Getting started with SLOS

Overview

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:

  1. use the pre-trained model (Section 2), or
  2. train a new model on your own data (Section 3).

In both cases, SLOS() then turns patient-level predictions into unit-level results (Section 4).

library(SLOS)
data(SampledData)   # loads `sampled_data`, 1000 anonymized admissions

Using the pre-trained model

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 plot

Patient-level predictions and error metrics are also available directly:

pred <- slos_predict_pretrained(sampled_data, outcome = "UnitLengthStay_trunc")
pred$RMSE

Training a model with your own data

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_completed

Step 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_vars

Step 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 R2

Step 4 – Evaluate the units with the new model.

res <- SLOS(test_imp, icu_column = "UnitCode", outcome = outcome, model = fit)

Interpreting the results

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.

References

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.