How to use DynForest with continuous outcome?

Illustrative dataset: data_simu1 and data_simu2 datasets

In this section, we present an illustration of DynForest with a continuous outcome. DynForest was used on a simulated dataset with 200 subjects and 10 predictors (6 time-dependent and 4 time-fixed predictors). The 6 longitudinal predictors were generated using a linear mixed model with linear trajectory according to time. We considered 6 measurements by subject (at baseline and then randomly drawn around theoretical annual visits up to 5 years). Then, we generated the continuous outcome using a linear regression with the random intercept of marker 1 and random slope of marker 2 as linear predictors. We generated two datasets (data_simu1 and data_simu2), one for each step (training and prediction). These datasets are available in the DynForest package.

Data management

library("DynForest")
data(data_simu1)
head(data_simu1)
#>   id     time cont_covar1 cont_covar2 bin_covar1 bin_covar2     marker1
#> 1  1 0.000000   -1.749785   -1.003791          0          0   2.2335135
#> 2  1 1.202289   -1.749785   -1.003791          0          0  -0.3215715
#> 3  1 2.347419   -1.749785   -1.003791          0          0  -4.7698038
#> 4  1 3.068631   -1.749785   -1.003791          0          0  -8.4552464
#> 5  1 4.233713   -1.749785   -1.003791          0          0 -13.8477022
#> 6  1 5.100951   -1.749785   -1.003791          0          0 -15.0028533
#>     marker2    marker3   marker4   marker5   marker6     Y_res
#> 1  5.223892  3.1760262  5.273218  5.812561  4.965971 -10.80882
#> 2  8.101329 -0.7506024  6.333628  7.297978  0.452370 -10.80882
#> 3 14.533583 -0.9671289  8.487332 10.833834 -1.986883 -10.80882
#> 4 16.789004 -0.3502227  9.695727 13.149591 -2.932720 -10.80882
#> 5 20.209609 -2.1694950 12.568234 15.442259 -6.615383 -10.80882
#> 6 25.862403 -7.8668698 11.540265 17.758242 -8.698395 -10.80882
data(data_simu2)
head(data_simu2)
#>   id     time cont_covar1 cont_covar2 bin_covar1 bin_covar2    marker1  marker2
#> 1  1 0.000000   -1.335695  -0.7260684          1          1  2.9558630 3.112574
#> 2  1 1.289630   -1.335695  -0.7260684          1          1  0.3125889 4.469843
#> 3  1 2.251666   -1.335695  -0.7260684          1          1 -2.3155137 5.426203
#> 4  1 3.119182   -1.335695  -0.7260684          1          1 -2.2991755 7.043745
#> 5  1 4.007813   -1.335695  -0.7260684          1          1 -3.0584749 7.905586
#> 6  1 5.075597   -1.335695  -0.7260684          1          1 -4.9786537 7.268507
#>     marker3   marker4     marker5   marker6    Y_res
#> 1 0.5024613  7.254809  3.27673801  1.722360 1.758953
#> 2 2.7152092  6.570345  1.12816586  5.113413 1.758953
#> 3 2.7096358  9.545796  0.08583794  5.006489 1.758953
#> 4 2.2938817  9.844153 -0.53377189  9.649994 1.758953
#> 5 1.8877637 12.749989 -1.80312531  9.900544 1.758953
#> 6 1.0021239 12.522054 -2.92588781 11.557506 1.758953

First of all, we load the data and we build the mandatory objects needed to execute dynforest() function that are timeData_train for time-dependent predictors and fixedData_train for time-fixed predictors. We specify the model for the longitudinal predictors in timeVarModel object. We considered linear trajectories over time for the 6 longitudinal predictors.

timeData_train <- data_simu1[,c("id","time",
                                paste0("marker",seq(6)))]
timeVarModel <- lapply(paste0("marker",seq(6)),
                       FUN = function(x){
                         fixed <- reformulate(termlabels = "time",
                                              response = x)
                         random <- ~ time
                         return(list(fixed = fixed, random = random))
                       })
fixedData_train <- unique(data_simu1[,c("id",
                                        "cont_covar1","cont_covar2",
                                        "bin_covar1","bin_covar2")])

To define the Y object for a continuous outcome, the type argument should be set to numeric to run the random forest in regression mode. The dataframe Y should include two columns with the unique identifier id and the continuous outcome, here Y_res.

