AI-Driven SaO2 prediction from pulse oximetry and electronic health records Authors: JiWon Woo, Orian Stapleton, Jay Luo, Chao Cheng Chuang, Yolanda Su, Sreenidhi Sankararaman, Esanika Mukherjee, Nikita Sivakumar, Katherine Calligy, Summer Duffy, Rebecca Mosier, Joseph Greenstein, Casey Overby Taylor, Danielle Gottlieb Sen Published: 2025 ABSTRACT AbstractObjectiveTo address limitations in pulse oximetry accuracy associated with low saturation and melanin content using machine learning (ML) and to compare our results to standard pulse oximetry and gold standard arterial blood gas readings.BackgroundOxygen saturation is traditionally measured through the gold standard arterial blood gas (SaO2) or pulse oximetry (SpO2), which approximates SaO2 using light absorption patterns. However, SpO2 has been shown to overestimate oxygen saturation, particularly in individuals with darker skin tone, leading to hidden hypoxemia and delayed medical interventions.MethodWe developed a machine learning (ML) model trained on the BOLD dataset, integrating patient data from eICU, MIMIC-III, and MIMIC-IV (n = 49,093). With 64 clinical features and 2 outcomes (SaO2 and hidden hypoxemia events), we trained regression ML models (linear regression, random forest, and XGBoost) to predict SaO2, minimizing the mean squared error between predicted SaO2 and ground truth SaO2. We used our test data to compare model performance to standard SaO2 measurement with accuracy root mean square error (Arms), R2, and change in hidden hypoxemia events. We used SHapley Additive exPlanations (SHAP) to rank important features for SaO2 prediction.ResultsThe XGBoost-Vanilla model improved Arms to 3.3% from a baseline of 4.1%. In low SaO2 patients (SaO2 < 90%), the accuracy of pulse oximetry was heavily compromised with Arms of 10.3%; the linear regression model with weighted loss was able to reduce Arms to 8.6%. We found that SpO2, creatinine level, mean corpuscular hemoglobin level, respiratory rate, and age at admission were the leading features driving the SaO2 prediction.ConclusionThese findings suggest that ML-based models can enhance the accuracy of the standard SpO2. Further investigation is warranted to address SpO₂ inaccuracies and bias.Supplementary InformationThe online version contains supplementary material available at 10.1186/s13040-025-00511-3.Keywords: Pulse oximetry, Machine learning, Hidden hypoxemia, Health equity Introduction Oxygen saturation can be measured directly from arterial blood (SaO2) to identify hypoxemic individuals. Although SaO2 is considered the ‘gold standard’ oxygen saturation measurement, obtaining arterial blood requires an invasive procedure. As a result, pulse oximetry (SpO2) was developed to noninvasively approximate the gold standard SaO2 using light absorption patterns [1]. SpO2 has played an essential role in the early detection and management of respiratory diseases. Normal SpO2 level ranges between 95% and 100% [2]. Hypoxemic conditions stemming from low oxygen saturation can increase the likelihood of organ damage and pose significant health risks. Recent studies reveal significant concerns regarding the accuracy of pulse oximetry, particularly among individuals with darker skin [3–5]. Research indicates that melanin, the pigment responsible for skin color, can scatter light used in pulse oximetry, potentially leading to the overestimation of SaO2 levels [5]. This overestimation can lead to hidden hypoxemia (HH), where SaO2 is less than 88%, yet SpO2 is more than or equal to 88% [3]. Such discrepancies pose grave risks as they may delay crucial medical treatments by masking severe oxygen deficits. A seminal study by the University of Michigan Hospital analyzed paired SpO2 and SaO2 measurements across diverse patient groups and uncovered a near-triple incidence of undetected hypoxemia in black patients compared to white patients after adjusting for factors like age, sex, and cardiovascular health [6]. This disparity emphasizes the need to improve pulse oximetry technology to accurately reflect oxygen saturation values across different skin tones, thereby mitigating healthcare inequities. As healthcare moves to smarter and less invasive monitoring systems, pulse oximetry remains in its original form. Without a reliable method to account for how variations in skin tone affect SpO2 measurements, a disproportionate number of people of color are at risk of receiving incorrect estimations of the severity of their conditions. Addressing this error is not only a matter of technological improvement but also one of providing equitable healthcare, emphasizing the critical need for diagnostic accuracy in modern era of precision medicine. Using the MIMIC-IV