Development and Validation of the Hypotensive Exposure Duration Index for Mortality Risk Prediction in Critically III Patients

Development and Validation of the Hypotensive Exposure Duration Index for Mortality Risk Prediction in Critically III Patients | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Development and Validation of the Hypotensive Exposure Duration Index for Mortality Risk Prediction in Critically III Patients Xiao-Yan Ding, Hai-Ping Xu, Jing-Ru Zhang, Han Chen This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7288238/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 9 You are reading this latest preprint version Abstract Background Blood pressure management is crucial in critical care, but relationships between pressure patterns and outcomes remain incompletely understood. We analyzed minute-by-minute blood pressure data to develop and validate a novel index quantifying hypotensive exposure burden. Methods In this retrospective study using the Salzburg Intensive Care Database, 11,059 ICU admissions with continuous invasive arterial monitoring were analyzed. Heatmaps were constructed from high-resolution hemodynamic data to visualize relationships between blood pressure thresholds (50–120 mmHg), exposure durations (5 minutes to 6 hours), and mortality. The Hypotensive Exposure Duration Index (HEDI) was developed to quantify cumulative hypotensive burden by integrating exposure across multiple MAP thresholds. HEDI's prognostic value was evaluated through ROC analysis and nine machine learning algorithms. External validation using the eICU database assessed HEDI's consistency across different populations. Results Non-survivors showed significantly higher HEDI compared to survivors (0.75 [-0.25-2.37] vs -0.16 [-0.65-0.53], p < 0.001). HEDI demonstrated increasing predictive capability, with AUC values rising from 0.622 at 24h to 0.686 at 72h post-admission. The Extra Trees classifier achieved exceptional performance (test AUC: 0.996), with HEDI ranking among the top predictive features. Both internal cross-validation and external validation confirmed the model's robustness (accuracy 0.871), demonstrating HEDI's consistent prognostic value across different patient populations regardless of vasopressor use. Conclusions HEDI effectively quantifies cumulative hypotensive burden in critically ill patients, demonstrating significant predictive ability for ICU mortality validated across diverse populations. High-Temporal Resolution Blood Pressure Exposure Hypotension Exposure Hypotensive Exposure Duration Index Machine Learning Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Maintaining adequate tissue perfusion is fundamental in critical care, with mean arterial pressure (MAP) serving as a crucial surrogate marker 1 – 3 . While MAP > 65 mmHg is widely accepted as a safe threshold 4 , 5 , emerging evidence suggests that the relationship between blood pressure exposure and patient outcomes is more complex than previously recognized 6 . Current approaches to evaluating blood pressure exposure in critically ill patients often focus on specific thresholds or predetermined patterns 7 – 11 , failing to capture the dynamic of hemodynamic variations. Our previous work using hourly blood pressure data from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC) database advanced this field by demonstrating that both the magnitude and duration of hypotensive episodes significantly impact patient outcomes, revealing distinct risk patterns through heatmap visualization 12 . However, the hourly data resolution has significant limitations in capturing the true hemodynamic burden experienced by critically ill patients. Rapid blood pressure fluctuations can occur within minutes, with potential cellular and tissue-level consequences even during brief hypotensive episodes 13 . Furthermore, the pattern, frequency, and recovery dynamics of these brief hypotensive episodes may contain important prognostic information that is entirely missed at hourly sampling intervals. To enhance our understanding of these relationships, we leveraged minute-by-minute blood pressure measurements in this study, enabling more precise characterization of exposure patterns. Building upon validated exposure-outcome relationships, we developed the Hypotensive Exposure Duration Index (HEDI), an integrated scoring system that quantifies the cumulative burden of hypotensive episodes in critically ill patients. We hypothesized that this comprehensive assessment would provide superior prognostic value compared to conventional approaches. Materials and Methods Study Design and Population This machine learning model was developed using data from the Salzburg Intensive Care Database (SICdb) v1.0.8 14,15 , a publicly accessible intensive care database collected from four specialized intensive care units at the University Hospital Salzburg between 2013 and 2021. The database contains deidentified clinical information from more than 27,000 ICU admissions, including patient characteristics, vital parameters, laboratory measurements, and medication records. A distinctive feature of SICdb is its dual temporal resolution, preserving both hourly aggregated summaries and minute-by-minute physiological measurements. For external validation of our model, we utilized the eICU Collaborative Research Database 16 , which includes data from over 200,000 ICU stays across multiple centers in the United States. The study was conducted in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines 17 , and database access was granted following completion of required institutional protocols. Data Extraction Data extraction was performed using structured query language (SQL). The following variables were collected: demographic characteristics (age, sex, height, and weight at admission), clinical severity scores; invasive arterial blood pressure measurements, laboratory results, comorbidities, vital signs, and survival status (ICU mortality). Dr. Han Chen have been authorized to extract the data from the SICdb database. Dr. Xiao-Yan Ding has been authorized to extract the data from the eICU Collaborative Research Database (database access certification number: XYD 55860595). The study protocol was approved by the Institutional Review Board of Provincial Hospital Affiliated to Fuzhou University. Missing Data Extreme MAP values exceeding the 99th percentile was replaced by the 99th percentile, and those below the 1st percentile were replaced by the 1st percentile value. This same outlier handling approach was applied to all variables. For baseline data (e.g., highest lactate, lowest hemoglobin), missing values were imputed using either the median or mean based on each variable's distribution (Supplementary File, Table S1 ). Inclusion and Exclusion Criteria The inclusion criteria were: 1) Available invasive MAP data; 2) Adult patients ( \(\:\ge\:\:\) 18 years old); 3) ICU stay of at least 24 hours. The exclusion criteria were: 1) No invasive MAP monitoring within the first 6 hours of ICU admission; 2) Outcome data unavailable. In addition, blood pressure data beyond 28 days after ICU stay were excluded from the analysis. Odds Ratio Deviation Heatmap The analysis of minute-by-minute invasive MAP was performed using predefined thresholds (50 to 120 mmHg with 2-mmHg intervals), exposure durations from 5 minutes to 6 hours were assessed at 5-minute intervals, which was modified from previous study 13 , 18 . Briefly, for each MAP threshold-duration combination, exposure events were identified and their frequencies were calculated. The odds ratios (OR) were computed as the ratio of exposure frequencies between non-survivors and survivors, and the OR deviations were calculated by subtracting the overall population OR from each specific combination's OR. A heatmap visualization was generated to display the OR deviations for each combination. Development of the HEDI HEDI was developed to quantify the cumulative burden of hypotensive episodes after ICU admission. As shown in Fig. 1 , the development of HEDI consisted of two steps. First, for each patient, exposure frequencies were calculated across all MAP threshold-duration combinations during their ICU stay. Then, individual HEDI scores were computed by multiplying these patient-specific exposure frequencies by their corresponding odds ratios from the population heatmap. The final HEDI scores were normalized by the total number of threshold-duration combinations to ensure comparability across patients. Different temporal HEDI scores were calculated using MAP recordings from corresponding time windows after ICU admission, ranging from 24 to 72 hours with 6-hour increments. Validation of HEDI We performed receiver operating characteristic (ROC) curve analysis to evaluate the discriminative ability of HEDI for predicting ICU mortality at different time points after admission (24h to 72h). This allowed us to identify the optimal time window for HEDI assessment and demonstrate the evolution of its predictive performance over time. Nine machine learning algorithms were applied to evaluate HEDI's predictive value for ICU mortality: Artificial Neural Network (ANN), Decision Tree Classifier (DT), Extra Trees Classifier (ET), Gradient Boosting Machine (GBM), K-Nearest Neighbors (KNN), Light Gradient Boosting Machine (LightGBM), Random Forest Classifier (RF), Support Vector Machine (SVM), and EXtreme Gradient Boosting (XGBoost). Data imbalance was addressed through synthetic minority oversampling technique (SMOTE) to create a balanced training dataset. Feature selection process involved: (1) initial ranking of variables using a RF classifier, (2) stepwise feature addition evaluating incremental performance improvement, and (3) termination when performance plateaued at 14 features as measured by area under the receiver operating characteristic curve (AUC-ROC). The dataset was randomly split into training (70%) and testing (30%) sets. Model hyperparameters were optimized using k-fold cross-validation combined with grid search and manual fine-tuning. All features were standardized before model training. Model performance was assessed using multiple metrics: AUC-ROC, F1 score, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Cohen's Kappa score. Feature importance was evaluated using mean decrease in impurity in the tree-based models to identify key predictors of mortality. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) values to quantify individual feature contributions to model predictions. SHAP analysis provided both global feature importance and patient-level explanations of model decisions. External Validation To assess the generalizability of our findings, we externally validated both the time-dependent performance of HEDI and our machine learning models using the eICU Collaborative Research Database. The eICU database includes high-granularity data including vital signs, laboratory measurements and outcomes. The inclusion criteria for the external validation cohort mirrored those of the development cohort: 1) Available invasive MAP data; 2) Adult patients (≥ 18 years old); 3) ICU stay of at least 24 hours. Similarly, exclusion criteria were: 1) No invasive MAP monitoring within the first 6 hours of ICU admission; 2) Outcome data unavailable; 3) MAP data unavailable or insufficient for accurate hypotension duration calculation, including patients without continuous MAP monitoring and those with monitoring gaps > 2h. Data extraction and missing data handling of the eICU database followed the same protocol used for the SICdb. For subgroup analysis in the external validation cohort, we examined HEDI's predictive performance across different patient populations stratified by vasopressor use. We conducted external validations separately for patients receiving vasopressors, those without vasopressor therapy, and the combined population. Statistical Analysis Continuous variables were presented as mean and standard deviation or median and interquartile ranges (IQR) and were compared using a Student's t -test or the Mann-Whitney U test, as appropriate. Categorical variables were presented as counts (percentages) and compared using the chi-square test or Fisher's exact test. All analyses were performed using Python (Version 3.12.7). Statistical analyses utilized pandas and scipy libraries. Data visualization was created with Matplotlib. Statistical significance was defined as p < 0.05. Results Baseline Characteristics A total of 11,059 patients were included, of whom 732 (6.6%) died in the ICU (Figure S1 ). Baseline characteristics of the study population are presented in Table 1 . Non-survivors were older (75 [65–80] vs 70 [60, 75], p < 0.001) and had lower BMI (25.2 [22.8, 29.3] vs 26.1 [23.2, 29.4], p < 0.001). The maximum SOFA score was significantly higher in non-survivors (7 [5, 9] vs 4 [3, 6], p < 0.001). Table 1 Comparison of Baseline Characteristics Between Survivors and Non-Survivors Survivors (n = 10327) Non-Survivors (n = 732) P value Age, (years) 70 [60, 75] 75 [65, 80] < 0.001 Male, (%) 6638 (64.3%) 459 (62.7%) 0.391 Weight, (kg) 75 [65, 90] 75 [65, 85] 0.002 Height, (m) 1.7 [1.65, 1.75] 1.7 [1.65, 1.75] 0.985 BMI, (kg/m 2 ) 26.1 [23.2, 29.4] 25.2 [22.8, 29.3] < 0.001 Minimum SpO2, (%) 93 [91, 94] 89 [84, 92] < 0.001 Minimum RR, (bpm) 17 [15, 19] 20 [17, 26] < 0.001 Maximum T, (℃) 37.8 [37.5, 38.2] 37.8 [37.3, 38.4] 0.718 Hypertension, n (%) 5136 (49.7%) 344 (47.0%) 0.163 Renal Dysfunction, n (%) 1450 (14.0%) 127 (17.3%) 0.016 Diabetes, n (%) 1744 (16.9%) 114 (15.6%) 0.386 Lung Disease, n (%) 1262 (12.2%) 132 (18.0%) < 0.001 Fluid Input, (ml) 10141 [7131, 13351] 13035 [7658, 19918] < 0.001 Blood Product Input, (ml) 0 [0, 250] 0 [0, 1250] < 0.001 Urine Output, (ml) 3960 [2435, 5810] 2495 [964, 4236] < 0.001 Fluid Balance, (ml) 5091 [2834, 8117] 9684 [4413, 17247] < 0.001 Minimum Hb, (g/L) 8.9 [7.8, 10.3] 8.5 [7.5, 10.1] < 0.001 Maximum WBC, (x10 9 /L) 12.7 [10.0, 16.3] 15.1 [11.1, 20.2] < 0.001 Minimum MCV 87.6 [84.4, 90.9] 87.3 [84.1, 91.4] 0.518 Minimum MCHC 33.6 [32.8, 34.3] 33.2 [32.3, 34.1] < 0.001 Minimum MCH, (g/L) 30.0 [28.8, 31.1] 29.9 [28.8, 30.9] 0.061 Maximum Lactate, (mmol/L) 2.49 [1.75, 3.44] 4.47 [2.68, 8.34] < 0.001 Maximum Cr, (mg/dL) 1.00 [0.80, 1.34] 1.62 [1.10, 2.50] < 0.001 Minimum PLT, (x10 9 /L) 148 [113, 194] 122 [72, 181] < 0.001 Maximum SOFA 4 [3, 6] 7 [5, 9] < 0.001 Maximum HCT, (%) 37.0 [32.0, 40.8] 35.0 [31.0, 40.0] < 0.001 Minimum BE -3.7 [-5.9, -1.6] -7.40 [-12.00, -3.90] < 0.001 Minimum Mg (mmol/L) 0.80 [0.74, 0.87] 0.85 [0.77, 0.96] < 0.001 Minimum Sodium, (mmol/L) 138 [136, 140] 138 [135, 141] 0.215 Minimum pH 7.33 [7.28, 7.37] 7.23 [7.13, 7.31] < 0.001 Blood Product Transfusion, n (%) 3283 (31.8%) 351 (48.0%) < 0.001 Positive Balance, n (%) 9734 (94.3%) 704 (96.2%) 0.036 Negative Balance, n (%) 590 (5.7%) 28 (3.8%) 0.039 Vasopressor Used, n (%) 7728 (74.8%) 638 (87.2%) < 0.001 Hyperkalemia, n (%) 1523 (14.7%) 232 (31.7%) < 0.001 Hyperthermia, n (%) 1262 (12.2%) 172 (23.5%) < 0.001 Fever, n (%) 9023 (87.4%) 601 (82.1%) < 0.001 Mean MAP, (mmHg) 75 [70., 81] 70 [66, 77] < 0.001 Minimum MAP, (mmHg) 49 [49, 49] 49 [49, 49] 0.032 Maximum MAP, (mmHg) 120 [120, 120] 120 [120, 120] 0.963 TWMAP, (mmHg) 75 [70, 81] 70 [66, 77] < 0.001 BE: Base Excess; BMI: body mass index; bpm: breath per minute; Cr: Creatinine; Hb: Hemoglobin; HCT: Hematocrit; HEDI: Hypotension Exposure Duration Index; MAP: Mean Artery Pressure; MCH: Mean Corpuscular Hemoglobin; MCHC: Mean Corpuscular Hemoglobin Concentration; MCV: Mean Corpuscular Volume; Mg: Magnesium; PLT: Platelet Count; RR: Respiratory Rate; SOFA: Sequential Organ Failure Assessment; SpO2: Oxygen Saturation; T: Temperature; TWMAP: Time Weighted Mean Artery Pressure; WBC: White Blood Cell Count. Non-survivors showed worse respiratory parameters, with lower minimum SpO₂ (89 [84, 92] vs 93 [91, 94], p < 0.001) and higher respiratory rates (20 [17, 26] vs 17 [15, 19], p < 0.001). Laboratory abnormalities in non-survivors included higher maximum lactate (4.47 [2.68, 8.34] vs 2.49 [1.75, 3.44], p < 0.001), higher maximum creatinine (1.62 [1.10, 2.50] vs 1.00 [0.80, 1.34], p < 0.001), and lower hemoglobin (8.5 [7.5, 10.1] vs 8.9 [7.8, 10.3], p < 0.001). Non-survivors had higher rates of lung disease (18.0% vs 12.2%, p < 0.001) and renal dysfunction (17.3% vs 14.0%, p = 0.016), more frequently required vasopressor support (87.2% vs 74.8%, p < 0.001), and received more fluid (13,035 [7,658, 19,918] vs 10,141 [7,131, 13,351], p < 0.001). No significant differences were observed in gender distribution or history of hypertension. MAP Exposure Pattern Analysis The relationship between MAP exposure patterns and ICU mortality risk was visualized through OR deviation heatmaps (Fig. 1 ). A white dashed line (OR deviation = 0) divided the heatmap into two distinct regions: the red region representing mortality risk and the blue region indicating survival benefit. The red region (positive OR deviation) extends further into higher MAP values as exposure duration increases, while the blue region (negative OR deviation) is most prominent at higher MAP values and shorter exposure durations. Time-dependent ROC Analysis of HEDI Based on these exposure patterns, we calculated the HEDI, which was significantly higher in non-survivors compared to survivors (0.75 [-0.25, 2.37] vs -0.16 [-0.65, 0.53], p < 0.001). The discriminative performance of HEDI for predicting ICU mortality was evaluated at sequential time points (Fig. 2 ). The analysis demonstrated a consistent pattern of increasing predictive capability over time. For the total population, the AUC values progressively increased from 0.622 at 24h to 0.686 at 72h after ICU admission (Fig. 2 A). A similar trend was observed in the validation cohort, with AUC values rising from 0.587 at 24h to 0.681 at 72h (Fig. 2 B). The predictive performance of HEDI improved with each consecutive time interval (Supplementary File, Table S2). Machine Learning Analysis Based on the time-dependent ROC analysis showing optimal discriminative performance at 72 hours post-admission (AUC = 0.686), the 72-hour HEDI values were selected for machine learning model development. This time point was chosen to maximize the predictive capability of the HEDI while maintaining clinical relevance for early intervention. Feature selection analysis identified 14 optimal predictors for model development (Supplementary File, Figure S2). The 72-hour HEDI ranked among the top ten predictors, demonstrating its clinical relevance in mortality prediction. The significant features included age, laboratory parameters (minimum sodium, maximum creatinine), and fluid management indicators. Beyond the top 14 features, additional variables contributed minimally to model performance improvement, justifying the final feature set selection. Nine machine learning algorithms were evaluated for ICU mortality prediction (Fig. 3 ). In both training (Fig. 3 A-B) and test sets (Fig. 3 C-D), the ET model demonstrated the best overall performance. While it showed high discriminative ability on the training set (Fig. 3 A-B), more importantly, it achieved superior generalization with the highest test set AUC of 0.971 (95% CI: 0.966–0.977) (Fig. 3 C-D). LightGBM and XGBoost followed as second and third best performers, respectively (Table S3). Feature importance analysis (Fig. 4 A) revealed that while traditional clinical parameters such as maximum lactate level and minimum peripheral oxygen saturation (SpO₂) demonstrated the strongest predictive power, the 72-hour HEDI contributed significantly to outcome prediction, ranking 8th among all features. Other influential predictors included minimum arterial pH, maximum SOFA score, fluid input, urine output, and minimum respiratory rate. SHAP summary plots revealed that the 72-hour HEDI exhibited a clear directional impact, with higher HEDI values (red) associated with positive SHAP values, indicating increased mortality predictions (Fig. 4 B). The magnitude of HEDI's influence, while not as pronounced as maximum lactate level or SpO₂, showed directionality aligned with its clinical rationale as a marker of hemodynamic dysfunction. Age showed a consistent positive correlation with mortality predictions, while features such as minimum sodium level and fluid input displayed more complex, bi-directional relationships with prediction outcomes, suggesting non-linear physiological effects. External Validation Following the same inclusion and exclusion criteria used for the training cohort, a total of 13,180 patients were identified for external validation (Supplementary File, Figure S3). The baseline characteristic of external validation cohort is presented in Table S4. The optimized model Extra Trees was evaluated using three subsets of the external validation cohort. In the total external validation cohort (n = 13,180), the model showed good overall accuracy (0.871 [0.865, 0.876]) and specificity (0.911 [0.906, 0.916]), with an AUC of 0.727 [0.715, 0.740], indicating acceptable discriminative ability. Among patients not receiving vasoactive medications (n = 9,204), the model achieved higher accuracy (0.901 [0.895, 0.907]) and specificity (0.938 [0.933, 0.944]). Whereas in patients receiving vasoactive medications (n = 3,976), while maintaining good accuracy (0.802 [0.788, 0.814]) and specificity (0.839 [0.827, 0.851]), the model demonstrated improved sensitivity (0.645 [0.611, 0.679]) and slightly better discriminative performance (AUC = 0.742 [0.725, 0.759]). Detailed performance metrics are presented in Table S5. ROC curve analysis confirmed that the model consistently outperformed random prediction across all validation sets (Fig. 5 ). Discussion The main findings of our study were: (1) In this large-scale analysis of high-temporal resolution arterial pressure data, we identified a dynamic relationship between blood pressure thresholds and exposure duration, with mortality risk increasing progressively at lower blood pressure levels and longer exposure durations; (2) The predictive value of HEDI for ICU mortality exhibited a time-dependent pattern, with its discriminative capability progressively strengthening as the MAP monitoring window extended from 24 hours to 72 hours after ICU admission. This consistent trend was observed in both training and validation cohorts, suggesting that HEDI provides enhanced prognostic information regardless of the population studied; (3) Machine learning analysis confirmed HEDI's contribution to ICU mortality prediction. The developed models incorporating HEDI achieved robust performance (AUC > 0.95) and maintained good discriminative capability during external validation, thus establishing HEDI's generalizable prognostic utility in critical care settings. Previous studies examining the relationship between hypotension and mortality have been limited by predefined blood pressure thresholds (60, 70, 75 mmHg, etc.) and coarse temporal resolutions (5, 15, or 120 minutes) 7 , 10 , 19 – 21 , which may not reflect the complex interplay between blood pressure and patient outcomes. Organ injury is more likely to accumulate progressively as blood pressure decreases, rather than occurring suddenly at a specific threshold 22 . While the 2023 PeriOperative Quality Initiative (POQI) international consensus statement emphasized the importance of considering hypotension duration 23 , the optimal approach to quantifying hypotensive burden remains unclear. The relationship between blood pressure and outcomes appears more complex, requiring more refined assessment methods 23 , 24 . Moreover, emerging evidence suggests that brief, frequent hypotensive episodes may significantly impact outcomes 24 . Most blood pressure measurements rely on intermittent noninvasive monitoring at 5-minute intervals, which may underestimate the true incidence of hypotension 25 . Our preliminary work attempted to better understand this relationship through more intensive measurements, but hourly intervals potentially masked important transient exposure events. Therefore, we analyzed the high-temporal resolution (minute by minute) MAP measurements using 5-minute intervals across a range of 5-360 minutes. The minute-resolution heatmap analysis revealed complex relationships between blood pressure exposure and ICU mortality risk, extending our previous findings with higher temporal precision. High-resolution data demonstrated that mortality risk increased non-linearly with both the intensity and duration of blood pressure deviations. Similar to our previous findings, we observed that achieving target blood pressure alone may be insufficient 13 . In other words, maintaining MAP stability within an optimal range appears crucial for favorable outcomes. Using this high-resolution analytical approach, we have further refined our understanding of the dose-response relationship between blood pressure exposure and clinical outcomes, demonstrating that both the magnitude and duration of hypotension may contribute significantly to patient risk. The HEDI was developed based on the minute-resolution heatmap. Unlike traditional approaches that use single blood pressure thresholds or fixed exposure durations (such as cumulative time with MAP < 65mmHg or 5-minute episodes) 26 , HEDI integrates exposure information across multiple blood pressure thresholds and multiple duration windows, providing a more comprehensive assessment of hypotension exposure burden. Univariate analysis demonstrated significantly higher HEDI values in the mortality group compared to survivors, confirming the association between severe hypotension exposure burden and adverse outcomes. Further ROC curve analysis revealed HEDI's predictive value for ICU mortality, with a progressive improvement in predictive capability as observation time prolonged (from 24 to 72h). This time-dependency suggests that the cumulative effect of hemodynamic fluctuations may better reflect tissue perfusion status than single measurements, aligning with recent trends in critical care medicine emphasizing "time-weighted" physiological parameter assessment 27 , 28 . While longer observation windows (72h) provide superior predictive performance due to the capture of more hemodynamic data, this comes with practical tradeoffs. Extended monitoring periods require more complete datasets, which may not be available for all patients, and delay the availability of predictive insights for clinical decision-making. After considering the balance between predictive accuracy and clinical utility, we selected the 72-hour timepoint for subsequent machine learning analysis, as it offered optimal discrimination while still allowing for interventions within a clinically relevant timeframe for most ICU patients. Multiple machine learning algorithms were evaluated, with Extra Trees identified as the optimal model for mortality prediction. Excellent discrimination ability (AUC > 0.95) was demonstrated in both training and testing sets, indicating strong classification and generalization capabilities. Feature importance analysis revealed that traditional markers of physiological derangement (maximum lactate, minimum SpO 2 , minimum pH, and maximum SOFA score) contributed most significantly to predictions. Although not among the top predictors, HEDI outperformed conventional blood pressure monitoring metrics including minimum MAP and time-weighted-MAP. This comparative advantage carries dual significance: first, it confirms that simultaneous consideration of exposure intensity and duration provides greater value than single-threshold approaches; second, it suggests that the cumulative impact of hemodynamic fluctuations is effectively captured, potentially better approximating the physiological process by which hypotension leads to organ injury. Additionally, the predictive capability of HEDI was evaluated across multiple time windows. The 48-hour HEDI model demonstrated excellent performance in both internal and external validation, achieving an AUC of 0.696, only marginally lower than the 72-hour model. Notably, in high-risk populations receiving