Association of oxygen saturation with delirium in hemorrhagic stroke: A retrospective cohort study using MIMIC-IV Authors: Penghui Du, Chunhui Che, Changyun Liu, Qian Chen, Jiyan Lin, Qingliu Zheng Published: 2026 ABSTRACT Delirium is a common complication in patients with hemorrhagic stroke (HS) admitted to intensive care, but its association with peripheral oxygen saturation (SpO2) remains uncertain. We analyzed 1935 adults with HS from the Medical Information Mart for Intensive Care IV (MIMIC-IV, v3.1). SpO2 was recorded and categorized as low (<95%), moderate (95–98%), or high (>98%) after intensive care unit (ICU) admission. Delirium was identified using the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU). Associations were assessed with multivariable logistic regression, restricted cubic spline (RCS) modeling, subgroup analyses, and receiver operating characteristic (ROC) curves for SpO2, Glasgow Coma Scale (GCS), and their combination. Delirium occurred most frequently in the high SpO2 group (29.6%), followed by the low (19.7%) and moderate groups (17.6%). Compared with the moderate group, high SpO2 was consistently associated with increased delirium risk (adjusted odds ratio: 1.855–1.998; all P < .001), while low SpO2 showed no independent association. RCS analysis demonstrated a U-shaped relationship with the lowest risk between 95% and 98%. Subgroup analyses confirmed consistency across patient characteristics. Combining SpO2 with GCS improved prediction (AUC = 0.702) compared with either alone. High SpO2 (>98%) was independently associated with increased delirium risk, while low SpO2 (<95%) showed a nonlinear trend but was not statistically significant in fully adjusted categorical models. Maintaining SpO2 around 95% to 98% and integrating SpO2 with GCS may aid early risk stratification. 1. Introduction Hemorrhagic stroke (HS), encompassing intracerebral hemorrhage (ICH) and subarachnoid hemorrhage (SAH), accounts for approximately 10% to 20% of all strokes and is associated with disproportionately higher mortality and morbidity rates.[1,2] Delirium is a frequent complication after acute stroke, with a pooled prevalence of about 24%, and hemorrhagic stroke has been identified as a risk factor for poststroke delirium.[3,4] Critically ill patients with HS admitted to the intensive care unit (ICU) often present with complex pathophysiological alterations, such as elevated intracranial pressure, blood–brain barrier (BBB) disruption, and cerebral hypoperfusion, which may predispose them to delirium.[5,6] Delirium has a multifactorial etiology, with impaired cerebral oxygen supply and brain injury representing major contributors to its pathogenesis, whereas arterial oxygen tension (PaO2) may serve as a partial indicator of cerebral oxygenation status.[4,7] As an indirect measure of PaO2, peripheral oxygen saturation (SpO2), a noninvasive indicator of systemic oxygen delivery, is routinely monitored in the ICU.[8] While SpO2 has been implicated in the development of delirium in stroke, its association with delirium in HS specifically has not been well established.[9-11] Given the limited treatment options, early detection and risk stratification are essential, and timely recognition and intervention may play a critical role in improving patient outcomes.[12,13] However, most existing evidence comes from mixed ICU populations or ischemic stroke cohorts, which underscores the need for studies specifically focused on hemorrhagic stroke. This study aims to examine the association between SpO2 levels during the first 24 hours of ICU admission and the incidence of delirium in patients with HS, using the Medical Information Mart for Intensive Care IV (MIMIC-IV) database. Additionally, we assess whether integrating SpO2 with the Glasgow Coma Scale (GCS) can improve early risk stratification for ICU delirium in these patients. 2. Materials and methods 2.1. Data source and study population This retrospective cohort study utilized data from the publicly available MIMIC-IV version 3.1 database, which contains over 70,000 ICU admissions at Beth Israel Deaconess Medical Center from 2008 to 2019.