database, previous efforts have been made to predict SaO2 from SpO2 using machine learning models, yielding mixed results [7]. The BOLD dataset was created to enable the analyses of paired arterial blood gas and pulse oximetry data in an effort to reduce racial disparities in these measurements [8, 9]. We aim to use this curated dataset to calibrate SpO2 with SaO2 and to evaluate their relationship between patient characteristics and estimation error in SpO2 using various machine learning models. This study was performed to test our calibrated SpO2 accuracy compared to the SaO2 and investigate the features associated with SaO2 error. Methods Source of data The BOLD dataset is publicly available on PhysioNet and contains 49,093 patient entries with 142 features across multiple patient demographics (Table 1). The patients originally participated in eICU-Collaborative Research Database (eICU-CRD), MIMIC-III, and MIMIC-IV. The eICU-CRD cohort is derived from multiple critical care units across the United States and comprises 43,438 patient entries. Patients from MIMIC-III/MIMIC-IV are drawn from the Beth Israel Deaconess Medical Center’s critical care unit in Massachusetts, comprising 5655 patient entries. Of importance for this analysis is that each patient has ABG SaO2 and paired-pulse oximeter SpO2 measurements taken within 5 min of each other. Here, we assume that transmittance pulse oximetry was used. We defined hidden hypoxemia as SaO2<88% and SpO2>88% [3]. Given inherent pulse oximetry measurement variability, some hidden hypoxemia cases, particularly those with small discordance, may reflect both systematic bias and measurement noise. Table 1Distribution and prevalence of hidden hypoxemia in the BOLD datasetRace and EthnicityNo HH(N = 38500)HH(N = 1337)All Patients(N = 39837)American Indian / Alaska Native293(0.761%)6(0.449%)299 (0.751%)Asian684(1.777%)19 (1.421%)703 (1.765%)Black3689(9.582%)158 (11.818%)3847 (9.657%)Hispanic OR Latino1701(4.418%)58 (4.338%)1759 (4.415%)More Than One Race3(0.008%)0(0.0%)3 (0.008%)Native Hawaiian / Pacific Islander8(0.021%)0(0.0%)8 (0.02%)Unknown2958(7.683%)79 (5.909%)3037 (7.624%)White29,164(75.751%)1017 (76.066%)30,181 (75.761%)*HH: Hidden Hypoxemia Distribution and prevalence of hidden hypoxemia in the BOLD dataset 293 (0.761%) 6 (0.449%) 684 (1.777%) 3689 (9.582%) 1701 (4.418%) 3 (0.008%) 0 (0.0%) 8 (0.021%) 0 (0.0%) 2958 (7.683%) 29,164 (75.751%) *HH: Hidden Hypoxemia Data processing The original BOLD dataset was further processed using the following hierarchical system: Feature Inclusion-Exclusion Criteria: We excluded features directly associated with outcome of hidden hypoxemia. We excluded other ABG-derived features, as well as SpO2 time stamp data. We excluded any SOFA score that happened after ABG measurement to prevent the knowledge of the future events. We excluded all IDs, source database tags, and admission characteristics. Features missing in more than 80% of the patients were removed. Comorbidity scores were included. Feature Inclusion-Exclusion Criteria: We excluded features directly associated with outcome of hidden hypoxemia. We excluded other ABG-derived features, as well as SpO2 time stamp data. We excluded any SOFA score that happened after ABG measurement to prevent the knowledge of the future events. We excluded all IDs, source database tags, and admission characteristics. Features missing in more than 80% of the patients were removed. Comorbidity scores were included. Patient Inclusion-Exclusion Criteria: We included the first entry for each unique patient ID for any patients with multiple entries. Patients with more than 50% missing features were excluded. Patient Inclusion-Exclusion Criteria: We included the first entry for each unique patient ID for any patients with multiple entries. Patients with more than 50% missing features were excluded. One-hot encoding: Region and race_ethnicity categorical features were one-hot encoded. One-hot encoding: Region and race_ethnicity categorical features were one-hot encoded. Ordinal encoding: Numbedscategory was ordinal encoded. Ordinal encoding: Numbedscategory was ordinal encoded. Miscellaneous: Height was standardized into centimeter values. Miscellaneous: Height was standardized into centimeter values. A total of 39,837 patients, 64 features, and 2 outcomes (SaO2 and hidden_hypoxemia) were analyzed. Statistical analysis We applied a chi-squared test of independence to elucidate the association between hidden hypoxemia and categorical features [10]. The Wilcoxon Rank-Sum test [11] was used to evaluate the association between hidden hypoxemia and numeric features. We used an alpha value of 0.05 to determine whether to reject the null hypothesis (no association between hidden hypoxemia and not-hidden hypoxemia). Furthermore, we applied false discovery rate (FDR) detection as proposed by Benjamini and Hochberg to calculate adjusted p-values [12]. Machine learning model training We further processed the dataset to prepare for machine learning: we split the dataset with 80/20 train/test. We imputed the continuous features using K-nearest-neighbor with 5-neighbors and normalized the continuous features. The training and test sets were imputed and normalized independently. This imputation was not performed for the XGBoost model. We trained and validated the machine learning model to perform regression on the outcome variable “SaO2,” utilizing the features in the dataset. We employed linear regression, random forest, and XGBoost algorithms. We optimized the models using Bayesian optimization in a 5-fold cross-validation on the train set to minimize the mean squared error. We employed optimized models on the test set to evaluate the model performance. We evaluated the model performances with the primary metric of accuracy root mean square error (Arms) between ground-truth SaO2 () and predicted SaO2 ()[Eq. 1], the performance metric outlined by the Food and Drug Administration (FDA) [13]. We further evaluated the models’ performance with two secondary metrics: (1) R-squared values comparing ground-truth SaO2 and predicted SaO2 (calibrated SpO2) and (2) incidence of hidden hypoxemia between ground-truth SaO2 and predicted SaO2. Oversampling and custom loss function We implemented either Synthetic Minority Over-Sampling Technique for Regression with Gaussian Noise (SMOGN) or a custom loss function to emphasize the minority cases (SpO2 < 90%). We utilized SMOGN to under-sample the majority cases (SpO2 > 90%) and to oversample the minority cases to augment the training set to be more balanced. We utilized the custom loss function to increase the weight of training for minority cases and decrease the weight of training for majority cases. These two techniques were used only in training, but not in testing. Feature importance We further explored feature importance in the trained machine learning models to better understand the prediction method. We employed SHapley Additive exPlanations (SHAP) analysis to calculate feature importance. Results Analysis of the BOLD dataset The overall incidence of hidden hypoxemia in the BOLD dataset was 3.2% (Table 1). We did not observe a significant trend between the time discrepancies of SpO2 and SaO2 measurements and the percentage of hidden hypoxemia (0 min: 3.55%; 1 min: 3.35%; 2 min 3.69%; 3 min: 3.55%; 4 min: 3.73%; 5 min 1.43%). We found a higher incidence of hidden hypoxemia in the black population compared to other races. We first performed a Chi-Square Test of Independence on the BOLD dataset to determine the statistical association between the categorical features and the incidence of hidden hypoxemia (Fig. 1A). We found a statistically significant association between the Midwest region, the Northeast region, and Black race with the incidence of hidden hypoxemia. Further, we performed the Wilcoxon Rank Sum test to evaluate the statistical significance between the numeric features against the incidence of hidden hypoxemia (Fig. 1B). Here, we removed SpO2 and SaO2 from our analysis as they directly relate to the outcome (i.e., hidden hypoxemia). We found that respiratory rate, RBC distribution width, mean corpuscular hemoglobin concentration, creatinine level, and blood urea nitrogen level were the leading features associated with the incidence of hidden hypoxemia. Clinical Features associated with hidden hypoxemia in the BOLD dataset. (A) Bar graph demonstrating the association between categorical features and incidence of hidden hypoxemia (Chi-squared test). (B) Bar graph demonstrating the association between continuous features and the incidence of hidden hypoxemia (Wilcoxon Rank Sum Test). Statistically significant features (p < 0.05) are shown Machine learning regression model estimating of SaO2 We developed a regression machine learning model to predict SaO2 from the SpO2 reading and other clinical features. Our initial exploratory data analysis demonstrated severe data imbalance where hidden hypoxemia events accounted for only 3.36% of the entire dataset. Thus, we developed a regression model instead of classifying oxygen status. By employing a regression model instead of a classification model, we minimize the negative impact of data imbalance on our model. However, on the continuous spectrum, the SpO2 and SaO2 were also imbalanced, where patients with SaO2 > 90% represented 92% of the sample population. Thus, we employed both SMOGN and weighted