A multicenter observational study of 3 years of monthly presentation data from six centers using time series analysis to measure the effect of the pandemic on TIA, ischemic stroke, IV thrombolysis (IVT), and endovascular thrombectomy (EVT) counts. This study was approved by the Monash Health Human Research Ethics Committee (RES-20-0000-915Q-72167).

The initial paper identified 16 clinics via a literature review of rapid TIA pathways in PubMed (9). Six of these centers agreed to provide 3 years of data. Three were from Australia, and one center each from England, Canada, and Italy, and these are listed in Table 1. The publication of each original rapid TIA pathway paper ranged from 2007 to 2019 (1, 2, 10–14). The pathways were primarily completed in the emergency department, with the exception of the primary care based OXVASC study (2), and the combined ED and primary care based Foothills Medical Center pathway (12). All accepted TIA referrals; Oxfordshire (2) also accepted minor stroke in addition to TIA. Monash Medical Center and John Radcliffe reported using ultrasound for vascular imaging pre-COVID (9) with the rest using CT angiography. All centers were utilizing CT angiography for vessel imaging during the COVID period (9). While each service used their own unique inclusion and exclusion criteria for the rapid TIA pathway, two went further by prioritizing time to review by mechanism (i.e., the presence of atrial fibrillation and/or >50% internal carotid artery stenosis) (1) or by timing of symptom onset and presence of other high-risk features (e.g., motor/speech deficit >5 min, ABCD2 ≥ 4, or atrial fibrillation) (12). All pathways commenced treatment immediately. The study duration included a baseline pre-COVID period (April 2018 to March 2020), and an intra-COVID period (April 2020 to March 2021), with the impact defined as the World Health Organization’s declaration of the pandemic on 11th March 2020 (15). Therefore, the first full intra-pandemic month was defined as April 2020. This changepoint was specified a priori to avoid exploratory analysis to discover a significant time point, as recommended by previous researchers (16). We did not include lockdown periods into the model given the heterogeneity in duration and intensity of stay-at-home orders. For example, Melbourne in Australia had multiple lockdowns, some only lasting for several days. We felt that this would not fit into a model of monthly counts.

Patients presenting from April 2018 to March 2021 with ischemic stroke or TIA were included. TIA presentations were recorded either as clinic attendances or hospital presentations and/or admissions. Case ascertainment was retrospective based on administrative coding. We acknowledge that this may introduce TIA mimics into our counts, but this was allowed given that our aim was to measure service activity despite the actual final diagnosis. The primary outcome measure was the monthly count of TIA, ischemic stroke, thrombolysis, and EVT occurrences (n/month). Time was measured in months. The proportion of reperfusion candidates was estimated as thrombolysis count or endovascular clot retrieval count as numerator, and ischemic stroke count as denominator.

Statistical analysis was performed in R (Statistical Programming Language, version 4.2.1) in two stages – (1) descriptive statistics, and (2) interrupted time series analysis (ITS). Firstly, comparing the pre-COVID and intra-COVID period: we used the Mann–Whitney U test (17) to compare mean monthly rates, and test of proportions compared percentage of stroke requiring IVT or EVT.

Secondly, Interrupted time series analysis was performed using the fpp3 package (18) to calculate β (absolute change in presentation rate), standard error, t value, and p-value using a previously published method (19). Interrupted time series analysis is a segmented regression approach to test an effect of an intervention on a time-dependent outcome using a defined impact model. A linear model was initially fitted, but if any residual autocorrelation existed [i.e., a linear relationship between lagged variables tested by portmanteau (Box Pierce) testing (18, 20)], autoregressive integrated moving average (ARIMA) errors were introduced (18). Therefore, time varying confounders were considered, and accounted for using time series regression techniques (21). ARIMA models were fitted with the Hyndman-Khandakar algorithm (22), which is a step-wise method of fitting time series data by (1) using unit root testing for stationarity (23), (2) applying a process called differencing if stationarity exists, and (3) automating model fitting based on the lowest possible AICc (corrected Akaike Information Criterion) value (22). Relative change from baseline was calculated as absolute change divided by the average baseline count per month.

