Using data mining to predict success in a weight loss trial.

摘要: Background Traditional methods for predicting weight loss success use regression approaches, which make the assumption that relationships between independent and dependent (or logit of dependent) variable are linear. The aim present study was to investigate relationship common demographic early variables predict at 12 months without making this assumption. Methods Data mining (decision trees, generalised additive models multivariate adaptive splines), in addition logistic regression, were employed predict: (i) (defined as ≥5%) end a 12-month dietary intervention using [body mass index (BMI), sex age]; percentage 1 month; (iii) difference actual predicted an energy balance model. compared by assessing model parsimony area under curve (AUC). Results The decision tree provided most clinically useful had good accuracy (AUC 0.720 95% confidence interval = 0.600–0.840). Percentage month (≥0.75%) strongest predictor successful loss. Within those individuals losing ≥0.75%, with BMI (≥27 kg m–2) more likely be than 25 27 m–2. Conclusions Data can provide accurate way when conventional assumptions not met. In study, parsimonious Given cannot before randomisation, incorporating information into post randomisation trial design may give better results.