Evaluation of a Multilevel Program to Improve Clinician Adherence to Management Guidelines for Acute Ischemic Stroke

This quality improvement study evaluates the outcomes associated with a program to improve clinician adherence to management guidelines for acute ischemic stroke in hospitals in China.

Personnel adherence is expressed as a composite measure and an all-or-none score as co-primary outcomes based on the 12 KPIs. The composite measure is defined as the total number of eligible KPIs divided by total number of KPIs given to each eligible patient. The all-or-none score is defined as the proportion of eligible patients who received all of the KPIs. In addition, the percentage of each individual KPI was also calculated as the performed total number of patients divided by the total number of KPIs for eligible patients. The KPI of intravenous rt-PA was calculated as the total number of patients receiving rt-PA treatment divided by the total number of patients within 7 days of symptom onset, as the number of patients eligible for rt-PA treatment within the 4.5 h therapeutic time window was not available from participants.
The secondary outcomes were the distribution of modified Rankin Scale (mRS) score at discharge and the severe disability or death defined as mRS score 5-6 at discharge.

Statistical analysis
Baseline characteristics of hospitals and patients were compared between the intervention and non-intervention groups. Continuous variables were summarized as median (interquartile range), and categorical variables as frequency (percentage). The Wilcoxon rank-sum test was used for continuous variables and the Chi-square test for categorical variables. Means or medians for the weekly data of outcomes were calculated in the intervention and non-intervention groups.
The primary outcomes included the composite measures and the all-or-none scores, and the secondary outcome was the rate of severe disability or death defined as mRS score 5-6 at discharge. All individual KPIs were also analyzed.
Interrupted time series analysis (ITS), a longitudinal quasi-experimental approach, was used in the intervention group to evaluate the effect of the intervention program on the KPIs. To further control time-varying confounders which may affect both intervention and non-intervention groups, a controlled, interrupted time series analysis (CITS) was performed by recruiting data from the non-intervention group. To control for biases at the baseline level and trend, data from a defined pre-intervention period were included. Due to the consideration that the effects of interventions may take time to manifest, the post-intervention period was divided into short-term and long-term periods. The rapid change of rates in KPIs was observed in the short-term of the intervention and the gradual change in KPIs was observed over the long-term period, which were defined as slope change and level change, respectively. Segmented linear regression models were used to estimate the changes in levels and trends of rates after the implementation of the intervention. The regression model is as follows: Y t = β 0 + β 1 × T + β 2 × Int + β 3 × Post + β 4 × G + β 5 × G × T + β 6 × G × Int + β 7 × G × Post + ε, Yt is the outcome variable at time t for the segmented linear regression model. T is a variable representing the time since the start of the study. Int (Intervention) is a binary variable indicating pre-(Int=0) or post-(Int=1) intervention period. Post is a continuous variable representing the time after the intervention and is coded 0 during the pre-intervention period. G represents the intervention (G=1) or non-intervention (G=0) group. β0-β3 are parameter estimations for the non-intervention group, where β0 represents the baseline level at the beginning of the time series, β1 is the underlying pre-intervention trend, β2 is the level change and β3 is the slope change following the intervention. β4-β7 represents the differences between the intervention and non-intervention group. β4 is the difference in baseline levels; β5, the slope difference in the pre-intervention period; β6, the difference in level changes; and β7, the difference in slope changes. ε, random error. Cochrane-Orcutt autoregression would be used if we detected first-order autocorrelation in the data, and Durbin Watson statistics were close to the preferred value of 2.
In addition to the ITS model, difference-in-difference (DID) model was also used to analyze the differences in outcomes over the 2-time periods, pre-intervention and post-intervention. A standard DID regression model is as follow: Y = β 1 × baseline + β 2 × group + β 3 × (baseline x group) Y is the outcome of interest, i.e. composite or all-or-none scores, or individual KPI scores; baseline is a dummy variable for before or after intervention; and group is another dummy variable for intervention and nonintervention group. β 1 , β 2 , and β 3 are coefficients. β 3 is the target value that represents differential results of the interest outcomes between intervention and non-intervention group.
Nevertheless, a double robust regression model was used in this study, in which weights were added to balance potential unbalanced group distributions between intervention and non-intervention group, and additionally, those unbalanced characteristics of hospitals and patients as covariates were added. In brief, the multinomial propensity score weighting (mnps) function in the R twang package with propensity score weighting was used to adjust group distribution balance. Both hospital and patient characters were used in the mnps function to extract weights. Then the svyglm regression function was used to examine the effect of intervention on the outcomes, which was adjusted using both hospital and patient characters as covariates.
In more detail for the DID analyses, composite and all-or-none scores, as well as all individual KPI scores (percentages) were calculated per day based on each hospital. To account for selection bias, a propensity score using the hospital and patient characteristics was created. We used the twang package in R to estimate the propensity scores and weighting of the comparison cases to estimate the average treatment effect on the treated (ATT). Multiple imputation (five times) was performed for missing values of all covariates by using the R package named mice. Only NIHSS and LDL-C had missing data at 10.2% and 3.8%, respectively. The data were divided into 4 groups: intervention group and non-intervention group during pre-intervention period and post-intervention period. According to these 4 groups, mnps (multinomial propensity scores) was calculated using characteristics of all hospital and patients by the R package named Twang, in which the parameter version was xgboost. The balance results were evaluated, and those characteristics of hospitals and patients (hospital grade, stroke unit, neurologist available at emergency department, number of stroke physicians, intravenous thrombolysis per year, annual stroke admission and hypertension) were not well balanced were selected again for the later DID regression model. Hospital characteristics include annual stroke admissions number, whether it is a teaching hospital, with stroke unit or not, neurologist available at emergency department, its grade being secondary or tertiary, bed numbers, neurological bed numbers, the capacity of performing endovascular therapy (EVT), stroke physicians number, intravenous thrombolysis treatment number, and EVT treatment number per year. Patient characteristics include age, sex, baseline NIHSS, LDL-C values, whether or not patients have hypertension, atrial fibrillation, diabetes, history of stroke, coronary heart disease, and smoking.

At discharge
Antithrombotics (  AIS patients who received anti-thrombotics at discharge/  the total number of AIS patients hospitalized in the same period) *100% Antihypertensive medication ((  AIS patients with hypertension who were given anti-hypertensive medication at discharge /  the total number of AIS patients with hypertension hospitalized in the same period) *100% Antidiabetic medication (  AIS patient with diabetes who were given hypoglycemic drugs at discharge/  the total number of AIS patients with diabetes hospitalized in the same period) *100% Lipid-lowering for LDL-C >100 mg/dL (  AIS patients with LDL-C >100 mg/dL, or treated with lipid lowering agent prior to admission, or LDL-C not documented who were given lipid lowering agent at discharge /  the total number of AIS patients with LDL-C >100 mg/dL, or treated with lipid lowering agent prior to admission, or LDL-C not documented hospitalized in the same period) *100%