[14] The creation and public release of this de-identified database were approved by the Institutional Review Boards of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center, with a waiver of individual informed consent. Our access to the database was authorized after completing the required training and signing the data use agreement (Record ID: 72969370). Adult patients (≥18 years) with a first ICU admission and a diagnosis of HS, including ICH and SAH, were identified from the MIMIC-IV database using ICD-9 and ICD-10 codes (ICH: ICD-9-CM 431; ICD-10-CM I610–I619, I62.9; SAH: ICD-9-CM 430; ICD-10-CM I60, I600–I6052). Patients with multiple ICU admissions were only included if they had their first ICU admission meeting the inclusion criteria. We excluded patients who did not have a recorded CAM-ICU assessment during their ICU stay (i.e., delirium status could not be ascertained) and those with a history of dementia to minimize confounding and misclassification. Patients who underwent CAM-ICU assessment and were consistently negative throughout the ICU stay were retained in the non-delirium group. After applying these criteria, 1935 patients were included in the final analysis (Fig. 1). Flowchart of patient selection from the MIMIC-IV database for the study of oxygen saturation and delirium in hemorrhagic stroke. A total of 1935 patients were included in the final analysis. HS = hemorrhagic stroke; ICU = intensive care unit, MIMIC-IV = Medical Information Mart for Intensive Care IV. 2.2. Exposure and outcome The primary exposure was SpO2 measured within the first 24 hours after ICU admission. Continuous SpO2 values were extracted and categorized into 3 groups based on commonly used clinical thresholds: low (<95%), moderate (95–98%), and high (>98%). The primary outcome was ICU delirium assessed using the Confusion Assessment Method for the Intensive Care Unit (CAM-ICU). Patients were classified as delirious if any recorded CAM-ICU assessment during the ICU stay was positive. Because SpO2 was summarized over the first 24 hours, whereas delirium could occur at any time during the ICU stay, we interpret findings as associations and address temporality as a limitation. 2.3. Covariates Demographic variables included age, sex, and race. Comorbidities such as chronic obstructive pulmonary disease (COPD), chronic kidney disease (CKD), type 2 diabetes mellitus (T2DM), myocardial infarction (MI), cancer, glucose and heart failure (HF) were identified using ICD codes. Clinical and laboratory variables included vital signs (heart rate [HR], respiratory rate [RR], noninvasive blood pressure [NBP], and body temperature), white blood cell count (WBC), red blood cell count (RBC), hemoglobin (Hb), platelet count (PLT), creatinine (Cr), sodium (Na), potassium (K), and calcium (Ca). Additionally, data on clinical management, including the use of vasopressors, antihypertensives, and glucocorticoids, were collected. Neurological status at ICU admission was assessed using the GCS. 2.4. Statistical analysis Baseline characteristics were summarized according to SpO2 categories. Categorical variables were compared using the chi-square test, while continuous variables were compared using the Kruskal–Wallis test. The association between SpO2 and delirium was examined using multivariable logistic regression with progressively adjusted models (model 1: demographics; model 2: +comorbidities; model 3: +laboratory variables; model 4: +vital signs; model 5: +medication/treatment-related variables). Restricted cubic spline (RCS) regression was used to explore potential nonlinear relationships, treating SpO2 as a continuous variable. Subgroup analyses were conducted according to age, sex, comorbidities, and treatment-related factors, with interaction terms tested to evaluate effect modification. Predictive performance was assessed using receiver operating characteristic (ROC) curve analysis, and the area under the curve (AUC) was calculated for SpO2 alone, GCS alone, and their combination. We evaluated potential multicollinearity among covariates using variance inflation factors and considered the risk of overadjustment when specifying fully adjusted models. Analyses were performed using complete cases for the covariates included in each model. All analyses were performed using R software (version 4.3.1; R Foundation for Statistical Computing, Vienna, Austria). A 2-tailed P value < .05 was considered statistically significant. Results were reported as odds ratios (OR) with corresponding 95% confidence intervals (CI). 