Splitting datas in three sets with balanced datas

EDIT : OK so now I have my train, validation and test sets with rows belonging to patients in same groups. But, using a plot test, I see that the original imbalanced data from the original dataset (from the outcome LesionResponse, 1: 70% and 0 : 30%) is not very respected...Indeed, in the training datas, I have a nearly 55/45 repartition and it's not really welcomed for me. How can I do to correct this ?

summary(train$LesionResponse)
#   0   1
# 159 487
summary(validation$LesionResponse)
#  0   1
# 33 170
summary(test$LesionResponse)
#  0   1
# 77 126

Hi guys,
I have my dataset (here an exemple) and I must build a predictive model for an outcome : "LesionResponse".
So I have in a first time split my datas in train (60%), validation and test (20% each) sets.
I have a huge problem, many rows of my table belong to same patients...so in order to dodge bias, I must divide my datas and take into account the PatientIDs...
I am here stuck because I don't know how to split my datas in three and keep the rows belonging to same patients together.

Here is my code :

structure(list(PatientID = c("P1", "P1", "P1", 
"P2", "P3", "P3", "P4", "P5", 
"P5", "P6"), LesionResponse = structure(c(2L, 2L, 1L, 2L, 2L, 2L, 2L, 
    2L, 1L, 2L), .Label = c("0", 
    "1"), class = "factor"), pyrad_tum_original_shape_LeastAxisLength = c(19.7842995242803, 
    15.0703960571122, 21.0652247652897, 11.804125918871, 27.3980336338908, 
    17.0584330264122, 4.90406343942677, 4.78480430022189, 6.2170232078547, 
    5.96309532740722, 5.30141540007441), pyrad_tum_original_shape_Sphericity = c(0.652056853392657, 
    0.773719977240238, 0.723869070051882, 0.715122964970338, 
    0.70796498824535, 0.811937882810929, 0.836458991713367, 0.863337931630415, 
    0.851654860256904, 0.746212862162174), pyrad_tum_log.sigma.5.0.mm.3D_firstorder_Skewness = c(0.367453961973625, 
    0.117673346718817, 0.0992025164349288, -0.174029385779302, 
    -0.863570016875989, -0.8482193060411, -0.425424618080682, 
    -0.492420174157913, 0.0105111292451967, 0.249865833210199), pyrad_tum_log.sigma.5.0.mm.3D_glcm_Contrast = c(0.376932105256115, 
    0.54885738172596, 0.267158344601612, 2.90094719958076, 0.322424096161189, 
    0.221356030145403, 1.90012334870722, 0.971638740404501, 0.31547550396399, 
    0.653999340294952), pyrad_tum_wavelet.LHH_glszm_GrayLevelNonUniformityNormalized = c(0.154973213866752, 
    0.176128379241556, 0.171129002059539, 0.218343919352019, 
    0.345985943932352, 0.164905080489496, 0.104536489151874, 
    0.1280276816609, 0.137912385073012, 0.133420904484894), pyrad_tum_wavelet.LHH_glszm_LargeAreaEmphasis = c(27390.2818110851, 
    11327.7931034483, 51566.7948885976, 7261.68702290076, 340383.536555142, 
    22724.7792207792, 45.974358974359, 142.588235294118, 266.744186046512, 
    1073.45205479452), pyrad_tum_wavelet.LHH_glszm_LargeAreaLowGrayLevelEmphasis = c(677.011907073653, 
    275.281153810458, 582.131636238695, 173.747506476692, 6140.73990175018, 
    558.277670638306, 1.81042257642817, 4.55724031114589, 6.51794350173746, 
    19.144924585586), pyrad_tum_wavelet.LHH_glszm_SizeZoneNonUniformityNormalized = c(0.411899490603372, 
    0.339216399209913, 0.425584323452468, 0.355165782879786, 
    0.294934042125209, 0.339208410636982, 0.351742274819198, 
    0.394463667820069, 0.360735532720389, 0.36911240382811)), row.names = c(NA, -10L), class = c("tbl_df", "tbl", 
"data.frame"))

i was thinking about a loop who would split a unique(PatientID) dataset in three with 60% in the train set, and if there is no balanced outcome in the sets, to do it again and again. I was thinking more of an interval to solve it...
How would you do guys ?

