Hosmer–Lemeshow goodness-of-fit test resulted in a p-value of 0.277 (>0.05), which confirmed the high consistency between the predicted delay risk and the actual probability of occurrence. The calibration curve approached the ideal curve, indicating that the model had high accuracy. DCA revealed a positive net clinical benefit for patients using this predictive model, with thresholds ranging from 0 to 75%. The included predictors (level of education, history of cerebral infarction, Place of onset, Subsequent measures, and NIHSS score) are readily obtainable by nurses during clinical practice without requiring complex equipment or specialized tests. This highlights the need for clinical nurses to increase their vigilance toward such high-risk populations, calculate the probability of prehospital delay, and implement effective preventive and stratified interventions based on risk assessment. Thus, the risk prediction model developed in this study is practical and feasible, providing a reference for nursing professionals in other countries to assess high-risk populations for AIS.