Critical Care Explorations

BackgroundProgressive hypoxemia is the predominant mode of deterioration in COVID-19. Among hypoxemia measures, the ratio of the partial pressure of arterial oxygen to the fraction of inspired oxygen (P/F ratio) has optimal construct validity but poor availability because it requires arterial blood sampling. Pulse oximetry reports oxygenation continuously, but occult hypoxemia can occur in Black patients because the technique is affected by skin color. Oxygen dissociation curves allow non-invasive estimation of P/F ratios (ePFR) but this approach remains unproven. Research QuestionCan ePFRs measure overt and occult hypoxemia? Study Design and methodsWe retrospectively studied COVID-19 hospital encounters (n=5319) at two academic centers (University of Virginia [UVA] and Emory University). We measured primary outcomes (death or ICU transfer within 24 hours), ePFR, conventional hypoxemia measures, baseline predictors (age, sex, race, comorbidity), and acute predictors (National Early Warning Score (NEWS) and Sepsis-3). We updated predictors every 15 minutes. We assessed predictive validity using adjusted odds ratios (AOR) and area under receiver operating characteristics curves (AUROC). We quantified disparities (Black vs non-Black) in empirical cumulative distributions using the Kolmogorov-Smirnov (K-S) two-sample test. ResultsOvert hypoxemia (low ePFR) predicted bad outcomes (AOR for a 100-point ePFR drop: 2.7 [UVA]; 1.7 [Emory]; p<0.01) with better discrimination (AUROC: 0.76 [UVA]; 0.71 [Emory]) than NEWS (AUROC: 0.70 [UVA]; 0.70 [Emory]) or Sepsis-3 (AUROC: 0.68 [UVA]; 0.65 [Emory]). We found racial differences consistent with occult hypoxemia. Black patients had better apparent oxygenation (K-S distance: 0.17 [both sites]; p<0.01) but, for comparable ePFRs, worse outcomes than other patients (AOR: 2.2 [UVA]; 1.2 [Emory], p<0.01). InterpretationThe ePFR was a valid measure of overt hypoxemia. In COVID-19, it may outperform multi-organ dysfunction models like NEWS and Sepsis-3. By accounting for biased oximetry as well as clinicians real-time responses to it (supplemental oxygen adjustment), ePFRs may enable statistical modelling of racial disparities in outcomes attributable to occult hypoxemia.

Background: The Therapeutic Distance framework (Paper 1) achieved AUC 0.61 for orbit-based mortality prediction in 11,627 sepsis patients. We hypothesised that incorporating state-dependent parameter relevance would substantially improve prediction. Methods: We extended the framework to 84,176 ICU patients from MIMIC-IV v3.1 across 16 clinical syndromes. Validation included full-population leave-one-out (n=59,362), head-to-head comparison against SAPS-II and logistic regression on 34,467 matched patients with bootstrap confidence intervals, temporal validation, outcome permutation, sensitivity analysis, and calibration assessment. Results: Full-population leave-one-out achieved AUC 0.832 (n=59,362). On 34,467 matched patients, Therapeutic Distance (AUC 0.841) significantly outperformed both SAPS-II (0.786; delta=+0.055, 95% CI +0.048 to +0.061, p<0.001) and logistic regression (0.788). Temporal validation showed stable performance (delta=-0.006). Outcome permutation confirmed genuine signal (AUC 0.859 to 0.498 with shuffled mortality). Sensitivity analysis demonstrated near-zero variation (delta 0.0006-0.003). The framework performed well for 8 of 16 syndromes (AUC >0.70) and failed for DKA and post-cardiac surgery (AUC <0.40). Conclusions: Therapeutic Distance provides therapy-specific risk stratification that exceeds both established severity scores and standard machine learning while remaining robust to hyperparameter choices, temporal drift, and outcome permutation.

BackgroundEarly lactate is widely used to risk-stratify septic shock, yet clinically actionable cut-offs for 28-day mortality remain uncertain. MethodsIn a single-centre study conducted across two intensive care units, we analysed 84 adults with septic shock identified within 24 hours of intensive care unit admission. The primary endpoint was 28-day mortality. Four lactate metrics obtained during the first 24 hours were evaluated: first (admission) lactate, last lactate, peak lactate, and lactate clearance from first to last. Associations were tested using logistic regression with and without adjustment for the Simplified Acute Physiology Score 3; discrimination was assessed by area under the receiver-operating characteristic curve (AUROC), and optimal cut-offs were defined by the Youden index. ResultsThirty-nine of 84 patients (46.4%) died by day 28. Higher absolute lactate values were independently associated with death (adjusted odds ratio (OR) per 1 mmol/L increase: First 1.47, p<0.001; Last 1.41, p=0.002; Peak 1.39, p<0.001), whereas Lactate clearance was not (OR 0.65, p=0.202). Discrimination was moderate to good for peak (AUROC 0.817), first (0.791), and last (0.757) lactate, and poor for clearance (0.577). Youden-derived thresholds provided pragmatic trade-offs: First 3.55 mmol/L (sensitivity 0.821, specificity 0.689), Last 3.15 mmol/L (0.567, 0.864), and Peak 3.55 mmol/L (0.973, 0.556). Kaplan-Meier curves using these cut-offs showed early and sustained separation. ConclusionsIn adults with septic shock, simple early lactate thresholds around 3.3- 3.6 mmol/L (first/peak) and approximately 3.15 mmol/L (last) identify 28-day mortality risk and outperform lactate clearance.