loss function to mitigate the overfit in the majority case in our regression model. The model was trained and optimized as outlined in Methods. Machine learning models were used to estimate SaO2 from SpO2 and other electronic health record (EHR) features (Table 2). R2 values for all models were significantly higher than the baseline R2 value of 0.040. XGBoost-Vanilla model had the highest absolute R2 value of 0.353 and the lowest Arms of 3.338, and the Random Forest (RF)-Vanilla model demonstrated comparable performances. However, both models showed an increase in the incidence of hidden hypoxemia from the baseline. In our XGBoost-Vanilla model, we observed the overall overestimation of SaO2 values, which contributes to the increase in the incidence of hidden hypoxemia (Fig. 2). The linear regression (LR)-Weighted Loss model had the greatest reduction in the incidence of hidden hypoxemia, though the absolute incidence was reduced by only 11 cases. The LR-Weighted Loss model had an Arms value of 3.805 and an R2 value of 0.159. While both residuals of the XGBoost-Vanilla and LR-Weighted Loss models are skewed to the left, the LR-Weighted Loss model shows a greater level of skewness. Table 2Summary of model performanceAll casesSaO2 > = 90%SaO2 < 90%ModelArmsR²ΔHH (%Change)ArmsR²ArmsR² Baseline 4.0650.0400 (0%)3.139-0.53910.297-3.432 XGBoost Vanilla 3.338 0.35353 (+ 18%)2.2000.2449.779-2.998 Weighted Loss3.5740.25815 (+ 5%)2.687-0.1289.338-2.645 SMOGN3.5140.2828 (+ 3%)2.666-0.1119.091-2.455 Random Forest Vanilla3.3410.35257 (+ 20%) 2.195 0.2479.809-3.022 Weighted Loss3.5250.2784 (+ 1%)2.765-0.1958.749-2.200 SMOGN3.4600.30420 (+ 7%)2.610-0.0649.011-2.394 Linear Regression Vanilla3.5230.27946 (+ 16%)2.3660.12510.185-3.336 Weighted Loss3.8050.159-11 (-4%)3.172-0.571 8.588 -2.083 SMOGN4.725-0.297-8 (-3%)4.284-1.8678.699-2.164* ΔHH: Change in Hidden Hypoxemia Incidence Summary of model performance * ΔHH: Change in Hidden Hypoxemia Incidence Machine learning models can successfully predict SaO2 measurement (A) Density plots of baseline, XGBoost-Vanilla, and Linear Regression (LR)-Weighted Loss models. Yellow reflects denser areas; purple reflects sparser areas. (B) Residual plots of baseline, XGBoost-Vanilla, and LR-Weighted Loss models (black = baseline, red = tested model) Model performance discrepancy in high SaO2 vs. low SaO2 patients Many studies have reported the inaccuracy of pulse oximetry in low saturation patients. Here, we further explored the regression model performance in high SaO2 (SaO2 > = 90%) and low SaO2 (SaO2 < 90%) patients to investigate if our models were able to increase the accuracy in low SaO2 patients as well (Table 2). As expected, most of our models had much better performance in high-SaO2 patients than low-SaO2 patients. RF-Vanilla had the lowest Arms with 2.195%, which was comparable to the Arms scored by XGBoost-Vanilla. Compared to the baseline, the two models increased the Arms by approximately 1%pt. The baseline Arms for low saturation patients was much higher at 10.297%. Further, our models were not able to effectively improve the accuracy in this low-saturation subgroup, though weighted loss and SMOGN approaches showed modest improvements over vanilla models. Feature importance and limited-feature model Here, we investigated which clinical features contribute the most to SaO2 estimation in our best-performing machine-learning models. We selected three best-performing models: (1) XGBoost-Vanilla for overall performance, (2) RF-Vanilla for high saturation performance, and (3) LR-Weighted Loss for low saturation performance. We utilized SHAP for XGBoost and RF models’ feature importances. We report the coefficients for the LR model for feature importance as our features were normalized. In the XGBoost-Vanilla model, the most important clinical features associated with SaO2 prediction included SpO2, bmp_creatinine, admission_age, sofa_past_cns_24hr, and vitals_resp_rate (Fig. 3a). In the RF-Vanilla model, the most important features included SpO2, vitals_resp_rate, bmp_creatinine, South, cbc_mchc, and admission_age (Fig. 3b). In the LR-weighted model, the most important features included SpO2, Midwest, admission_age, vitals_resp_rate, and hfp_albumin (Fig. 4). List of the 20 most important features in predicting SaO2 determined by SHAP. (A) SHAP feature importance of the XGBoost-Vanilla model. (B) SHAP feature importance of the RF-Vanilla model. Higher absolute SHAP value notes higher importance. Features are listed in descending order. The value of individual features is noted in red (higher feature value) to blue (lower feature value). The positive and negative values of the SHAP value represent predicting higher SaO2 