Collab-CXR Teaching Dataset: Data Dictionary (DASH Format)

STA/GLHLTH 198L

Author

George

Published

September 18, 2026

Warning: package 'tibble' was built under R version 4.3.3

Warning: package 'knitr' was built under R version 4.3.3

A DASH-format data dictionary for the three simplified teaching files in cxr_data.RData (built via a reproducible R pipeline from the full Collab-CXR research file). See simplification_memo_shared_Jul21.qmd for the reasoning behind every simplification choice.

Study_Variables sheet

Code

Dataset_Name <- c(rep("cxr_readings", 9), rep("cxr_cases", 7), rep("cxr_combined", 12))

Variable_Name <- c(
  "patient_id", "pathology", "radiologist_id", "had_ai_help", "had_history",
  "radiologist_prob", "ai_prob", "truth", "time_seconds",
  "patient_id", "pathology", "truth", "ai_prob", "mean_human_prob", "human_accuracy", "n_reads",
  "patient_id", "pathology", "radiologist_id", "had_ai_help", "had_history",
  "radiologist_prob", "ai_prob", "truth", "time_seconds", "mean_human_prob", "human_accuracy", "n_reads"
)

Variable_Label <- c(
  "Patient case identifier", "Pathology being rated", "Radiologist identifier",
  "AI assistance available", "Clinical history available",
  "Radiologist's probability rating", "AI's probability rating",
  "Ground truth label", "Active reading time",
  "Patient case identifier", "Pathology being rated", "Ground truth label",
  "AI's probability rating", "Average radiologist probability",
  "Share of correct human reads", "Number of readings",
  "Patient case identifier", "Pathology being rated", "Radiologist identifier",
  "AI assistance available", "Clinical history available",
  "Radiologist's probability rating", "AI's probability rating",
  "Ground truth label", "Active reading time",
  "Average radiologist probability", "Share of correct human reads", "Number of readings"
)

Variable_Label_Description <- c(
  "Links to the Stanford AIMI CheXmultimodal dataset; patient_id N corresponds to image file N.jpg. One patient appears many times (once per pathology, per radiologist who read that case).",
  "Which of 4 top-level chest X-ray findings this row rates. Simplified from the original 104-code hierarchy down to the 4 top-level ('level 0') categories that also have an AI prediction available.",
  "Opaque ID for the study participant who produced this reading. 227 distinct radiologists appear across the file. Same person can appear under different had_ai_help / had_history combinations.",
  "TRUE if the radiologist was shown the AI's prediction (alg_pred) while making this reading; FALSE if they read unassisted. Derived from the source file's with_ai boolean, not by parsing the treatment code string.",
  "TRUE if the radiologist was given the patient's clinical history/indication when making this reading; FALSE otherwise. Derived from the source file's with_ch boolean.",
  "The radiologist's own probability estimate (0-1) that this pathology is present, reported via an interface slider.",
  "The CheXpert model's probability estimate (0-1) that this pathology is present, based on the image alone. Identical for every radiologist reading the same patient/pathology, whether or not had_ai_help is TRUE for that reading.",
  "Binary (0/1) diagnostic gold standard, from the 5 Mount-Sinai-based expert labelers' averaged rating thresholded at 0.5 (source column gt_binary_simple_us). Constant across every radiologist who read the same case.",
  "Seconds the radiologist spent actively working on this case (source column active_time). Excludes idle/away time.",
  "Same identifier and meaning as in cxr_readings, but each row here is one patient x pathology combination, not one individual reading.",
  "Same 4 categories as in cxr_readings.",
  "Same binary ground truth as cxr_readings; one value per case since it does not depend on who read it.",
  "Same AI probability as cxr_readings; included once per case since it is identical across all readers of that case.",
  "Average of radiologist_prob across all radiologists who read this patient/pathology combination (n_reads readings), rounded to 4 decimal places.",
  "Proportion of individual radiologist reads for this case that were correct, using a 0.5 cutoff on radiologist_prob against truth.",
  "How many individual radiologist readings (rows in cxr_readings) this case/pathology combination has. Varies 35-106 per case because of the study's mixed designs - some cases were read once, others up to 4 times under a within-subject washout design.",
  "Same as cxr_readings.", "Same as cxr_readings.", "Same as cxr_readings.",
  "Same as cxr_readings.", "Same as cxr_readings.", "Same as cxr_readings.", "Same as cxr_readings.",
  "Same as cxr_readings.", "Same as cxr_readings.",
  "Same as cxr_cases's mean_human_prob, joined onto every reading of this case (so the value repeats n_reads times across the readings for a given patient/pathology).",
  "Same as cxr_cases's human_accuracy, joined onto every reading of this case (repeats n_reads times).",
  "Same as cxr_cases's n_reads, joined onto every reading of this case (repeats n_reads times)."
)

