Appendix A — Project Proposal

Undetected and Undertreated: Racial and Gender Disparities in Hidden Hypoxemia and Their Economic Consequences

Team Members

Nayla Trigueros Ortiz (Style Lead)
Patricia Escobar Contreras (Point of Contact)
William Acosta Lora (Technical Lead)

Project Description

We are interested in exploring racial and gender disparities in access, treatment, and healthcare outcomes in the medical industry for patients. Specifically, we focus on hidden hypoxemia, a condition where pulse oximeters fail to accurately detect low blood oxygen levels in patients with darker skin tones. This makes it a measurable, concrete case study of medical device bias in action. We use it to examine how that bias translates into unequal care and unequal economic burden for patients of color.

Hidden hypoxemia through pulse oximeter bias is a particularly powerful entry point because it is not a question of patient behavior or lifestyle — it is a direct, measurable failure of medical technology. Research has documented that Black and Hispanic patients are significantly more likely to experience undetected hypoxemia, which has been linked to delayed eligibility for critical treatments and worse outcomes. The economic consequences of this — higher costs, longer stays, greater morbidity — fall disproportionately on communities of color, making this both a health equity and an economic justice issue.

Research Questions

To guide our analysis, we will answer the following research questions:

How does pulse oximetry accuracy vary across skin tone, race, and gender, and what is the magnitude of hidden hypoxemia disparities?

What are the economic consequences of respiratory misdiagnosis and delayed treatment for patients of color in hospital settings?

Do race, gender, and insurance status intersect to compound economic disparities in respiratory care?

Motivation and Inspiration

As international students and people of color from different countries across Latin America, we bring perspectives shaped by healthcare systems outside the United States. Having navigated both our home countries’ medical institutions and the U.S. system, we have noticed firsthand how race and skin tone influence the quality of care patients receive, and how those dynamics shift depending on the country and context. Moving to a predominantly white country has made these disparities more visible to us, not less. The experience of being a patient, or watching family members be patients, in systems that were not designed with us in mind gives this project a personal dimension alongside its academic one. We want to use data to examine what we have observed anecdotally, and to measure the real consequences that medical device bias and systemic inequity have on patients of color in U.S. hospital settings.

Dataset

Our project combines two datasets: the OpenOximetry dataset (UCSF Hypoxia Lab), which gives us paired pulse oximeter and blood gas measurements stratified by skin tone and race, and the NY SPARCS Inpatient Discharge dataset, which provides hospital-level economic outcomes (total charges, length of stay, costs, mortality) broken down by race, gender, and diagnosis. Together, they let us connect the clinical disparity to its downstream economic consequences.

We use these two complementary datasets that together let us measure both the clinical disparity (who gets misdiagnosed) and the economic consequences (who pays more and stays longer).

Primary Dataset: NY SPARCS Hospital Inpatient Discharges (2021)

The New York State Statewide Planning and Research Cooperative System (SPARCS) [https://health.data.ny.gov/resource/tg3i-cinn.csv?$limit=10000] is a free, publicly available hospital discharge database from the NY Department of Health. The 2021 de-identified file contains over 2 million patient-level discharge records with race, ethnicity, gender, diagnosis, total charges, total costs, length of stay, insurance type, and severity of illness. The dataset is downloaded directly from the NY Health Data open portal, no login or account needed. We load it locally and filter to respiratory diagnoses only using the ccsr_diagnosis_code column, which uses CCSR (Clinical Classifications Software Refined) codes. All respiratory diagnoses begin with “RSP”.

The following output shows the structure of the filtered respiratory dataset, including variable names, types, and a preview of values.

Show code

library(tidyverse)

# Source: https://health.data.ny.gov/resource/tg3i-cinn.csv?$limit=10000
# Downloaded from NY Health Data open portal 

sparcs_raw <- read_csv("../data/raw/tg3i-cinn.csv", show_col_types = FALSE)

# Filter to respiratory diagnoses only (all CCSR codes starting with RSP)
# and clean numeric columns
sparcs <- sparcs_raw |>
  filter(str_starts(ccsr_diagnosis_code, "RSP")) |>
  mutate(
    length_of_stay = as.numeric(length_of_stay),
    total_charges  = as.numeric(total_charges),
    total_costs    = as.numeric(total_costs)
  )

glimpse(sparcs)

Rows: 278
Columns: 33
$ hospital_service_area          <chr> "New York City", "New York City", "New …
$ hospital_county                <chr> "Bronx", "Bronx", "Bronx", "Bronx", "Br…
$ operating_certificate_number   <dbl> 7000006, 7000006, 7000008, 7000008, 700…
$ permanent_facility_id          <chr> "001168", "003058", "001172", "001172",…
$ facility_name                  <chr> "Montefiore Medical Center-Wakefield Ho…
$ age_group                      <chr> "50 to 69", "70 or Older", "50 to 69", …
$ zip_code_3_digits              <chr> "104", "104", "104", "104", "104", "104…
$ gender                         <chr> "F", "F", "M", "M", "F", "M", "M", "M",…
$ race                           <chr> "Other Race", "Other Race", "Other Race…
$ ethnicity                      <chr> "Spanish/Hispanic", "Spanish/Hispanic",…
$ length_of_stay                 <dbl> 3, 11, 1, 2, 3, 1, 2, 1, 2, 1, 2, 3, 2,…
$ type_of_admission              <chr> "Emergency", "Emergency", "Emergency", …
$ patient_disposition            <chr> "Short-term Hospital", "Hospice - Medic…
$ discharge_year                 <dbl> 2021, 2021, 2021, 2021, 2021, 2021, 202…
$ ccsr_diagnosis_code            <chr> "RSP009", "RSP010", "RSP008", "RSP002",…
$ ccsr_diagnosis_description     <chr> "Asthma", "Aspiration pneumonitis", "Ch…
$ ccsr_procedure_code            <chr> "ADM017", "ESA004", "ESA004", NA, "ESA0…
$ ccsr_procedure_description     <chr> "ADMINISTRATION OF NUTRITIONAL AND ELEC…
$ apr_drg_code                   <chr> "141", "137", "140", "139", "140", "139…
$ apr_drg_description            <chr> "ASTHMA", "MAJOR RESPIRATORY INFECTIONS…
$ apr_mdc_code                   <chr> "04", "04", "04", "04", "04", "04", "04…
$ apr_mdc_description            <chr> "DISEASES AND DISORDERS OF THE RESPIRAT…
$ apr_severity_of_illness_code   <dbl> 3, 4, 3, 3, 3, 3, 1, 4, 2, 1, 3, 2, 1, …
$ apr_severity_of_illness        <chr> "Major", "Extreme", "Major", "Major", "…
$ apr_risk_of_mortality          <chr> "Major", "Extreme", "Major", "Moderate"…
$ apr_medical_surgical           <chr> "Medical", "Medical", "Medical", "Medic…
$ payment_typology_1             <chr> "Medicare", "Medicare", "Medicaid", "Me…
$ payment_typology_2             <chr> "Medicaid", "Private Health Insurance",…
$ payment_typology_3             <chr> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,…
$ birth_weight                   <chr> NA, NA, NA, NA, NA, NA, NA, NA, NA, NA,…
$ emergency_department_indicator <chr> "Y", "Y", "Y", "Y", "Y", "Y", "Y", "Y",…
$ total_charges                  <dbl> 77183.27, 219857.19, 19207.88, 15297.87…
$ total_costs                    <dbl> 12057.91, 38291.05, 12276.24, 9777.25, …

The table below lists all respiratory diagnosis codes present in the data, sorted by how frequently each appears in the dataset

Show code

# Respiratory diagnosis codes present in the data
sparcs |>
  count(ccsr_diagnosis_code, ccsr_diagnosis_description) |>
  arrange(desc(n))

# A tibble: 14 × 3
   ccsr_diagnosis_code ccsr_diagnosis_description                              n
   <chr>               <chr>                                               <int>
 1 RSP008              Chronic obstructive pulmonary disease and bronchie…    57
 2 RSP009              Asthma                                                 51
 3 RSP002              Pneumonia (except that caused by tuberculosis)         49
 4 RSP012              Respiratory failure; insufficiency; arrest             49
 5 RSP010              Aspiration pneumonitis                                 20
 6 RSP014              Pneumothorax                                           10
 7 RSP006              Other specified upper respiratory infections            9
 8 RSP011              Pleurisy, pleural effusion and pulmonary collapse       6
 9 RSP016              Other specified and unspecified lower respiratory …     6
10 RSP004              Acute and chronic tonsillitis                           5
11 RSP005              Acute bronchitis                                        5
12 RSP007              Other specified and unspecified upper respiratory …     5
13 RSP017              Postprocedural or postoperative respiratory system…     4
14 RSP015              Mediastinal disorders                                   2

The table below shows the racial and ethnic breakdown of respiratory patients in the dataset, sorted by count.

Show code

# Racial and ethnic breakdown of respiratory patients
sparcs |>
  count(race, ethnicity) |>
  arrange(desc(n))

# A tibble: 12 × 3
   race                   ethnicity             n
   <chr>                  <chr>             <int>
 1 Black/African American Not Span/Hispanic    77
 2 White                  Not Span/Hispanic    71
 3 Other Race             Spanish/Hispanic     67
 4 Other Race             Not Span/Hispanic    18
 5 Other Race             Unknown              17
 6 Black/African American Unknown              11
 7 Black/African American Spanish/Hispanic      4
 8 Multi-racial           Not Span/Hispanic     4
 9 White                  Unknown               4
10 White                  Spanish/Hispanic      3
11 Black/African American Multi-ethnic          1
12 Multi-racial           Unknown               1

Key variables:

Variable	Description
`race` / `ethnicity`	Patient self-reported race and Hispanic/Latino ethnicity
`gender`	Patient sex
`length_of_stay`	Days hospitalized
`total_charges`	Amount billed by the hospital
`total_costs`	Estimated actual cost of care
`payment_typology_1`	Insurance type (Medicaid, Medicare, Private Health Insurance, Self-Pay)
`apr_severity_of_illness_code`	1–4 scale of illness severity at admission
`apr_severity_of_illness`	Severity label: Minor, Moderate, Major, Extreme
`apr_risk_of_mortality`	Risk label: Minor, Moderate, Major, Extreme
`ccsr_diagnosis_code`	CCSR diagnosis code — RSP codes = respiratory diagnoses
`ccsr_diagnosis_description`	Plain-text diagnosis name
`patient_disposition`	Outcome: home, expired, transferred, etc.
`hospital_county`	County of the treating hospital
`emergency_department_indicator`	Whether patient came through the ER

Source: NY State Department of Health — Health Data NY
Download URL: https://health.data.ny.gov/resource/tg3i-cinn.csv?$limit=10000
License: Public domain / open government data — no login required

Secondary Dataset: OpenOximetry Dataset (UCSF Hypoxia Lab)

The OpenOximetry dataset [https://physionet.org/content/openox-repo/1.1.1/] is a prospective dataset from the University of California San Francisco Hypoxia Lab. It contains paired SpO₂ (pulse oximeter) and SaO₂ (arterial blood gas gold standard) measurements alongside objective skin tone measurements using the Fitzpatrick scale, Monk scale, and spectrophotometry. This dataset is stored locally as five relational CSVs. The following output shows the structure of the merged oximetry dataset after joining all five relational tables and computing key outcome variables.

Show code

# Load locally stored OpenOximetry tables (place CSVs in /data folder)
patients          <- read_csv("data/patient.csv")
encounter         <- read_csv("data/encounter.csv")
pulseoximeter     <- read_csv("data/pulseoximeter.csv")
bloodgas          <- read_csv("data/bloodgas.csv")
spectrophotometer <- read_csv("data/spectrophotometer.csv")

# Merge tables and compute key outcome variables
oximetry <- pulseoximeter |>
  inner_join(
    bloodgas |> select(encounter_id, sample, so2, ph, pco2, po2, thb),
    by = c("encounter_id", "sample_number" = "sample")
  ) |>
  rename(SpO2 = saturation, SaO2 = so2) |>
  left_join(
    encounter |> select(
      encounter_id, patient_id, age_at_encounter,
      fitzpatrick, monk_forehead, monk_dorsal,
      monk_palmar, monk_upper_arm
    ),
    by = "encounter_id"
  ) |>
  left_join(
    patients |> select(patient_id, assigned_sex, race, ethnicity),
    by = "patient_id"
  ) |>
  mutate(
    bias             = SpO2 - SaO2,
    occult_hypoxemia = SpO2 >= 88 & SaO2 < 88,
    skin_group       = cut(
      fitzpatrick,
      breaks         = c(0, 2, 4, 6),
      labels         = c("Light", "Medium", "Dark"),
      include.lowest = TRUE
    )
  )

glimpse(oximetry)

Key variables:

Variable	Description
`SpO2`	Blood oxygen saturation measured by pulse oximeter
`SaO2`	True blood oxygen saturation from arterial blood gas
`bias`	Difference between SpO2 and SaO2 (SpO2 minus SaO2)
`occult_hypoxemia`	Whether SpO2 is at or above 88% while SaO2 is below 88%
`fitzpatrick`	Fitzpatrick skin tone scale score (1 to 6)
`monk_forehead`	Monk skin tone score at forehead
`monk_dorsal`	Monk skin tone score at dorsal hand
`monk_palmar`	Monk skin tone score at palm
`monk_upper_arm`	Monk skin tone score at upper arm
`race`	Patient self-reported race
`ethnicity`	Patient self-reported ethnicity
`assigned_sex`	Patient sex assigned at birth
`age_at_encounter`	Patient age at time of measurement

Source: OpenOximetry Project, UCSF Hypoxia Lab
Access: Data Use Agreement required at the OpenOximetry portal

Communication Plan

Primary channels: WhatsApp and Google Chat
Expected response time: Within 1 business day on weekdays
Standing meeting: Thursdays, 3:00–4:30 PM (as needed throughout the project)

Roles

Role	Team Member
Point of Contact — emails instructor and preceptors	Patricia Escobar Contreras
Technical Lead — oversees codebase and Quarto website	William Acosta Lora
Style Lead — enforces COMP/STAT 212 Code Styling Guide	Nayla Trigueros Ortiz

All three team members are expected to contribute equally to data analysis, visualization, and writing.

Conflict Resolution

Communication will be key. If any of the members will be late to a commitment or misses a deadline, they should feel comfortable communicating that to the team and we will redistribute the workload accordingly. If anyone has to take on more work for a deadline due to a re-adjustment, we will adjust back for the next deadline so that the responsibilities remain balanced.

For disagreements where the team is split, we will discuss openly during our Thursday meeting. Ties on project direction will be resolved by a short async team vote with a 24-hour window, deferring to the lead most relevant to the decision (e.g., technical disagreements go to the Technical Lead, formatting to the Style Lead).

Implementation Plan

The following tasks will also be added as GitHub Issues in the project repository, with role and priority labels.

#	Task	Due	Lead
1	Set up GitHub repo and Quarto project website skeleton	Week 1	William
2	Submit this proposal	Week 2	Patricia
3	EDA on SPARCS: charges, LOS, and mortality by race, gender, and insurance	Week 3	Patricia + William
4	EDA on OpenOximetry: SpO₂ bias and occult hypoxemia rates by skin tone and race	Week 3	William + Nayla
5	Literature review write-up: clinical and economic evidence for hidden hypoxemia disparities	Week 3–4	Nayla
6	Visualizations: cost and LOS disparities by race and gender from SPARCS	Week 4	Patricia
7	Visualizations: bias distributions and occult hypoxemia rates by skin tone from OpenOximetry	Week 4	William
8	Intersectional analysis: race x gender x insurance interaction effects on costs	Week 5	All
9	Draft results and interpretation sections	Week 6	All
10	Style pass: code review against COMP/STAT 212 Code Styling Guide	Week 7	Nayla
11	Final write-up, website polish, and presentation prep	Week 8	All