RationaleHeterogeneity of sepsis limits discovery and targeting of treatments. Clustering approaches in critical illness have identified patient groups who may respond differently to therapies. These include in acute respiratory distress syndrome (ARDS) two inflammatory sub-phenotypes, using latent class analysis (LCA), and in sepsis two Sepsis Response Signatures (SRS), based on transcriptome profiling. It is unknown if inflammatory sub-phenotypes such as those identified in ARDS are present in sepsis and how sub-phenotypes defined with different techniques compare. ObjectivesTo identify inflammatory sub-phenotypes in sepsis using LCA and assess if these show differential treatment responses. These sub-phenotypes were compared to hierarchical clusters based on inflammatory mediators and to SRS sub-phenotypes. MethodsLCA was applied to clinical and biomarker data from two septic shock randomized trials. VANISH compared norepinephrine to vasopressin and hydrocortisone to placebo and LeoPARDS compared levosimendan to placebo. Hierarchical cluster analysis (HCA) was applied to 65, 21 and 11 inflammatory mediators measured in patients from the GAinS (n=124), VANISH (n=155) and LeoPARDS (n=484) studies. Measurements and Main ResultsLCA and HCA identified a sub-phenotype of patients with high cytokine levels and worse organ dysfunction and survival, with no interaction between LCA classes and trial treatment responses. Comparison of inflammatory and transcriptomic sub-phenotypes revealed some similarities but without sufficient overlap that they are interchangeable. ConclusionsA sub-phenotype with high levels of inflammation and increased disease severity is consistently identifiable in sepsis, with similarities to that described in ARDS. There was limited overlap with the transcriptomic sub-phenotypes.

BackgroundSepsis-related organ dysfunction results from complex interactions between systemic hemodynamics, microcirculatory alterations, and cellular metabolic failure. Conventional resuscitation strategies guided by global parameters may miss persistent tissue hypoperfusion, a phenomenon termed "hemodynamic incoherence." The PRISM trial was designed to determine whether individualized management guided by advanced multimodal circulatory and perfusion monitoring improves outcomes in septic shock. MethodsThe PRISM trial is a multicenter, randomized, controlled, open-label study with blinded outcome assessment. Adults with septic shock (Sepsis-3 criteria) are randomized (1:1) to structured multimodal monitoring versus standard care. The intervention integrates advanced systemic hemodynamic indices --including mean circulatory filling pressure analogue and other determinants of venous return, heart efficiency, cardiac power output, power efficiency, and volume efficiency-- with a comprehensive perfusion panel (capillary refill time, mottling score, temperature gradients, lactate kinetics, central venous oxygen saturation, venous-arterial carbon dioxide difference, near-infrared spectroscopy-derived skeletal muscle tissue oxygen saturation, and arterial-interstitial glucose gradients). A predefined treatment algorithm links abnormal thresholds to therapeutic interventions. The primary endpoint is change in SOFA and SAPS II scores from baseline to 72 hours. Secondary endpoints include 28-day mortality, ICU and hospital length of stay, ventilator- and vasopressor-free days, lactate clearance, and safety outcomes. DiscussionBy combining advanced hemodynamic physiology with structured multimodal perfusion monitoring, the PRISM trial tests whether individualized, pathophysiology-guided resuscitation can overcome hemodynamic incoherence and improve patient-centered outcomes in septic shock.

Background: Severity scoring systems such as SOFA, NEWS2, and qSOFA effectively identify deteriorating ICU patients by aggregating physiological parameters into composite indices that trigger clinical alerts. However, these systems evaluate patient state at discrete time points and do not model the temporal dynamics of organ deterioration or the pharmacokinetic constraints that govern whether a given intervention can achieve therapeutic effect before an organ trajectory crosses an irreversible threshold. This limitation is consequential because interventions across critical care span pharmacokinetic onset times from seconds (vasopressors) to hours (metabolic corrections, blood products, enzymatic cofactors), yet no existing framework quantifies timing adequacy as a function of these intervention-specific pharmacokinetic properties. Methods: We developed the Multi-Organ Intervention State Space (MOISS), a collision geometry framework that classifies intervention timing adequacy by computing the temporal relationship between the predicted time for a biomarker trajectory to reach a critical threshold and the time required for the administered intervention to achieve peak therapeutic effect. Biomarker trajectories were estimated using the Kunche Adaptive Estimator (KAE), a reliability-adaptive Kalman filter that provides continuous position and velocity estimates from intermittent laboratory measurements. MOISS assigns each intervention event to one of six ordinal categories: PROPHYLACTIC, ON_TIME, PARTIAL, MARGINAL, FUTILE, or TOO_LATE. We applied this framework to 301,470 ICU patients across three databases (eICU-CRD, MIMIC-IV, MIMIC-III), evaluating 65 distinct intervention-organ pairs spanning 10 organ systems: Cardiovascular, Metabolic, Respiratory, Renal, Hematologic, Hepatic, Gastrointestinal, Infection, Endocrine, and Neurological. Results: Timing-mortality associations were identified across all 10 organ systems, with 87 intervention-database combinations achieving statistical significance (p<0.05). The highest timing sensitivity was observed in metabolic corrections: thiamine supplementation for metabolic acidosis (OR 5.76; 95% CI 4.86-6.83 in MIMIC-IV), sodium bicarbonate (OR 4.99; 95% CI 4.27-5.82 in MIMIC-IV). Respiratory interventions showed comparable magnitude: mechanical ventilation initiation (OR 5.03; 95% CI 4.42-5.73 in MIMIC-IV). Hematologic interventions demonstrated strong timing dependency: platelet transfusion (OR 4.25; 95% CI 3.68-4.90), fresh frozen plasma (OR 3.41; 95% CI 2.94-3.95). Cardiovascular agents ranged from OR 1.40 for norepinephrine (consistent with its rapid 1-2 minute onset providing a forgiving therapeutic window) to OR 2.23 for milrinone. Infection-directed therapies, hepatic support, renal replacement, endocrine correction, gastrointestinal interventions, and neurological agents all contained timing-sensitive members. Cross-database consistency was demonstrated for 29 of 52 testable interventions (55.8%), with 6 interventions achieving significance across all three databases. Conclusions: Intervention timing sensitivity is pervasive across the entire spectrum of critical care therapeutics, spanning all 10 organ systems and all pharmacokinetic classes evaluated. MOISS provides a systematic framework for quantifying this timing adequacy that complements existing severity scoring by adding the pharmacokinetic timing dimension: where SOFA, NEWS2, and qSOFA identify that a patient is deteriorating, MOISS computes whether the specific planned intervention can still achieve its intended effect given the current organ trajectory and pharmacokinetic constraints. The universality of timing sensitivity across organ systems argues for multi-organ trajectory monitoring as the foundation for next-generation clinical decision support.

RationaleAutonomic dysfunction is a hallmark of sepsis pathophysiology, yet its quantification remains challenging. Multiscale entropy (MSE) derived from heart rate variability (HRV) offers a dynamic measure of physiological complexity and may serve as a biomarker of early deterioration associated with subsequent organ failure, vasopressor escalation, or mortality. ObjectiveTo determine whether MSE computed across multiple temporal scales during the first 24 hours of Intensive Care Unit (ICU) admission is associated with short-term mortality and longer-term organ dysfunction in patients with sepsis, and whether these relationships vary across vasopressor exposure. Unlike prior studies that focused on short-term HRV metrics, we applied MSE across multiple temporal scales and incorporated these features into machine learning models to evaluate their prognostic utility in septic shock. MethodsThis retrospective cohort study included adult ICU sepsis patients at Emory University Hospital from January 2016 to December 2019. Of 2,076 eligible patients, 958 were propensity matched into two cohorts: fluids-only and fluids-plus-vasopressor, with norepinephrine as the primary vasopressor. High-resolution electrocardiogram (ECG) waveforms were analyzed to compute MSE across 20 temporal scales. Machine learning models using (1) MSE features alone and (2) MSE combined with demographic and vital sign data (MSE-DV) were compared against traditional HRV measures based model and severity of illness scores for predicting outcomes. Model performance was assessed using the area under the receiver operating characteristic curve (AUROC), with a primary outcome of mortality at day 7 and secondary outcome of persistent organ dysfunction at day 28. ResultsIn the fluids-plus-vasopressor cohort, MSE-based models demonstrated superior predictive performance for 7-day mortality (AUROC 0.84) compared to severity of illness scores (AUROC 0.64). MSE-DV models also predicted organ dysfunction including 28-day renal (AUROC 0.75), neurological (AUROC 0.79), and respiratory (AUROC 0.71) dysfunction. Patients receiving second-line and third-line vasopressors and corticosteroids exhibited progressively lower MSE values, particularly at mid-range and long-range scales. ConclusionMSE features in the first 24 hours of ICU stay predict mortality and organ dysfunction with higher discrimination than traditional severity of illness scores. Future work should validate these findings, assess longitudinal MSE trends, and race-specific autonomic patterns to refine predictive models.

BackgroundDelirium during acute respiratory failure is common and morbid. Pharmacologic sedation is a major risk factor for delirium, but some sedation is often necessary for the provision of safe care of mechanically ventilated patients. A simple, transparent model that predicts sedative-associated delirium in mechanically ventilated ICU patients could be used to guide decisions about personalized sedation. Research QuestionCan the risk of sedative-associated delirium be estimated in mechanically-ventilated ICU patients? Study Design and MethodsUsing the subset of patients in a previously-published ICU cohort who received mechanical ventilation, we performed backward stepwise logistic regression to derive a model predictive of sedative-associated delirium. We validated this model internally using hundredfold bootstrapping. We then validated this model externally in a separate prospective cohort of mechanically ventilated ICU patients. Results836 patients comprised the derivation cohort. Backwards stepwise regression produced a model with age, BMI, sepsis, SOFA, malignancy, COPD, stroke, sex, and doses of sedatives (opioids, propofol, and/or benzodiazepines) as predictors of sedative-associated delirium. The model had very good discriminative power, with an area under the receiver-operator curve (AUROC) of 0.83. Internal validation via bootstrapping showed preserved discriminatory function with an AUROC of 0.81 and graphical evidence of good calibration. External validation in a separate set of 340 patients showed good discrimination, with AUROC of 0.70. InterpretationSedative-associated delirium during acute respiratory failure requiring mechanical ventilation can be predicted using a simple, transparent model, which can now be validated in a prospective study.

ImportanceDifferential use of therapies for respiratory failure according to patient race/ethnicity may represent health inequity and could impact patient survival. ObjectiveMeasure the association between patient race/ethnicity and the use of invasive ventilation, and the impact of any association on survival. DesignRetrospective cohort analysis using a Bayesian multistate model that adjusted for baseline covariates and time-varying severity. SettingMulticenter study using the Medical Information Mart for Intensive Care IV (MIMIC-IV) and Phillips eICU (eICU) databases from the USA. ParticipantsNon-intubated adults receiving oxygen within the first 24 hours of ICU admission. ExposurePatient race/ethnicity (Asian, Black, Hispanic, white). Main outcomes and measuresPrimary output was the cause-specific hazard ratio (HR) of invasive ventilation for patient race/ethnicity. Secondary output was change in 28-day survival mediated by differences in invasive ventilation rate. We reported posterior means and 95% credible intervals (CrI). ResultsWe studied 38,263 patients, 52% (20,033) from MIMIC-IV and 48% (18,230) from eICU, 2% Asian (892), 11% Black (4,289), 5% Hispanic (1,964), and 81% white (31,923). Invasive ventilation occurred in 3,511 (9.2%), and 2,869 (7.5%) died. The rate of invasive ventilation was lower in Asian (HR 0.82, CrI 0.70 to 0.95), Black (HR 0.78, CrI 0.71 to 0.86), and Hispanic (HR 0.70, CrI 0.61 to 0.79) patients as compared to white patients. For the average patient, lower rates of invasive ventilation did not mediate differences in survival. For a reference patient with inspired oxygen (FiO2) varied from 0.5 to 1.0, the change in survival mediated by lower rates of invasive ventilation ranged from probable benefit (probability 0.82 for Asian patients, 0.91 for Black patients, and 0.93 for Hispanic patients) at FiO2 0.5 to probable harm (probability 0.87 for Asian patients, 0.92 for Black patients, and 0.97 for Hispanic patients) at FiO2 1.0, although the mean absolute changes in mortality were all less than 1.5%. ConclusionsAsian, Black, and Hispanic patients had a lower rate of invasive ventilation than white patients. The changes in 28-day survival mediated by this difference ranged from slight benefit at lower inspired oxygen fractions to slight harm at inspired oxygen fraction of 1.0, and there was no difference in survival for the average patient. Key PointsO_ST_ABSQuestionC_ST_ABSWhat is the association between patient race/ethnicity and the use of invasive ventilation in hypoxemic respiratory failure, and what is the impact of any differences on survival? FindingsWe studied 38,263 patients from two US databases, who were 2% Asian (892), 11% Black (4,289), 5% Hispanic (1,964), and 81% white (31,118). Invasive ventilation occurred in 3,511 (9.2%), and 2,869 (7.5%) died. The hazard ratio (HR) for invasive ventilation was lower in Asian (HR 0.82, CrI 0.70 to 0.95), Black (HR 0.78, CrI 0.71 to 0.86), and Hispanic (HR 0.70, CrI 0.61 to 0.79) patients as compared to white patients. For the average patient, race/ethnicity differences in invasive ventilation rates did not mediate differences in 28-day survival. For the reference patient, at inspired oxygen fractions up to 0.9, lower invasive ventilation rates mediated a modest survival benefit, whereas at inspired oxygen fraction of 1.0, the lower invasive ventilation rates mediated a modest survival decrease, although the absolute changes were all less than 1.5%. MeaningAsian, Black, and Hispanic patients had a lower rate of invasive ventilation than white patients. Although this difference had no impact on 28-day survival for the average patient, the change in survival mediated by lower rates of invasive ventilation could range from slight benefit at lower inspired oxygen fractions to slight harm at inspired oxygen fraction of 1.0.

RATIONALEThe optimal treatment for early hypoxemic respiratory failure is unclear, and both high-flow nasal cannula and non-invasive ventilation are used. Determining clinically relevant outcomes for evaluating non-invasive respiratory support modalities remains a challenge. OBJECTIVESTo compare the effectiveness of initial treatment with high-flow nasal cannula versus non-invasive ventilation for acute hypoxemic respiratory failure. METHODSWe conducted a retrospective cohort study of patients with acute hypoxemic respiratory failure treated with high-flow nasal cannula or non-invasive ventilation within 24 hours of Emergency Department arrival (1/2018-12/2022). We matched patients 1:1 using a propensity score for odds of receiving non-invasive ventilation. The primary outcome was major adverse pulmonary events (28-day mortality, ventilator-free days, non-invasive respiratory support hours) calculated using a Win Ratio. MEASUREMENTS AND MAIN RESULTS1,265 patients met inclusion criteria. 795 (62.8%) received high-flow oxygen and 470 (37.2%) received non-invasive ventilation. We propensity score matched 736/1,265 (58.2%) patients. There was no difference between non-invasive ventilation vs high-flow nasal cannula in 28-day mortality (17.7% vs 23.1%, p=0.08) or ventilator-free days (median [Interquartile Range]: 28 [25, 28] vs 28 [13, 28], p=0.50), but patients on non-invasive ventilation required treatment for fewer hours (median 7 vs 13, p< 0.001). Win Ratio for composite major adverse pulmonary events favored non-invasive ventilation (1.26, 95%CI 1.06-1.49, p< 0.001). CONCLUSIONSIn this observational study of patients with acute hypoxemic respiratory failure, initial treatment with non-invasive ventilation was superior to high-flow nasal cannula for major pulmonary adverse events. Evaluation of composite outcomes is important in the assessment of respiratory support modalities.

ObjectiveTo develop and validate a predictive model incorporating behavioral telemetry signals--documentation pattern anomalies derived from routine EHR charting--alongside clinical variables for ICU mortality prediction in patients with low acute physiologic derangement. Materials and MethodsRetrospective cohort study of 46,002 adult ICU stays from MIMIC-IV v3.1 (2008-2022) with SOFA scores 0-2, excluding neurological units. We extracted 66 variables spanning demographics, acuity, behavioral telemetry, clinical enrichment, and temporal factors. Progressive logistic regression models (M1-M7) were compared using cross-validation, DeLong tests, net reclassification improvement, and calibration analysis. ResultsOverall mortality was 9.34% (4,295 deaths). The clinical model (M5) achieved cross-validated AUROC 0.691 versus 0.639 for demographics alone (M2; {Delta}AUROC = 0.052, DeLong p = 4.41x10-47). NRI was 24.3%. Discordant care patients received 30.5% more chart events than concordant patients, with the sole deficit in neurological assessments (-15.4%), refuting the neglect hypothesis. Kaplan-Meier analysis confirmed survival separation (log-rank {chi}2 = 138.6, p = 5.32x10-32). In the most conservative subgroup (SOFA 0, no sedation, no ventilation, N = 11,158), orientation omission remained associated with mortality (adjusted OR 1.52, p = 0.027). DiscussionDeep sedation and mechanical ventilation function as mediators on the causal pathway rather than traditional confounders; the discordant care signal retains significance after full sedation adjustment. ConclusionDocumentation pattern analysis adds measurable predictive value for ICU mortality risk stratification and represents a novel signal for real-time EHR-based clinical decision support.

BackgroundSystemic hemodynamics and specific ventilator settings have been shown to predict survival during venoarterial extracorporeal membrane oxygenation (VA ECMO). While these factors are intertwined with right ventricular (RV) function, the independent relationship between RV function and survival during VA ECMO is unknown. ObjectivesTo identify the relationship between RV function with mortality and duration of ECMO support. MethodsCardiac ECMO runs in adults from the Extracorporeal Life Support Organization (ELSO) Registry between 2010 and 2022 were queried. RV function was quantified via pulmonary artery pulse pressure (PAPP) for pre-ECMO and on-ECMO periods. A multivariable model was adjusted for Society for Cardiovascular Angiography and Interventions (SCAI) stage, age, gender, and concurrent clinical data (i.e., pulmonary vasodilators and systemic pulse pressure). The primary outcome was in-hospital mortality. ResultsA total of 4,442 ECMO runs met inclusion criteria and had documentation of hemodynamic and illness severity variables. The mortality rate was 55%; non-survivors were more likely to be older, have a worse SCAI stage, and have longer pre-ECMO endotracheal intubation times (P < 0.05 for all) than survivors. Improving PAPP from pre-ECMO to on-ECMO time ({Delta} PAPP) was associated with reduced mortality per 10 mm Hg increase (OR: 0.91 [95% CI: 0.86-0.96]; P=0.002). Increasing on-ECMO PAPP was associated with longer time on ECMO per 10 mm Hg (Beta: 15 [95% CI: 7.7-21]; P<0.001). ConclusionsEarly improvements in RV function from pre-ECMO values were associated with mortality reduction during cardiac ECMO. Incorporation of {Delta} PAPP into risk prediction models should be considered.

BackgroundPatients with severe SARS-CoV-2 pneumonia experience longer durations of critical illness yet similar mortality rates compared to patients with severe pneumonia secondary to other etiologies. As secondary bacterial infection is common in SARS-CoV-2 pneumonia, we hypothesized that unresolving ventilator-associated pneumonia (VAP) drives the apparent disconnect between length-of-stay and mortality rate among these patients. MethodsWe analyzed VAP in a prospective single-center observational study of 585 mechanically ventilated patients with suspected pneumonia, including 190 patients with severe SARS-CoV-2 pneumonia. We developed CarpeDiem, a novel machine learning approach based on the practice of daily ICU team rounds to identify clinical states for each of the 12,495 ICU patient-days in the cohort. We used the CarpeDiem approach to evaluate the effect of VAP and its resolution on clinical trajectories. FindingsPatients underwent a median [IQR] of 4 [2,7] transitions between 14 clinical states during their ICU stays. Clinical states were associated with differential hospital mortality. The long length-of-stay among patients with severe SARS-CoV-2 pneumonia was associated with prolonged stays in clinical states defined by severe respiratory failure and with a lower frequency of transitions between clinical states. In all patients, including those with COVID-19, unresolving VAP episodes were associated with transitions to unfavorable states and hospital mortality. InterpretationCarpeDiem offers a machine learning approach to examine the effect of VAP on clinical outcomes. Our findings suggest an underappreciated contribution of unresolving secondary bacterial pneumonia to outcomes in mechanically ventilated patients with pneumonia, including due to SARS-CoV-2. O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=110 SRC="FIGDIR/small/22280118v1_ufig1.gif" ALT="Figure 1"> View larger version (13K): org.highwire.dtl.DTLVardef@1b4d28eorg.highwire.dtl.DTLVardef@6b8220org.highwire.dtl.DTLVardef@2c2f33org.highwire.dtl.DTLVardef@ced3ed_HPS_FORMAT_FIGEXP M_FIG Graphical abstract Disentangling the contributions of ICU complications and interventions to ICU outcomes. (A) Traditional approaches evaluate the ICU stay as a black box with severity of illness measured on presentation and dichotomized survival at an arbitrary time point (e.g., day 28) or on ICU or hospital discharge. Hence, the effect of intercurrent complications and interventions cannot be easily measured, a problem that is compounded when ICU stays are long or significantly differ between groups. (B) Defining the ICU course by clinical features during each day in the ICU permits the association of a complication or intervention with transitions toward clinical states associated with favorable or unfavorable outcomes. C_FIG