vasopressors, the 48-hour HEDI model maintained an AUC of 0.718, indicating that HEDI provides robust predictive information even during shorter observation periods. Importantly, HEDI consistently maintained its high position in feature importance rankings (Supplementary File, Figures S4-8), validating its robustness as a prognostic predictor. These findings support HEDI's potential as an early risk assessment tool providing valuable predictive information within 48 hours of ICU admission, while confirming that extended observation periods allow HEDI to capture more comprehensive hemodynamic burden information, further improving predictive accuracy. As a dynamic indicator that can be automatically calculated and displayed in real-time through integration with modern ICU monitoring systems, HEDI enables continuous risk reassessment. In clinical practice, this characteristic allows HEDI to serve not only for early risk stratification but also for treatment effect evaluation and dynamic prognosis monitoring. Similar to our observations with univariate HEDI analysis, the selection between 48-hour and 72-hour timepoints for machine learning model development represents a fundamental tradeoff between feasibility and accuracy. While the 72-hour model achieved marginally superior discriminative performance, the 48-hour model maintained good predictive capability with the significant advantage of earlier availability. This tradeoff mirrors clinical decision-making processes, where the balance between obtaining more comprehensive data and providing timely interventions must be carefully considered. In practical implementation, healthcare systems might select different timepoints based on their specific patient populations, clinical workflow, and resource availability. This flexibility in application timeframes enhances HEDI's clinical utility across diverse critical care settings. External validation results further support HEDI's clinical utility. In a validation cohort comprising 13,180 patients, our model demonstrated robust predictive performance, particularly in the critically ill subgroup receiving vasopressors. These patients typically represent the most hemodynamically unstable high-risk population with complex treatment strategies and physiological responses. The model maintained performance in this subgroup suggests that the hemodynamic information captured by HEDI exhibits consistency and reliability across populations, further confirming its potential value as a hemodynamic assessment tool. The high accuracy observed in both internal and external validation indicates promising prospects for clinical application. Several limitations should be considered. First, as an observational study, causal relationships between HEDI and mortality cannot be established. The observed associations may be influenced by unmeasured confounding factors, including variations in treatment strategies, or individualized therapeutic targets. Second, our analysis was restricted to patients with arterial catheter monitoring, who typically represent a more critically ill population, potentially limiting the applicability of our findings to the general ICU population. Third, despite external validation, our validation was confined to data from a specific healthcare system, necessitating broader multicenter validation in future studies. Finally, our study did not evaluate interactions between HEDI and other dynamic physiological parameters (e.g., heart rate variability, vascular resistance changes), which might provide additional prognostic information. Despite these limitations, our study uniquely features minute-by-minute hemodynamic data, enabling comprehensive analysis of hypotensive exposures across multiple thresholds and durations. Based on these high-resolution measurements, we developed HEDI, which captures the cumulative impact of varying hypotensive intensities on patient outcomes. Through rigorous internal and external validation, our study confirms HEDI's independent value in predicting mortality risk, particularly among hemodynamically unstable high-risk patients. This novel index surpasses both single-threshold blood pressure monitoring and fixed exposure duration assessments, more accurately reflecting the physiological mechanisms through which hypotensive burden affects outcomes. HEDI demonstrates potential as both a risk stratification tool and a dynamic monitoring indicator, potentially facilitating individualized blood pressure management in critically ill patients and establishing a foundation for precision-medicine in critical care. Conclusion In conclusion, HEDI, as a hemodynamic indicator that simultaneously considers both exposure intensity and duration, demonstrates significant value as a predictor variable in prognostic modeling. Our high-resolution analysis of blood pressure exposure patterns provides new insights into the relationship between hypotension exposure and mortality in critical illness. The HEDI offers a promising tool for early risk stratification, though its clinical utility ultimately needs to be validated in prospective interventional studies. Abbreviations ANN: Artificial Neural Network AUC: Area Under Curve BMI: Body Mass Index DT: Decision Tree ET: Extra Trees GBM: Gradient Boosting Machine HEDI: Hypotensive Exposure Duration Index ICU: Intensive Care Unit KNN: K-Nearest Neighbors LightGBM: Light Gradient Boosting Machine MAP: Mean Arterial Pressure MIMIC: Multiparameter Intelligent Monitoring in Intensive Care NPV: Negative Predictive Value OR: Odds Ratio POQI: PeriOperative Quality Initiative PPV: Positive Predictive Value RF: Random Forest ROC: Receiver Operating Characteristic SICdb: Salzburg Intensive Care Database SMOTE: Synthetic Minority Oversampling Technique SpO₂: Peripheral Oxygen Saturation SQL: Structured Query Language SOFA: Sequential Organ Failure Assessment STROBE: Strengthening the Reporting of Observational Studies in Epidemiology SVM: Support Vector Machine XGBoost: EXtreme Gradient Boosting Declarations Ethics approval and consent to participate This study utilized data from two databases: (1) the eICU Collaborative Research Database, which is exempt from institutional review board approval due to the retrospective design, lack of direct patient intervention, and the security schema, for which the re-identification risk was certified as meeting safe harbor standards by an independent privacy expert (Privacert, Cambridge, MA) (Health Insurance Portability and Accountability Act Certification no. 1031219-2); and (2) the SICdb, which is fully approved by the local ethical commission of the Land Salzburg, Austria (EK Nr: 1115/2021). Consent for publication Not applicable Availability of data and materials The data that support the findings of this study are available from the publicly available critical care database (eICU database and SICdb version 1.0.8), but restrictions apply to the availability of these data. However, available from the authors upon reasonable request and with permission of the holder of the database. Competing interests All authors declare that they have no competing interests. Funding HC is supported by the Youth Top Talent Project of Fujian Provincial Foal Eagle Program. XYD is supported by the Startup Fund for Scientific Research, Fujian Medical University (Grant number: 2021QH1290). HPX was supported by the Natural Science Foundation of Fujian Province(2023J05250) and the Fuzhou Science and Technology Project (2024-S-003). Author's contributions HC has significantly shaped the conceptual framework of this work. Throughout the processes of data extraction, analysis, and manuscript drafting, HC offered valuable suggestions and critical feedback, ensuring that the final version was thoroughly reviewed and approved for publication. XYD played a crucial role in data extraction and manuscript drafting, and contributed to the data analysis and interpretation. HPX primarily performed the data analysis and interpretation. JRZ assisted with literature review and reference collection. Acknowledgements Not applicable. References Malakar S, Singh SK, Usman K. Optimizing Blood Pressure Management in Type 2 Diabetes: A Comparative Investigation of One-Time Versus Periodic Lifestyle Modification Counseling. Cureus . Jun 2024;16(6):e61607. doi:10.7759/cureus.61607 Mistry EA, Hart KW, Davis LT, et al. Blood Pressure Management After Endovascular Therapy for Acute Ischemic Stroke: The BEST-II Randomized Clinical Trial. JAMA . Sep 5 2023;330(9):821-831. doi:10.1001/jama.2023.14330 Welte M, Saugel B, Reuter DA. [Perioperative blood pressure management : What is the optimal pressure?]. Anaesthesist . Sep 2020;69(9):611-622. Perioperatives Blutdruckmanagement : Was ist der optimale Druck? doi:10.1007/s00101-020-00767-w Weiss SL, Peters MJ, Alhazzani W, et al. Surviving Sepsis Campaign International Guidelines for the Management of Septic Shock and Sepsis-Associated Organ Dysfunction in Children. Pediatr Crit Care Med . Feb 2020;21(2):e52-e106. doi:10.1097/pcc.0000000000002198 Evans L, Rhodes A, Alhazzani W, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021. Intensive Care Med . 2021;47(11):1181-1247. doi:10.1007/s00134-021-06506-y Cecconi M, Evans L, Levy M, Rhodes A. Sepsis and septic shock. Lancet . Jul 7 2018;392(10141):75-87. doi:10.1016/s0140-6736(18)30696-2 Kouz K, Weidemann F, Naebian A, et al. Continuous Finger-cuff versus Intermittent Oscillometric Arterial Pressure Monitoring and Hypotension during Induction of Anesthesia and Noncardiac Surgery: The DETECT Randomized Trial. Anesthesiology . 2023;139(3):298-308. doi:10.1097/ALN.0000000000004629 Zuin M, Rigatelli G, Bongarzoni A, et al. Mean arterial pressure predicts 48 h clinical deterioration in intermediate-high risk patients with acute pulmonary embolism. Eur Heart J Acute Cardiovasc Care . 2023;12(2):80-86. doi:10.1093/ehjacc/zuac169 Marshall JC. Choosing the Best Blood Pressure Target for Vasopressor Therapy. JAMA . 2020;323(10):931-933. doi:10.1001/jama.2019.22526 Chen J, Lin J, Wu D, Guo X, Li X, Shi S. Optimal Mean Arterial Pressure Within 24 Hours of Admission for Patients With Intermediate-Risk and High-Risk Pulmonary Embolism. Clin Appl Thromb Hemost . Jan-Dec 2020;26:1076029620933944. doi:10.1177/1076029620933944 Griffin BR, Vaughan-Sarrazin M, Shi Q, et al. Blood Pressure, Readmission, and Mortality Among Patients Hospitalized With Acute Kidney Injury. JAMA Netw Open . May 1 2024;7(5):e2410824. doi:10.1001/jamanetworkopen.2024.10824 Johnson AEW, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. Scientific Data . 2023/01/03 2023;10(1):1. doi:10.1038/s41597-022-01899-x Ding X-Y, Chen Z-Z, Chen H. Both intensity and duration of arterial blood pressure exposure are associated with mortality in critically ill patients: a retrospective database study. British Journal of Anaesthesia . 2025/01/30/ 2025;doi:https://doi.org/10.1016/j.bja.2024.12.020 Rodemund N, Wernly B, Jung C, Cozowicz C, Koköfer A. The Salzburg Intensive Care database (SICdb): an openly available critical care dataset. Intensive Care Med . Jun 2023;49(6):700-702. doi:10.1007/s00134-023-07046-3 Rodemund N, Wernly B, Jung C, Cozowicz C, Koköfer A. Harnessing Big Data in Critical Care: Exploring a new European Dataset. Scientific Data . 2024/03/28 2024;11(1):320. doi:10.1038/s41597-024-03164-9 Pollard TJ, Johnson AEW, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. Scientific Data . 2018/09/11 2018;5(1):180178. doi:10.1038/sdata.2018.178 von Elm E, Altman DG, Egger M, Pocock SJ, Gøtzsche PC, Vandenbroucke JP. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement: Guidelines for Reporting Observational Studies. Annals of Internal Medicine . 2007/10/16 2007;147(8):573-577. doi:10.7326/0003-4819-147-8-200710160-00010 Ding XY, Chen ZZ, Chen H. Visualizing ICP "Dose" of neurological critical care patients. Intensive Care Med . May 2024;50(5):781-783. doi:10.1007/s00134-024-07424-5 Gregory A, Stapelfeldt WH, Khanna AK, et al. Intraoperative Hypotension Is Associated With Adverse Clinical Outcomes After Noncardiac Surgery. Anesth Analg . Jun 1 2021;132(6):1654-1665. doi:10.1213/ane.0000000000005250 Khanna AK, Kinoshita T, Natarajan A, et al. Association of systolic, diastolic, mean, and pulse pressure with morbidity and mortality in septic ICU patients: a nationwide observational study. Ann Intensive Care . Feb 20 2023;13(1):9. doi:10.1186/s13613-023-01101-4 Knight J, Hill A, Melnyk V, et al. Intraoperative Hypoxia Independently Associated With the Development of Acute Kidney Injury Following Bilateral Orthotopic Lung Transplantation. Transplantation . Apr 1 2022;106(4):879-886. doi:10.1097/tp.0000000000003814 Wesselink EM, Wagemakers SH, van Waes JAR, Wanderer JP, van Klei WA, Kappen TH. Associations between intraoperative hypotension, duration of surgery and postoperative myocardial injury after noncardiac surgery: a retrospective single-centre cohort study. Br J Anaesth . Oct 2022;129(4):487-496. doi:10.1016/j.bja.2022.06.034 Saugel B, Fletcher N, Gan TJ, Grocott MPW, Myles PS, Sessler DI. PeriOperative Quality Initiative (POQI) international consensus statement on perioperative arterial pressure management. Br J Anaesth . Aug 2024;133(2):264-276. doi:10.1016/j.bja.2024.04.046 Saab R, Wu BP, Rivas E, et al. Failure to detect ward hypoxaemia and hypotension: contributions of insufficient assessment frequency and patient arousal during nursing assessments. Br J Anaesth . Nov 2021;127(5):760-768. doi:10.1016/j.bja.2021.06.014 Szrama J, Gradys A, Bartkowiak T, Woźniak A, Kusza K, Molnar Z. Intraoperative Hypotension Prediction—A Proactive Perioperative Hemodynamic Management—A Literature Review. Medicina . 2023;59(3). doi:10.3390/medicina59030491 Dupont V, Bonnet-Lebrun AS, Boileve A, et al. Impact of early mean arterial pressure level on severe acute kidney injury occurrence after out-of-hospital cardiac arrest. Ann Intensive Care . Jul 18 2022;12(1):69. doi:10.1186/s13613-022-01045-1 Li Z, Zhao X, Wang D, et al. Reliability and accuracy analysis of time-weighted average exposure to heavy metals based on personal exposure. Sci Total Environ . Aug 10 2022;833:155209. doi:10.1016/j.scitotenv.2022.155209 Maheshwari K, Shimada T, Yang D, et al. Hypotension Prediction Index for Prevention of Hypotension during Moderate- to High-risk Noncardiac Surgery. Anesthesiology . Dec 1 2020;133(6):1214-1222. doi:10.1097/aln.0000000000003557 Additional Declarations No competing interests reported. Supplementary Files SUPPLYMENTARYFile.docx Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 03 Sep, 2025 Reviews received at journal 03 Sep, 2025 Reviews received at journal 21 Aug, 2025 Reviewers agreed at journal 18 Aug, 2025 Reviewers agreed at journal 11 Aug, 2025 Reviewers invited by journal 11 Aug, 2025 Editor assigned by journal 05 Aug, 2025 Submission checks completed at journal 05 Aug, 2025 First submitted to journal 04 Aug, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-7288238","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":500162028,"identity":"febdc71b-d83c-4537-97e9-5259a9abee70","order_by":0,"name":"Xiao-Yan Ding","email":"","orcid":"","institution":"Fujian Provincial Hospital","correspondingAuthor":false,"prefix":"","firstName":"Xiao-Yan","middleName":"","lastName":"Ding","suffix":""},{"id":500162029,"identity":"c0f415e8-62b5-4420-b55f-12299bd299ac","order_by":1,"name":"Hai-Ping Xu","email":"","orcid":"","institution":"Minjiang University","correspondingAuthor":false,"prefix":"","firstName":"Hai-Ping","middleName":"","lastName":"Xu","suffix":""},{"id":500162030,"identity":"1b1299f1-573d-423a-b897-b45f731d8079","order_by":2,"name":"Jing-Ru Zhang","email":"","orcid":"","institution":"Fujian Medical University","correspondingAuthor":false,"prefix":"","firstName":"Jing-Ru","middleName":"","lastName":"Zhang","suffix":""},{"id":500162031,"identity":"090a4723-007e-4081-9bcd-0b9395e6fab5","order_by":3,"name":"Han Chen","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAzklEQVRIiWNgGAWjYBACAyBmhjCZDxxgKADSB4jXwpZwAMwlQQuPAQNRWszZzx78XFBxh4F/ds/HAz8MbBL7DjA/fHQDjxbLnrxk6RlnnjFI3Dm74WCPQVrizANsxsY5+Bx2IMeMmbftMIOBRO4GIHk4ccMBHjZpvFrOvwFq+QfSkvOASC03QLY0gLUwEKvljbE0z7HDPBI30gxAfjGeeZiQX87nGH7mqTksxz8j+fGHHxU2sn3Hmx8+xqcFBngQTGYilI+CUTAKRsEowA8Ag8xO2jew5pUAAAAASUVORK5CYII=","orcid":"","institution":"Fujian Provincial Hospital","correspondingAuthor":true,"prefix":"","firstName":"Han","middleName":"","lastName":"Chen","suffix":""}],"badges":[],"createdAt":"2025-08-04 07:38:38","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7288238/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7288238/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":89276207,"identity":"f0e8ad8f-ec29-4103-b2c7-4544cae835e5","added_by":"auto","created_at":"2025-08-18 09:30:54","extension":"jpg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":437297,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMethodology for calculating the hypotension exposure duration index.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe figure illustrated the calculation process of hypotension exposure duration index (HEDI) based on individualized blood pressure-duration exposure patterns. (A) Heat map showing odds ratio (OR) deviations for different combinations of mean arterial pressure (MAP, ranging from 5 to 360 mmHg in 5-mmHg increments) and exposure durations (ranging from 50 to 120 minutes in 2-minute increments). The heat map exhibited distinct zones where blue regions represented survival benefit while red regions indicated increased mortality risk. A white contour line was fitted to highlight areas with OR deviation equal to zero, demonstrating the transition boundary between beneficial and harmful hemodynamic exposures. (B) Examples of HEDI calculation. Three representative patient exposure patterns with varying counts of blood pressure-duration combinations were presented. For each patient, the exposure count at each MAP-duration combination was multiplied by the corresponding OR deviation value from the heat map to generate a weighted score. The sum of all weighted scores was calculated and normalized by dividing by the total number of possible MAP-duration combinations (2,592), resulting in the final HEDI score. This normalization facilitated score interpretation and clinical application. Higher HEDI scores indicated greater cumulative exposure to hemodynamic patterns associated with adverse outcomes. Note: The numbers used in these examples were arbitrary values for illustration purposes only. Empty cells in the patient examples did not indicate absence of values but were simplified for demonstration purposes.\u003c/p\u003e","description":"","filename":"Figure1.jpg","url":"https://assets-eu.researchsquare.com/files/rs-7288238/v1/50e9c7c9fe4ff0c83e837bb3.jpg"},{"id":89276208,"identity":"daa9be7a-1afd-47c1-9deb-df042fda8934","added_by":"auto","created_at":"2025-08-18 09:30:54","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":265613,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eTime-Dependent Predictive Performance of HEDI for ICU Mortality.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eROC curves demonstrating the predictive performance of HEDI for ICU mortality at different time points after ICU admission. Penal A: Training cohort (SICdb dataset): AUC values progressively increase from 0.622 at 2 h to 0.686 at 72h. Penal B: Validation cohort (eICU dataset): Similar time-dependent improvement in predictive performance is observed, with AUC values increasing from 0.587 at 24h to 0.681 at 72h.\u003c/p\u003e","description":"","filename":"Figure2.png","url":"https://assets-eu.researchsquare.com/files/rs-7288238/v1/fb097eb3e01df1b8801ebffe.png"},{"id":89276213,"identity":"151ede25-1feb-42cf-8166-04ae8c0d190e","added_by":"auto","created_at":"2025-08-18 09:30:54","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":244053,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePerformance Comparison of Machine Learning Models for Predicting ICU Mortality.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePenal A: ROC curves for nine different machine learning models on the training set, all demonstrating excellent discriminative ability with AUC values exceeding 0.95. Penal B: Comprehensive evaluation metrics for all models on the training set, including accuracy, sensitivity, specificity, positive and negative predictive values, F1 score, Kappa score, and ROC AUC. Penal C: ROC curves for the same models applied to the internal test set, maintaining robust performance. Penal D: Evaluation metrics for the internal test set, showing consistent performance across all assessment criteria.\u003c/p\u003e","description":"","filename":"Figure3.png","url":"https://assets-eu.researchsquare.com/files/rs-7288238/v1/43f9efdf98d6d1b7bf74b2da.png"},{"id":89277144,"identity":"c6b6a91a-2cce-4bbe-a0bc-2330014b8cc1","added_by":"auto","created_at":"2025-08-18 09:38:54","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":184543,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eFeature Analysis for ICU Mortality Prediction.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003ePenal A: Feature importance rankings from the Extra Trees model. Maximum lactate level and minimum peripheral oxygen saturation (SpO₂) demonstrate the highest predictive value, while the HEDI ranks 8th among the 14 clinical parameters evaluated, confirming its independent clinical value as a mortality predictor. Penal B: SHAP analysis revealing the magnitude and direction of feature contributions to ICU mortality predictions. Results demonstrate that maximum lactate level has the strongest positive impact on mortality prediction, while the SHAP distribution pattern for HEDI shows that higher values are significantly associated with increased mortality risk. This analysis also reveals complex non-linear relationships between certain parameters (such as urine output and fluid input) and outcomes.\u003c/p\u003e","description":"","filename":"Figure4.png","url":"https://assets-eu.researchsquare.com/files/rs-7288238/v1/b979c885b0bf50c94c3eefb7.png"},{"id":89277143,"identity":"fa9c3fe7-e14b-4ded-8278-7cc583e825ed","added_by":"auto","created_at":"2025-08-18 09:38:54","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":116203,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eExtra Trees Model Performance in External Validation.\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eROC curves demonstrating model performance in training (yellow), test (blue), and external validation cohorts (A: overall eICU population, n = 13,180; B: patients receiving vasopressors, n = 3,976; C: patients without vasopressors, n = 9,204)\u003c/p\u003e","description":"","filename":"Figure5.png","url":"https://assets-eu.researchsquare.com/files/rs-7288238/v1/81750fca724c7a8359a619c4.png"},{"id":89278372,"identity":"7f92162d-55be-47a5-aa94-83a92829de82","added_by":"auto","created_at":"2025-08-18 09:54:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2377235,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7288238/v1/e5a4700a-b5ae-4a4a-9a16-e3f48d45b753.pdf"},{"id":89276210,"identity":"8918cc00-41ce-4b7f-a3e4-99e535cdcdc4","added_by":"auto","created_at":"2025-08-18 09:30:54","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":1191025,"visible":true,"origin":"","legend":"","description":"","filename":"SUPPLYMENTARYFile.docx","url":"https://assets-eu.researchsquare.com/files/rs-7288238/v1/ef63ebc7358eb9b91dafd308.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Development and Validation of the Hypotensive Exposure Duration Index for Mortality Risk Prediction in Critically III Patients","fulltext":[{"header":"Introduction","content":"\u003cp\u003eMaintaining adequate tissue perfusion is fundamental in critical care, with mean arterial pressure (MAP) serving as a crucial surrogate marker \u003csup\u003e\u003cspan additionalcitationids=\"CR2\" citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u003c/sup\u003e. While MAP\u0026thinsp;\u0026gt;\u0026thinsp;65 mmHg is widely accepted as a safe threshold \u003csup\u003e\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e,\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e, emerging evidence suggests that the relationship between blood pressure exposure and patient outcomes is more complex than previously recognized \u003csup\u003e\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u003c/sup\u003e.\u003c/p\u003e\u003cp\u003eCurrent approaches to evaluating blood pressure exposure in critically ill patients often focus on specific thresholds or predetermined patterns \u003csup\u003e\u003cspan additionalcitationids=\"CR8 CR9 CR10\" citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, failing to capture the dynamic of hemodynamic variations. Our previous work using hourly blood pressure data from the Multiparameter Intelligent Monitoring in Intensive Care (MIMIC) database advanced this field by demonstrating that both the magnitude and duration of hypotensive episodes significantly impact patient outcomes, revealing distinct risk patterns through heatmap visualization \u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u003c/sup\u003e. However, the hourly data resolution has significant limitations in capturing the true hemodynamic burden experienced by critically ill patients. Rapid blood pressure fluctuations can occur within minutes, with potential cellular and tissue-level consequences even during brief hypotensive episodes \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. Furthermore, the pattern, frequency, and recovery dynamics of these brief hypotensive episodes may contain important prognostic information that is entirely missed at hourly sampling intervals. To enhance our understanding of these relationships, we leveraged minute-by-minute blood pressure measurements in this study, enabling more precise characterization of exposure patterns. Building upon validated exposure-outcome relationships, we developed the Hypotensive Exposure Duration Index (HEDI), an integrated scoring system that quantifies the cumulative burden of hypotensive episodes in critically ill patients. We hypothesized that this comprehensive assessment would provide superior prognostic value compared to conventional approaches.\u003c/p\u003e"},{"header":"Materials and Methods","content":"\u003cp\u003e\u003cb\u003eStudy Design and Population\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThis machine learning model was developed using data from the Salzburg Intensive Care Database (SICdb) v1.0.8 \u003csup\u003e14,15\u003c/sup\u003e, a publicly accessible intensive care database collected from four specialized intensive care units at the University Hospital Salzburg between 2013 and 2021. The database contains deidentified clinical information from more than 27,000 ICU admissions, including patient characteristics, vital parameters, laboratory measurements, and medication records. A distinctive feature of SICdb is its dual temporal resolution, preserving both hourly aggregated summaries and minute-by-minute physiological measurements. For external validation of our model, we utilized the eICU Collaborative Research Database \u003csup\u003e\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, which includes data from over 200,000 ICU stays across multiple centers in the United States. The study was conducted in accordance with the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) guidelines \u003csup\u003e\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u003c/sup\u003e, and database access was granted following completion of required institutional protocols.\u003c/p\u003e\u003cp\u003e\u003cb\u003eData Extraction\u003c/b\u003e\u003c/p\u003e\u003cp\u003eData extraction was performed using structured query language (SQL). The following variables were collected: demographic characteristics (age, sex, height, and weight at admission), clinical severity scores; invasive arterial blood pressure measurements, laboratory results, comorbidities, vital signs, and survival status (ICU mortality). Dr. Han Chen have been authorized to extract the data from the SICdb database. Dr. Xiao-Yan Ding has been authorized to extract the data from the eICU Collaborative Research Database (database access certification number: XYD 55860595). The study protocol was approved by the Institutional Review Board of Provincial Hospital Affiliated to Fuzhou University.