Abstract Details

Title Machine Learning-based Prediction of Unfavorable Outcomes in Embolic Stroke of Undetermined Source (ESUS)

Topic Cerebrovascular Disease and Interventional Neurology

Presentation(s) P5 - Poster Session 5 (11:45 AM-12:45 PM)

Poster/Presentation
Number 4-018

Objective This study aimed to develop machine learning (ML) models to predict unfavorable outcomes at 3 months and 1 year after stroke, and to identify relevant features in ESUS patients.

Background Evidence on functional outcomes and their predictors in patients with ESUS remains limited.

Design/Methods A retrospective analysis was conducted using data from the Real World Study of Embolic Stroke of Undetermined Sources (ROS-ESUS), a multicenter prospective study encompassing 19 tertiary medical centers in South Korea from 2014 to 2019. Patients with ESUS were included, with demographic, clinical, laboratory variables, electrocardiographic, Holter monitoring, and transthoracic echocardiographic parameters as predictors. A baseline logistic regression (LR) model and two ML models—tree-based gradient boosting model (LightGBM) and Multi-Layer Perceptron (MLP)—were developed. Unfavorable outcomes were defined as a modified Rankin Scale score ≥3 at 3 months and 1-year. The primary metric for model performance was the area under the receiver operating characteristic curve (AUROC).

Results Among 5,712 patients analyzed for 3-month outcomes, 1,450(25.4%) had unfavorable outcomes, while 4,599 were included for 1-year predictions, with 1,133(24.6%) experiencing poor outcomes. For a 3-month prediction, both ML models outperformed LR (AUROCs: 0.854 [95% CI, 0.837–0.872] for LightGBM and 0.854 [95% CI, 0.836–0.871] for MLP vs. 0.821 [95% CI, 0.802–0.840] P<.001). For a 1-year prediction, LightGBM demonstrated the highest predictive performance (0.835 [95% CI, 0.814–0.856]), significantly surpassing LR (0.812 [95% CI, 0.789–0.835] P=.012), while MLP showed no significant difference. Key predictors across both timepoints included initial NIHSS, age, and D-dimer. For a 3-month prediction, mitral E/e' ratio and minimum heart rate on Holter, while for a 1-year prediction, mitral A-wave velocity and hemoglobin became more influential.

Conclusions ML approaches demonstrate their capability to capture the complex interplay of clinical variables and prognosis unique to ESUS.

Authors/Disclosures
Seung Min Kim PRESENTER	Seung Min Kim has received research support from Veterans Health Service Medical Center Research Grant.
JIN-MAN JUNG, MD, PhD	The institution of Prof. JUNG has received research support from National Research Foundation of Korea (NRF) grant funded by the Korean government.

��ɫ��

��ɫ��