Variable_Type <- c(
  "Numeric", "Coded", "Character", "Coded", "Coded", "Numeric", "Numeric", "Coded", "Numeric",
  "Numeric", "Coded", "Coded", "Numeric", "Numeric", "Numeric", "Numeric",
  "Numeric", "Coded", "Character", "Coded", "Coded", "Numeric", "Numeric", "Coded", "Numeric", "Numeric", "Numeric", "Numeric"
)

Variable_Code_List <- c(
  NA, "Pathology_Selected", NA, "Boolean_TrueFalse", "Boolean_TrueFalse", NA, NA, "Binary_01", NA,
  NA, "Pathology_Selected", "Binary_01", NA, NA, NA, NA,
  NA, "Pathology_Selected", NA, "Boolean_TrueFalse", "Boolean_TrueFalse", NA, NA, "Binary_01", NA, NA, NA, NA
)

Value_Range_Minimum <- c(NA, NA, NA, NA, NA, 0, 0, 0, 0, NA, NA, 0, 0, 0, 0, 35, NA, NA, NA, NA, NA, 0, 0, 0, 0, 0, 0, 35)
Value_Range_Maximum <- c(NA, NA, NA, NA, NA, 1, 1, 1, NA, NA, NA, 1, 1, 1, 1, 106, NA, NA, NA, NA, NA, 1, 1, 1, NA, 1, 1, 106)

all_vecs <- list(Variable_Name, Variable_Label, Variable_Label_Description, Variable_Type,
                  Variable_Code_List, Value_Range_Minimum, Value_Range_Maximum, Dataset_Name)
stopifnot(length(unique(sapply(all_vecs, length))) == 1)
stopifnot(!anyNA(Variable_Name), !anyNA(Variable_Label))

study_variables <- tibble(
  `Dataset Name` = Dataset_Name,
  `Variable Name` = Variable_Name,
  `Variable Label` = Variable_Label,
  `Variable Label Description` = Variable_Label_Description,
  `Variable Type` = Variable_Type,
  `Variable Code List` = Variable_Code_List,
  `Value Range Minimum` = Value_Range_Minimum,
  `Value Range Maximum` = Value_Range_Maximum
)

kable(study_variables, caption = "Study_Variables sheet (DASH Codebook format)")

Study_Variables sheet (DASH Codebook format)
Dataset Name	Variable Name	Variable Label	Variable Label Description	Variable Type	Variable Code List	Value Range Minimum	Value Range Maximum
cxr_readings	patient_id	Patient case identifier	Links to the Stanford AIMI CheXmultimodal dataset; patient_id N corresponds to image file N.jpg. One patient appears many times (once per pathology, per radiologist who read that case).	Numeric	NA	NA	NA
cxr_readings	pathology	Pathology being rated	Which of 4 top-level chest X-ray findings this row rates. Simplified from the original 104-code hierarchy down to the 4 top-level (‘level 0’) categories that also have an AI prediction available.	Coded	Pathology_Selected	NA	NA
cxr_readings	radiologist_id	Radiologist identifier	Opaque ID for the study participant who produced this reading. 227 distinct radiologists appear across the file. Same person can appear under different had_ai_help / had_history combinations.	Character	NA	NA	NA
cxr_readings	had_ai_help	AI assistance available	TRUE if the radiologist was shown the AI’s prediction (alg_pred) while making this reading; FALSE if they read unassisted. Derived from the source file’s with_ai boolean, not by parsing the treatment code string.	Coded	Boolean_TrueFalse	NA	NA
cxr_readings	had_history	Clinical history available	TRUE if the radiologist was given the patient’s clinical history/indication when making this reading; FALSE otherwise. Derived from the source file’s with_ch boolean.	Coded	Boolean_TrueFalse	NA	NA
cxr_readings	radiologist_prob	Radiologist’s probability rating	The radiologist’s own probability estimate (0-1) that this pathology is present, reported via an interface slider.	Numeric	NA	0	1
cxr_readings	ai_prob	AI’s probability rating	The CheXpert model’s probability estimate (0-1) that this pathology is present, based on the image alone. Identical for every radiologist reading the same patient/pathology, whether or not had_ai_help is TRUE for that reading.	Numeric	NA	0	1
cxr_readings	truth	Ground truth label	Binary (0/1) diagnostic gold standard, from the 5 Mount-Sinai-based expert labelers’ averaged rating thresholded at 0.5 (source column gt_binary_simple_us). Constant across every radiologist who read the same case.	Coded	Binary_01	0	1
cxr_readings	time_seconds	Active reading time	Seconds the radiologist spent actively working on this case (source column active_time). Excludes idle/away time.	Numeric	NA	0	NA
cxr_cases	patient_id	Patient case identifier	Same identifier and meaning as in cxr_readings, but each row here is one patient x pathology combination, not one individual reading.	Numeric	NA	NA	NA
cxr_cases	pathology	Pathology being rated	Same 4 categories as in cxr_readings.	Coded	Pathology_Selected	NA	NA
cxr_cases	truth	Ground truth label	Same binary ground truth as cxr_readings; one value per case since it does not depend on who read it.	Coded	Binary_01	0	1
cxr_cases	ai_prob	AI’s probability rating	Same AI probability as cxr_readings; included once per case since it is identical across all readers of that case.	Numeric	NA	0	1
cxr_cases	mean_human_prob	Average radiologist probability	Average of radiologist_prob across all radiologists who read this patient/pathology combination (n_reads readings), rounded to 4 decimal places.	Numeric	NA	0	1
cxr_cases	human_accuracy	Share of correct human reads	Proportion of individual radiologist reads for this case that were correct, using a 0.5 cutoff on radiologist_prob against truth.	Numeric	NA	0	1
cxr_cases	n_reads	Number of readings	How many individual radiologist readings (rows in cxr_readings) this case/pathology combination has. Varies 35-106 per case because of the study’s mixed designs - some cases were read once, others up to 4 times under a within-subject washout design.	Numeric	NA	35	106
cxr_combined	patient_id	Patient case identifier	Same as cxr_readings.	Numeric	NA	NA	NA
cxr_combined	pathology	Pathology being rated	Same as cxr_readings.	Coded	Pathology_Selected	NA	NA
cxr_combined	radiologist_id	Radiologist identifier	Same as cxr_readings.	Character	NA	NA	NA
cxr_combined	had_ai_help	AI assistance available	Same as cxr_readings.	Coded	Boolean_TrueFalse	NA	NA
cxr_combined	had_history	Clinical history available	Same as cxr_readings.	Coded	Boolean_TrueFalse	NA	NA
cxr_combined	radiologist_prob	Radiologist’s probability rating	Same as cxr_readings.	Numeric	NA	0	1
cxr_combined	ai_prob	AI’s probability rating	Same as cxr_readings.	Numeric	NA	0	1
cxr_combined	truth	Ground truth label	Same as cxr_readings.	Coded	Binary_01	0	1
cxr_combined	time_seconds	Active reading time	Same as cxr_readings.	Numeric	NA	0	NA
cxr_combined	mean_human_prob	Average radiologist probability	Same as cxr_cases’s mean_human_prob, joined onto every reading of this case (so the value repeats n_reads times across the readings for a given patient/pathology).	Numeric	NA	0	1
cxr_combined	human_accuracy	Share of correct human reads	Same as cxr_cases’s human_accuracy, joined onto every reading of this case (repeats n_reads times).	Numeric	NA	0	1
cxr_combined	n_reads	Number of readings	Same as cxr_cases’s n_reads, joined onto every reading of this case (repeats n_reads times).	Numeric	NA	35	106

Code_Lists sheet

Code

Code_List_Name <- c(
  "Pathology_Selected", "Pathology_Selected", "Pathology_Selected", "Pathology_Selected",
  "Boolean_TrueFalse", "Boolean_TrueFalse",
  "Binary_01", "Binary_01"
)
Value <- c(
  "Any abnormality", "Airspace opacity", "Cardiomediastinal abnormality", "Support device or hardware",
  "TRUE", "FALSE",
  "0", "1"
)
Value_Label <- c(
  "Top-level flag: is any finding present on this image at all? (source pathology code 106_abnormal)",
  "Any airspace opacity finding, e.g. consolidation, infiltrate (source pathology code 2_airspace_opacity)",
  "Any abnormality of the heart/mediastinum silhouette (source pathology code 34_cardiomediastinal_abnormality)",
  "Any tube, line, or other medical hardware finding (source pathology code 78_support_device_hardware)",
  "Yes",
  "No",
  "No / Negative",
  "Yes / Positive"
)
stopifnot(length(Code_List_Name) == length(Value), length(Value) == length(Value_Label))

code_lists <- tibble(
  `Code List Name` = Code_List_Name,
  Value = Value,
  `Value Label` = Value_Label
)

kable(code_lists, caption = "Code_Lists sheet (DASH Codebook format)")

Code_Lists sheet (DASH Codebook format)
Code List Name	Value	Value Label
Pathology_Selected	Any abnormality	Top-level flag: is any finding present on this image at all? (source pathology code 106_abnormal)
Pathology_Selected	Airspace opacity	Any airspace opacity finding, e.g. consolidation, infiltrate (source pathology code 2_airspace_opacity)
Pathology_Selected	Cardiomediastinal abnormality	Any abnormality of the heart/mediastinum silhouette (source pathology code 34_cardiomediastinal_abnormality)
Pathology_Selected	Support device or hardware	Any tube, line, or other medical hardware finding (source pathology code 78_support_device_hardware)
Boolean_TrueFalse	TRUE	Yes
Boolean_TrueFalse	FALSE	No
Binary_01	0	No / Negative
Binary_01	1	Yes / Positive

Cross-check against the live RData file

Code

load("cxr_data.RData")

dict_readings_vars <- study_variables$`Variable Name`[study_variables$`Dataset Name` == "cxr_readings"]
dict_cases_vars <- study_variables$`Variable Name`[study_variables$`Dataset Name` == "cxr_cases"]
dict_combined_vars <- study_variables$`Variable Name`[study_variables$`Dataset Name` == "cxr_combined"]

check_match <- function(dict_vars, data_vars, label) {
  missing_from_dict <- setdiff(data_vars, dict_vars)
  missing_from_data <- setdiff(dict_vars, data_vars)
  if (length(missing_from_dict) > 0) {
    message(label, ": columns in data but not in dictionary: ", paste(missing_from_dict, collapse = ", "))
  }
  if (length(missing_from_data) > 0) {
    message(label, ": columns in dictionary but not in data: ", paste(missing_from_data, collapse = ", "))
  }
  if (length(missing_from_dict) == 0 && length(missing_from_data) == 0) {
    message(label, ": dictionary and data columns match exactly (", length(data_vars), " variables).")
  }
}

check_match(dict_readings_vars, colnames(cxr_readings), "cxr_readings")

cxr_readings: dictionary and data columns match exactly (9 variables).

Code

check_match(dict_cases_vars, colnames(cxr_cases), "cxr_cases")

cxr_cases: dictionary and data columns match exactly (7 variables).

Code

check_match(dict_combined_vars, colnames(cxr_combined), "cxr_combined")

cxr_combined: dictionary and data columns match exactly (12 variables).