In a cohort of adult patients with a NIH Stroke Scale ≤ 3 admitted to EDs’ observation units for an accelerated diagnostic protocol for suspected TIA or stroke, we prospectively recorded clinical and imaging results. Non-mydriatic fundus photographs were systematically obtained in the ED and reviewed for retinal microvascular findings (retinal hemorrhages, cotton wool spots, retinal emboli or occlusions, hard exudates, or microaneurysms). A stroke neurologist rated the probability each patient’s presentation (masked to fundus photography) represented a CVE on a visual analogue scale. Linear regression modeling was used for analysis.

We included 395 patients (median age: 57 years (interquartile range [IQR]: 50–66); 34 (9%) retinal microvascular findings present).  Comparing regression coefficients suggested replacing diabetes in the ABCD² with microvascular findings, but when using the original ABCD² point cutoffs (e.g., age ≥ 60: 1 point), both models had poor calibration (predicted vs. neurologist-determined probability of TIA). By incorporating DWI positivity and either microvascular findings or diabetes and by using the full regression coefficients, excellent calibration was achieved (bootstrap bias corrected mean absolute error ~1%) and slightly favored microvascular findings over diabetes.

Although we have reported that retinal microvascular findings among patients with suspected TIA in the ED are an independent factor differentiating TIA and stroke from mimics, incorporating them into ABCD² does not markedly improve rule performance. However, slight improvements obtained by switching from diabetes to retinal microvascular findings suggests presence of diabetic ophthalmic microvascular disease better predicts CVE than simply the presence of diabetes.