Presenter of 2 Presentations
LATE ANTICOAGULATION IS ASSOCIATED WITH INCREASED NUMBER OF ISCHEMIC LESIONS IN ATRIAL FIBRILLATION RELATED STROKE: THE ATTUNE TRIAL
Abstract
Background and Aims
The optimal time to commence anticoagulation in patients with atrial fibrillation (AF) after ischemic stroke is unclear. Early anticoagulation potentially decreases recurrent ischemic events but is offset by increased risk of hemorrhagic transformation. We aimed to quantify new ischemic lesions and hemorrhages on repeat MRI at 30 days in stroke patients with AF commenced on early versus late anticoagulation. We hypothesized that late anticoagulation was associated with increased new ischemic lesions.
Methods
This was a prospective multicentre study. Inclusion criteria were acute ischemic stroke within 14 days and AF. Anticoagulation timing was categorized as early (<4 days) or late (>/=4 days). MRI brain was obtained at baseline and day 30. The primary outcome was new ischemic lesions on follow up MRI. Multivariate regression analysis was used to analyse the association between anticoagulation timing and new infarcts.
Results
Two hundred and eight patients were included. Median age was 75 (IQR 68 -82), median baseline NIHSS was 5 (IQR 1-11.5) and 40% were female. There were 107 patients in the early and 101 in the late anticoagulation group. Late anticoagulation was associated with increased new infarcts (OR 1.57, 95% CI 1.01- 2.43, p=0.04).
Conclusions
Late anticoagulation (>4 days) is associated with increased ischemic lesions on MRI at follow up. Anticoagulation should not be delayed in patients with AF after ischemic stroke.
ISCHAEMIC STROKE AND CARDIOEMBOLISM: DIAGNOSTIC UTILITY OF A PREDICTIVE DEEP LEARNING MODEL
Abstract
Background and Aims
Cardioembolism, including atrial fibrillation (AF), is a common cause of stroke. AF is often occult and can go undiagnosed despite extensive investigation. There is evidence that cardioembolism is associated with specific topographical patterns on MRI. Human observation however is inadequate to diagnose cardioembolic stroke based on neuroimaging alone. We aimed to develop a predictive deep learning model that can accurately differentiate stroke infarct patterns on MRI, due to a cardioembolic or non-cardioembolic source.
Methods
A random sample of 605 patients with acute ischaemic stroke was analysed from the Royal Melbourne Hospital database. After detailed analysis of patient records and investigations, stroke aetiology was determined as being due to cardioembolism (200) or large vessel atherosclerosis (190). Patients with a mixed or unknown aetiology were excluded. Diffusion-weighted MRI images from patients were de-identified, normalised and labelled. A 3D convolutional neural network (CNN) was employed with 320 patients in the training cohort and 70 in the validation cohort. Various data uniforming and augmentation echniques were then utilised to improve model accuracy.
Results
The median age of the patient cohort was 74 (IQR 64-83). Forty-three percent of patients were female. The best training accuracy obtained by the model was 80% at epoch 28.
Conclusions
In patients with acute ischaemic stroke, a deep learning model has good accuracy in classifying infarct patterns on MRI as due to cardioembolism or other cause. Future studies should be performed with additional clinical context on larger, prospective datasets.