ImportanceDelay in transfer to Intensive Care Unit (ICU) is associated with known adverse clinical and economic outcomes. There are several early warning systems (EWS) that help identify patients that could benefit from earlier ICU transfer but are fraught with challenges when used to measure delays. An objective time stamped blood test metric, such as a blood gas analysis (BGA), could be a valuable adjunct in identifying patients and measuring delays in who require intrahospital transfer to the ICU. BackgroundDelays in transferring critically ill patients to the ICU are linked to increased mortality, organ failure, prolonged recovery, and higher hospital costs. While EWS systems like MEWS, NEWS2, and eCART aim to detect deterioration using vital signs and medical data, they often rely on intermittent and/or subjective inputs. Despite advances, including AI-driven models, most systems still lack accuracy to detect and quantify transfer delays, an important operational metric. ObjectiveThis review explores the clinical and operational impact of ICU transfer delays and evaluates the potential role of BGA as an objective, time-stamped adjunct biomarker for early identification of high-risk patients. We also assess whether BGA could be integrated into EWS tools to enhance predictive accuracy. MethodsWe conducted a systematic literature review of studies published between 1994 and 2024 using PubMed, EMBASE, Cochrane, and NIH databases. Inclusion criteria focused on studies that examined ICU transfer delays, BGA parameters (e.g., lactate, pH, base excess), and clinical outcomes in adult or pediatric patients. Studies were excluded if they had small sample sizes (n < 50), lacked outcome data, or were not published in English. ResultsThe review found that delays in ICU transfer are consistently linked to worse clinical outcomes and higher healthcare costs. While EWS tools have improved early recognition of patient deterioration, they still lack objective, time-stamped markers to measure delays. Approximately one-third of the included studies specifically examined BGA parameters in relation to ICU transfer or outcomes. Elevated lactate levels and abnormal pH values correlated with increased ICU admission, adverse prognosis and mortality risk. Despite this, BGA is not currently integrated into most clinical decision-making tools used for ICU triage. ConclusionBGA represents a promising, underutilized tool that could fill a critical gap in current ICU triage systems. As a time-stamped, objective measure of physiological instability, BGA could enhance the accuracy, timeliness and measurability of ICU transfer decisions--especially when combined with electronic medical records and modern EWS platforms. Future research should focus on evaluating BGA as a predictive input within next-generation EWS tools, with the goal of reducing ICU transfer delays, improving patient outcomes, and optimizing hospital resource use.

ImportanceAcute kidney injury (AKI) affects one-third of patients after cardiac surgery and increases morbidity and mortality. AKI lasting over 48 hours, known as persistent AKI (pAKI), has much worse outcomes. Hemodynamic optimization is cornerstone of AKI management, however, current strategies rely on bundled care interventions that are inconsistently implemented, underscoring the need for personalized hemodynamic optimization. ObjectiveTo develop and validate a reinforcement learning (RL) model to guide individualized dosing of intravenous (IV) fluids, vasopressors, and inotropes for prevention of pAKI after cardiac surgery. DesignCohort study. Model development and internal validation were performed retrospectively in MIMIC-IV, with external validation in SICdb, a European database (retrospective), and then in Mount Sinai Health System cohort using data from Jan 1-Aug 18, 2025). SettingMulticenter retrospective cohort study. ParticipantsAdmissions to ICU after cardiac surgery. ExposuresPostoperative hemodynamic management during first 72 hours of ICU stay using IV fluids, vasopressors, and inotropes. Main Outcomes and MeasuresPrimary outcome was pAKI within 5 days after surgery. The RL model optimized treatment policies through reward-based learning, where higher rewards reflected improved outcome. We assessed model performance relative to clinicians using Fitted Q Evaluation and adjusted weighted pooled logistic regression. ResultsThere were 6,643 adult ICU admissions following cardiac surgery in MIMIC-IV, 2,254 in SICdb, and 846 in MSHS. Median age was 70 years in MIMIC-IV, 70.0 years in SICdb, and 64 years in MSHS cohort with 72%, 73%, and 70% males respectively. AKI occurred in 41.4%, 19.7%, and 22.5% of admissions, with pAKI in 30.5%, 43.0%, and 33.7% of AKI cases, respectively. RL model achieved higher cumulative rewards than clinicians across all cohorts. Concordance between clinician actions and RL models recommendations was associated with lower adjusted odds of pAKI (OR, 0.92 [0.89-0.96] in SICdb; 0.91 [0.86-0.96] in MSHS). RL model favored smaller IV fluid volumes, moderate vasopressor dosing, and greater inotrope use. Conclusions and RelevanceIn this study, personalization of early postoperative hemodynamic management using an RL model was associated with decreased risk of pAKI. These findings suggest that AI guided hemodynamic strategies may enhance postoperative care after cardiac surgery. Key PointsO_ST_ABSQuestionC_ST_ABSCan reinforcement learning (RL) personalize early postoperative hemodynamic management to prevent persistent AKI (pAKI) after cardiac surgery? FindingsIn 9,743 postoperative cardiac surgery ICU admissions across 3 cohorts (MIMIC-IV, SICdb, and Mount Sinai Health System), the RL model achieved higher cumulative rewards than clinician policies and was associated with lower adjusted odds of developing pAKI when clinician actions aligned with model recommendations. The RL model favored smaller intravenous fluid volumes and earlier, graded adjustments in vasopressor and inotrope dosing compared with standard practice. MeaningRL guided individualized hemodynamic management after cardiac surgery shows promise in reducing the risk of persistent AKI and should be tested in randomized clinical trials.

RationaleWhile clinical criteria are used to diagnose and treat pneumonia in critically ill patients, rates of concordance between a clinicians suspicion for pneumonia and a confirmed diagnosis using bronchoalveolar lavage (BAL) results are undefined. Factors that contribute to diagnostic concordance, and clinical outcomes associated with diagnostic discordance, are unknown. Objective(s)To assess rates of diagnostic concordance between clinicians pre-test probability of pneumonia and BAL-confirmed diagnosis, and to identify clinical factors and outcomes associated with diagnostic discordance in an intensive care unit (ICU) population. MethodsThis was a single-center, prospective observational study of intubated, mechanically ventilated patients. From 2018 to 2022, clinicians were asked to provide a pre-test probability of pneumonia on the same day they performed a bronchoalveolar lavage for the patient. ResultsAmong 659 patients, 84% (553/659) had pneumonia. Diagnostic concordance occurred in 80% (445/553) of these cases. Clinicians assigned a low pre-test probability for pneumonia to 20% (109/553) of patients with confirmed pneumonia. Clinicians assigned a high pre-test probability for pneumonia in 28% (30/106) of patients without pneumonia. Therefore, overdiagnosis in the setting of no pneumonia occurred more often than a missed diagnosis in the setting of true pneumonia (28% vs 20%, p = 0.05). Amongst patients with pneumonia, there were no significant differences in vital signs or laboratory values between those assigned a low pre-test probability of pneumonia and those assigned a high pre-test probability of pneumonia. In patients with culture negative pneumonia (n = 117), those assigned a low pre-test probability of pneumonia, compared to those assigned a high pre-test probability of pneumonia, had a longer length of stay in the hospital (36 days vs 18 days, p = 0.02) and the ICU (21 days vs 9 days, p = 0.01). ConclusionsOver-diagnosis, rather than a missed diagnosis, is the more frequent cause of diagnostic discordance. In culture-negative pneumonia, a low-pretest probability is associated with longer lengths of stay in the hospital and ICU. Future research should explore alternative approaches to improve diagnostic accuracy in critically ill patients.

BackgroundPulmonary vascular pressures are freq[ABS]suently elevated in critically ill patients and are associated with worse outcomes. However, whether elevated pressures persist and their impact on outcomes after critical illness is unknown. Research QuestionsWhat factors asre associated with persistently elevated pulmonary vascular pressures? What are the outcomes associated with persistently elevated pulmonary vascular pressures? Study Design and MethodsThis is a single center retrospective cohort study of critically ill patients during the year 2021. Adult patients with a measured tricuspid regurgitant velocity[&ge;]2.8 m/s during critical illness and had a repeat echocardiogram done after hospital discharge were included. Kaplan Meier and logistic regression were used for mortality and multivariate analysis. ResultsOf 540 patients, 257 (47.6%) had an elevated TRV. Of 51 patients with a repeat echocardiogram, 33 (64.7%) had an elevated repeat TRV. These patients had higher heart rates (91{+/-}23 vs 73{+/-}14 bpm, p<0.01), lower hemoglobin levels (8.42{+/-}1.76 vs 10.0{+/-}2.16, p=0.02), decreased TAPSE (1.90 {+/-} 0.52 mm vs 2.23 {+/-} 0.43, p = 0.03), increased RV middle diameter(3.27{+/-}0.85 vs 2.72{+/-}0.78, p=0.04) and decreased left ventricular stroke volume (61.76{+/-}15.10 vs 84.36{+/-}27.09, p=0.01) compared to those with a normal repeat TRV. Hemoglobin (p=0.03, 95% CI: 0.30-0.90) and SVI (p=0.03, 95% CI: 0.77-0.98) were associated with elevated repeat TRV levels. Elevated TRV on repeat echocardiogram was not associated with worse survival (log-rank test, p=0.33). InterpretationElevated pulmonary vascular pressures persisted after critical illness in a large number of patients, although the impact of persistently elevated pulmonary vascular pressures is uncertain.

BackgroundMixed trial results suggest that some ventilated patients with acute respiratory distress syndrome (ARDS) benefit from high PEEP while others may be harmed, indicating heterogeneity of treatment effect (HTE). This study applies data-driven predictive approaches to uncover HTE and re-examines previously hypothesized HTE. This manuscript serves as a pre-registration of planned external validation of our trained models. MethodsWe identified eight randomized trials, and obtained individual patient data (IPD) from three of them (ALVEOLI, LOVS, EXPRESS), as our train cohort. We used effect modelling to predict individualized treatment effects (predicted 28-day mortality risk difference between PEEP strategies) across patient subgroups stratified by observed tertiles ([&le;]8 cmH2O, 9-11 cmH2O, [&ge;]12 cmH2O). Candidate effect modelling methods included meta-learners and technique-specific methods. Optimal methods were selected through leave-one-trial-out cross-validation, evaluating the methods performances in each PEEP tertile using AUC-benefit. We trained final models using the best performing methods implemented with or without forward selection (which yielded sufficient AUC-benefit), and additional final models by selecting the variables that yielded consistency in the forward selections performed in the cross validation, if any. We further evaluated earlier hypothesized HTE comparing (1) patients with baseline PaO2/FiO2 [&le;] 200 versus > 200 mmHg, and (2) patients with hypoinflammatory versus hyperinflammatory subphenotypes. Preliminary findingsIn the lower PEEP tertile ([&le;]8 cmH2O), an X-learner implemented without, and an S-learner implemented with forward selection (both with flexible base learners), yielded the highest AUC benefits and were used to train final models. In the high PEEP tertile ([&ge;]12 cmH2O), only the causal forest implemented with forward selection yielded an AUC benefit exceeding zero. Respiratory-system compliance (CRS) was consistently selected in the forward selections of cross validation, and was used to train an extra final causal forest model, with predicted effects shifting from harm to benefit for CRS 26.5 mL/cmH2O or higher. Higher PEEP benefited patients with baseline PaO2/FiO2 [&le;]200 mmHg (OR 0.80, 95% CI 0.66-0.98), incurred harm among those with PaO2/FiO2 >200 mmHg (OR 1.74, 95% CI 1.02-2.98; interaction P=0.01). This HTE was strongest when PaO2/FiO2 was measured at low PEEP ([&le;]8 cmH2O), reduced at mid-level PEEP (9-11 cmH2O), and negligible at high PEEP ([&ge;]12 cmH2O). A second-order interaction showed significant heterogeneity of HTE (ie, second-order heterogeneity) across PEEP tertiles (P=0.03). Preliminary ConclusionsOur preliminary findings indicated that baseline CRS [&ge;] 26.5 mL/cmH2O predicts benefit, while CRS < 26.5 mL/cmH2O predicts harm from high PEEP when CRS is measured at high baseline PEEP ([&ge;]12 cmH2O). Similarly, baseline PaO2/FiO2 [&le;] 200 mmHg predicts benefit, while PaO2/FiO2 > 200 mmHg predicts harm from high PEEP when PaO2/FiO2 is measured at a low baseline PEEP ([&le;]8 cmH2O). Using data from the LOVS trial, we investigated HTE for high PEEP between hypo- and hyperinflammatory subphenotypes but found none, despite significant HTE observed earlier in the ALVEOLI trial.

