Predicting mortality and incident immobility in older Belgian men by characteristics related to sarcopenia and frailty

Summary There is an increasing awareness of sarcopenia in older people. We applied machine learning principles to predict mortality and incident immobility in older Belgian men through sarcopenia and frailty characteristics. Mortality could be predicted with good accuracy. Serum 25-hydroxyvitamin D and bone mineral density scores were the most important predictors. Introduction Machine learning principles were used to predict 5-year mortality and 3-year incident severe immobility in a population of older men by frailty and sarcopenia characteristics. Methods Using prospective data from 1997 on 264 older Belgian men ( n = 152 predictors), 29 statistical models were developed and tuned on 75% of data points then validated on the remaining 25%. The model with the highest test area under the curve (AUC) was chosen as the best. From these, ranked predictor importance was extracted. Results Five-year mortality could be predicted with good accuracy (test AUC of .85 [.73; .97], sensitivity 78%, specificity 89% at a probability cut-off of 22.3%) using a Bayesian generalized linear model. Three-year incident severe immobility could be predicted with fair accuracy (test AUC .74 [.57; .91], sensitivity 67%, specificity 78% at a probability cut-off of 14.2%) using a multivariate adaptive regression splines model. Serum 25-hydroxyvitamin D levels and hip bone mineral density scores were the most important predictors of mortality, while biochemical androgen markers and Short-Form 36 Physical Domain questions were the most important predictors of immobility. Sarcopenia assessed by lean mass estimates was relevant to mortality prediction but not immobility prediction. Conclusions Using advanced statistical models and a machine learning approach 5-year mortality can be predicted with good accuracy using a Bayesian generalized linear model and 3-year incident severe immobility with fair accuracy using a multivariate adaptive regression splines model.