Y <- list(type = "numeric",
          Y = unique(data_simu1[,c("id","Y_res")]))

The random forest building

To build the random forest, we chose default hyperparameters (i.e., ntree = 200 and nodesize = 1), except for mtry which was fixed at its maximum (i.e., mtry = 10). We ran dynforest() function with the following code:

res_dyn <- dynforest(timeData = timeData_train, 
                     fixedData = fixedData_train,
                     timeVar = "time", idVar = "id", 
                     timeVarModel = timeVarModel,
                     mtry = 10, Y = Y, 
                     ncores = 7, seed = 1234)

Out-Of-Bag error

For continuous outcome, the OOB prediction error is evaluated using the mean square error (MSE). We used compute_ooberror() function to compute the OOB prediction error and we provided overall results with summary() function as shown below:

res_dyn_OOB <- compute_ooberror(dynforest_obj = res_dyn)
summary(res_dyn_OOB)

dynforest executed for continuous outcome 
    Splitting rule: Minimize weighted within-group variance 
    Out-of-bag error type: Mean square error 
    Leaf statistic: Mean 
---------------- 
Input 
    Number of subjects: 200 
    Longitudinal: 6 predictor(s) 
    Numeric: 2 predictor(s) 
    Factor: 2 predictor(s) 
---------------- 
Tuning parameters 
    mtry: 10 
    nodesize: 1 
    ntree: 200 
---------------- 
---------------- 
dynforest summary 
    Average depth per tree: 9.06 
    Average number of leaves per tree: 126.47 
    Average number of subjects per leaf: 1 
---------------- 
Out-of-bag error based on Mean square error 
    Out-of-bag error: 4.3713 
---------------- 
Computation time 
    Number of cores used: 7 
    Time difference of 8.261093 mins
----------------

The random forest was executed in regression mode (for a continuous outcome). The splitting rule aimed to minimize the weighted within-group variance. We built the random forest using 200 subjects and 10 predictors (6 time-dependent and 4 time-fixed predictors) with hyperparameters ntree = 200, mtry = 10 and nodesize = 1. As we can see, nodesize = 1 leads to deeper trees (the average depth by tree is 9.1) and a single subject by leaf. We obtained 4.4 for the MSE. This quantity can be minimized by tuning hyperparameters mtry and nodesize.

Prediction of the outcome

In regression mode, the tree and leaf-specific means are averaged across the trees to get a unique prediction over the random forest. predict() function provides the individual predictions. We first created the timeData and fixedData from the testing sample data_simu2. We then predicted the continuous outcome by running predict() function:

timeData_pred <- data_simu2[,c("id","time",
                               paste0("marker",seq(6)))]
fixedData_pred <- unique(data_simu2[,c("id","cont_covar1","cont_covar2",
                                       "bin_covar1","bin_covar2")])
pred_dyn <- predict(object = res_dyn,
                    timeData = timeData_pred, 
                    fixedData = fixedData_pred,
                    idVar = "id", timeVar = "time")

Individual predictions can be extracted using print() function:

head(print(pred_dyn))

         1          2          3          4          5          6 
 5.2184031 -1.2786887  0.8591368  1.5115312  5.2984117  7.9073981

For instance, we predicted 5.22 for subject 1, -1.28 for subject 2 and 0.86 for subject 3.

Predictiveness of the variables

In this illustration, we want to evaluate if can identify the true predictors (i.e., random intercept of marker1 and random slope of marker2). To do this, we used the minimal depth which allows to understand the random forest at the feature level.

Minimal depth information can be extracted using compute_vardepth() function and can be displayed with plot() function. For the purpose of this illustration, we displayed the minimal depth in Figure 1 by predictor and by feature.

depth_dyn <- compute_vardepth(dynforest_obj = res_dyn)
p1 <- plot(depth_dyn, plot_level = "predictor")
p2 <- plot(depth_dyn, plot_level = "feature")

We observe in figure @ref(fig:DynForestRdepthscalar)A that marker2 and marker1 have the lowest minimal depth, as expected. To go further, we also looked into the minimal depth computed on features. We perfectly identified the random slope of marker2 (i.e., marker2.bi1) and the random intercept of marker1 (i.e., marker1.bi0) as the predictors in this simulated dataset.

plot_grid(p1, p2, labels = c("A", "B"))
Figure 1: Average minimal depth level by predictor (A) and by feature (B).

Figure 1: Average minimal depth level by predictor (A) and by feature (B).