key: cord-0767277-gbrsg37c authors: Nogueira, Raul G.; Davies, Jason M.; Gupta, Rishi; Hassan, Ameer E.; Devlin, Thomas; Haussen, Diogo C.; Mohammaden, Mahmoud H.; Kellner, Christopher P.; Arthur, Adam; Elijovich, Lucas; Owada, Kumiko; Begun, Dina; Narayan, Mukund; Mordenfeld, Nadia; Tekle, Wondwossen G.; Nahab, Fadi; Jovin, Tudor G.; Frei, Don; Siddiqui, Adnan H.; Frankel, Michael R.; Mocco, J title: Epidemiological Surveillance of the Impact of the COVID-19 Pandemic on Stroke Care Using Artificial Intelligence date: 2021-03-04 journal: Stroke DOI: 10.1161/strokeaha.120.031960 sha: 3e3d41b8ad154079660232cd1279fe56577a2330 doc_id: 767277 cord_uid: gbrsg37c BACKGROUND AND PURPOSE: The degree to which the coronavirus disease 2019 (COVID-19) pandemic has affected systems of care, in particular, those for time-sensitive conditions such as stroke, remains poorly quantified. We sought to evaluate the impact of COVID-19 in the overall screening for acute stroke utilizing a commercial clinical artificial intelligence platform. METHODS: Data were derived from the Viz Platform, an artificial intelligence application designed to optimize the workflow of patients with acute stroke. Neuroimaging data on suspected patients with stroke across 97 hospitals in 20 US states were collected in real time and retrospectively analyzed with the number of patients undergoing imaging screening serving as a surrogate for the amount of stroke care. The main outcome measures were the number of computed tomography (CT) angiography, CT perfusion, large vessel occlusions (defined according to the automated software detection), and severe strokes on CT perfusion (defined as those with hypoperfusion volumes >70 mL) normalized as number of patients per day per hospital. Data from the prepandemic (November 4, 2019 to February 29, 2020) and pandemic (March 1 to May 10, 2020) periods were compared at national and state levels. Correlations were made between the inter-period changes in imaging screening, stroke hospitalizations, and thrombectomy procedures using state-specific sampling. RESULTS: A total of 23 223 patients were included. The incidence of large vessel occlusion on CT angiography and severe strokes on CT perfusion were 11.2% (n=2602) and 14.7% (n=1229/8328), respectively. There were significant declines in the overall number of CT angiographies (−22.8%; 1.39–1.07 patients/day per hospital, P<0.001) and CT perfusion (−26.1%; 0.50–0.37 patients/day per hospital, P<0.001) as well as in the incidence of large vessel occlusion (−17.1%; 0.15–0.13 patients/day per hospital, P<0.001) and severe strokes on CT perfusion (−16.7%; 0.12–0.10 patients/day per hospital, P<0.005). The sampled cohort showed similar declines in the rates of large vessel occlusions versus thrombectomy (18.8% versus 19.5%, P=0.9) and comprehensive stroke center hospitalizations (18.8% versus 11.0%, P=0.4). CONCLUSIONS: A significant decline in stroke imaging screening has occurred during the COVID-19 pandemic. This analysis underscores the broader application of artificial intelligence neuroimaging platforms for the real-time monitoring of stroke systems of care. disease has resulted in over 2.5 million deaths worldwide, with recent surges portending a grim outlook. has been linked to thrombotic events, including a higher risk of cerebrovascular events. [1] [2] [3] [4] Although a causal association between stroke and COVID-19 is highly plausible, the majority of patients who may require care for acute stroke during the COVID-19 outbreak are unlikely to be infected with SARS-CoV2, and yet the social, psychological, and health care impact of the pandemic seems to indirectly affect even those patients with stroke not infected by the virus. Indeed, an increasing number of reports describe dramatic reductions in stroke admissions, delays in hospital arrival times, and significant declines in the rates of intravenous thrombolysis and mechanical thrombectomy (MT) across broad geographic locations. 2, [4] [5] [6] [7] [8] [9] [10] [11] See related article, p 1691 The degree to which COVID-19 pandemic has impacted the systems of care, in particular, those for timesensitive conditions requiring acute response such as stroke, remains poorly measured. We sought to quantify the impact of COVID-19 on screening for acute stroke by analyzing a large neuroimaging database of suspected patients with stroke across 97 hospitals in 20 US states. The data that support the findings of the study are available from the corresponding author on reasonable request. The Viz Neuroimaging Platform (Vizai, Palo Alto, CA) was designed to optimize the assessment and workflow for patients with acute ischemic stroke to facilitate acute reperfusion treatments. The system is based on a proprietary artificial intelligence convolutional neural network that automates neuroimaging processing and interpretation. The clinical tool consists of a HIPAA-compliant mobile interactive module encompassing (1) a Digital Imaging and Communications in Medicine viewer, (2) a communication system including text messaging and audiovideo calling capabilities, (3) an artificial intelligence algorithm that automatically identifies suspected large vessel occlusion (LVO) strokes on computed tomography (CT) angiography (CTA) and triggers an alert to the on-call team, and (4) a second artificial intelligence algorithm that automatically analyzes CT perfusion (CTP) images using motion correction, artifacts rejection, and arterial input function/venous output function extraction to build an imaging output based on deconvolution models. The CTP outputs include threshold-based maps for relative cerebral blood flow, relative cerebral blood volume, and time-to-maximum of the residue function. Baseline predicted infarct core volume is calculated based on the relative cerebral blood flow <30% threshold (relative to the mean contrast concentration on the contralateral side) and hypoperfused tissue at risk is calculated based on time-to-maximum of the residue function >6 seconds. Perfusion mismatch is defined in both relative and absolute terms based on the ratio and absolute difference between the time-to-maximum of the residue function >6 seconds and relative cerebral blood flow <30% lesion volumes, respectively. 