A total of 6,231 TIA and 13,191 ischemic stroke presentations were included. Of the ischemic strokes, there were 1,664 (12.6%) instances of IVT and 2,145 (16.3%) instances of EVT. More than half of the study participants were from Australia (52.1%), followed by Canada (35.0%), Italy (7.6%) and England (5.4%).

Monthly counts of TIA and ischemic stroke presentations and instances of IVT and EVT are depicted in Figure 1 by country. The results of the univariable analysis are described in Table 2 and summarized next. When comparing the average monthly counts (n/month) for the pre-COVID and intra-COVID period, there was a higher rate of TIA presentations intra-COVID (171–176, p = 0.19), lower rate of ischemic stroke presentations (370 to 360, p = 0.36), lower rate of monthly IVT counts (49–41, p < 0.05), and higher rate of monthly EVT counts (59–60, p = 0.80). Only the change in IVT counts was statistically significant. Percentage of ischemic strokes requiring IVT decreased overall from 13.2 to 11.4% (p < 0.05). Percentage of ischemic strokes requiring EVT increased overall but this was not significant (16.0–16.7, p = 0.33).

Results of the ITS and the calculated relative change from baseline are presented in Table 3 and summarized next. During the pandemic, TIA presentations remained the same (2.9, 95% CI –1.8 to 7.6, p = 0.24), as did ischemic stroke (2.9, 95% CI –7.8 to 1.9, p = 0.25), IVT (−14.3, 95% CI –36. to 6.1, p < 0.01), and EVT (0, 95% CI –16.9 to 16.9, p = 0.98) counts. Linear, ARIMA, and seasonal ARIMA models were needed to fit the data to adjust for time varying confounders such as seasonality and autocorrelation, and the exact model used for each of the 24 analyses is listed in Table 3. Visualization of the pooled data and associated fitted models are described in Figure 2.

Results of the ITS by country are also presented in Table 3 and summarized next. The centers from Australia reported an increase in TIA presentations of 17.3% (95% CI 12.0 to 22.7, p < 0.00001), while the center from UK (−21.4, 95% CI -42.9 to 0, p < 0.05) and Italy (−55.6, 95% CI –88.9 to 33.3, p < 0.001) reported decreases, leading to an overall non-significant change in rate. The fall in ischemic stroke counts appeared to be largely due to reduction of −29.7% (−43.2 to −13.5, p < 0.001) in the center from Italy. Regarding reperfusion, only EVT counts (−50, 95% CI –100.0 to −50.0, p < 0.0001) from Bologna changed significantly. Results of the ITS by center are presented in Supplementary Table S1.

Our study demonstrated that TIA clinic volume has remained constant over 3 years, inclusive of a 1-year intra-COVID period, and this provides evidence against initial concerns regarding patients with transient symptoms not presenting for healthcare. Previous authors have hypothesized about this reluctance (6, 24, 25), but these were made during an earlier phase of the COVID-19 pandemic. Our study adds value to these initial reports as it includes (1) an extended period of observation and (2) includes hospitals from four different countries to allow initial investigation of regional differences. It is important to note that the experience within our group was varied, with the three centers from Australia seeing an increase in TIA presentations of 17.3% as opposed to the decrease of 21.4% at the center from UK and 55.6% at the center from Italy. Regional differences may explain this contrast, including differences in COVID-19 incidence and mortality, governmental policy, public health awareness and campaigns, and cultural differences in health-seeking behavior, however these were not captured by this study. Similar single-country reports however were consistent with our data, with three hyperacute stroke centers in the UK describing an immediate reduction in TIA referrals by 28% after the first UK COVID death on 9th March 2020 (26) and a tertiary stroke care center in Italy describing a 51% decrease in TIA incidence when comparing 8th March to 2nd May 2020 to the corresponding period in 2019 (6). In contrast, a center in Hong Kong reported no increase in referrals to TIA clinic in January to March 2020, despite fewer TIA cases being admitted to their hospital when compared to the same period in 2019 (8). Together, the existing evidence combined with our results demonstrates region-specific differences in TIA activity in response to COVID-19 and highlights the complexity in a patient’s decision to seek care for transient and/or minor symptoms. Nevertheless, the overall result of constant TIA clinic activity throughout the pandemic suggests that the overall impact of COVID-19 may have only been temporary, and that patients are still accessing care via rapid TIA clinics. Overall, we believe that the previously documented adaptations of each clinic in response to the pandemic (9) have contributed to meeting service demand.

