Abstract TMP49: Accuracy Of Toast Subtype Classification In Clinical Practice: Implications For The Get With The Guidelines National Stroke Registry

Abstract

Background: TOAST is the most commonly used ischemic stroke subtype classification system worldwide and a required field in the US National GWTG-Stroke registry. However, stroke diagnostics have advanced substantially since TOAST was designed 30 years ago, potentially making it difficult to now apply reliably. Methods: We analyzed consecutive ischemic stroke patients admitted to a CSC between July-Oct 2021. Clinical practice TOAST diagnoses rendered by the stroke team in the EMR at discharge were retrieved from GWTG-Stroke registry and compared to a reference (“gold”) standard diagnosis derived from agreement between 2 expert raters after review of the EMR and patient imaging. Results: Among 49 patients, age was 72.3y (±12.1), 53% female, and presenting NIHSS median 3 (IQR 1-11). Work-up included: brain imaging in 100%; cardiac rhythm assessment in 100%; cervical/cerebral vessel imaging in 98%; TTE±TEE in 92%; and TCD emboli eval in 51%. Reference standard diagnoses were: LAA-6%, SVD-14%, CE-41%, OTH-8%, UND-more than one cause-18%, and UND-cryptogenic-12%. The Figure shows a bubble plot of congruence between EMR/GWTG-Stroke and reference standard diagnoses. GWTG-Stroke TOAST diagnoses agreed with reference standard diagnoses in 29/49 (59%). Among the 6 subtype diagnoses singly, specificity was generally high (83%-100%), but sensitivity suboptimal for LAA (33%), OTH (75%), UND-MT1 (0%), and UND-CRYPT (17%). Positive predictive value was suboptimal for all: LAA (13%), SVD (58%), CE (77%), OTH (75%), SVD OTH (75%), UND-MT1 (0%), and UND-CRYPT (33%). Conclusion: Clinical practice subtype diagnoses using TOAST and entered into GWTG-Stroke were accurate in only 59% of patients, hampering the ability of the national registry to provide dependable insights into subtype-related aspects of care. Development of an updated ischemic stroke subtype classification system, with algorithmic logic directly embedded in electronic medical records, is desirable.