ENCoDE, mEasuring skiN Color to correct pulse Oximetry DisparitiEs: skin tone and clinical data from a prospective trial on acute care patients.

Abstract

The wide adoption of pulse oximeters has given clinicians an easy, non-invasive way to measure arterial oxygen saturation. However, evidence suggests that pulse oximeter measurements have a deeper discrepancy in patients of darker skin tones compared to their lighter counterparts. It has been hypothesized that skin tone is the root cause of this phenomenon. However, skin tone as a medical concept has not been extensively studied in acute care.

Our study included patients admitted to Duke University Hospital with pulse oximetry recorded up to 5 minutes prior to arterial blood gas (ABG) measurements. Skin tone was measured across sixteen body locations using administered visual scales (Fitzpatrick, Monk Skin Tone, and Von Luschan), reflectance colorimetry (Delfin SkinColorCatch), and reflectance spectrophotometry (Konica Minolta CM-700D, Variable Spectro 1). IPhone SE 2020 and Google Pixel 4 (Android) image data are available for non-biometric body locations.

One hundred twenty-eight patients are enrolled in this study. A total of 167 skin tone variables and two temperature variables are collected per body location, excluding images, together with ten non-biometric body location images per patient and the associated electronic health record (EHR) data. The ENCoDE project is a comprehensive EHR-linked skin tone database to combat skin tone associate disparities.

Background

Pulse oximetry is a standardized multidisciplinary medical device contributing to therapeutic interventions in patient care. Commonly, pulse oximeters obtain non-invasive measurements through the use of two light-emitting diodes (LEDs) that emit light at the 660 nm (red) and the 940 nm (infrared) wavelengths to measure oxygen saturation of the blood and pulse rate by showing differences in absorption by oxyhemoglobin and deoxyhemoglobin. Pulse oximetry estimates the peripheral oxygen saturation (SpO2) as a proxy for arterial oxygen saturation (SaO2) by analyzing the ratio of absorbance levels at different wavelengths and displaying waveforms to improve the identification of errors and true signals [1].

Ultimately, the primary purpose of pulse oximetry is to deliver prompt and accurate measurements to assist in guiding interventions effectively. However, it has gained incredible attention due to the lack of reliability in SaO₂ and SpO₂ predictions among critically ill patients. Racial and ethnic discrepancies secondary to pulse oximetry inaccuracies limit the identification of disease severity and ultimately hinder patient health outcomes [2-8]. Particularly, inadequate identification of hypoxemia (low oxygen levels or SaO2 less than 88%) is known to cause organ dysfunction and death [9]. This increases the risk of hidden hypoxemia, in which racial and ethnic discrepancies are only discoverable using an arterial blood gas (ABG) to measure SaO₂.

Patients’ skin tone has been hypothesized to be the root cause of such disparity [10-13]. However, skin tone as a medical concept has not been investigated deeply, especially in critically ill patients. To address this problem, we designed a study to prospectively collect skin tone data from critically ill patients with linked electronic health record (EHR) data on different body locations, including all the pulse oximeter locations. Skin tone was measured across sixteen body locations using administered visual scales (Fitzpatrick, Monk Skin Tone, and Von Luschan) [14-15], reflectance colorimetry (Delfin SkinColorCatch), and reflectance spectrophotometry (Konica Minolta CM-700D, Variable Spectro 1).

Methods

The ENCoDE project enrolled patients admitted to inpatient care at Duke University Hospital (Durham, NC, USA) with hospital encounters. 