BackgroundPulse oximeters are typically validated on cohorts of 200-500 subjects under controlled conditions. Whether these cohorts capture the demographic heterogeneity of national clinical practice -- and whether measurement error is associated with patient outcomes -- has not been established at scale. MethodsWe analyzed paired SpO2/SaO2 readings from three independent sources spanning 209 U.S. hospitals: MIMIC-IV (1 hospital; 12,934 ICU stays), eICU-CRD (208 hospitals; 55,178 stays), and the Open Oximetry Repository (PhysioNet; 52.4 million readings with continuous melanin and perfusion indices). Bias was defined as SpO2 - SaO2. Hidden hypoxemia (SpO2 [&ge;] 94% with SaO2 < 88%) was assessed per ICU stay. Mortality was compared between hidden-hypoxemia-positive and -negative stays with multivariable logistic regression adjusting for age, sex, race, and four laboratory severity markers (cluster-robust SEs by hospital). Sensitivity analyses included landmark restriction (first 48 hours), lactate stratification, alternate thresholds, and patient-level aggregation. PPG signal quality was assessed in 125 ICU patients with demographic-linked waveform data. ResultsBias was minimal at normal perfusion but amplified under low perfusion in high-melanin patients, consistent with known optics: at very low perfusion x high melanin x severe hypoxia, mean bias reached +12.8% (n = 458,571), with 47% of readings constituting hidden severe hypoxemia. National bias in African American patients was +2.76% (n = 529,541; 208 hospitals), 62% higher than academic estimates. Across 55,178 eICU stays, hidden hypoxemia was associated with approximately doubled mortality after adjustment for age, sex, race, and illness severity (adjusted OR 1.86, 95% CI 1.69-2.04, p < 0.001), consistent across all racial groups. Hidden hypoxemia was not a pre-terminal phenomenon: 63% of events occurred >48 hours before death (median first event: 15.3 hours; mean time to death: 151 hours), and the association persisted in landmark analysis (first 48 hours only), in patients with normal lactate (adjusted OR 1.87, 95% CI 1.61-2.16), and when both restrictions were applied simultaneously (16.5% vs. 11.1%). Waveform analysis (n = 125) showed no fixed racial difference in baseline PPG AC/DC ratio (Black: 0.299, White: 0.273), suggesting the signal deficit is conditional on perfusion state. Full extraction (n = 1,545) is in progress. ConclusionsIn this multicenter retrospective analysis, national pulse oximetry variance exceeded published benchmarks and was associated with approximately doubled ICU mortality, replicated across 209 U.S. hospitals. Hidden hypoxemia was not a pre-terminal artifact: events occurred throughout the ICU stay at a constant rate, and mortality associations persisted in landmark and lactate-stratified analyses. These findings suggest that current regulatory validation standards may underestimate the real-world prevalence of demographic bias in pulse oximetry, and that perfusion-dependent mechanisms may offer a target for algorithmic correction.

BackgroundWe conducted a scoping review of machine learning systems that inform individualised cardiovascular resuscitation of adults in hospital with sepsis. Our study reviews the resuscitation tasks that the systems aim to assist with, system robustness and potential to improve patient care, and progress towards deployment in clinical practice. We assume no expertise in machine learning from the reader and introduce technical concepts where relevant. MethodsThis study followed the Preferred Reporting Items for Systematic reviews and Meta-Analyses extension for Scoping Reviews guidance. MEDLINE, EMBASE, Scopus, ClinicalTrials.gov, arXiv, bioRxiv and medRxiv were systematically searched up to September 2021. We present a narrative synthesis of the included studies, which also aims to equip clinicians with an understanding of the foundational machine learning concepts necessary to interpret them. Results73 studies were included with 80% published after 2018. Supervised learning systems were often used to predict septic shock onset. Reinforcement learning systems were increasingly popular in the last five years, and were used to guide specific dosing of fluids and vasopressors. A minority of studies proposed systems containing biological models augmented with machine learning. Sepsis and septic shock were heterogeneously defined and 63% of studies derived their systems using a single dataset. Most studies performed only retrospective internal validation, with no further steps taken towards translating their proposed systems into clinical practice. ConclusionsMachine learning systems can theoretically match, or even exceed, human performance when predicting patient outcomes and choosing the most suitable cardiovascular treatment strategy in sepsis. However, with some notable exceptions, the vast majority of systems to date exist only as proof of concept, with significant barriers to translation.