value and lower SaO2 value, respectively. Out of the top 20 most important features, 14 features appeared in both models List of the 20 highest coefficients in the LR-weight loss model. Red notes positive coefficient values and blue notes negative coefficient values We observed that there was overlap in the top 20 most important features of the SaO2 prediction across the three models. Notably, SpO2, bmp_creatinine, vitals_resp_rate, admission_age, and cbc_mchc scored high across all three models. SpO2 scored, by a significant margin, the highest in all three models, which is expected as SpO2 inherently estimates SaO2. Bmp_creatinine measures the creatinine level, which reflects kidney function. Higher creatinine levels may indicate kidney dysfunction or injury. Creatinine was inversely related to predicted SaO2. Vitals_resp_rate corresponds to the respiratory rate, and we observed that a higher respiratory rate was related to a lower predicted SaO2 value. Admission_age notes the age of the patient at the time of admission. We observed that older age was correlated with higher predicted SaO2. Cbc_mchc measures the mean corpuscular hemoglobin concentration (MCHC), which measure the average amount of hemoglobin in red blood cells. We observed that a higher level of MCHC correlated to a higher predicted SaO2 value. Further, black race was ranked 20th in both XGBoost-Vanilla and RF-Vanilla. Black race was associated with lower predicted SaO2 value. Discussion In this study, we investigated the feasibility of predicting SaO₂ using machine learning with SpO₂ and other EHR data. We found that machine learning models improved prediction accuracy compared to baseline SpO₂, though performance varied substantially across patient subgroups. While this represents potential advancement in the reliability of SpO₂ as a noninvasive measure of oxygen saturation in critically ill patients, substantial work remains before clinical deployment standards are achieved. These findings provide proof-of-concept for ML-based approaches to reduce measurement inaccuracies and race-associated bias in pulse oximetry. Feature analysis and model performance We first investigated which clinical features were correlated with hidden hypoxemia. From categorical features, the region Midwest, region Northeast, and the race black were correlated with the incidence of hidden hypoxemia. Variables related to the region may demonstrate residual confounding, as this association cannot be otherwise explained. The correlation between hidden hypoxemia and the race category of black was identified, as previously shown [3]. In our analysis, an increase in respiratory rate corresponded with lower SaO2, likely reflecting the known association between hypoxemia and respiratory difficulty [14]. Additionally, red blood cell distribution width (RDW), which reflects variability in the volume and size of red blood cells, has been identified as another significant feature linked to hypoxemia. Hypoxemia stimulates substantial increases in erythropoietin (EPO), which induces unusually large red blood cells. These larger cells disrupt the normal range of RDW, making it a relevant marker in the context of hypoxemia [15]. We further developed a regression machine learning model to predict SaO₂. We expected a high correlation of the original SpO₂ with SaO₂, as previously shown in other literature [16], but this was not the case, as reflected by the extremely low R² value in the original test set (Fig. 3). This extremely low R² value resulted from poor linearity in the low-SaO₂ patients (SaO₂ < 90%), as noted by the high R² in high-SaO₂ patients and low R² in low-SaO₂ patients. In the low-SaO2 subgroup, R² values were negative, ranging from − 2.083 to -3.432, depending on the model. These negative R² values occur when prediction errors (residual sum of squares) far exceed the variance around the mean (total sum of squares). This arises because this low-SaO2 subgroup, which only represents 7% of patients, has limited variance, while prediction errors remain large due to measurement noise in both SpO₂ and ABG readings, temporal misalignment, and physiological variability. In such scenarios, R² becomes highly sensitive to individual errors and fails to capture clinically meaningful improvements [17]. We therefore prioritized Arms, the FDA standard for pulse oximetry accuracy, and ΔHH, which quantifies misdiagnosis rates, as more appropriate metrics. Using these metrics, our machine learning models demonstrated meaningful improvements over baseline SpO₂ measurements. The best-performing model achieved R² of 0.35 and Arms of 3.338%, compared to baseline R² of 0.040 and Arms of 4.065%, with the baseline Arms exceeding the