Abstract
Pulse oximeters measure peripheral arterial oxygen saturation (SpO<sub>2</sub>) noninvasively, while the gold standard (SaO<sub>2</sub>) involves arterial blood gas measurement. There are known racial and ethnic disparities in their performance. BOLD is a dataset that aims to underscore the importance of addressing biases in pulse oximetry accuracy, which disproportionately affect darker-skinned patients. The dataset was created by harmonizing three Electronic Health Record databases (MIMIC-III, MIMIC-IV, eICU-CRD) comprising Intensive Care Unit stays of US patients. Paired SpO<sub>2</sub> and SaO<sub>2</sub> measurements were time-aligned and combined with various other sociodemographic and parameters to provide a detailed representation of each patient. BOLD includes 49,099 paired measurements, within a 5-minute window and with oxygen saturation levels between 70-100%. Minority racial and ethnic groups account for ~25% of the data - a proportion seldom achieved in previous studies. The codebase is publicly available. Given the prevalent use of pulse oximeters in the hospital and at home, we hope that BOLD will be leveraged to develop debiasing algorithms that can result in more equitable healthcare solutions.
Introduction: Addressing "Hidden Hypoxemia"
Pulse oximeters are fundamental tools in clinical practice, yet their reliance on light absorption introduces a critical flaw: they are often less accurate for patients with darker skin tones. Higher melanin levels can lead to an overestimation of oxygen saturation, creating a state of "hidden hypoxemia." In this scenario, a device may report a safe oxygen level while the patient's actual arterial blood oxygen is dangerously low.
The primary obstacle to solving this problem has been the lack of large-scale, standardized datasets that pair non-invasive $SpO_2$ readings with the "gold standard" invasive arterial blood gas ($SaO_2$) measurements across diverse populations. The BOLD (Blood-gas and Oximetry Linked Dataset) project addresses this gap by harmonizing data from three major ICU databases into a single, open-access resource.
Methodology and Data Harmonization
The creation of BOLD involved a rigorous pipeline to extract and synchronize data from MIMIC-III, MIMIC-IV, and eICU-CRD. A central challenge in Electronic Health Record (EHR) research is the temporal misalignment of different measurements. To ensure physiological relevance, the authors implemented a strict 5-minute window for pairing $SpO_2$ and $SaO_2$ readings.
Beyond oxygen levels, the dataset includes a comprehensive set of "past covariates" (e.g., vitals, lab values, and SOFA scores from the previous 24 hours) and "outcome covariates" (e.g., future SOFA scores, in-hospital mortality, and length of stay). This holistic approach allows researchers to investigate not just the measurement error itself, but also its impact on clinical outcomes.
Dataset Composition and Validation
The final BOLD dataset comprises 49,099 unique measurement pairs. Notably, minority racial and ethnic groups represent approximately 25% of the data, including 5,467 Black patients, 2,436 Hispanic/Latino patients, and 1,061 Asian patients. This level of representation provides the statistical power necessary for meaningful subgroup analysis.
To ensure technical integrity, the authors validated the dataset against physiological models like the oxyhemoglobin dissociation curve. They also utilized Bland-Altman plots to analyze the systematic bias across different saturation levels and demographic groups.
Key Insights
- Optimal Pairing Windows: Analysis of Root Mean Squared Error (RMSE) confirmed that a 5-minute window minimizes noise while maintaining a sufficient sample size. Larger windows were found to introduce significant error, particularly for minority groups.
- Operationalizing Bias Study: By providing high-fidelity pairs alongside demographic data, BOLD allows for the direct quantification of hidden hypoxemia under real-world ICU conditions.
- Open Science Commitment: The entire processing pipeline—from BigQuery SQL queries to Python-based harmonization—is publicly available on GitHub, ensuring reproducibility and encouraging further community-driven research.
Limitations and Future Directions
While BOLD is a significant step forward, it relies on self-reported race and ethnicity as a proxy for skin tone, as objective measures like the Fitzpatrick scale are rarely recorded in clinical settings. Additionally, there may be inherent sampling imbalances in how often different patient groups receive arterial blood gas draws.
Future work aims to expand BOLD by including raw waveform data from oximeters and extending the harmonization methodology to other biometric parameters, such as body temperature and blood pressure, to further improve the equity of medical monitoring technologies.
Reference: Matos, J., et al. (2024). BOLD: Blood-gas and Oximetry Linked Dataset. Scientific Data. https://doi.org/10.1038/s41597-024-03225-z
Summary updated