Unsupervised learning based on CT imaging in identifying subtypes of acute ischemic stroke severity

Abstract

To explore unsupervised learning based on CT imaging for identifying severity subtypes of acute ischemic stroke (AIS) in the middle cerebral artery territory. One hundred seventy-three AIS patients were retrospectively enrolled. K-means clustering was performed on clinical and CT features to identify severity subtypes. NIHSS scores were compared between subtypes. Hemorrhagic transformation (HT) served as an adverse outcome to compare associations of cluster subtypes versus NIHSS scores. Two subtypes were identified. Subtype 1 (severe) showed larger infarct volumes, higher rates of basal ganglia involvement and dense middle cerebral artery sign, longer onset time, and younger age than Subtype 2 (non-severe) (all p < 0.001). Median NIHSS score was higher in Subtype 1 [17.0 (14.0,19.0) vs. 12.0 (8.0,17.0), p < 0.001]. HT identification accuracy was 70% for cluster subtypes vs. 67% for NIHSS scores. AUC for HT was 0.692 (subtypes), 0.679 (NIHSS), and 0.741 (combined). For poor 90-day functional outcome (mRS ≥ 3), AUC was 0.71 (subtypes), 0.69 (NIHSS), and 0.76 (combined). Unsupervised learning based on CT imaging effectively identifies AIS severity subtypes and may serve as an objective method for assessing post-stroke neurological damage.