Secondarily, our study demonstrated reductions in ischemic stroke presentations and a disproportionate decrease in IVT during the pandemic period, and this contrasts with TIA clinic activity described earlier. This group may be a separate demographic to TIA patients, perhaps more advanced in age, and perhaps more reluctant to seek care that the TIA group. The data reverses the pattern of larger reductions in TIA incidence than ischemic stroke incidence in both primary care (5) and hospital (6), and further highlights this group of patients as a distinct subgroup. Nevertheless, the reductions in ischemic stroke presentations remain consistent with published studies, with papers reporting reductions in stroke codes (7, 24, 27), stroke presentations (28, 29), new stroke diagnoses (25), stroke admissions (7, 26, 30–34), and stroke discharges (35). Similarly, our reduction in IV thrombolysis numbers was consistent with data from France (27), China (30), and Iran (32), but data from Japan (34) and US (36) did not report changes in IV thrombolysis numbers. This could be because patients were presenting later and outside the IVT window, leading to our observed 14.3% reduction in thrombolysis counts. It is not clear whether patients may be presenting elsewhere or whether they are simply presenting later, but a study of 1,262 general practices in Germany noted a 10% reduction in stroke incidence in April to June 2020 when compared to the corresponding period in 2019, suggesting that patients may not be presenting to their general practitioners either (5). Lastly, we did not observe any significant change in endovascular clot retrieval activity, similar to the Japan (34) and US (36) studies, but note that the centers from France (27) and China (30) did experience significant reductions. In conclusion, delays in seeking treatment in what appears to be a separate subgroup of patient behavior from those with transient and/or minor symptoms could be a valuable target for public health intervention during periods of public concern, especially pandemics.

Various methods have been used to describe service-level changes as a result of COVID-19, including cohort comparisons to the corresponding period in 2019 (8, 27, 29, 30, 35, 37), comparisons to monthly averages over years (28), or as before and after t-testing (25). These previous approaches make certain a priori assumptions about the timing of the impact of an event and allow only two groups: ‘before’ and ‘after.’ However, interrupted time series analysis is a useful study design to evaluate the effectiveness of population-level interventions because (1) there is flexibility in modelling the impact (e.g., level and/or slope change, temporary or permanent change) and (2) consideration of time varying confounders such as seasonal trends and adjustment for autocorrelation of data (21). Consecutive observations over time can tend to be more similar to closer months than further months, a phenomenon known as autocorrelation, and this violates the assumption of non-ITS regression models that observations need to be independent (21). Not many studies choose this advanced statistical approach to analyze the impact of COVID on service-level data, which highlights the methodological strength of our analysis (19, 26, 31–34, 38).

The follow-up duration was longer than most existing studies, providing a more longitudinal view of pandemic health service impact beyond the initial acute phase. Sampling from multiple countries is a strength, but variations in COVID-19 community levels, durations of Governmental lockdown, and presence of public awareness campaigns remain that may influence results from each center. Furthermore, the proportion of mimics was not recorded in the data, and it is known that up to 38% of TIA clinic attendees are ultimately diagnosed as mimics (1). Lastly, more than half of the study population was from centers in Australia, which is not representative of rapid TIA clinics in general.

These findings are expected to be valid in stroke centers that offer rapid TIA clinics, as well as systemic and/or mechanical reperfusion. Hospitals that intend to apply these findings to current and future planning could use this data as an overall direction of effect but would benefit from an individualized time series prediction of their own presentation numbers.