12 Following standard imaging acquisition, Digital Imaging and Communications in Medicine data are transmitted to the server for cloud-based processing, and the generated output is shared with the clinical teams via a mobile application. In parallel, an automated deidentification process creates an anonymous patient identifier and removes any protected health information. The deidentified output along with its associated metadata (including patient's age and sex as well as date, time, and site of image acquisition) is subsequently transferred to Viz Analytics, a clinical intelligence platform comprised of a data warehouse containing a software package that empowers large data analysis in real time. Since imaging is automatically collected for each patient with stroke, we assessed stroke screening rates over time by interrogating Viz Analytics for the number of newly captured patients. Data were analyzed from 23 223 patients with CTA performed across 97 US hospitals distributed across 20 states (Arizona, California, Colorado, Florida, Georgia, Kansas, Kentucky, Mississippi, Nebraska, New Hampshire, New Jersey, New York, North Carolina, Ohio, South Carolina, Tennessee, Texas, Utah, Virginia, Wyoming) utilizing the Vizai software over an entire period of 27 weeks (November 4, 2019 to May 10, 2020). Data were analyzed as a national aggregate. A detailed state level analysis was also performed in the 5 states with the broadest utilization of the Vizai software (3451 patients in New York; 2903 in Georgia; 2822 in Tennessee; 2108 in Ohio; 1074 in South Carolina). In addition, given the unique aspects of the software utilization in the State of Georgia, the Viz Analytics data were correlated with the combined multi-institutional data of the 2 stroke networks that employ the software in the state. These networks include 48 primary stroke centers along with the 3 comprehensive stroke centers (CSCs) performing MT for the entire region covered by the Viz application. As this represents a nearly completely closed system, it provides the opportunity to evaluate the epidemiological surveillance function of Viz Analytics by comparing the satellite macroenvironment application assessment with the ground view from the CSCs microenvironment. Specifically, comparisons were made to assess the ability of the proportional changes in the number of overall CTAs, LVOs on CTA, and severe strokes on CTP on the Viz system to predict changes in the number of stroke admissions and MT procedures at the CSCs. Finally, we assessed at the relationship between CTAs/hospital per day versus the number of overall COVID-19 cases and COVID-19 deaths per 1 000 000 at each state as an attempt to explore whether or not the heterogeneity of COVID-19 incidence across the different states could have impacted our results. LVO strokes on CTA were defined according to the automated software detection. Severe strokes on CTP were defined as those with hypoperfusion lesion volumes >70 mL. The prepandemic and pandemic periods were defined using March 1, 2020, the day that New York City, the US epicenter of the pandemic, reported its first COVID-19 hospitalization, as the cutoff date. We hypothesized that, given the pandemic-related psychosocial impact on the population and overall burden on the systems of health care, (1) a significant decline would be seen in imaging stroke screening across the 2 periods, (2) this would more prominently affect the milder strokes but would still affect patients with the more severe presentations (ie, LVO and hypoperfusion deficits >70 mL), (3) as patients would be potentially arriving in a delayed manner, the infarct volumes would be higher and the mismatch volumes/ratios would be lower across the 2 epochs, (4) the age and sex distribution of the decline would potentially differ. This was an investigator-initiated project. The first author wrote the first draft of the article with subsequent input of all co-authors. Vizai supplied the data and analytic support, but the company was not involved in the study design or in the preparation of the article. The institutional review board from the coordinating center (Emory University) considered that the investigators did not have access to protected health information and thus the study did not meet the federal description of human subject research. Outcome data were assessed from daily measurements of patients undergoing neuroimaging. Values were plotted versus time, relative to the incidence of COVID-19, including the total number of CTA/CTPs and LVOs, normalized as number of patients/day per hospital. Summary statistics included age, sex, LVO detection, core volume, hypoperfused volume and mismatch volume/ratio, and were reported as means with 95% CIs, or medians and interquartile range (IQR). Values were compared across the prepandemic and pandemic periods. Two-sample t-tests and Kolmogorov-Smirnov tests were used for singlecategory comparisons and Poisson regression for multicategory comparisons, with P<0.05 noted as statistically significant. Volumes for stroke admissions and MT at the CSCs and Vizai CTAs, LVOs, and severe strokes on CTP in Georgia were computed as weekly aggregates, and P were calculated by fitting a Poisson regression model to the count data, and testing for the significance of the interaction term between different count variables (such as the number MT versus Viz LVOs). A nonsignificant interaction indicates that the relationship between the 2 counts is consistent both before and during the pandemic, and therefore one variable is a good proxy for the other. Analyses were performed using Python-based SciPy v1.3.1 and R. A total of 23 223 patients with CTA were included. The median age was 67 years (IQR, 54-78) and 45.6% were male. The incidence of LVO on CTA was 11.2% (n=2602; median age, 71 [IQR, 59-82] years; 41.3% male; 10.7% prepandemic versus 11.5% post-pandemic, P=0.13). There were 8328 patients with CTP. The mean CTP core volume was 8±28 mL in the overall population and 28±50 mL in the LVO cohort. The incidence of patients with severe strokes on CTP was 14 (Table 2) . Statistically significant reductions in the number of CTAs and CTPs acquired, as well as in the incidence of LVO detected on CTA, were seen at an individual level in 4 of the 5 analyzed states (New York, Georgia, Tennessee, and Ohio). The changes were not significant in South Carolina, which had a considerably smaller sample size compared with the other states. The decline in CTA screening was numerically higher in New York, Georgia, and Tennessee, the 3 states with the highest per capita rates of COVID-19 infection ( Table 2) , but there was no linear relationship between the magnitude of decline in CTA screening and either the number of overall COVID-19 cases or COVID-19 deaths per 1 000 000 at each state ( Our study adds to the fast-growing body of evidence underlining the collateral damage done to stroke systems of care by the COVID-19 pandemic. In the current analysis comprising 23 223 patients across 20 US states, we have demonstrated a drop of 22.8% in the rate of stroke imaging screening during the outbreak. This drop was more evident across the milder imaging patterns, but the rates of LVO on CTA and severe strokes on CTP still showed a significant decline of about 17%. This agrees with recent reports describing a greater reduction in the rates of TIA and mild strokes as compared with more severe stroke presentations. 3, 8, 13 Interestingly, there was a significant reduction (−23.3%; P<0.001) in the mismatch ratio in patients with LVO. This might suggest greater penumbral consumption presumably related to delayed hospital arrival times. However, despite the directional changes in favor of larger ischemic cores and smaller absolute mismatch volumes for both LVO on CTA and severe strokes on CTP, these changes were of small magnitude and did not reach statistical significance. Therefore, we could not substantiate the hypothesis that more patients presented to stroke centers with imaging patterns suggestive of delayed hospital arrival. Notably, the findings about stroke presentation times during the pandemic have been controversial. In fact, some reports have demonstrated a relative preservation of prehospital workflow times despite significant reductions in overall stroke volumes and rates of reperfusion treatments. This has been described even in some of the epicenters of the disease including Barcelona and France. 2, 5 In contrast, significant delays in presentation times have been reported in other systems. [7] [8] [9] 14 These discrepancies might simply reflect regional variations in prehospital workflow over the course of the outbreak. Conspicuously, South Carolina demonstrated marked decreases in the absolute mismatch volumes in both patients with severe strokes on CTP (−28.1%; P=0.01) and LVO on CTA (−36.2%; P=0.24) in the face of a concurrent report showing longer times to hospital arrival in the region. 9 This finding supports the concept that big data imaging surveillance may potentially allow for the early identification of workflow problems. The significantly younger age seen in male patients with LVO could theoretically reflect a greater hesitation from elderly male patients to seek medical care in face of their higher vulnerability to COVID-19 or possibly represent a byproduct of the strokes directly related to the SARS-CoV2 infection, which tend to affect younger patients. [1] [2] [3] [4] Given the exploratory nature of our analysis, these findings may simply represent chance and have only hypothesis generating value. Our analysis shares some similarities in terms of both methodology and findings with a recent study that utilized another neuroimaging software platform (RAPID, iSchemaView) where the authors collected data across 856 US hospitals and demonstrated that the number of patients who underwent imaging decreased by 39%, from 1.18 to 0.72 patients/day per hospital, from the pre- (February 1-29 , 2020) to early-(March 26 to April 8, 2020) pandemic periods. 