ABSTRACT
Background & Objective:
Acute respiratory distress syndrome (ARDS) frequently complicates sepsis, leading to significant morbidity and mortality. This study aimed to identify factors that predict short-term (30-day) and long-term (one-year) mortality in ARDS patients and to develop robust predictive models.
Methodology:
This retrospective study, conducted from August 2024 to October 2024 in XinHua Hospital Affiliated with Shanghai Jiao Tong University, used data from the MIMIC database—specifically MIMIC IV 2.2—to identify sepsis patients who were diagnosed with ARDS within 24 hours of ICU admission. Univariate logistic regression was used to explore associations between respiratory parameters and the Glasgow coma scale (GCS) score (low group ≤ 12). Mortality at 30 days and one-year post-ICU admission was used as outcome measures. The dataset was balanced via synthetic minority over-sampling technique and split into training (70%) and validation (30%) sets. Variable selection was performed via the best subset, least absolute shrinkage and selection operator (LASSO), random forest, and boruta methods. Predictive models were developed and validated via calibration and decision curve analyses.
Results:
The cohort included 3,158 patients (58% female). Significant differences in PaCO2 levels were detected between 30-day survivors and non-survivors (p>0.05), but not detected at one year (p>0.05). More patients with low GCS scores died within one year (20.9%) than survivors (17.1%; p=0.01), but no such association was found for 30-day mortality (p>0.05). The predictive models perform well for predicating short-term and long-term mortality and had AUCs of 0.820 and 0.790, respectively.
Conclusions:
GCS scores were significantly associated with one-year mortality but not with 30-day mortality or with respiratory-related parameters. The developed predictive models demonstrated good performance. These findings aid in the treatment of ARDS patients with sepsis.
KEYWORDS: Glasgow Coma Scale, Intensive Care Units, Mortality, Respiratory Distress Syndrome, Sepsis
INTRODUCTION
Sepsis is a life-threatening condition caused by a dysregulated host response to infection.1 Acute respiratory distress syndrome (ARDS) and acute respiratory failure (ARF) are common complications of sepsis and can lead to severe organ damage and even death.2 Additionally, the incidence rate of sepsis-associated encephalopathy (SAE) can reach 70%.3 A growing body of research shows that lung injury can promote brain damage through complex network processes that involve mechanical and chemical factors, as well as various inflammatory mediators.4 The impact of these lung and brain parameters on the short-term and long-term mortality of patients with sepsis is still unclear.
This study aimed to use the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database of patients with sepsis complicated by ARDS to identify risk factors for short-term and long-term mortality and to establish a predictive model. Our results may help in estimating the risk of short- and long-term mortality in patients with sepsis accompanied by pulmonary conditions, detecting high-risk patients in a timely manner, and further improving the prognosis of this vulnerable group.
METHODOLOGY
In the present study, conducted from August 2024 to October 2024 in XinHua Hospital Affiliated with Shanghai Jiao Tong University, the authors followed all applicable regulations, and all study data were sourced from the MIMIC database. First, the Institutional Review Boards (IRBs) of Beth Israel Deaconess Medical Center and the Massachusetts Institute of Technology granted the MIMIC-IV database a waiver of ethics approval and informed consent (Supplementary Fig.1). Second, the author (Record ID 62454298) obtained the necessary data-use permissions and completed Collaborative Institutional Training Initiative (CITI) certification prior to accessing the database.
Supplementary Fig.1.
CITI completion certificate.
Inclusion criteria:
The study population in MIMIC_IV comprised 9,064 adult patients diagnosed with sepsis (based on the diagnostic guidelines of sepsis-3) and with ARDS (PaO2/FiO2 ratio less than 300 mmHg), received by the ICU.
Exclusion criteria:
Pulmonary neoplasms.
Neurological disorders.
More than 80% missing data.
A final sample of 3,158 patients was included in the study (Supplementary Fig.2). Patients were stratified into 30-day survivors and non-survivors, and, in parallel, into one-year survivors and non-survivors, according to vital status at 30 days and one year, respectively. Only data from the initial ICU admission were considered for patients admitted multiple times.
Supplementary Fig.2.
Flowchart illustrating the screening criteria for patient inclusion in this study.
Data collection and analysis:
Demographics, medical history, vital signs, laboratory tests, and scores within the first 24 hours post-ICU admission were recorded, and cases with more than 80% missing data were removed. Univariate logistic regression was used to examine the correlation between GCS scores and ventilator parameters, then assessed the associations of GCS scores with 30 days and one-year mortality, and subsequently constructed predictive models for mortality at both time points.
Statistical analysis:
Continuous variables were described using mean (standard deviation [SD]) or median (interquartile range [IQR]). Group comparisons were performed using the Mann-Whitney U test or Student’s t-test. Categorical variables were presented as frequencies and percentages (%), and group comparisons were conducted using Fisher’s exact test or the chi-square test. Univariate logistic regression was used to explore the relationships between variables.
Predictive model development and Validation:
Variables with more than 20% missing data were excluded from the MIMIC database analysis. To improve model accuracy, data underwent SMOTE balancing to equalize the proportions of deceased and surviving patients. The cohort was then randomly divided into a training set (70%) and a validation set (30%). All modeling procedures were performed separately for 30-day and one-year mortality. The selection methods included best subset, least absolute shrinkage and selection operator (LASSO) regression, random forest (RF), and Boruta. The selected variables were included in a multivariable logistic regression model. The best models were chosen on the basis of the BIC, AIC, and ROC analyses, with performance illustrated through nomograms, calibration curves, Hosmer–Lemeshow (HL) testing, and decision curve analysis (DCA). Model robustness was assessed via unbalanced datasets and across patient groups with initial 24-hour SOFA scores ≥2. Statistical significance was set at p < 0.05. Data analysis was conducted via R software (version 4.4.0).
RESULTS
Baseline and association analysis:
Baseline characteristics according to 30-day mortality are presented in Table-I. Corresponding data stratified by one-year survival status are shown in Table-II. No relationships were detected between the GCS score and respiratory parameters (p > 0.05) (Supplementary Table-I). The percentage of patients with GCS scores ≤12 was greater in the one-year mortality group (20.9%) than in the long-term survival group (17.1%, p = 0.01), but no difference was found in the 30-day mortality group (p > 0.05).
Table-I.
Baseline characteristics of patients stratified by 30-day mortality status.
| