Biostat 203B Homework 5

Due Mar 22 @ 11:59PM

Author

Your Name and UID

Predicting ICU duration

Using the ICU cohort mimiciv_icu_cohort.rds you built in Homework 4, develop at least three machine learning approaches (logistic regression with enet regularization, random forest, boosting, SVM, MLP, etc) plus a model stacking approach for predicting whether a patient’s ICU stay will be longer than 2 days. You should use the los_long variable as the outcome. You algorithms can use patient demographic information (gender, age at ICU intime, marital status, race), ICU admission information (first care unit), the last lab measurements before the ICU stay, and first vital measurements during ICU stay as features. You are welcome to use any feature engineering techniques you think are appropriate; but make sure to not use features that are not available at an ICU stay’s intime. For instance, last_careunit cannot be used in your algorithms.

  1. Data preprocessing and feature engineering.

  2. Partition data into 50% training set and 50% test set. Stratify partitioning according to los_long. For grading purpose, sort the data by subject_id, hadm_id, and stay_id and use the seed 203 for the initial data split. Below is the sample code.

set.seed(203)

# sort
mimiciv_icu_cohort <- mimiciv_icu_cohort |>
  arrange(subject_id, hadm_id, stay_id)

data_split <- initial_split(
  mimiciv_icu_cohort, 
  # stratify by los_long
  strata = "los_long", 
  prop = 0.5
  )

  1. Train and tune the models using the training set.

  2. Compare model classification performance on the test set. Report both the area under ROC curve and accuracy for each machine learning algorithm and the model stacking. Interpret the results. What are the most important features in predicting long ICU stays? How do the models compare in terms of performance and interpretability?