Multiple Imputation and Multivariate Outliers Detection and Correction Integrated to Machine Learning: A Prediction of Functional Outcome in Post-Stroke Survivors
摘要
Health data analysis faces the challenge of missing data and multivariate outliers, making Multiple Imputation (MI) techniques and multivariate outlier detection and correction (MODC) essential to avoid discarding patients and to ensure data quality. This research proposes an integrated pipeline to predict functional outcomes from post-stroke rehabilitation, combining MI and MODC with cross-validation of machine learning (ML) algorithm, using predictors collected at admission. The analysis involved 220 patients undergoing intensive rehabilitation. With random forest, a median absolute error of 9.72 points on the estimated modified Barthel Index was achieved. Compared to single imputation and MI methods integrated to ML, MI + MODC improved prognostic performances by14.82% and 8.65%, respectively.