Edit I misunderstood how you wished to handle PatientIDs. The original answer is at the bottom, however note stratification will aim to put equivalent proportions of each PatientID in each split. You should use the group_ splitting function indicated by @Rui Barradas.

library(tidymodels)

set.seed(217)
df_split <- group_initial_split(df, PatientID, prop = 4/5)
df_training <- training(df_split)
df_testing <- testing(df_split)
df_validation <- group_validation_split(df_training, PatientID, prop = 3/4)

Original reply
In the tidymodels framework you can opt to stratify the sampling using your PatientID variable. The resulting resamples will have equivalent proportions.

To create your desired splits you could first split the data 80:20 training:testing, then split the training set 75:25 into training:validation.

library(tidymodels)

set.seed(217)
df_split <- initial_split(df, prop = 4/5, strata = PatientID)
df_training <- training(df_split)
df_testing <- testing(df_split)
df_validation <- validation_split(df_training, prop = 3/4, strata = PatientID)

Here is a way with package rsample.
First split in test and other data (named train in the code below) keeping all PatientID in the same subsets, then split train.

library(rsample)

set.seed(2023)
g <- group_initial_split(df1, group = PatientID, prop = 0.8)
train <- training(g)
test <- testing(g)
g <- group_initial_split(train, group = PatientID, prop = 3/4)
train <- training(g)
validation <- testing(g)

# check data split proportions
df_list <- list(train = train, validation = validation, test = test)
sapply(df_list, nrow)
#>      train validation       test 
#>        600        199        201

# this shows that all groups belong to one subset only
lapply(df_list, \(x) unique(x[[1]]))
#> $train
#> [1] "P5"  "P9"  "P8"  "P3"  "P10" "P4" 
#> 
#> $validation
#> [1] "P2" "P7"
#> 
#> $test
#> [1] "P1" "P6"

^{Created on 2023-02-17 with reprex v2.0.2}

Test data

set.seed(2023)
p <- sprintf("P%d", 1:10)
n <- 1e3
df1 <- data.frame(
  PatientID = sample(p, n, TRUE),
  x = rnorm(n)
)

^{Created on 2023-02-17 with reprex v2.0.2}

You could use a one-liner that samples one of 1:3 for unique patient IDs and splits df by that.

set.seed(42)
res <- split(df, with(df, ave(id, id, FUN=\(x) sample.int(3, 1, prob=c(.6, .2, .2)))))

Tests:

## test proportions (should approx. be [.6, .2, .2])
proportions(sapply(res, \(x) length(unique(x$id)))) |> round(2)
#    1    2    3 
# 0.53 0.25 0.22 

## test uniqueness
stopifnot(length(Reduce(intersect, lapply(res, `[[`, 'id'))) == 0)

Update

To get more stable proportions, we could use fixed group sizes by repeating 1:3 by vector p.

len <- length(u <- unique(df$id))
p1 <- c(.2, .2)
rlp <- round(len*p1)
p <- c(len - sum(rlp), rlp)
set.seed(42)
a <- setNames(rep.int(1:3, p), sample(u))

res <- split(df, a[match(df$id, names(a))])  ## this line splits the df

proportions(sapply(res, \(x) length(unique(x$id))))
#   1   2   3 
# 0.6 0.2 0.2 

## test uniqueness
stopifnot(length(Reduce(intersect, lapply(res, `[[`, 'id'))) == 0)

Data:

set.seed(42)
n <- 200; np <- 100
df <- data.frame(id=paste0('P', as.integer(as.factor(sort(sample.int(np, n, replace=TRUE))))),
                 les=sample(0:1, n, replace=TRUE),
                 pyr=runif(n))