Abstract
Introduction Accurate early prediction of outcome destination of patients presenting with acute stroke (AS) could usefully inform management decisions. Machine learning (ML) has the potential to assist medical management in many areas of medicine. We investigated the ability of ML to predict the discharge destination of patients presenting with AS to determine which ML algorithm would generate the best-performing predictive model and which clinical features were most predictive of discharge destination.
Method 299 patients presenting with AS were included. 17 variables were extracted retrospectively for every patient, comprising 16 possible predictors and one outcome variable (discharge destination). The predictive power of these variables was assessed at each of four stages during the early patient journey, using Relief for feature selection followed by both k-Nearest Neighbour (kNN) and ensemble-based classification algorithms to predict outcome and determine overall and per-class accuracy.
Result Of the ML models employed, the AdaBoost ensemble algorithm generated the most accurate prediction of death outcome by stage 4 (90%), though accuracy for home and rehabilitation destinations were only 84.5% and 56.6%, respectively. 24-hour Scandinavian Stroke Scale (SSS), 24-hour National Institutes of Health Stroke Scale (NIHSS), dyslipidaemia, hypertension, and premorbid mRS were the most important contributing features. There was a higher correlation between initial and 24-hour scores for SSS (0.93) compared to NIHSS (0.81).
Conclusion This study suggests that, for predicting discharge destination, AdaBoost appears to be the most useful algorithm based on the associated ML model’s predictive accuracy. SSS appeared to have slightly higher feature importance than NIHSS.