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMissing Data\u003c/b\u003e\u003c/p\u003e\u003cp\u003eExtreme MAP values exceeding the 99th percentile was replaced by the 99th percentile, and those below the 1st percentile were replaced by the 1st percentile value. This same outlier handling approach was applied to all variables. For baseline data (e.g., highest lactate, lowest hemoglobin), missing values were imputed using either the median or mean based on each variable's distribution (Supplementary File, Table \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003cb\u003eInclusion and Exclusion Criteria\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe inclusion criteria were: 1) Available invasive MAP data; 2) Adult patients (\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\ge\\:\\:\\)\u003c/span\u003e\u003c/span\u003e18 years old); 3) ICU stay of at least 24 hours.\u003c/p\u003e\u003cp\u003eThe exclusion criteria were: 1) No invasive MAP monitoring within the first 6 hours of ICU admission; 2) Outcome data unavailable. In addition, blood pressure data beyond 28 days after ICU stay were excluded from the analysis.\u003c/p\u003e\u003cp\u003e\u003cb\u003eOdds Ratio Deviation Heatmap\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe analysis of minute-by-minute invasive MAP was performed using predefined thresholds (50 to 120 mmHg with 2-mmHg intervals), exposure durations from 5 minutes to 6 hours were assessed at 5-minute intervals, which was modified from previous study \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e,\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e\u003c/sup\u003e. Briefly, for each MAP threshold-duration combination, exposure events were identified and their frequencies were calculated. The odds ratios (OR) were computed as the ratio of exposure frequencies between non-survivors and survivors, and the OR deviations were calculated by subtracting the overall population OR from each specific combination's OR. A heatmap visualization was generated to display the OR deviations for each combination.\u003c/p\u003e\u003cp\u003e\u003cb\u003eDevelopment of the HEDI\u003c/b\u003e\u003c/p\u003e\u003cp\u003eHEDI was developed to quantify the cumulative burden of hypotensive episodes after ICU admission. As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e, the development of HEDI consisted of two steps. First, for each patient, exposure frequencies were calculated across all MAP threshold-duration combinations during their ICU stay. Then, individual HEDI scores were computed by multiplying these patient-specific exposure frequencies by their corresponding odds ratios from the population heatmap. The final HEDI scores were normalized by the total number of threshold-duration combinations to ensure comparability across patients. Different temporal HEDI scores were calculated using MAP recordings from corresponding time windows after ICU admission, ranging from 24 to 72 hours with 6-hour increments.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eValidation of HEDI\u003c/b\u003e\u003c/p\u003e\u003cp\u003eWe performed receiver operating characteristic (ROC) curve analysis to evaluate the discriminative ability of HEDI for predicting ICU mortality at different time points after admission (24h to 72h). This allowed us to identify the optimal time window for HEDI assessment and demonstrate the evolution of its predictive performance over time.\u003c/p\u003e\u003cp\u003eNine machine learning algorithms were applied to evaluate HEDI's predictive value for ICU mortality: Artificial Neural Network (ANN), Decision Tree Classifier (DT), Extra Trees Classifier (ET), Gradient Boosting Machine (GBM), K-Nearest Neighbors (KNN), Light Gradient Boosting Machine (LightGBM), Random Forest Classifier (RF), Support Vector Machine (SVM), and EXtreme Gradient Boosting (XGBoost). Data imbalance was addressed through synthetic minority oversampling technique (SMOTE) to create a balanced training dataset. Feature selection process involved: (1) initial ranking of variables using a RF classifier, (2) stepwise feature addition evaluating incremental performance improvement, and (3) termination when performance plateaued at 14 features as measured by area under the receiver operating characteristic curve (AUC-ROC). The dataset was randomly split into training (70%) and testing (30%) sets. Model hyperparameters were optimized using k-fold cross-validation combined with grid search and manual fine-tuning. All features were standardized before model training.\u003c/p\u003e\u003cp\u003eModel performance was assessed using multiple metrics: AUC-ROC, F1 score, sensitivity, specificity, positive predictive value (PPV), negative predictive value (NPV), and Cohen's Kappa score. Feature importance was evaluated using mean decrease in impurity in the tree-based models to identify key predictors of mortality. Model interpretability was assessed using SHapley Additive exPlanations (SHAP) values to quantify individual feature contributions to model predictions. SHAP analysis provided both global feature importance and patient-level explanations of model decisions.\u003c/p\u003e\u003cp\u003e\u003cb\u003eExternal Validation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eTo assess the generalizability of our findings, we externally validated both the time-dependent performance of HEDI and our machine learning models using the eICU Collaborative Research Database. The eICU database includes high-granularity data including vital signs, laboratory measurements and outcomes. The inclusion criteria for the external validation cohort mirrored those of the development cohort: 1) Available invasive MAP data; 2) Adult patients (\u0026ge;\u0026thinsp;18 years old); 3) ICU stay of at least 24 hours. Similarly, exclusion criteria were: 1) No invasive MAP monitoring within the first 6 hours of ICU admission; 2) Outcome data unavailable; 3) MAP data unavailable or insufficient for accurate hypotension duration calculation, including patients without continuous MAP monitoring and those with monitoring gaps\u0026thinsp;\u0026gt;\u0026thinsp;2h. Data extraction and missing data handling of the eICU database followed the same protocol used for the SICdb.\u003c/p\u003e\u003cp\u003eFor subgroup analysis in the external validation cohort, we examined HEDI's predictive performance across different patient populations stratified by vasopressor use. We conducted external validations separately for patients receiving vasopressors, those without vasopressor therapy, and the combined population.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003eStatistical Analysis\u003c/h2\u003e\u003cp\u003eContinuous variables were presented as mean and standard deviation or median and interquartile ranges (IQR) and were compared using a Student's \u003cem\u003et\u003c/em\u003e-test or the Mann-Whitney \u003cem\u003eU\u003c/em\u003e test, as appropriate. Categorical variables were presented as counts (percentages) and compared using the chi-square test or Fisher's exact test. All analyses were performed using Python (Version 3.12.7). Statistical analyses utilized pandas and scipy libraries. Data visualization was created with Matplotlib. Statistical significance was defined as \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.05.\u003c/p\u003e\u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cb\u003eBaseline Characteristics\u003c/b\u003e\u003c/p\u003e\u003cp\u003eA total of 11,059 patients were included, of whom 732 (6.6%) died in the ICU (Figure \u003cspan refid=\"MOESM1\" class=\"InternalRef\"\u003eS1\u003c/span\u003e). Baseline characteristics of the study population are presented in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e. Non-survivors were older (75 [65\u0026ndash;80] vs 70 [60, 75], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and had lower BMI (25.2 [22.8, 29.3] vs 26.1 [23.2, 29.4], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). The maximum SOFA score was significantly higher in non-survivors (7 [5, 9] vs 4 [3, 6], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eComparison of Baseline Characteristics Between Survivors and Non-Survivors\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u0026nbsp;\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eSurvivors\u003c/p\u003e\u003cp\u003e(n\u0026thinsp;=\u0026thinsp;10327)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eNon-Survivors\u003c/p\u003e\u003cp\u003e(n\u0026thinsp;=\u0026thinsp;732)\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eP value\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eAge, (years)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e70 [60, 75]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e75 [65, 80]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMale, (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e6638 (64.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e459 (62.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.391\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eWeight, (kg)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e75 [65, 90]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e75 [65, 85]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.002\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eHeight, (m)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1.7 [1.65, 1.75]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.7 [1.65, 1.75]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.985\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBMI, (kg/m\u003c/b\u003e\u003csup\u003e\u003cb\u003e2\u003c/b\u003e\u003c/sup\u003e\u003cb\u003e)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e26.1 [23.2, 29.4]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e25.2 [22.8, 29.3]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum SpO2, (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e93 [91, 94]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e89 [84, 92]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum RR, (bpm)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e17 [15, 19]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e20 [17, 26]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMaximum T, (℃)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e37.8 [37.5, 38.2]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e37.8 [37.3, 38.4]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.718\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eHypertension, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e5136 (49.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e344 (47.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.163\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eRenal Dysfunction, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1450 (14.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e127 (17.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.016\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eDiabetes, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1744 (16.9%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e114 (15.6%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.386\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLung Disease, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1262 (12.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e132 (18.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eFluid Input, (ml)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e10141 [7131, 13351]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e13035 [7658, 19918]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBlood Product Input, (ml)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0 [0, 250]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0 [0, 1250]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eUrine Output, (ml)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3960 [2435, 5810]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e2495 [964, 4236]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eFluid Balance, (ml)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e5091 [2834, 8117]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e9684 [4413, 17247]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum Hb, (g/L)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e8.9 [7.8, 10.3]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e8.5 [7.5, 10.1]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMaximum WBC, (x10\u003c/b\u003e\u003csup\u003e\u003cb\u003e9\u003c/b\u003e\u003c/sup\u003e\u003cb\u003e/L)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e12.7 [10.0, 16.3]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e15.1 [11.1, 20.2]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum MCV\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e87.6 [84.4, 90.9]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e87.3 [84.1, 91.4]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.518\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum MCHC\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e33.6 [32.8, 34.3]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e33.2 [32.3, 34.1]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum MCH, (g/L)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e30.0 [28.8, 31.1]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e29.9 [28.8, 30.9]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.061\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMaximum Lactate, (mmol/L)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e2.49 [1.75, 3.44]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e4.47 [2.68, 8.34]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMaximum Cr, (mg/dL)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1.00 [0.80, 1.34]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e1.62 [1.10, 2.50]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum PLT, (x10\u003c/b\u003e\u003csup\u003e\u003cb\u003e9\u003c/b\u003e\u003c/sup\u003e\u003cb\u003e/L)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e148 [113, 194]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e122 [72, 181]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMaximum SOFA\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e4 [3, 6]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e7 [5, 9]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMaximum HCT, (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e37.0 [32.0, 40.8]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e35.0 [31.0, 40.0]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum BE\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e-3.7 [-5.9, -1.6]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e-7.40 [-12.00, -3.90]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum Mg (mmol/L)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e0.80 [0.74, 0.87]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e0.85 [0.77, 0.96]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum Sodium, (mmol/L)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e138 [136, 140]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e138 [135, 141]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.215\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum pH\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7.33 [7.28, 7.37]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e7.23 [7.13, 7.31]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eBlood Product Transfusion, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e3283 (31.