The standardized length of stay (SLOS) of an intensive care unit (ICU) is the ratio between the length of stay observed in the unit and the length of stay expected for its patients:
\[\mathrm{SLOS} = \frac{\sum \text{observed LoS}}{\sum \text{expected LoS}}\]
A SLOS below 1 means that patients stay in the ICU for less time than expected given their case mix; above 1, for longer. The expected length of stay comes from a prediction model. SLOS offers two ways of getting one:
In both cases, SLOS() then turns patient-level
predictions into unit-level results (Section 4).
SLOS() downloads the pre-trained model on first use (it
needs an internet connection; the file is large, about 1.1 GB, so it can
take a long time; the model is kept for the rest of the R session).
sampled_data identifies each ICU in the column
UnitCode and has the observed length of stay in
UnitLengthStay_trunc.
res <- SLOS(
data = sampled_data,
icu_column = "UnitCode",
outcome = "UnitLengthStay_trunc"
)
head(res$df_unit_slos) # one row per ICU
res$plot_SLOS_obs_prev # observed vs. predicted LoS per ICU
plot(res$funnel_plot) # funnel plotPatient-level predictions and error metrics are also available directly:
The pre-trained model reflects the case mix and the data of the
hospitals it was trained on. For other settings, build a new model. The
workflow has four steps; a complete script is available in
system.file("examples", "retrain_workflow.R", package = "SLOS").
Step 1 – Pre-processing. Removes near-zero variance predictors and highly correlated numeric or categorical predictors, splits the data into training and test sets and imputes missing values with mice.
outcome <- "UnitLengthStay_trunc"
non_predictors <- c("UnitLengthStay_days", "admissionid", "UnitCode")
pp <- slos_pre_processing(
data = sampled_data, outcome = outcome,
corr_cutoff_num = 0.8, corr_cutoff_cat = 0.5,
do_impute = TRUE, include_outcome = TRUE, m = 1, maxit = 5, seed = 123
)
pp$removed # what was removed and why
train_imp <- pp$imputation$train_completed
test_imp <- pp$imputation$test_completedStep 2 – Feature selection (recursive feature elimination).
res_fs <- slos_feature_select_rfe(
data = train_imp[, setdiff(names(train_imp), non_predictors)],
outcome = outcome, method = "treebag", cv_folds = 5, verbose = FALSE
)
res_fs$selected_varsStep 3 – Training. A linear model and a random forest are stacked with a random-forest meta-learner.
keep_cols <- union(res_fs$selected_vars, outcome)
fit <- slos_train_new_model(
train = train_imp[, keep_cols],
test = test_imp,
outcome = outcome,
icu_column = "UnitCode"
)
fit$metrics # patient-level and ICU-level RMSE, MAE and R2Step 4 – Evaluate the units with the new model.
res$df_unit_slos contains, for each ICU, the number of
admissions, the sums of observed and expected length of stay and the
SLOS. Summaries of the distribution across units are straightforward to
compute:
slos_values <- res$df_unit_slos$SLOS
sum(res$df_unit_slos$soma_los_obs) / sum(res$df_unit_slos$soma_los_esp) # overall
median(slos_values)
quantile(slos_values, c(0.25, 0.75))The funnel plot shows each unit’s SLOS against its number of admissions with control limits around the overall SLOS; units outside the limits differ from the overall performance by more than chance would explain.
Peres et al. (2022) https://doi.org/10.1016/j.accpm.2022.101142. Peres et al. (2023) https://doi.org/10.1007/s00134-023-07255-w.