Abstract TP174: Clustering and Predicting Functional Recovery Patterns of the First-ever Ischemic Stroke Using Artificial Intelligence: The KOSCO Study

Abstract

Objective: The objective of this study was to apply the clustering approach of multi-facet functional recovery pattern with big data of in the Korean Stroke Cohort for Functioning and Rehabilitation (KOSCO) using artificial intelligence, and to provide valuable prediction models for clinically use. Methods: This study was an interim analysis of the KOSCO designed as 10 years long-term follow-up study of stroke patients. We analyzed data of participants who completed functional assessments from 7 days to 12 months after ischemic stroke onset. Functional assessments included Korean modified Barthel Index (K-MBI), Korean Mini-Mental State Examination (K-MMSE), Fugl-Meyer Assessment (FMA), Functional Ambulatory Category (FAC), the American Speech-Language-Hearing Association National Outcome Measurement System Swallowing Scale (ASHA-NOMS), and Short Korean Version of Frenchay Aphasia Screening Test (Short K-FAST). The cluster analysis using artificial intelligence was performed for multi-facet functional recovery patterns of independency, motor, ambulation, cognition, language, and swallowing functions. After the cluster analysis, a group of rehabilitation specialists reviewed the clinical meaningfulness with clustered population, whether the groups had high homogeneity and representativeness of the clinical stroke recovery patterns. After these clustering approaches, a prediction model using machine learning was performed. The accuracy of classification of this prediction model was evaluated by comparing how much the prediction was equivalent to the actual clustering result. Results: After the machine learning in supervised manners on artificial intelligence, recovery patterns after stroke could be classified into ten groups. Each group showed a different multi-facet functional recovery pattern from 7 day to 12 months, and this clustering showed a clinically acceptance. In addition, the accuracy in classification with clinical characteristics at 7 days showed more than 73.0%. Conclusion: The results of this study demonstrated the potentials of the clustering and predicting functional recovery patterns of stroke patients using artificial intelligence.