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e351 (48.0%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003ePositive Balance, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e9734 (94.3%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e704 (96.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.036\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eNegative Balance, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e590 (5.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e28 (3.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.039\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eVasopressor Used, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e7728 (74.8%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e638 (87.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eHyperkalemia, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1523 (14.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e232 (31.7%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eHyperthermia, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e1262 (12.2%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e172 (23.5%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eFever, n (%)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e9023 (87.4%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e601 (82.1%)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMean MAP, (mmHg)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e75 [70., 81]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e70 [66, 77]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMinimum MAP, (mmHg)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e49 [49, 49]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e49 [49, 49]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.032\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMaximum MAP, (mmHg)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e120 [120, 120]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e120 [120, 120]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.963\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eTWMAP, (mmHg)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003e75 [70, 81]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003e70 [66, 77]\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u0026lt;\u0026thinsp;0.001\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003ctfoot\u003e\u003ctr\u003e\u003ctd colspan=\"4\"\u003eBE: Base Excess; BMI: body mass index; bpm: breath per minute; Cr: Creatinine; Hb: Hemoglobin; HCT: Hematocrit; HEDI: Hypotension Exposure Duration Index; MAP: Mean Artery Pressure; MCH: Mean Corpuscular Hemoglobin; MCHC: Mean Corpuscular Hemoglobin Concentration; MCV: Mean Corpuscular Volume; Mg: Magnesium; PLT: Platelet Count; RR: Respiratory Rate; SOFA: Sequential Organ Failure Assessment; SpO2: Oxygen Saturation; T: Temperature; TWMAP: Time Weighted Mean Artery Pressure; WBC: White Blood Cell Count.\u003c/td\u003e\u003c/tr\u003e\u003c/tfoot\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eNon-survivors showed worse respiratory parameters, with lower minimum SpO₂ (89 [84, 92] vs 93 [91, 94], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and higher respiratory rates (20 [17, 26] vs 17 [15, 19], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). Laboratory abnormalities in non-survivors included higher maximum lactate (4.47 [2.68, 8.34] vs 2.49 [1.75, 3.44], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), higher maximum creatinine (1.62 [1.10, 2.50] vs 1.00 [0.80, 1.34], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), and lower hemoglobin (8.5 [7.5, 10.1] vs 8.9 [7.8, 10.3], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001).\u003c/p\u003e\u003cp\u003eNon-survivors had higher rates of lung disease (18.0% vs 12.2%, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001) and renal dysfunction (17.3% vs 14.0%, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;=\u0026thinsp;0.016), more frequently required vasopressor support (87.2% vs 74.8%, \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001), and received more fluid (13,035 [7,658, 19,918] vs 10,141 [7,131, 13,351], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). No significant differences were observed in gender distribution or history of hypertension.\u003c/p\u003e\u003cp\u003e\u003cb\u003eMAP Exposure Pattern Analysis\u003c/b\u003e\u003c/p\u003e\u003cp\u003eThe relationship between MAP exposure patterns and ICU mortality risk was visualized through OR deviation heatmaps (Fig.\u0026nbsp;\u003cspan refid=\"Fig1\" class=\"InternalRef\"\u003e1\u003c/span\u003e). A white dashed line (OR deviation\u0026thinsp;=\u0026thinsp;0) divided the heatmap into two distinct regions: the red region representing mortality risk and the blue region indicating survival benefit. The red region (positive OR deviation) extends further into higher MAP values as exposure duration increases, while the blue region (negative OR deviation) is most prominent at higher MAP values and shorter exposure durations.\u003c/p\u003e\u003cp\u003e\u003cb\u003eTime-dependent ROC Analysis of HEDI\u003c/b\u003e\u003c/p\u003e\u003cp\u003eBased on these exposure patterns, we calculated the HEDI, which was significantly higher in non-survivors compared to survivors (0.75 [-0.25, 2.37] vs -0.16 [-0.65, 0.53], \u003cem\u003ep\u003c/em\u003e\u0026thinsp;\u0026lt;\u0026thinsp;0.001). The discriminative performance of HEDI for predicting ICU mortality was evaluated at sequential time points (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). The analysis demonstrated a consistent pattern of increasing predictive capability over time. For the total population, the AUC values progressively increased from 0.622 at 24h to 0.686 at 72h after ICU admission (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eA). A similar trend was observed in the validation cohort, with AUC values rising from 0.587 at 24h to 0.681 at 72h (Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003eB). The predictive performance of HEDI improved with each consecutive time interval (Supplementary File, Table S2).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eMachine Learning Analysis\u003c/b\u003e\u003c/p\u003e\u003cp\u003eBased on the time-dependent ROC analysis showing optimal discriminative performance at 72 hours post-admission (AUC\u0026thinsp;=\u0026thinsp;0.686), the 72-hour HEDI values were selected for machine learning model development. This time point was chosen to maximize the predictive capability of the HEDI while maintaining clinical relevance for early intervention.\u003c/p\u003e\u003cp\u003eFeature selection analysis identified 14 optimal predictors for model development (Supplementary File, Figure S2). The 72-hour HEDI ranked among the top ten predictors, demonstrating its clinical relevance in mortality prediction. The significant features included age, laboratory parameters (minimum sodium, maximum creatinine), and fluid management indicators. Beyond the top 14 features, additional variables contributed minimally to model performance improvement, justifying the final feature set selection.\u003c/p\u003e\u003cp\u003eNine machine learning algorithms were evaluated for ICU mortality prediction (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e). In both training (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA-B) and test sets (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC-D), the ET model demonstrated the best overall performance. While it showed high discriminative ability on the training set (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eA-B), more importantly, it achieved superior generalization with the highest test set AUC of 0.971 (95% CI: 0.966\u0026ndash;0.977) (Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003eC-D). LightGBM and XGBoost followed as second and third best performers, respectively (Table S3).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eFeature importance analysis (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eA) revealed that while traditional clinical parameters such as maximum lactate level and minimum peripheral oxygen saturation (SpO₂) demonstrated the strongest predictive power, the 72-hour HEDI contributed significantly to outcome prediction, ranking 8th among all features. Other influential predictors included minimum arterial pH, maximum SOFA score, fluid input, urine output, and minimum respiratory rate.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003cp\u003eSHAP summary plots revealed that the 72-hour HEDI exhibited a clear directional impact, with higher HEDI values (red) associated with positive SHAP values, indicating increased mortality predictions (Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e4\u003c/span\u003eB). The magnitude of HEDI's influence, while not as pronounced as maximum lactate level or SpO₂, showed directionality aligned with its clinical rationale as a marker of hemodynamic dysfunction. Age showed a consistent positive correlation with mortality predictions, while features such as minimum sodium level and fluid input displayed more complex, bi-directional relationships with prediction outcomes, suggesting non-linear physiological effects.\u003c/p\u003e\u003cp\u003e\u003cb\u003eExternal Validation\u003c/b\u003e\u003c/p\u003e\u003cp\u003eFollowing the same inclusion and exclusion criteria used for the training cohort, a total of 13,180 patients were identified for external validation (Supplementary File, Figure S3). The baseline characteristic of external validation cohort is presented in Table S4.\u003c/p\u003e\u003cp\u003eThe optimized model Extra Trees was evaluated using three subsets of the external validation cohort. In the total external validation cohort (n\u0026thinsp;=\u0026thinsp;13,180), the model showed good overall accuracy (0.871 [0.865, 0.876]) and specificity (0.911 [0.906, 0.916]), with an AUC of 0.727 [0.715, 0.740], indicating acceptable discriminative ability. Among patients not receiving vasoactive medications (n\u0026thinsp;=\u0026thinsp;9,204), the model achieved higher accuracy (0.901 [0.895, 0.907]) and specificity (0.938 [0.933, 0.944]). Whereas in patients receiving vasoactive medications (n\u0026thinsp;=\u0026thinsp;3,976), while maintaining good accuracy (0.802 [0.788, 0.814]) and specificity (0.839 [0.827, 0.851]), the model demonstrated improved sensitivity (0.645 [0.611, 0.679]) and slightly better discriminative performance (AUC\u0026thinsp;=\u0026thinsp;0.742 [0.725, 0.759]). Detailed performance metrics are presented in Table S5. ROC curve analysis confirmed that the model consistently outperformed random prediction across all validation sets (Fig.\u0026nbsp;\u003cspan refid=\"Fig5\" class=\"InternalRef\"\u003e5\u003c/span\u003e).\u003c/p\u003e\u003cp\u003e\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eThe main findings of our study were: (1) In this large-scale analysis of high-temporal resolution arterial pressure data, we identified a dynamic relationship between blood pressure thresholds and exposure duration, with mortality risk increasing progressively at lower blood pressure levels and longer exposure durations; (2) The predictive value of HEDI for ICU mortality exhibited a time-dependent pattern, with its discriminative capability progressively strengthening as the MAP monitoring window extended from 24 hours to 72 hours after ICU admission. This consistent trend was observed in both training and validation cohorts, suggesting that HEDI provides enhanced prognostic information regardless of the population studied; (3) Machine learning analysis confirmed HEDI's contribution to ICU mortality prediction. The developed models incorporating HEDI achieved robust performance (AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.95) and maintained good discriminative capability during external validation, thus establishing HEDI's generalizable prognostic utility in critical care settings.\u003c/p\u003e\u003cp\u003ePrevious studies examining the relationship between hypotension and mortality have been limited by predefined blood pressure thresholds (60, 70, 75 mmHg, etc.) and coarse temporal resolutions (5, 15, or 120 minutes) \u003csup\u003e\u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e,\u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e,\u003cspan additionalcitationids=\"CR20\" citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e,\u003c/sup\u003e which may not reflect the complex interplay between blood pressure and patient outcomes. Organ injury is more likely to accumulate progressively as blood pressure decreases, rather than occurring suddenly at a specific threshold \u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e\u003c/sup\u003e. While the 2023 PeriOperative Quality Initiative (POQI) international consensus statement emphasized the importance of considering hypotension duration \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, the optimal approach to quantifying hypotensive burden remains unclear. The relationship between blood pressure and outcomes appears more complex, requiring more refined assessment methods \u003csup\u003e\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e,\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Moreover, emerging evidence suggests that brief, frequent hypotensive episodes may significantly impact outcomes \u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Most blood pressure measurements rely on intermittent noninvasive monitoring at 5-minute intervals, which may underestimate the true incidence of hypotension \u003csup\u003e\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u003c/sup\u003e. Our preliminary work attempted to better understand this relationship through more intensive measurements, but hourly intervals potentially masked important transient exposure events. Therefore, we analyzed the high-temporal resolution (minute by minute) MAP measurements using 5-minute intervals across a range of 5-360 minutes.\u003c/p\u003e\u003cp\u003eThe minute-resolution heatmap analysis revealed complex relationships between blood pressure exposure and ICU mortality risk, extending our previous findings with higher temporal precision. High-resolution data demonstrated that mortality risk increased non-linearly with both the intensity and duration of blood pressure deviations. Similar to our previous findings, we observed that achieving target blood pressure alone may be insufficient \u003csup\u003e\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e. In other words, maintaining MAP stability within an optimal range appears crucial for favorable outcomes. Using this high-resolution analytical approach, we have further refined our understanding of the dose-response relationship between blood pressure exposure and clinical outcomes, demonstrating that both the magnitude and duration of hypotension may contribute significantly to patient risk.\u003c/p\u003e\u003cp\u003eThe HEDI was developed based on the minute-resolution heatmap. Unlike traditional approaches that use single blood pressure thresholds or fixed exposure durations (such as cumulative time with MAP\u0026thinsp;\u0026lt;\u0026thinsp;65mmHg or 5-minute episodes) \u003csup\u003e\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e\u003c/sup\u003e, HEDI integrates exposure information across multiple blood pressure thresholds and multiple duration windows, providing a more comprehensive assessment of hypotension exposure burden. Univariate analysis demonstrated significantly higher HEDI values in the mortality group compared to survivors, confirming the association between severe hypotension exposure burden and adverse outcomes. Further ROC curve analysis revealed HEDI's predictive value for ICU mortality, with a progressive improvement in predictive capability as observation time prolonged (from 24 to 72h). This time-dependency suggests that the cumulative effect of hemodynamic fluctuations may better reflect tissue perfusion status than single measurements, aligning with recent trends in critical care medicine emphasizing \"time-weighted\" physiological parameter assessment \u003csup\u003e\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e,\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e\u003c/sup\u003e. While longer observation windows (72h) provide superior predictive performance due to the capture of more hemodynamic data, this comes with practical tradeoffs. Extended monitoring periods require more complete datasets, which may not be available for all patients, and delay the availability of predictive insights for clinical decision-making. After considering the balance between predictive accuracy and clinical utility, we selected the 72-hour timepoint for subsequent machine learning analysis, as it offered optimal discrimination while still allowing for interventions within a clinically relevant timeframe for most ICU patients.\u003c/p\u003e\u003cp\u003eMultiple machine learning algorithms were evaluated, with Extra Trees identified as the optimal model for mortality prediction. Excellent discrimination ability (AUC\u0026thinsp;\u0026gt;\u0026thinsp;0.95) was demonstrated in both training and testing sets, indicating strong classification and generalization capabilities. Feature importance analysis revealed that traditional markers of physiological derangement (maximum lactate, minimum SpO\u003csub\u003e2\u003c/sub\u003e, minimum pH, and maximum SOFA score) contributed most significantly to predictions. Although not among the top predictors, HEDI outperformed conventional blood pressure monitoring metrics including minimum MAP and time-weighted-MAP. This comparative advantage carries dual significance: first, it confirms that simultaneous consideration of exposure intensity and duration provides greater value than single-threshold approaches; second, it suggests that the cumulative impact of hemodynamic fluctuations is effectively captured, potentially better approximating the physiological process by which hypotension leads to organ injury.\u003c/p\u003e\u003cp\u003eAdditionally, the predictive capability of HEDI was evaluated across multiple time windows. The 48-hour HEDI model demonstrated excellent performance in both internal and external validation, achieving an AUC of 0.696, only marginally lower than the 72-hour model. Notably, in high-risk populations receiving vasopressors, the 48-hour HEDI model maintained an AUC of 0.718, indicating that HEDI provides robust predictive information even during shorter observation periods. Importantly, HEDI consistently maintained its high position in feature importance rankings (Supplementary File, Figures S4-8), validating its robustness as a prognostic predictor. These findings support HEDI's potential as an early risk assessment tool providing valuable predictive information within 48 hours of ICU admission, while confirming that extended observation periods allow HEDI to capture more comprehensive hemodynamic burden information, further improving predictive accuracy. As a dynamic indicator that can be automatically calculated and displayed in real-time through integration with modern ICU monitoring systems, HEDI enables continuous risk reassessment. In clinical practice, this characteristic allows HEDI to serve not only for early risk stratification but also for treatment effect evaluation and dynamic prognosis monitoring.\u003c/p\u003e\u003cp\u003eSimilar to our observations with univariate HEDI analysis, the selection between 48-hour and 72-hour timepoints for machine learning model development represents a fundamental tradeoff between feasibility and accuracy. While the 72-hour model achieved marginally superior discriminative performance, the 48-hour model maintained good predictive capability with the significant advantage of earlier availability. This tradeoff mirrors clinical decision-making processes, where the balance between obtaining more comprehensive data and providing timely interventions must be carefully considered. In practical implementation, healthcare systems might select different timepoints based on their specific patient populations, clinical workflow, and resource availability. This flexibility in application timeframes enhances HEDI's clinical utility across diverse critical care settings.\u003c/p\u003e\u003cp\u003eExternal validation results further support HEDI's clinical utility. In a validation cohort comprising 13,180 patients, our model demonstrated robust predictive performance, particularly in the critically ill subgroup receiving vasopressors. These patients typically represent the most hemodynamically unstable high-risk population with complex treatment strategies and physiological responses. The model maintained performance in this subgroup suggests that the hemodynamic information captured by HEDI exhibits consistency and reliability across populations, further confirming its potential value as a hemodynamic assessment tool. The high accuracy observed in both internal and external validation indicates promising prospects for clinical application.\u003c/p\u003e\u003cp\u003eSeveral limitations should be considered. First, as an observational study, causal relationships between HEDI and mortality cannot be established. The observed associations may be influenced by unmeasured confounding factors, including variations in treatment strategies, or individualized therapeutic targets. Second, our analysis was restricted to patients with arterial catheter monitoring, who typically represent a more critically ill population, potentially limiting the applicability of our findings to the general ICU population. Third, despite external validation, our validation was confined to data from a specific healthcare system, necessitating broader multicenter validation in future studies. Finally, our study did not evaluate interactions between HEDI and other dynamic physiological parameters (e.g., heart rate variability, vascular resistance changes), which might provide additional prognostic information.\u003c/p\u003e\u003cp\u003eDespite these limitations, our study uniquely features minute-by-minute hemodynamic data, enabling comprehensive analysis of hypotensive exposures across multiple thresholds and durations. Based on these high-resolution measurements, we developed HEDI, which captures the cumulative impact of varying hypotensive intensities on patient outcomes. Through rigorous internal and external validation, our study confirms HEDI's independent value in predicting mortality risk, particularly among hemodynamically unstable high-risk patients. This novel index surpasses both single-threshold blood pressure monitoring and fixed exposure duration assessments, more accurately reflecting the physiological mechanisms through which hypotensive burden affects outcomes. HEDI demonstrates potential as both a risk stratification tool and a dynamic monitoring indicator, potentially facilitating individualized blood pressure management in critically ill patients and establishing a foundation for precision-medicine in critical care.\u003c/p\u003e"},{"header":"Conclusion","content":"\u003cp\u003eIn conclusion, HEDI, as a hemodynamic indicator that simultaneously considers both exposure intensity and duration, demonstrates significant value as a predictor variable in prognostic modeling. Our high-resolution analysis of blood pressure exposure patterns provides new insights into the relationship between hypotension exposure and mortality in critical illness. The HEDI offers a promising tool for early risk stratification, though its clinical utility ultimately needs to be validated in prospective interventional studies.\u003c/p\u003e"},{"header":"Abbreviations","content":"\u003cp\u003eANN: Artificial Neural Network\u003c/p\u003e\n\u003cp\u003eAUC: Area Under Curve\u003c/p\u003e\n\u003cp\u003eBMI: Body Mass Index\u003c/p\u003e\n\u003cp\u003eDT: Decision Tree\u003c/p\u003e\n\u003cp\u003eET: Extra Trees\u003c/p\u003e\n\u003cp\u003eGBM: Gradient Boosting Machine\u003c/p\u003e\n\u003cp\u003eHEDI: Hypotensive Exposure Duration Index\u003c/p\u003e\n\u003cp\u003eICU: Intensive Care Unit\u003c/p\u003e\n\u003cp\u003eKNN: K-Nearest Neighbors\u003c/p\u003e\n\u003cp\u003eLightGBM: Light Gradient Boosting Machine\u003c/p\u003e\n\u003cp\u003eMAP: Mean Arterial Pressure\u003c/p\u003e\n\u003cp\u003eMIMIC: Multiparameter Intelligent Monitoring in Intensive Care\u003c/p\u003e\n\u003cp\u003eNPV: Negative Predictive Value\u003c/p\u003e\n\u003cp\u003eOR: Odds Ratio\u003c/p\u003e\n\u003cp\u003ePOQI: PeriOperative Quality Initiative\u003c/p\u003e\n\u003cp\u003ePPV: Positive Predictive Value\u003c/p\u003e\n\u003cp\u003eRF: Random Forest\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eROC: Receiver Operating Characteristic\u003c/p\u003e\n\u003cp\u003eSICdb: Salzburg Intensive Care Database\u003c/p\u003e\n\u003cp\u003eSMOTE: Synthetic Minority Oversampling Technique\u003c/p\u003e\n\u003cp\u003eSpO₂: Peripheral Oxygen Saturation\u003c/p\u003e\n\u003cp\u003eSQL: Structured Query Language\u003c/p\u003e\n\u003cp\u003eSOFA: Sequential Organ Failure Assessment\u003c/p\u003e\n\u003cp\u003eSTROBE: Strengthening the Reporting of Observational Studies in Epidemiology\u003c/p\u003e\n\u003cp\u003eSVM: Support Vector Machine\u003c/p\u003e\n\u003cp\u003eXGBoost: EXtreme Gradient Boosting\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study utilized data from two databases: (1) the eICU Collaborative Research Database, which is exempt from institutional review board approval due to the retrospective design, lack of direct patient intervention, and the security schema, for which the re-identification risk was certified as meeting safe harbor standards by an independent privacy expert (Privacert, Cambridge, MA) (Health Insurance Portability and Accountability Act Certification no. 1031219-2); and (2) the SICdb, which is fully approved by the local ethical commission of the Land Salzburg, Austria (EK Nr: 1115/2021).\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data that support the findings of this study are available from the publicly available critical care database (eICU database and SICdb version 1.0.8), but restrictions apply to the availability of these data. However, available from the authors upon reasonable request and with permission of the holder of the database.