1. Eligibility. Synchronized ABG-pulse oximetry measurements (as defined by ≥1 pulse oximetry value(s) up to 5 minutes prior to an ABG value captured in Epic Clarity (Epic Systems, Verona, WI), the clinical data warehouse, referenced hereafter as SaO₂-SpO₂ pairs) were required for eligibility. All data, including patients’ consent, measurements, and EHR data, were stored in REDCap [16].
2. Exclusion criteria.  Exclusion criteria included unremovable fingernail polish, admission for a vascular complication (e.g., grafting or stenting), any limb amputations, and other causes of skin discoloration as a result of vitiligo, jaundice, and wounds/bruising.  The exclusion criteria were developed to ensure data quality in all patient locations by avoiding arterial insufficiency or cytopenias, which have the potential to affect skin tone.  
3. Skin tone and temperature measurement tools. In all patients, four skin assessments were conducted: infrared temperature (“clinical”: HoMedics HTD8813C, “general”: IDEAL Model #61-847), administered visual scales (Fitzpatrick Skin Type, Monk Skin Tone, and Von Luschan), colorimetric (Delfin SkinColorCatch), spectrophotometric (Konica-Minolta CM-700d, Variable Spectro 1 Pro), and photography via mobile phone cameras (Google Pixel 4a, iPhone SE 2020) were used. Visual skin scales were printed on 4”x 6” photo paper for reference. 
4. Measurement location. Measurements were taken at sixteen different locations: 8 in the left and right upper extremities (dorsal and ventral finger pad, dorsal and ventral palm), 3 in the head (forehead, inner and outer surface of an earlobe), 1 in the sternum, and 4 in the left and right lower extremities (dorsal and ventral toe). Measurements were collected from patients lying down or in a seated position. For earlobes (and fingers if needed), a black card was placed on the opposite side. 
5. Considerations when collecting measurement.  Considerations are largely dependent on the sample population and location of measurements.  For this particular study, measurements were collected among patients admitted to an intensive care unit (ICU) as well as floor units.  Due to the variability of the hospital environment, the study took into account the criticality of the patient (i.e., patient trajectory), standard care treatments and procedures, and extended lengths of stay in the hospital that may result in loss of skin tone coloration. Additionally, the study utilized two trained personnel to collect measurements to improve the timeliness of data collection and to create an efficient workflow in an environment that is dynamic, multidisciplinary, and fast-paced.
   1. Patients with data monitoring committee 
   2. Criticality of the patient
   3. If the patient was on a temp protocol
   4. Patients who had an extended ICU stay may be less pigmented
   5. Measurements were taken in the ICU as well as floor units
   6. Two people to collect measurements to improve timeliness and to work around the unique workflow of the ICU environment (it is dynamic and fast-paced)
   7. Discoloration of nails (blue, green, purple)
6. Pulse oximeter location. The location of the pulse oximeter was reported as both directly observed by the clinical research coordinator at the time of data collection and surveyed by clinical staff regarding location at the time of the ABG.
7. Image data processing. Each image from mobile devices is filtered for the brightest part by selecting the largest contour above median brightness in greyscale. Next, a mask was created to remove extraneous values. The average and standard deviation on each RGB and L*C*H channel are extracted from the masked images as features associated with the patient’s skin tone. 
8. Patient EHR data. Patients’ hospital encounter information, demographic information, laboratory measurements, and flowsheets were extracted from the EHR system of Duke University Hospital. All data, including image records, are linked to the patient’s encounter using the unique contact serial number (CSN). 
9. De-identification. Data are de-identified strictly following the Health Insurance Portability and Accountability Act (HIPAA) Safe Harbor provision. Each patient's date-time information is randomly shifted with a center at 50 years in the future. Patient encounter numbers and medical records numbers are randomly re-mapped to visit_occurrence_id and patient_id. All the measurement values are first evaluated as strings and then whitelisted before deidentification. Image data for patients’ potential biometric locations are excluded (ventral side of both left and right finger, palm, and toe). Processed features (e.g., mean or standard deviation RGB values) from those images are not considered biometric information and are therefore not excluded. 
10. Missingness. Some patients refused further measurements or had their measurements interrupted by clinical care, leading to some partial missingness in the skin tone and temperature data. This data was not imputed to preserve missingness. 

OMOP conversion

Structured tables are converted into Observational Medical Outcomes Partnership (OMOP) format following OMOP Common Data Model (CDM) v5.4 [17]. Concepts are manually mapped to the OMOP concepts, and two clinically trained persons have confirmed them. 

The Observational Health Data Science and Informatics (OHDSI) community has developed a suite of tools for interacting with and characterizing data in OMOP format. A subset of these tools - namely Achilles, DQD, and Ares - were applied to this dataset and executed a host of targeted SQL queries to calculate high-level metrics about the data (e.g., age distributions) and its quality (e.g., plausibility and model conformance). Results from those queries were then visualized in an interactive web application that enabled iterative feedback and subsequent updates of the dataset.

Data Description

Skin Tone and Temperature Measurements

A total of 167 skin tone features across three administered visual scales, a colorimetry device (Delfin Technologies, SkinColorCatch), two spectrophotometer devices (Konica Minolta CM700d; Variable Inc, Spectro1Pro), and two types of mobile phone cameras (iPhone SE 2020; Google Pixel 4a), are collected. Skin tone measurements as a medical concept have yet to be deeply investigated. Since concepts for skin tone at locations do not yet exist as a standard OMOP vocabulary, we mapped it to a temporary concept ID that can be found in the CONCEPT table. We also measured skin temperature in the same skin location with one clinical temperature measurement device and one general temperature measurement device.  In our datasets, to fully utilize the information on skin tone measurement in different body locations with varying scales of measurement within the OMOP format, we ordered all the skin tone and skin temperature measurements in a long format and joined as part of the MEASUREMENT table. 