15 The numerical differences in the decline of stroke screening across the 2 analyses (39% versus 22.8%) might be explained by the fact that the RAPID study excluded the earlier pandemic phases due to an apparent increase in neuroimaging screening at that time. This represents a selection bias that may have led to an overestimation of the reported decline. Indeed, the duration of both the prepandemic (17 versus 4 weeks) and pandemic (10 versus 2 weeks) periods included in our analysis was 4 to 5 times longer than in the previous report, which presumably makes our study less vulnerable to random influences and more representative of the actual impact of COVID-19 on stroke care. Moreover, our study distinguished itself by the inclusion of a specific subgroup analysis of patients with automated LVO detection on CTA, the ideal target population to investigate the pandemic impact on endovascular stroke treatment. Both reports used the volumes of imaging screening as a surrogate for the amount of care provided, an assumption that was corroborated by the sampling of the State of Georgia where similar declines were seen across the number of LVOs detected by the Viz system at all (primary and comprehensive) stroke centers and the number of MTs performed. Additionally, similar declines were observed across the numbers of LVOs on CTA and severe strokes on CTP at all centers versus CSCs hospital admissions. The proportionally higher change in the number of overall CTAs as compared with the change in CSC admissions is justified by the fact that the reduction in screening was more significant for milder strokes. Since the more severe strokes are either preferentially referred by EMS to CSCs or are more likely to be subsequently transferred to CSCs (if initially taken to primary centers), they tend to be over-represented in CSCs. As such, LVOs on CTA and severe strokes on CTP are presumably better surrogates for CSCs volumes than the overall number of CTAs. This analysis demonstrates the utility of large imaging and analytic platforms to aggregate and expedite the systematic collection, analysis, and broadcasting of data in support of planning, implementing, and appraising stroke systems of care. The capability for real-time assessment represents a major breakthrough in stroke epidemiological surveillance since data from currently employed mechanisms typically suffer from significant recognition time lags that considerably limit their ability to issue swift responses during time-sensitive crises. Despite its prospective data collection, our study carries all the limitations inherent to any retrospective analysis. In addition, as both the utilization of the Viz system and the volume of MT procedures are increasing over time, we could not compare the pandemic findings with those from the corresponding period the prior year. Thus, it remains possible that seasonal variations may have contributed to some of the changes. The Viz LVO system has been trained to detect proximal anterior circulation occlusions only. Therefore, it likely underestimated the total LVOs by failing to compute vertebrobasilar occlusions and some of the more distal middle cerebral artery LVOs (ie, distal M2s). Details about stroke onset time and clinical severity are not available. Magnetic resonance imaging selection was not included in the study and it is known that diffusion weighted imaging provides more reliable quantification of infarct size/severity than CTP. While the inclusion of a relatively long period for the COVID-19 epoch in our study may have reduced random influences and helped capturing the highest impact of the pandemic in most studied regions, there is also the possibility that it might have diluted its real impact since the study period may have also included weeks in which the pandemic had not yet affected some specific regions. Finally, as close stroke referral networks were required to properly validate the Viz surveillance function, we had to limit our correlations to a pool of CSCs in a single state. This large neuroimaging database analysis encompassing data on 23 223 suspected patients with stroke across 97 hospitals in 20 US states demonstrated a 22.8% drop in the number of patients undergoing stroke imaging screening during the COVID-19 pandemic. Although the decline was more dramatic in patients with milder imaging patterns, it also significantly impacted more severe strokes including those with LVO, and sampling data correlated these findings with corresponding reductions in the number of MT procedures and hospitalizations to CSCs. Our results underscore the urgent need to devise Large-vessel stroke as a presenting feature of Covid-19 in the young Acute stroke care is at risk in the era of COVID-19: experience at a Comprehensive Stroke Center in Barcelona SARS-CoV-2 and stroke in a New York Healthcare System Underutilization of health care for strokes during the COVID-19 outbreak Mechanical thrombectomy for acute ischemic stroke amid the COVID-19 outbreak: decreased activity, and increased care delays Impact of the COVID-19 epidemic on stroke care and potential solutions Delays in stroke onset to hospital arrival time during COVID-19 Break in the stroke chain of survival due to COVID-19 Delayed presentation of acute ischemic strokes during the COVID-19 crisis COVID-19: stroke admissions, emergency department visits, and prevention clinic referrals The curious case of the missing strokes during the covid-19 pandemic EXTEND-IA Investigators. Endovascular therapy for ischemic stroke with perfusion-imaging selection Falling stroke rates during COVID-19 pandemic at a comprehensive stroke center Acute stroke management pathway during coronavirus-19 pandemic Collateral effect of Covid-19 on stroke evaluation in the United States None.