FDA standard of 3% [13]. Importantly, despite negative R² values, our weighted loss model reduced Arms from 10.3% to 8.6% in low-SaO₂ patients, demonstrating that bias correction can occur even when R² suggests poor performance. Clinical insights and feature importance Further, we examined the feature importance of the 3 best-performing models using SHAP. Many clinical features were present in all three important feature lists. Notably, creatinine level, respiratory rate, admission_age, and cbc_mchc were identified in all three lists of features. Creatinine level, respiratory rate, and black race negatively contributed to the predicted SaO2 value, while admission age and MCHC positively contributed to the predicted SaO2 values. While these clinical variables may demonstrate residual confounding, it is important to note that these clinical variables were the most important features in predicting SpO2 values in the machine learning models. In clinical settings, it would be beneficial to closely monitor patients with high creatinine level and high respiratory rate as these features may contribute to overestimation of SaO2 from SpO2. It is also important to note that Black race was ranked as the 20th most important feature in both the XGBoost-Vanilla and RF-Vanilla models. This ranking is somewhat surprising given that Black race showed statistically significant association with hidden hypoxemia in our analysis and previous literature has demonstrated racial disparities in pulse oximetry accuracy [3, 6]. This relatively modest feature importance may reflect limitations of categorical race as a model input. Race is a qualitative social construct rather than a quantitative biological measure [17]; as a categorical variable, it lacks the granularity to adequately capture the continuous variation in skin pigmentation associated with pulse oximetry bias. A continuous skin tone measure would likely provide greater predictive power for addressing the physiological mechanism underlying SpO₂ overestimation. Despite this limitation, our SHAP analysis showed that Black race was associated with lower predicted SaO₂ values, suggesting overestimation of SaO₂ in Black patients. This may be explained by two potential mechanisms: Black patients show lower SaO₂ in ICU and/or Black patients experience overestimation of SpO₂ in ICU. Further study is needed to identify which mechanism drives this pattern. Challenges and limitations One of the primary challenges of this work is overcoming significant data imbalances. Only 7% of patients have SaO2 values below 90%. This scarcity makes it challenging for models to capture and understand patterns affecting lower SaO2 levels accurately. This challenge is evident in low error in high-saturation patients and high error in low-saturation patients (Table 2). Predominantly, the data represents higher SaO2 values, leading models to predict these values more frequently and precisely [18]. This bias arises because machine learning algorithms aim to minimize the overall error across the dataset. Consequently, by adapting predictions to better align with the majority class, models inherently reduce total error but at the risk of inaccurate estimates in the minority class. Despite our effort to mitigate the bias from the imbalanced dataset with SMOGN and custom loss function, our models’ performance deteriorates as the SpO2 values drop below 90%. Additionally, our dataset does not perfectly represent US racial demographics, with underrepresentation of minority populations and overrepresentation of White patients. Validation in more racially diverse cohorts would strengthen confidence in generalizability. Another primary challenge in developing ML-based clinical tools is the trade-off between predictive accuracy and physiological interpretability. While more complex models (e.g., neural networks) may achieve better performance, their reduced interpretability limits clinical trust and the ability to verify predictions align with known physiology, a critical consideration for clinical risk mitigation. We addressed this partially through SHAP analysis, which provides post-hoc interpretability by identifying key features contributing to predictions. However, future work may benefit from exploring inherently interpretable models or hybrid approaches that balance accuracy with interpretability for clinical implementation. There are also concerns regarding the validity of using retrospective EHR data in the context of pulse oximetry bias [19]. These concerns stem from the limited synchronization between pulse oximetry and arterial blood gas data, where the timing of measurements may be delayed or inaccurately recorded in the EHR. While the