Table 1 Example of skin tone data 

The table below is an example of how skin tone measurements are stored in the MEASUREMENT Table and how the concept name can be found in the CONCEPT Table. We mapped each of the skin tone concepts to a 10-digit temporary concept ID starting with “2,” and the concept name can be found on the concept table. 

Body location information and skin tone/skin temperature measurements

Two string values were used to represent the measurement of skin tone and skin temperature at specific body locations. MEASUREMENT_CONCEPT_ID stores skin tone measurements at specific locations, and UNIT_CONCEPT_ID stores information regarding measurement devices with manufacturer information and specific measurement scales. Please see the sample MEASUREMENT table above. 

Image data extraction

Similar to skin tone measurements, image data extractions, such as the average channel intensity of the color red, also lack existing OMOP concepts. Following skin tone and skin temperature measurements, we stored extracted image data in the MEASUREMENT table. The path of the original image can be found in the MEASUREMENT_SOURCE_VALUE.

Table Descriptions

Detailed information can be found at https://github.com/OHDSI/CommonDataModel.

• PERSON.csv: The person table contains person-level demographic data, such as a patient's birth year, race, sex, and ethnicity.
• VISIT_OCCURRENCE.csv: This table contains details about the patient’s hospital visit, including admission and discharge time and the patient’s age when the patient was admitted to the hospital. Note that since all patients recruited only had a single admission, the number of visits is the same as the number of people.
• MEASUREMENT.csv: The measurement table contains EHR data such as labs and vitals. Also, skin tone measurements and skin temperature are taken at different body locations. It can be linked with the VISIT_OCCURRENCE.csv table and the PERSON.csv table with “person_id” and “visit_occurrences_id.” Non-standard concepts can be linked to CONCEPT.csv
• OBSERVATION.csv: The observation table contains information about the pulse oximeter site.
• DEVICE_EXPOSURE.csv: The device_exposure table contains the oxygen delivery devices of the patient.
• PROCEDURE_OCCURRENCE.csv: The procedure_occurrence table contains the record of whether a patient is currently receiving any oxygen therapy treatment.
• OBSERVATION_PERIOD.csv: The observation_period table documents the time period of each patient’s record.
• CONCEPT.csv: This table contains all the skin tone and temperature measurements that we couldn’t map to standard OMOP concepts. We have mapped them to a temporary ID (2, xxx, xxx, xxx) and stored them in this table.

Image records

This folder contains all the processed images, excluding the 6 locations for possible biometric leakage. Within the Image record folder, processed iPhone images and Android images are stored separately. The naming conversion follows the pattern of <person_id>_<location_id>. The table below shows the reference of locaiont_id and the location names. 

Table 2 Skin tone location name reference chart

The table below is a reference chart between Location ID and Location name for all the image records.  

Usage Notes

Reuse Potential 

This dataset has significant reuse potential for skin tone evaluation. To our knowledge, the ENCoDE project will be the first open-source dataset that combines patients’ EHR records with skin tone measurements from multiple devices, including images from smartphones. This dataset could potentially support health disparity research investigating skin tone and health outcomes. The OMOP dataset's data format makes it earlier joinable with similar datasets. 

Limitations

The ENCoDE project was conducted at a single medical center, limiting the generalizability of our findings. Although our cohort included over 40% Black patients, the representation of the darkest skin tones was limited due to demographics in our community.  

GitHub Tutorial Code 

We have created an official GitHub repository for the ENCoDE project. All the complementary tutorial codes and user support for technical problems will be stored here [18]. https://github.com/aiwonglab/ENCoDE_tutorial

Ethics

The Institutional Review Board at Duke University Medical Center reviewed patient information collection and creation of the research resource, which granted a waiver of informed consent and approved the data-sharing initiative under Pro00110842.

Conflicts of Interest

AIW holds equity and management roles in Ataia Medical. AIW is supported by REACH Equity under the National Institute on Minority Health and Health Disparities (NIMHD) of the National Institutes of Health under U54MD012530.

