1257 Artificial Intelligence-Based Agglomerative Clustering of Patient Phenotypes and Disease Factors in Spontaneous Intracerebral Hemorrhage

INTRODUCTION: The role of surgical interventions to improve functional outcome in spontaneous intracerebral hemorrhage (SICH) remains controversial. Stratification of heterogeneous SICH patients into distinct phenotypes may enable personalized management to improve long-term functional outcomes. Unsupervised deep-learning clustering may present a new approach towards solving this age-old problem. METHODS: We recruited a retrospective cohort of 1214 consecutive SICH patients from a tertiary center in Singapore. Selected variables were pre-processed, and the resulting dataset was used in 3 different pipelines of hierarchical clustering: 1) “Combined”; 2) “Nested”; 3) “Separated”. 12 subclusters were generated for each pipeline. Cluster quality of each pipeline was evaluated using interclass and intraclass variance, cluster cardinality, and statistical differences between 30-day mortality and 90-day poor functional outcome (PFO) measures. RESULTS: Of 1214 patients included, 14.2% and 67.1% experienced 30-day mortality and 90-day PFO respectively. Among the 3 pipelines, the “Separated” pipeline demonstrated superior intraclass and interclass variances of 0.2484±0.0381 and 0.4189±0.0247 respectively, as well as the best cluster cardinality of 0.2265-0.2970. Two predominant patient phenotypes were identified: old with antiplatelet and anticoagulants (Oapac, n=352); and young with few risk factors (Yh, n=808). Three disease phenotypes were identified: small bleeds without ventricular extension (small/-VE, n=602); large bleeds with ventricular extension (large/+VE, n=440); and small lobar bleeds without ventricular extension (Lobar, n=172). Amongst the 12 subclusters in the “Separated” pipeline, 30-day mortality ranged from 9.6% to 34.6% and 90-day PFO ranged from 50.0% to 89.7%. CONCLUSIONS: Unsupervised hierarchical clustering can identify patient and disease phenotypes with significantly different outcomes. Such findings may be used to augment preoperative decision-making and serve as a step towards defining a classification scheme for SICH and an AI driven approach towards personalized neurosurgery.