Skip to contents

BenchmarkEHRShot: Benchmark predictive tasks using EHRShot

Source: R/Task_BenchmarkEHRShot.R
BenchmarkEHRShot.Rd

Task class for running benchmark evaluations on the EHRShot dataset. Supports multiple categories of prediction tasks including operational outcomes, lab value predictions, new diagnoses, and medical imaging findings.

Details

The BenchmarkEHRShot task class implements predictive modeling tasks from the EHRShot benchmark. Each task uses clinical codes from the ehrshot table as features and generates predictions based on task-specific labels.

Operational Outcomes (Binary Classification):

  • guo_los: Predicts if length of stay exceeds a threshold

  • guo_readmission: Predicts hospital readmission

  • guo_icu: Predicts ICU admission

Lab Values (Multiclass Classification):

  • lab_thrombocytopenia: Predicts low platelet count severity

  • lab_hyperkalemia: Predicts high potassium level severity

  • lab_hypoglycemia: Predicts low blood sugar severity

  • lab_hyponatremia: Predicts low sodium level severity

  • lab_anemia: Predicts anemia severity

New Diagnoses (Binary Classification):

  • new_hypertension: Predicts new hypertension diagnosis

  • new_hyperlipidemia: Predicts new hyperlipidemia diagnosis

  • new_pancan: Predicts new pancreatic cancer diagnosis

  • new_celiac: Predicts new celiac disease diagnosis

  • new_lupus: Predicts new lupus diagnosis

  • new_acutemi: Predicts new acute myocardial infarction diagnosis

Medical Imaging (Multilabel Classification):

  • chexpert: Predicts multiple chest X-ray findings simultaneously (14 possible findings from the CheXpert dataset)

Features

The task uses clinical codes from the ehrshot table as features. Optionally, you can filter events by OMOP table type using the omop_tables parameter to focus on specific types of clinical data (e.g., conditions, procedures, drugs).

Data Split

The task automatically assigns samples to train/validation/test splits based on the splits table in the EHRShot dataset.

See also

EHRShotDataset, BaseTask

Super class

RHealth::BaseTask -> BenchmarkEHRShot

Public fields

task

Name of the specific benchmark task.

omop_tables

Optional vector of OMOP table names to filter events.

max_seq_length

Maximum sequence length for codes.

truncation_count

Counter for truncated sequences.

task_name

Full task name (BenchmarkEHRShot/task).

input_schema

Input schema specification.

output_schema

Output schema specification.

tasks_by_category

List of available tasks organized by category.

Methods

Public methods

Method new()

Initialize the BenchmarkEHRShot task.

Usage

BenchmarkEHRShot$new(task, omop_tables = NULL, max_seq_length = 2000)

Arguments

task

Character. The specific benchmark task to run. Must be one of:

  • Operational outcomes: "guo_los", "guo_readmission", "guo_icu"

  • Lab values: "lab_thrombocytopenia", "lab_hyperkalemia", "lab_hypoglycemia", "lab_hyponatremia", "lab_anemia"

  • New diagnoses: "new_hypertension", "new_hyperlipidemia", "new_pancan", "new_celiac", "new_lupus", "new_acutemi"

  • Medical imaging: "chexpert"

omop_tables

Optional character vector. Names of OMOP tables to filter input events. If specified, only events from ehrshot with matching omop_table values will be included as features. Common values include: "condition_occurrence", "procedure_occurrence", "drug_exposure", "measurement", "observation".

max_seq_length

Integer. Maximum sequence length for clinical codes. Sequences longer than this will be truncated to the most recent codes. Default is 2000. Set to NULL for no limit (not recommended for large datasets).

Returns

A new BenchmarkEHRShot task object.

Examples

\dontrun{
# Basic task initialization
task <- BenchmarkEHRShot$new(task = "guo_los")

# With OMOP table filtering
task <- BenchmarkEHRShot$new(
  task = "new_hypertension",
  omop_tables = c("condition_occurrence", "drug_exposure")
)

# With custom max sequence length
task <- BenchmarkEHRShot$new(task = "guo_los", max_seq_length = 5000)
}

Method pre_filter()

Pre-filter hook to retain only necessary columns and optionally filter by OMOP tables.

Usage

BenchmarkEHRShot$pre_filter(df)

Arguments

df

A lazy query or data frame containing all events.

Returns

A filtered data frame.

Method call()

Process a single patient to generate samples.

Usage

BenchmarkEHRShot$call(patient)

Arguments

patient

A Patient object.

Returns

A list of samples, where each sample contains:

  • patient_id: Patient identifier

  • feature: Vector of clinical codes

  • label: Label value (type depends on task)

  • split: Data split assignment ("train", "val", or "test")

Method clone()

The objects of this class are cloneable with this method.

Usage

BenchmarkEHRShot$clone(deep = FALSE)

Arguments

deep

Whether to make a deep clone.

Examples

if (FALSE) { # \dontrun{
library(RHealth)

# Example 1: Binary classification task (operational outcome)
dataset <- EHRShotDataset$new(
  root = "/path/to/ehrshot",
  tables = c("ehrshot", "splits", "guo_los"),
  dev = TRUE
)
task <- BenchmarkEHRShot$new(task = "guo_los")
samples <- dataset$set_task(task = task)

# Example 2: Multiclass classification task (lab values)
dataset <- EHRShotDataset$new(
  root = "/path/to/ehrshot",
  tables = c("ehrshot", "splits", "lab_thrombocytopenia"),
  dev = TRUE
)
task <- BenchmarkEHRShot$new(task = "lab_thrombocytopenia")
samples <- dataset$set_task(task = task)

# Example 3: Multilabel classification task (medical imaging)
dataset <- EHRShotDataset$new(
  root = "/path/to/ehrshot",
  tables = c("ehrshot", "splits", "chexpert"),
  dev = FALSE
)
task <- BenchmarkEHRShot$new(task = "chexpert")
samples <- dataset$set_task(task = task)

# Example 4: Filter by specific OMOP tables
task <- BenchmarkEHRShot$new(
  task = "new_hypertension",
  omop_tables = c("condition_occurrence", "drug_exposure")
)
samples <- dataset$set_task(task = task)
} # }


## ------------------------------------------------
## Method `BenchmarkEHRShot$new`
## ------------------------------------------------

if (FALSE) { # \dontrun{
# Basic task initialization
task <- BenchmarkEHRShot$new(task = "guo_los")

# With OMOP table filtering
task <- BenchmarkEHRShot$new(
  task = "new_hypertension",
  omop_tables = c("condition_occurrence", "drug_exposure")
)

# With custom max sequence length
task <- BenchmarkEHRShot$new(task = "guo_los", max_seq_length = 5000)
} # }