References

1. Jubran A. Pulse oximetry. Crit Care. 2015;19: 272. doi:10.1186/s13054-015-0984-8
2. Wong A-KI, Charpignon M, Kim H, Josef C, de Hond AAH, Fojas JJ, et al. Analysis of Discrepancies Between Pulse Oximetry and Arterial Oxygen Saturation Measurements by Race and Ethnicity and Association With Organ Dysfunction and Mortality. JAMA Network Open. 2021;4: e2131674–e2131674. doi:10.1001/jamanetworkopen.2021.31674
3. Fawzy A, Wu TD, Wang K, Robinson ML, Farha J, Bradke A, et al. Racial and Ethnic Discrepancy in Pulse Oximetry and Delayed Identification of Treatment Eligibility Among Patients With COVID-19. JAMA Intern Med. 2022;182: 730–738. doi:10.1001/jamainternmed.2022.1906
4. Valbuena VSM, Barbaro RP, Claar D, Valley TS, Dickson RP, Gay SE, et al. Racial Bias in Pulse Oximetry Measurement Among Patients About to Undergo Extracorporeal Membrane Oxygenation in 2019-2020: A Retrospective Cohort Study. Chest. 2022;161: 971–978. doi:10.1016/j.chest.2021.09.025
5. Henry NR, Hanson AC, Schulte PJ, Warner NS, Manento MN, Weister TJ, et al. Disparities in Hypoxemia Detection by Pulse Oximetry Across Self-Identified Racial Groups and Associations With Clinical Outcomes. Crit Care Med. 2022;50: 204–211. doi:10.1097/CCM.0000000000005394
6. Jamali H, Castillo LT, Morgan CC, Coult J, Muhammad JL, Osobamiro OO, et al. Racial Disparity in Oxygen Saturation Measurements by Pulse Oximetry: Evidence and Implications. Ann Am Thorac Soc. 2022;19: 1951–1964. doi:10.1513/AnnalsATS.202203-270CME
7. Chesley CF, Lane-Fall MB, Panchanadam V, Harhay MO, Wani AA, Mikkelsen ME, et al. Racial Disparities in Occult Hypoxemia and Clinically Based Mitigation Strategies to Apply in Advance of Technological Advancements. Respir Care. 2022;67: 1499–1507. doi:10.4187/respcare.09769
8. Ward E, Katz MH. Confronting the Clinical Implications of Racial and Ethnic Discrepancy in Pulse Oximetry. JAMA internal medicine. 2022. p. 858. doi:10.1001/jamainternmed.2022.2581
9. Ketcham SW, Sedhai YR, Miller HC, Bolig TC, Ludwig A, Co I, et al. Causes and characteristics of death in patients with acute hypoxemic respiratory failure and acute respiratory distress syndrome: a retrospective cohort study. Crit Care. 2020;24: 391. doi:10.1186/s13054-020-03108-w
10. Sjoding MW, Dickson RP, Iwashyna TJ, Gay SE, Valley TS. Racial Bias in Pulse Oximetry Measurement. N Engl J Med. 2020;383: 2477–2478. doi:10.1056/NEJMc2029240
11. Bickler PE, Feiner JR, Severinghaus JW. Effects of skin pigmentation on pulse oximeter accuracy at low saturation. Anesthesiology. 2005;102: 715–719. doi:10.1097/00000542-200504000-00004
12. Monk E. The Monk Skin Tone Scale. 2023. doi:10.31235/osf.io/pdf4c
13. Shi C, Goodall M, Dumville J, Hill J, Norman G, Hamer O, et al. The accuracy of pulse oximetry in measuring oxygen saturation by levels of skin pigmentation: a systematic review and meta-analysis. BMC Med. 2022;20: 267. doi:10.1186/s12916-022-02452-8
14. Wikipedia contributors. Von Luschan’s chromatic scale. In: Wikipedia, The Free Encyclopedia [Internet]. 14 Jun 2024. Available: https://en.wikipedia.org/w/index.php?title=Von_Luschan%27s_chromatic_scale&oldid=1229032312
15. Hao S, Dempsey K, Matos J, Cox CE, Rotemberg V, Gichoya JW, et al. Utility of skin tone on pulse oximetry in critically ill patients: a prospective cohort study. medRxiv. 2024. doi:10.1101/2024.02.24.24303291
16. Harris PA, Taylor R, Thielke R, Payne J, Gonzalez N, Conde JG. Research electronic data capture (REDCap) - a metadata-driven methodology and workflow process for providing translational research informatics support. J Biomed Inform. 2009;42: 377–381. doi:10.1016/j.jbi.2008.08.010
17. OMOP CDM v5.4. Available: https://ohdsi.github.io/CommonDataModel/cdm54.html
18. GitHub repository for the ENCoDE project. https://github.com/aiwonglab/ENCoDE_tutorial