Mining acute stroke patients’ data using supervised machine learning

Abstract

Analysis of data for identifying patterns and building models has been used as a strong tool in different domains, including medical domains. In this paper, we analyse the registry of brain stroke patients collected over fifteen years in south London hospitals, known as South London Stroke Register. Our attempt is to identify the similar patterns between patients’ background and living conditions, their cognitive ability, the treatments they received, and the speed of their cognitive recovery; based on which most effective treatment can be predicted for new admitted patients. We designed a novel strategy which takes into account two different approaches. First is to predict, for each of the potential intervention treatments, whether that particular treatment would lead to recovery of a new patient or not. Second is to suggest a treatment (treatments) for the patient based on those that were given to the patients who have recovered and are most similar to the new patient. We built different classifiers using various state of the art machine learning algorithms. These algorithms were evaluated and compared based on three performance metrics, defined in this paper. Given that time is very crucial for stroke patients, main motivation of this research work is identifying the most effective treatment immediately for a new patient, and potentially increase the probability of their cognitive recovery.