3. Results 3.1. Baseline characteristics by SpO2 category A total of 1935 patients with HS were included in this study, comprising 720 (37.2%) in the moderate SpO2 group, 309 (16.0%) in the low group, and 906 (46.8%) in the high group. The median SpO2 values were 96.0%, 93.0%, and 99.0%, respectively (P < .001). Patients in the low SpO2 group were older (median 70.0 vs 67.0 and 64.5 years, P < .001) and had higher rates of COPD and HF (both P < .001). Significant intergroup differences were observed in race (P < .001), RBC, HB, Cr, Ca, HR, and RR (all P < .05). Vasopressor use was most frequent in the high SpO2 group (23.8%), whereas antihypertensive use was more common in the low SpO2 group (64.4%, P = .001). The incidence of delirium was highest in the high SpO2 group (29.6%), followed by the low (19.7%) and moderate groups (17.6%; P < .001). Detailed baseline characteristics are presented in Table 1. Table 1Baseline characteristics of critically ill patients with hemorrhagic stroke, stratified by peripheral oxygen saturation levels.Moderate SpO2 group (95–98%)Low SpO2 group (<95%)High SpO2 group (>98%)P valuen = 720n = 309n = 906Spo2 (M, IQR)96.0 [95.0;97.0]93.0 [92.0;94.0]99.0 [98.0;100]<.001Age (M, IQR)67.0 [56.0;78.0]70.0 [59.0;80.0]64.5 [52.0;76.0]<.001Gender (n, %):.075 Female335 (46.5%)154 (49.8%)473 (52.2%) Male385 (53.5%)155 (50.2%)433 (47.8%)Race (n, %):<.001 Asian20 (2.78%)8 (2.59%)36 (3.97%) Black45 (6.25%)19 (6.15%)90 (9.93%) Hispanic25 (3.47%)6 (1.94%)36 (3.97%) Other149 (20.7%)61 (19.7%)274 (30.2%) White481 (66.8%)215 (69.6%)470 (51.9%)WBC (M, IQR)9.50 [7.50;12.2]9.80 [7.80;12.2]9.90 [7.50;12.7].426RBC (M, IQR)4.21 [3.82;4.66]4.09 [3.62;4.54]4.06 [3.63;4.48]<.001Hb ((M, IQR)12.7 [11.5;14.0]12.4 [10.9;13.7]12.2 [10.9;13.4]<.001Cr (M, IQR)0.80 [0.70;1.00]0.90 [0.70;1.10]0.80 [0.60;1.00].027Na (M, IQR)139 [137;141]139 [137;142]139 [137;141].993K (M, IQR)3.90 [3.70;4.20]3.90 [3.60;4.30]3.90 [3.60;4.20].520Ca (M, IQR)8.90 [8.57;9.20]8.80 [8.40;9.20]8.80 [8.33;9.20]<.001PLT (M, IQR)208 [170;254]210 [162;271]208 [166;255].929Glucose (M, IQR)120 [103;149]123 [105;148]121 [103;146].366Cancer (n, %).106 No605 (84.0%)268 (86.7%)794 (87.6%) Yes115 (16.0%)41 (13.3%)112 (12.4%)CKD (n, %).769 No655 (91.0%)277 (89.6%)817 (90.2%) Yes65 (9.03%)32 (10.4%)89 (9.82%)T2DM (n, %).059 No561 (77.9%)235 (76.1%)739 (81.6%) Yes159 (22.1%)74 (23.9%)167 (18.4%)COPD (n, %)<.001 No675 (93.8%)266 (86.1%)857 (94.6%) Yes45 (6.25%)43 (13.9%)49 (5.41%)MI (n, %).020 No701 (97.4%)291 (94.2%)880 (97.1%) Yes19 (2.64%)18 (5.83%)26 (2.87%)HF (n, %)<.001 No650 (90.3%)243 (78.6%)834 (92.1%) Yes70 (9.72%)66 (21.4%)72 (7.95%)HR (M, IQR)81.0 [71.0;93.0]84.0 [72.0;97.0]79.0 [69.0;89.0]<.001NBP (M, IQR)90.0 [82.0;101]88.0 [80.0;100]90.0 [79.0;100].169RR (M, IQR)18.0 [15.0;21.0]19.0 [16.0;23.0]17.0 [15.0;20.0]<.001Temperature (SD)98.4 (0.78)98.5 (0.91)98.3 (3.38).410Vasopressor (n, %)<.001 No614 (85.3%)247 (79.9%)690 (76.2%) Yes106 (14.7%)62 (20.1%)216 (23.8%)Glucocorticoid (n, %).873 No530 (73.6%)227 (73.5%)657 (72.5%) Yes190 (26.4%)82 (26.5%)249 (27.5%)Antihypertensive (n, %).001 No322 (44.7%)110 (35.6%)431 (47.6%) Yes398 (55.3%)199 (64.4%)475 (52.4%)Delirium (n, %)<.001 No593 (82.36)248 (80.26)638 (70.42) Yes127 (17.64)61 (19.74)268 (29.58)Data are presented as median (interquartile range, Q1–Q3) for continuous variables and number (percentage) for categorical variables. Patients were categorized into 3 groups based on the initial recorded SpO2 level. The P value was derived from the Kruskal–Wallis test for continuous