BOLD dataset addresses this limitation by excluding data with more than a 5-minute gap between SpO2 and SaO2 measurements, temporal mismatch may still introduce measurement error, particularly in critically ill patients who may experience rapid physiological changes during this interval. Additionally, this error may be amplified due to external events (e.g., clinical intervention) that may be present during the delay. To mitigate potential errors associated with retrospective EHR data and leverage the full potential of machine learning, a large multicenter prospective study is warranted. Clinical applicability and future directions While our models demonstrated improvements over baseline, current performance in the low-SaO₂ subgroup, where accuracy is most critical for clinical decision-making, remains insufficient for deployment (Arms 8.6% vs. FDA standard of 3%). However, this work provides value as a discovery tool and proof-of-concept. The SHAP analysis identified physiological features (creatinine, respiratory rate, MCHC, race) associated with pulse oximetry bias, offering mechanistic insights into measurement inaccuracies and racial disparities. These findings can inform future model development, guide clinical monitoring of high-risk patients, and quantify bias magnitude that must be addressed. Rather than a deployment-ready solution, our contribution is demonstrating ML-based correction feasibility while highlighting the data and methodological challenges that must be overcome for clinical implementation. Future research should focus on several key areas to translate this proof-of-concept into clinically deployable models. First, overcoming data imbalances will be essential for improving model performance in hypoxemic patients. While this can be addressed partially by employing methods like SMOGN and feature engineering, increasing the absolute number of samples in the low-SaO₂ subgroup will be more critical. Second, skin tone measurement may improve future regression models by addressing the limitations of using race as a proxy for melanin concentration. Using a quantitative skin tone index could further reduce racial bias in pulse oximetry without dependence on categorical race assignments. Third, access to raw pulse oximetry waveform data (absorbance patterns from red and infrared emitter/detector) may improve predictive accuracy and generalizability. The raw waveform data likely contains latent features lost during mathematical summarization to SpO₂, and utilizing unprocessed optical waveform data to predict SaO₂ may yield optimal results. Finally, prospective multicenter validation studies with simultaneously collected measurements would be necessary to eliminate temporal mismatch and validate performance in real clinical settings. Conclusion In this study, we developed machine learning models to predict SaO₂ from SpO₂ and EHR data, addressing pulse oximetry inaccuracies and racial disparities. Our models demonstrated improvements over baseline, though performance in the critical low-SaO₂ range remains insufficient for clinical deployment. This work provides proof-of-concept for ML-based correction and identifies key physiological features associated with measurement bias. Future work should focus on acquiring larger hypoxemic patient datasets and conducting prospective validation studies to enable clinical implementation. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Acknowledgements Not applicable. Author contributions JWW, OS, and DGS conceptualized the study. JWW and OS developed the methodology. JWW, OS, JL, CC, YS, SS, EM, NS, and SD conducted data curation and formal analysis. JWW, OS, and RM developed and validated the software. KC, JG, COT, and DGS supervised the study. JWW, OS, and RM prepared the figures. JWW, OS, JG, COT, and DGS wrote the main manuscript text. All authors reviewed the manuscript. Funding Not applicable. No funding to declare. Data availability The dataset analyzed during the current study is available on PhysioNet as a credentialled database, [https://physionet.org/content/blood-gas-oximetry/1.0/]. Declarations Ethics approval and consent to participate Not applicable. The dataset analyzed during the current study is available on PhysioNet as a credentialled database, [https://physionet.org/content/blood-gas-oximetry/1.0/] [8]. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. JiWon Woo and Orian Stapleton contributed equally to this work. References Associated Data Supplementary Materials Supplementary Material 1 Data Availability Statement The dataset analyzed during the current study is available on PhysioNet as a credentialled database, [https://physionet.org/content/blood-gas-oximetry/1.0/].