Neural network-based clustering model of ischemic stroke patients with a maximally distinct distribution of 1-year vascular outcomes

Abstract

Clustering stroke patients with similar characteristics to predict subsequent vascular outcome events is critical. This study aimed to compare several clustering methods, particularly a deep neural network-based model, and identify the best clustering method with a maximally distinct 1-year outcome in patients with ischemic stroke. Prospective stroke registry data from a comprehensive stroke center from January 2011 to July 2018 were retrospectively analyzed. Patients with acute ischemic stroke within 7 days of onset were included. The primary outcomes were the composite of all strokes (either hemorrhagic or ischemic), myocardial infarction, and all-cause mortality within one year. Neural network-based clustering models (deep lifetime clustering) were compared with other clustering models (k-prototype and semi-supervised clustering, SSC) and a conventional risk score (Stroke Prognostic Instrument-II, SPI-II) to obtain a distinct distribution of 1-year vascular events. Ultimately, 7,650 patients were included, and the 1-year primary outcome event occurred in 13.1%. The DLC-Kuiper UB model had a significantly higher C-index (0.674), log-rank score (153.1), and Brier score (0.08) than the other cluster models (SSC and DLC-MMD) and the SPI-II score. There were significant differences in primary outcome events among the 3 clusters (41.7%, 13.4%, and 6.5% in clusters 0, 1, and 2, respectively) when the DLC-Kuiper UB model was used. A neural network-based clustering model, the DLC-Kuiper UB model, can improve the clustering of stroke patients with a maximally distinct distribution of 1-year vascular outcomes among each cluster. Further studies are warranted to validate this deep neural network-based clustering model in ischemic stroke.