variables and the chi-square test (or Fisher exact test) for categorical variables, as appropriate.Ca = calcium, CKD = chronic kidney disease, COPD = chronic obstructive pulmonary disease, Cr = creatinine, Hb = hemoglobin, HF = heart failure, HR = heart rate, IQR = interquartile range, K = potassium, MI = myocardial infarction, Na = sodium, NBP = noninvasive blood pressure, PLT = platelet count, RBC = red blood cell count, RR = respiratory rate, SD = standard deviation, SpO2 = peripheral oxygen saturation, T2DM = type 2 diabetes mellitus, WBC = white blood cell count. Baseline characteristics of critically ill patients with hemorrhagic stroke, stratified by peripheral oxygen saturation levels. Data are presented as median (interquartile range, Q1–Q3) for continuous variables and number (percentage) for categorical variables. Patients were categorized into 3 groups based on the initial recorded SpO2 level. The P value was derived from the Kruskal–Wallis test for continuous variables and the chi-square test (or Fisher exact test) for categorical variables, as appropriate. Ca = calcium, CKD = chronic kidney disease, COPD = chronic obstructive pulmonary disease, Cr = creatinine, Hb = hemoglobin, HF = heart failure, HR = heart rate, IQR = interquartile range, K = potassium, MI = myocardial infarction, Na = sodium, NBP = noninvasive blood pressure, PLT = platelet count, RBC = red blood cell count, RR = respiratory rate, SD = standard deviation, SpO2 = peripheral oxygen saturation, T2DM = type 2 diabetes mellitus, WBC = white blood cell count. 3.2. Association between SpO2 levels and delirium risk Results of the multivariable logistic regression analyses are presented in Table 2. A significant positive association was observed between SpO2 (modeled as a continuous variable) and delirium risk. Specifically, each 1% increase in SpO2 was associated with an approximately 11% to 15% increase in the odds of delirium, and this association remained statistically significant (P < .001) through sequential adjustment for demographics, comorbidities, laboratory markers, vital signs, and medication use (models 1–5). Table 2Multivariable logistic regression analysis of the association between peripheral oxygen saturation levels and the risk of delirium in critically ill patients with hemorrhagic stroke.ModelSpo2 as continuous (per 1% increase)Moderate SpO2 group (95–98%)Low SpO2 group (<95%)High SpO2 group (>98%)P for trendOR (95% CI)P valueOR (95% CI)P valueOR (95% CI)P valueUnadjusted model1.119 (1.071–1.169)<.001Reference1.148 (0.818–1.613).4241.961 (1.545–2.491)<.001<.001Model 11.127 (1.077–1.179)<.001Reference1.101 (0.781–1.553).5821.998 (1.562–2.554)<.001<.001Model 21.133 (1.083–1.186)<.001Reference1.034 (0.729–1.465).8531.971 (1.540–2.525)<.001<.001Model 31.128 (1.078–1.181)<.001Reference0.959 (0.674–1.366).8181.855 (1.443–2.384)<.001<.001Model 41.149 (1.096–1.204)<.001Reference0.920 (0.644–1.314).6471.973 (1.530–2.544)<.001<.001Model 51.145 (1.093–1.200)<.001Reference0.893 (0.622–1.282).5401.910 (1.476–2.473)<.001<.001Data are presented as odds ratio (OR) with 95% confidence interval (CI). The moderate SpO2 group (95–98%) is set as the reference category in all models. Models are progressively adjusted as follows: Unadjusted model – no covariates; model 1 – adjusted for demographics; model 2 – adjusted for demographics and comorbidities; model 3 – adjusted for demographics, comorbidities, and laboratory markers; model 4 – adjusted for demographics, comorbidities, laboratory markers, and vital signs; model 5 – adjusted for demographics, comorbidities, laboratory markers, vital signs, and medication use.CI = confidence interval, OR = odds ratio, SpO2 = peripheral oxygen saturation. Multivariable logistic regression analysis of the association between peripheral oxygen saturation levels and the risk of delirium in critically ill patients with hemorrhagic stroke. Data are presented as odds ratio (OR) with 95% confidence interval (CI). The moderate SpO2 group (95–98%) is set as the reference category in all models. Models are progressively