Acquired brain injury (ABI) can be a life changing condition, affecting housing, independence, and employment. Machine learning (ML) is increasingly used as a method to predict ABI outcomes, however improper model evaluation poses a potential bias to initially promising findings (Chapter One). This study aimed to evaluate, with transparent reporting, three common ML classification methods. Regularised logistic regression with elastic net, random forest and linear kernel support vector machine

more »

... re compared with unregularised logistic regression to predict good psychosocial outcomes after discharge from ABI inpatient neurorehabilitation using routine cognitive, psychometric and clinical admission assessments. Outcomes were selected on the basis of decision making for care packages: accommodation status, functional participation, supervision needs, occupation and quality of life. The primary outcome was accommodation (n = 164), with models internally validated using repeated nested cross-validation. Random forest was statistically superior to logistic regression for every outcome with areas under the receiver operating characteristic curve (AUC) ranging from 0.81 (95% confidence interval 0.77-0.85) for the primary outcome of accommodation, to its lowest performance for predicting occupation status with an AUC of 0.72 (0.69-0.76). The worst performing ML algorithm was support vector machine, only having statistically superior performance to logistic regression for one outcome, supervision needs, with an AUC of 0.75 (0.71-0.80). Unregularised logistic regression models were poorly calibrated compared to ML indicating severe overfitting, unlikely to perform well in new samples. Overall, ML can predict psychosocial outcomes using routine psychosocial admission data better than other statistical methods typically used by psychologists.