We calculated the mean μ with the output of a Poisson Lasso (least absolute shrinkage and selection operator) regression model [31,32]. The dispersion coefficient θ was calculated from the observed number of falls Y and the predictions μ issued by the regression model, using maximum likelihood and assuming the number of falls as drawn from a negative binomial distribution with mean equal to μ (R function theta.ml from package MASS[33]). Fitting and evaluation were performed with 10-fold cross-validation. Samples were split into 10 folds in such a way that all the samples from the same subject were assigned to the same fold. In turn, nine folds were used to fit a Poisson Lasso regression model [32] and calculate a dispersion coefficient. This regression model and the dispersion coefficient were used to issue the probabilistic predictions on the samples of the test fold (Fig 1). In the following, we refer to this approach as the “unconstrained Lasso model”, or “Lasso model” for brevity.