adjusted as follows: Unadjusted model – no covariates; model 1 – adjusted for demographics; model 2 – adjusted for demographics and comorbidities; model 3 – adjusted for demographics, comorbidities, and laboratory markers; model 4 – adjusted for demographics, comorbidities, laboratory markers, and vital signs; model 5 – adjusted for demographics, comorbidities, laboratory markers, vital signs, and medication use. CI = confidence interval, OR = odds ratio, SpO2 = peripheral oxygen saturation. When SpO2 was analyzed categorically, using the moderate range (95–98%) as the reference, a clear pattern emerged. The high SpO2 group (>98%) was consistently associated with a markedly elevated risk of delirium. The strength of this association was robust across all adjusted models, with OR ranging from 1.855 to 1.998 (all P < .001). Conversely, no statistically significant association was found for the low SpO2 group (<95%) compared to the reference group in any model (OR range: 0.893–1.148; all P > .05). A significant dose–response relationship across the SpO2 categories was confirmed in all analyses (P for trend < .001). Restricted cubic spline analyses further demonstrated a significant nonlinear association between SpO2 and delirium risk (overall P < .001; nonlinearity P < .001). The curves consistently showed a U-shaped relationship across unadjusted and fully adjusted models, with the lowest risk observed in the range of 95% to 98%. At SpO2 values below 95%, delirium risk tended to rise slightly, although confidence intervals were wider and estimates less stable. In contrast, SpO2 values above 98% were associated with a marked increase in risk, with odds ratios approaching 2.5 to 3.0 depending on the model (Fig. 2). Nonlinear association between oxygen saturation and the risk of delirium as analyzed by restricted cubic spline regression. The solid line represents the odds ratio, and the shaded area indicates the 95% confidence interval. Results are shown for the unadjusted model (panel A) and the fully adjusted models (panels B–F). Panel (A) shows the unadjusted model; panel (B) shows model 1 (demographics); panel (C) shows model 2 (demographics + comorbidities); panel (D) shows model 3 (demographics + comorbidities + laboratory variables); panel (E) shows model 4 (demographics + comorbidities + laboratory variables + vital signs); panel (F) shows model 5 (demographics + comorbidities + laboratory variables + vital signs + medication/treatment-related variables). CI = confidence interval, SpO2 = peripheral oxygen saturation. 3.3. Subgroup analyses of the SpO2–delirium association Subgroup analyses (Fig. 3) demonstrated that the association between higher SpO2 levels and increased risk of ICU-onset delirium was consistent across most patient subgroups. The effect estimates remained significant in both non-elderly (OR: 1.155, 95% CI: 1.048–1.273) and elderly patients (OR: 1.137, 95% CI: 1.077–1.199; P for interaction = .427), as well as in males (OR: 1.102, 95% CI: 1.034–1.174) and females (OR: 1.179, 95% CI: 1.100–1.265; P for interaction = .204). Similar associations were observed regardless of the presence of Ca, CKD, or T2DM, and regardless of antihypertensive or glucocorticoid use. Although the effect size was slightly attenuated among patients with MI, COPD, or vasopressor use, none of the interactions reached statistical significance (all P for interaction > .05). Overall, no significant effect modification was detected, indicating that the relationship between SpO2 and delirium risk was robust across diverse clinical subgroups. Subgroup analysis of the association between oxygen saturation levels and delirium risk in patients with hemorrhagic stroke. The forest plot displays odds ratios and 95% confidence intervals for each subgroup. CI = confidence interval, CKD = chronic kidney disease, COPD = chronic obstructive pulmonary disease, GCS = Glasgow Coma Scale, MI = myocardial infarction, OR = odds ratio, T2DM = type 2 diabetes mellitus. 