External validation of stroke mimic prediction scales in the emergency department

Abstract

Background Acute ischemic stroke is a time-sensitive emergency where accurate diagnosis is required promptly. Due to time pressures, stroke mimics who present with similar signs and symptoms as acute ischemic stroke, pose a diagnostic challenge to the emergency physician. With limited access to investigative tools, clinical prediction, tools based only on clinical features, may be useful to identify stroke mimics. We aim to externally validate the performance of 4 stroke mimic prediction scales, and derive a novel decision tree, to improve identification of stroke mimics. Methods We performed a retrospective cross-sectional study at a primary stroke centre, served by a telestroke hub. We included consecutive patients who were administered intravenous thrombolysis for suspected acute ischemic stroke from January 2015 to October 2017. Four stroke mimic prediction tools (FABS, simplified FABS, Telestroke Mimic Score and Khan Score) were rated simultaneously, using only clinical information prior to administration of thrombolysis. The final diagnosis was ascertained by an independent stroke neurologist. Area under receiver operating curve (AUROC) analysis was performed. A classification tree analysis was also conducted using variables which were found to be significant in the univariate analysis. Results Telestroke Mimic Score had the highest discrimination for stroke mimics among the 4 scores tested (AUROC = 0.75, 95% CI = 0.63–0.87). However, all 4 scores performed similarly (DeLong p > 0.05). Telestroke Mimic Score had the highest sensitivity (91.3%), while Khan score had the highest specificity (88.2%). All 4 scores had high positive predictive value (88.1 to 97.5%) and low negative predictive values (4.7 to 32.3%). A novel decision tree, using only age, presence of migraine and psychiatric history, had a higher prediction performance (AUROC = 0.80). Conclusion Four tested stroke mimic prediction scales performed similarly to identify stroke mimics in the emergency setting. A novel decision tree may improve the identification of stroke mimics.

Telestroke Assessment With Perfusion CT Improves the Diagnostic Accuracy of Stroke vs. Mimic

“…In contrast, symptoms which occurred more commonly in mimic, such as vertigo, decreased level of consciousness, confusion, headache, and seizure, were less likely to be strokes and tended to be the common presentation of mimic conditions, such as migraine, seizures, and metabolic disturbances ( 11 ). The diagnostic accuracy of our model using clinical features to differentiate between stroke and mimic was similar to that of Ali et al ( 12 ) who had a result of 0.72, using age, NIHSS, history of atrial fibrillation, hypertension, and facial weakness as predictors in their model. Similar to other studies looking at the primary presentation of stroke, the rate of mimics was about 24% ( 4 , 5 , 8 , 11 ).…”

“…Whilst there are no comparable studies using a telehealth population and CTP in their regression models, the accuracy of our model is as least as accurate as studies where face-to-face clinical assessment occurred and included CTA findings (but not CTP) in their models ( 6 ). Our model using CTP in addition to clinical features is considerably more accurate than that of previous stroke mimic prediction scales used in the telestroke context, which have only included clinical features ( 12 ). Our findings support the use of CTP in improving diagnostic confidence when attempting to distinguish stroke from mimics in the telehealth setting.…”

“…In a recent study by Tu et al (13), which externally validated four existing stroke mimic prediction scales, the scale with the highest discrimination for stroke mimics had an area under the receiver operator curve value of 0.75. Similar to the studies which focused on face-toface clinical evaluation in the Emergency Department, studies of telestroke assessment found that mimic patients had lower NIHSS scores, were younger in age, usually had a history of seizure, migraine or psychiatric illness, were less likely to have atrial fibrillation or hypertension, and had a lack of localising symptoms (10)(11)(12). However, none of these studies evaluated the utility of including imaging in predictive models.…”

“…There have been a few studies looking at clinical predictors to develop risk prediction tools for distinguishing between stroke and mimic in the telestroke population (10)(11)(12). In a recent study by Tu et al (13), which externally validated four existing stroke mimic prediction scales, the scale with the highest discrimination for stroke mimics had an area under the receiver operator curve value of 0.75.…”

Artificial intelligence in clinical decision support and outcome prediction – applications in stroke

“…31 Various clinical assessment tools, including stroke mimic prediction scales, have been developed to improve the early recognition of stroke. 32 However, to facilitate ease and rapidity of assessment, the requirement for brevity of these tools consequently limits the inclusion of a broad range of stroke-related parameters. AI implementations, with their ability to consider a comprehensive range of data, including medical history, family history and laboratory results, are specialised to address this problem.…”