Aim
To provide patients with a reliable tool to predict the risk of AR based on clinical data, and a score model for predicting AR risk was constructed.
Study design and participants
Data selection and analysis were performed using a retrospective cross-sectional design. The data were collected between January 2021 and April 2023.
The study participants were patients who experienced ischemic stroke and were admitted to the Department of Neurology and Stroke at the Affiliated Hospital of Beihua University in the Jilin Province between January 2021 and April 2023. The governing principle involved at least 10 times the number of logistically independent variables [24]. The incidence of AR in ischemic stroke can reach 65% [25], so the sample size should be increased by 20% and not less than (number of independent variables×10)/incidence. This study included a retrospective analysis of the demographic and laboratory examination characteristics of 285 patients. The hospital ethics committee reviewed and approved this study (BHFS-2021-009). Patients who met the following criteria were eligible for inclusion: (1) had an ischemic stroke diagnosed by head magnetic resonance imaging or computed tomography [26]; (2) had taken aspirin regularly, with a daily loading dose of 100 mg and a continuous duration of at least 7 days; and (3) had undergone thromboelastography [27]. The exclusion criteria were as follows: (1) use of nonsteroidal anti-inflammatory drugs, anticoagulants, or additional antiplatelet medications within 4 weeks of enrollment; (2) history of surgery within 1 week before admission; and (3) blood disease or bleeding tendency.
Data grouping
AR group: According to the inpatient medical records, the medication needed to be changed, the results of the AR risk genetic test were attenuated, or the AA inhibition rate was above 50%. Aspirin non-resistant group: Inpatient records demonstrated a moderate or good response to AR risk genetic testing and aspirin intolerance or AA inhibition rate of 50% ≤ AA ≤ 100%. For 199 cases (70%) in the training group and 86 cases (30%) in the validation group, a random division into the prediction model and verification model effects was made based on a ratio of 7:3.
Candidate predictor variables
Clinical and laboratory variables such as sex, age, body mass index, smoking, hypertension (HTN), coronary artery disease (CAD), diabetes mellitus (DM), hyperlipidemia (HLP), history of stroke, and fibrinogen (FBI) were included in the data collection. Other variables included platelet count (PLT), fasting blood glucose (FBG), low-density lipoprotein cholesterol (LDH), homocysteine (HCY), and glycosylated hemoglobin (HbA1c). Information collected at the time of the initial diagnosis from computerized and printed medical records served as the baseline. Using standardized data collection and quality control procedures, trained staff retrospectively collected relevant information from the medical data platform of the Affiliated Hospital of Beihua University. General data and risk factor indicators of the participants were collected through a two-person verification process and arranged based on the contents of the collection table.
Risk scoring and stratification
The AR ischemic stroke risk prediction model was evaluated based on the value of the partial regression coefficient (β) of each variable in the binary logistic regression analysis. Two methods were used to establish the risk assessment criteria:
(1) The basic score × grade score method was used to determine the scoring criteria for AR risk based on the β value of each risk factor and its variable types. For a given risk factor, a point was assigned to the lowest β value, and the division value (integer) of the factor was used to determine the baseline score of other factors. If a risk factor was a dichotomous variable, it was assigned a value of 0 or 1.
(2) The risk factors were ranked based on the partial regression coefficient (β value) in the logistic regression model using the partial regression coefficient method β × 4 (rounded to an integer). The risk factor score was determined using β × 4 (rounded to an integer), with the reference category for each variable being 0. This was accomplished by considering previous research and the special circumstances of this study.
(3) Prediction and evaluation of the AR risk prediction model for ischemic stroke. The prediction accuracy = (true positive + true negative)/total number of cases was used to evaluate the predictive capacity of the model in the validation set. Two evaluation criteria were used to compare and examine the frequency of AR prediction.
Statistical analysis
SPSS software (version 24.0; IBM; Armonk, NY, USA) was used to analyze all data. The frequency, rate, and percentage were used to represent the counting data. The risk factors were analyzed using the chi-square test or Fisher’s exact probability method to compare the resistant and non-resistant groups. P-values less than 0.05 indicated statistical significance.
Through systematic sampling, the selected patients with ischemic stroke were randomly assigned to training and validation groups, with 70% (199/285) of patients in the training group and 30% (86/285) in the validation group. The training group data were selected to build a prediction model, and single-factor analysis was used to eliminate the risk factors associated with AR in patients with ischemic stroke. AR was considered the dependent variable, and the risk factors associated with AR were considered independent variables. This analysis focused on determining the likelihood of AR when multiple independent variables were present. The backward likelihood ratio method is a special method of logistic regression. After determining the β value of the partial regression coefficient of each risk factor, the inclusion criteria were set at 0 points05, the exclusion criteria were set at 0 points1, and the prediction model was constructed.
The receiver operating characteristic (ROC) curve was drawn. The validity of the identification was better when the area under the line was greater than 0.7. The calibration level of the model was tested using the Hosmer–Lemeshow (H–L) test, and P > 0.05 was considered a good calibration level.