3.4. Predictive performance of SpO2 and GCS for delirium As shown in Figure 4, ROC curve analysis was conducted in the overall cohort (n = 1935) to evaluate the predictive value of SpO2, GCS, and their combination for ICU-onset delirium. The (AUC for SpO2 alone was 0.598, indicating limited discriminatory ability, while the AUC for GCS alone was 0.640). When SpO2 was combined with GCS in a single model, predictive performance improved to an AUC of 0.702, which was higher than either predictor alone. These findings demonstrate that the combined model provided superior classification performance relative to models using SpO2 or GCS individually. Receiver operating characteristic curves for predicting delirium risk using oxygen saturation alone, Glasgow Coma Scale score alone, and their combination in patients with hemorrhagic stroke. The area under the curve for each predictor is displayed in the legend. The diagonal line represents a model with no discriminatory ability (area under the curve = 0.5). AUC, area under the curve; GCS, Glasgow Coma Scale score; SpO2, oxygen saturation. 4. Discussion In this large retrospective cohort study of 1935 critically ill patients with HS from the MIMIC-IV database, we identified a U-shaped association between SpO2 and the risk of ICU-onset delirium. The lowest risk was observed at SpO2 levels of 95% to 98%, while hyperoxia (>98%) was consistently associated with a significantly increased risk after multivariable adjustment. In the fully adjusted categorical model, hypoxemia (<95%) did not exhibit a statistically significant independent association with delirium, although spline analyses suggested a trend toward increased risk at lower levels.[15] These results were robust across logistic regression, restricted cubic spline (RCS) analyses, and subgroup evaluations, indicating that the findings were consistent across different patient characteristics. Furthermore, the combination of SpO2 and the GCS improved predictive accuracy compared with either measure alone, with an AUC of 0.702, suggesting moderate discriminatory performance. These findings suggest that both oxygenation status and neurological function are associated with delirium risk in HS patients. Previous studies have shown that hypoxemia is associated with delirium in general ICU populations, with potential mechanisms involving neuronal ischemia, impaired cerebral energy metabolism, and neurotransmitter dysregulation.[16,17] More recent evidence indicates that hyperoxia may also contribute to delirium and poorer outcomes.[8,10] Previous studies have demonstrated that, in patients with cerebral infarction and underlying comorbidities, maintaining an appropriate range of oxygen saturation may help improve clinical outcome.[18] Similarly, in patients with acute brain injury, both hypoxemia and hyperoxia have been linked to increased risks of adverse outcomes, underscoring the importance of cautious oxygen management.[19-21] However, the optimal SpO2 target range remains uncertain, as findings have varied across studies, and few investigations have specifically addressed this issue in patients with HS.[11,22] Our findings align with evidence from mixed ICU cohorts indicating that hyperoxia may be as harmful as hypoxemia in neurocritical care.[23,24] Unlike most prior studies, we applied both categorical and continuous modeling to assess nonlinear associations, confirming a U-shaped relationship with the lowest delirium risk at 95% to 98% and higher risk at both extremes. The lack of significance in the low SpO2 group is likely due to smaller sample size and wider confidence intervals. Overall, these findings suggest that both hypoxemia and hyperoxia may elevate delirium risk, with hyperoxia exerting the stronger effect, and subgroup analyses confirmed the robustness of this association across diverse patient subgroups.[25] Several mechanisms may underlie the association between elevated SpO2 and increased delirium risk. Hyperoxia can induce oxidative stress and neuroinflammation, promoting ROS accumulation and upregulation of inflammatory cytokines such as IL-6.