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAll authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHC is supported by the Youth Top Talent Project of Fujian Provincial Foal Eagle Program.\u003c/p\u003e\n\u003cp\u003eXYD is supported by the Startup Fund for Scientific Research, Fujian Medical University (Grant number: 2021QH1290).\u003c/p\u003e\n\u003cp\u003eHPX was supported by the Natural Science Foundation of Fujian Province(2023J05250) and the Fuzhou Science and Technology Project (2024-S-003).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor\u0026apos;s contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eHC has significantly shaped the conceptual framework of this work. Throughout the processes of data extraction, analysis, and manuscript drafting, HC offered valuable suggestions and critical feedback, ensuring that the final version was thoroughly reviewed and approved for publication. XYD played a crucial role in data extraction and manuscript drafting, and contributed to the data analysis and interpretation. HPX primarily performed the data analysis and interpretation. JRZ assisted with literature review and reference collection.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eMalakar S, Singh SK, Usman K. Optimizing Blood Pressure Management in Type 2 Diabetes: A Comparative Investigation of One-Time Versus Periodic Lifestyle Modification Counseling. \u003cem\u003eCureus\u003c/em\u003e. Jun 2024;16(6):e61607. doi:10.7759/cureus.61607\u003c/li\u003e\n\u003cli\u003eMistry EA, Hart KW, Davis LT, et al. Blood Pressure Management After Endovascular Therapy for Acute Ischemic Stroke: The BEST-II Randomized Clinical Trial. \u003cem\u003eJAMA\u003c/em\u003e. Sep 5 2023;330(9):821-831. doi:10.1001/jama.2023.14330\u003c/li\u003e\n\u003cli\u003eWelte M, Saugel B, Reuter DA. [Perioperative blood pressure management : What is the optimal pressure?]. \u003cem\u003eAnaesthesist\u003c/em\u003e. Sep 2020;69(9):611-622. Perioperatives Blutdruckmanagement : Was ist der optimale Druck? doi:10.1007/s00101-020-00767-w\u003c/li\u003e\n\u003cli\u003eWeiss SL, Peters MJ, Alhazzani W, et al. Surviving Sepsis Campaign International Guidelines for the Management of Septic Shock and Sepsis-Associated Organ Dysfunction in Children. \u003cem\u003ePediatr Crit Care Med\u003c/em\u003e. Feb 2020;21(2):e52-e106. doi:10.1097/pcc.0000000000002198\u003c/li\u003e\n\u003cli\u003eEvans L, Rhodes A, Alhazzani W, et al. Surviving sepsis campaign: international guidelines for management of sepsis and septic shock 2021. \u003cem\u003eIntensive Care Med\u003c/em\u003e. 2021;47(11):1181-1247. doi:10.1007/s00134-021-06506-y\u003c/li\u003e\n\u003cli\u003eCecconi M, Evans L, Levy M, Rhodes A. Sepsis and septic shock. \u003cem\u003eLancet\u003c/em\u003e. Jul 7 2018;392(10141):75-87. doi:10.1016/s0140-6736(18)30696-2\u003c/li\u003e\n\u003cli\u003eKouz K, Weidemann F, Naebian A, et al. Continuous Finger-cuff versus Intermittent Oscillometric Arterial Pressure Monitoring and Hypotension during Induction of Anesthesia and Noncardiac Surgery: The DETECT Randomized Trial. \u003cem\u003eAnesthesiology\u003c/em\u003e. 2023;139(3):298-308. doi:10.1097/ALN.0000000000004629\u003c/li\u003e\n\u003cli\u003eZuin M, Rigatelli G, Bongarzoni A, et al. Mean arterial pressure predicts 48 h clinical deterioration in intermediate-high risk patients with acute pulmonary embolism. \u003cem\u003eEur Heart J Acute Cardiovasc Care\u003c/em\u003e. 2023;12(2):80-86. doi:10.1093/ehjacc/zuac169\u003c/li\u003e\n\u003cli\u003eMarshall JC. Choosing the Best Blood Pressure Target for Vasopressor Therapy. \u003cem\u003eJAMA\u003c/em\u003e. 2020;323(10):931-933. doi:10.1001/jama.2019.22526\u003c/li\u003e\n\u003cli\u003eChen J, Lin J, Wu D, Guo X, Li X, Shi S. Optimal Mean Arterial Pressure Within 24 Hours of Admission for Patients With Intermediate-Risk and High-Risk Pulmonary Embolism. \u003cem\u003eClin Appl Thromb Hemost\u003c/em\u003e. Jan-Dec 2020;26:1076029620933944. doi:10.1177/1076029620933944\u003c/li\u003e\n\u003cli\u003eGriffin BR, Vaughan-Sarrazin M, Shi Q, et al. Blood Pressure, Readmission, and Mortality Among Patients Hospitalized With Acute Kidney Injury. \u003cem\u003eJAMA Netw Open\u003c/em\u003e. May 1 2024;7(5):e2410824. doi:10.1001/jamanetworkopen.2024.10824\u003c/li\u003e\n\u003cli\u003eJohnson AEW, Bulgarelli L, Shen L, et al. MIMIC-IV, a freely accessible electronic health record dataset. \u003cem\u003eScientific Data\u003c/em\u003e. 2023/01/03 2023;10(1):1. doi:10.1038/s41597-022-01899-x\u003c/li\u003e\n\u003cli\u003eDing X-Y, Chen Z-Z, Chen H. Both intensity and duration of arterial blood pressure exposure are associated with mortality in critically ill patients: a retrospective database study. \u003cem\u003eBritish Journal of Anaesthesia\u003c/em\u003e. 2025/01/30/ 2025;doi:https://doi.org/10.1016/j.bja.2024.12.020\u003c/li\u003e\n\u003cli\u003eRodemund N, Wernly B, Jung C, Cozowicz C, Kok\u0026ouml;fer A. The Salzburg Intensive Care database (SICdb): an openly available critical care dataset. \u003cem\u003eIntensive Care Med\u003c/em\u003e. Jun 2023;49(6):700-702. doi:10.1007/s00134-023-07046-3\u003c/li\u003e\n\u003cli\u003eRodemund N, Wernly B, Jung C, Cozowicz C, Kok\u0026ouml;fer A. Harnessing Big Data in Critical Care: Exploring a new European Dataset. \u003cem\u003eScientific Data\u003c/em\u003e. 2024/03/28 2024;11(1):320. doi:10.1038/s41597-024-03164-9\u003c/li\u003e\n\u003cli\u003ePollard TJ, Johnson AEW, Raffa JD, Celi LA, Mark RG, Badawi O. The eICU Collaborative Research Database, a freely available multi-center database for critical care research. \u003cem\u003eScientific Data\u003c/em\u003e. 2018/09/11 2018;5(1):180178. doi:10.1038/sdata.2018.178\u003c/li\u003e\n\u003cli\u003evon Elm E, Altman DG, Egger M, Pocock SJ, G\u0026oslash;tzsche PC, Vandenbroucke JP. The Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) Statement: Guidelines for Reporting Observational Studies. \u003cem\u003eAnnals of Internal Medicine\u003c/em\u003e. 2007/10/16 2007;147(8):573-577. doi:10.7326/0003-4819-147-8-200710160-00010\u003c/li\u003e\n\u003cli\u003eDing XY, Chen ZZ, Chen H. Visualizing ICP \u0026quot;Dose\u0026quot; of neurological critical care patients. \u003cem\u003eIntensive Care Med\u003c/em\u003e. May 2024;50(5):781-783. doi:10.1007/s00134-024-07424-5\u003c/li\u003e\n\u003cli\u003eGregory A, Stapelfeldt WH, Khanna AK, et al. Intraoperative Hypotension Is Associated With Adverse Clinical Outcomes After Noncardiac Surgery. \u003cem\u003eAnesth Analg\u003c/em\u003e. Jun 1 2021;132(6):1654-1665. doi:10.1213/ane.0000000000005250\u003c/li\u003e\n\u003cli\u003eKhanna AK, Kinoshita T, Natarajan A, et al. Association of systolic, diastolic, mean, and pulse pressure with morbidity and mortality in septic ICU patients: a nationwide observational study. \u003cem\u003eAnn Intensive Care\u003c/em\u003e. Feb 20 2023;13(1):9. doi:10.1186/s13613-023-01101-4\u003c/li\u003e\n\u003cli\u003eKnight J, Hill A, Melnyk V, et al. Intraoperative Hypoxia Independently Associated With the Development of Acute Kidney Injury Following Bilateral Orthotopic Lung Transplantation. \u003cem\u003eTransplantation\u003c/em\u003e. Apr 1 2022;106(4):879-886. doi:10.1097/tp.0000000000003814\u003c/li\u003e\n\u003cli\u003eWesselink EM, Wagemakers SH, van Waes JAR, Wanderer JP, van Klei WA, Kappen TH. Associations between intraoperative hypotension, duration of surgery and postoperative myocardial injury after noncardiac surgery: a retrospective single-centre cohort study. \u003cem\u003eBr J Anaesth\u003c/em\u003e. Oct 2022;129(4):487-496. doi:10.1016/j.bja.2022.06.034\u003c/li\u003e\n\u003cli\u003eSaugel B, Fletcher N, Gan TJ, Grocott MPW, Myles PS, Sessler DI. PeriOperative Quality Initiative (POQI) international consensus statement on perioperative arterial pressure management. \u003cem\u003eBr J Anaesth\u003c/em\u003e. Aug 2024;133(2):264-276. doi:10.1016/j.bja.2024.04.046\u003c/li\u003e\n\u003cli\u003eSaab R, Wu BP, Rivas E, et al. Failure to detect ward hypoxaemia and hypotension: contributions of insufficient assessment frequency and patient arousal during nursing assessments. \u003cem\u003eBr J Anaesth\u003c/em\u003e. Nov 2021;127(5):760-768. doi:10.1016/j.bja.2021.06.014\u003c/li\u003e\n\u003cli\u003eSzrama J, Gradys A, Bartkowiak T, Woźniak A, Kusza K, Molnar Z. Intraoperative Hypotension Prediction\u0026mdash;A Proactive Perioperative Hemodynamic Management\u0026mdash;A Literature Review. \u003cem\u003eMedicina\u003c/em\u003e. 2023;59(3). doi:10.3390/medicina59030491 \u003c/li\u003e\n\u003cli\u003eDupont V, Bonnet-Lebrun AS, Boileve A, et al. Impact of early mean arterial pressure level on severe acute kidney injury occurrence after out-of-hospital cardiac arrest. \u003cem\u003eAnn Intensive Care\u003c/em\u003e. Jul 18 2022;12(1):69. doi:10.1186/s13613-022-01045-1\u003c/li\u003e\n\u003cli\u003eLi Z, Zhao X, Wang D, et al. Reliability and accuracy analysis of time-weighted average exposure to heavy metals based on personal exposure. \u003cem\u003eSci Total Environ\u003c/em\u003e. Aug 10 2022;833:155209. doi:10.1016/j.scitotenv.2022.155209\u003c/li\u003e\n\u003cli\u003eMaheshwari K, Shimada T, Yang D, et al. Hypotension Prediction Index for Prevention of Hypotension during Moderate- to High-risk Noncardiac Surgery. \u003cem\u003eAnesthesiology\u003c/em\u003e. Dec 1 2020;133(6):1214-1222. doi:10.1097/aln.0000000000003557\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":false,"email":"","identity":"journal-of-intensive-care","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Journal of Intensive Care","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"VoR Journals","inReviewEnabled":false,"inReviewRevisionsEnabled":false},"keywords":"High-Temporal Resolution, Blood Pressure Exposure, Hypotension Exposure, Hypotensive Exposure Duration Index, Machine Learning","lastPublishedDoi":"10.21203/rs.3.rs-7288238/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7288238/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e\u003cp\u003eBlood pressure management is crucial in critical care, but relationships between pressure patterns and outcomes remain incompletely understood. We analyzed minute-by-minute blood pressure data to develop and validate a novel index quantifying hypotensive exposure burden.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e\u003cp\u003eIn this retrospective study using the Salzburg Intensive Care Database, 11,059 ICU admissions with continuous invasive arterial monitoring were analyzed. Heatmaps were constructed from high-resolution hemodynamic data to visualize relationships between blood pressure thresholds (50\u0026ndash;120 mmHg), exposure durations (5 minutes to 6 hours), and mortality. The Hypotensive Exposure Duration Index (HEDI) was developed to quantify cumulative hypotensive burden by integrating exposure across multiple MAP thresholds. HEDI's prognostic value was evaluated through ROC analysis and nine machine learning algorithms. External validation using the eICU database assessed HEDI's consistency across different populations.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e\u003cp\u003eNon-survivors showed significantly higher HEDI compared to survivors (0.75 [-0.25-2.37] vs -0.16 [-0.65-0.53], p\u0026thinsp;\u0026lt;\u0026thinsp;0.001). HEDI demonstrated increasing predictive capability, with AUC values rising from 0.622 at 24h to 0.686 at 72h post-admission. The Extra Trees classifier achieved exceptional performance (test AUC: 0.996), with HEDI ranking among the top predictive features. Both internal cross-validation and external validation confirmed the model's robustness (accuracy 0.871), demonstrating HEDI's consistent prognostic value across different patient populations regardless of vasopressor use.\u003c/p\u003e\u003ch2\u003eConclusions\u003c/h2\u003e\u003cp\u003eHEDI effectively quantifies cumulative hypotensive burden in critically ill patients, demonstrating significant predictive ability for ICU mortality validated across diverse populations.\u003c/p\u003e","manuscriptTitle":"Development and Validation of the Hypotensive Exposure Duration Index for Mortality Risk Prediction in Critically III Patients","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-08-18 09:30:50","doi":"10.21203/rs.3.rs-7288238/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2025-09-04T03:31:36+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-09-03T15:01:07+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-08-22T02:58:59+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"337294928506411624623882995135973002696","date":"2025-08-18T10:07:26+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"130520705779330380169018636115081136720","date":"2025-08-12T01:22:25+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-08-11T05:24:32+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-08-05T09:59:02+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-08-05T09:58:27+00:00","index":"","fulltext":""},{"type":"submitted","content":"Journal of Intensive Care","date":"2025-08-04T07:32:59+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":false,"email":"","identity":"journal-of-intensive-care","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"","title":"Journal of Intensive Care","twitterHandle":"","acdcEnabled":false,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"VoR Journals","inReviewEnabled":false,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"165bca21-9048-4798-a281-45bbcf809d38","owner":[],"postedDate":"August 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[],"tags":[],"updatedAt":"2025-11-03T11:08:50+00:00","versionOfRecord":[],"versionCreatedAt":"2025-08-18 09:30:50","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7288238","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7288238","identity":"rs-7288238","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}