Shap values with Tidymodels

[Moalu]

Hi!
My shap values seems to be backwards when using xgboost classification in tidymodels. The results implies that a high blood glucose is correlated with lower diabetes risk. I can't make sense of it. Using other frameworks (ex standard xgboost-package) the shap values are logical, but not when using tidymodels. I suppose there is some issue with how I extract the engine or the training data when prepping for shap?

library(tidyverse)
library(tidymodels)
library(SHAPforxgboost)
library(mlbench)

#Loading data
data(PimaIndiansDiabetes)
data <- PimaIndiansDiabetes

#Converting target variable to factor
data$diabetes <- ifelse(data$diabetes == "pos", 1, 0)
data$diabetes <- as.factor(data$diabetes)

#Train/test split
set.seed(1992)
dm_split <- initial_split(data, strata = diabetes, prop = 0.8)
dm_train <- training(dm_split)
dm_test <- testing(dm_split)

#Recipe
dm_rec <- recipe(diabetes ~., data = dm_train) %>%
    step_zv(all_numeric()) 
dm_prep <- prep(dm_rec)

#Model specification
dm_spec <- boost_tree()%>%
  set_engine("xgboost") %>%
  set_mode("classification")

#Workflow
dm_wf <- workflow(
  dm_rec,
  dm_spec
)

#CV-folds
set.seed(1992)
dm_folds <- vfold_cv(data = dm_train, strata = diabetes, v = 5)

dm_res <- 
  dm_wf %>% 
  fit_resamples(
    resamples = dm_folds, 
    metrics = metric_set(
      recall, precision, f_meas, 
      accuracy, kap,
      roc_auc, sens, spec),
    control = control_resamples(save_pred = TRUE)
    ) 

dm_res %>% collect_metrics(summarize = TRUE)
best_auc <- select_best(dm_res, "roc_auc")

#Finalize workflow
dm_fit <- dm_wf %>%
  finalize_workflow(best_auc) %>%
  last_fit(dm_split)

#Preparing data for shap
dm_shap <-
  shap.prep(
    xgb_model = extract_fit_engine(dm_fit),
    X_train = bake(dm_prep,
      has_role("predictor"),
      new_data = NULL,
      composition = "matrix"
    ))

#Shap summary
shap.plot.summary(dm_shap)