[25-27] It may also trigger cerebral vasoconstriction or reperfusion injury, aggravating edema and ischemia, particularly in patients with impaired autoregulation.[20,28] In addition, oxidative stress may disrupt neurotransmitter balance by impairing acetylcholine synthesis and enhancing dopaminergic activity, thereby predisposing to delirium.[15,29] These findings suggest that maintaining SpO2 within an appropriate range could potentially reduce the risk of delirium in critically ill patients.[24,30] Neurological status, as reflected by the GCS, may also play a critical role.[31,32] In our cohort, ROC analysis showed that SpO2 alone offered limited discrimination for delirium, while GCS performed somewhat better. Importantly, their combination improved predictive accuracy, suggesting that oxygenation status and neurological function may act as complementary factors associated with delirium risk. This finding highlights the potential value of integrating both physiological and neurological markers into risk stratification strategies, which could guide more effective preventive interventions in HS patients.[33,34] Our findings carry important clinical implications. The observed U-shaped association emphasizes the need to avoid both hypoxemia and hyperoxia, with an optimal SpO2 range around 95% to 98%. In practice, this suggests that oxygen therapy should be carefully titrated, and that integrating SpO2 with neurological assessments such as the GCS may enhance early risk stratification and guide preventive interventions. Several limitations should be noted. First, the retrospective single-database design precludes causal inference, and unmeasured confounders cannot be fully excluded. Second, delirium was identified using recorded CAM-ICU assessments, and the frequency and completeness of CAM-ICU screening may vary across patients; therefore, under-recognition or measurement variability is possible. To reduce misclassification, we excluded patients without any recorded CAM-ICU assessment, but selection bias may still exist. Third, because SpO2 was summarized over the first 24 hours whereas delirium could occur earlier or later during the ICU stay, temporality cannot be fully established. Fourth, important clinical factors such as sedation depth, mechanical ventilation parameters, and detailed oxygen therapy data (e.g., FiO2 and PaO2) were not available for inclusion, raising the possibility of residual confounding. Finally, oxygenation was evaluated using SpO2 rather than PaO2; in the upper range of the oxyhemoglobin dissociation curve, SpO2 values above 98% may correspond to a wide range of PaO2, so using SpO2 > 98% as a proxy for hyperoxemia has conceptual limitations and may affect biological interpretation and clinical generalizability. Future prospective studies with standardized delirium screening, time-resolved oxygenation measures (including PaO2/FiO2), and detailed ventilatory/sedation data are needed to validate these findings and refine oxygenation targets in HS. 5. Conclusions In critically ill patients with HS, higher SpO2 (>98%) was consistently associated with a higher risk of ICU delirium, whereas low SpO2 (<95%) showed a U-shaped trend in spline analyses but was not independently significant in fully adjusted categorical models. The lowest risk was observed at SpO2 levels of 95% to 98%. Incorporating SpO2 with the GCS modestly improved delirium risk prediction, supporting the potential value of combining oxygenation monitoring with neurological assessment for early risk stratification. Acknowledgments We are grateful to the teams at the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center for the creation, maintenance, and public sharing of the MIMIC-IV database, which made this study possible. Author contributions Conceptualization: Penghui Du, Qingliu Zheng. Data curation: Chunhui Che. Formal analysis: Chunhui Che, Jiyan Lin. Investigation: Qian Chen. Methodology: Changyun Liu, Jiyan Lin. Supervision: Penghui Du, Qingliu Zheng. Validation: Changyun Liu. Visualization: Qian Chen. Writing – original draft: Penghui Du. Writing – review & editing: Qingliu Zheng.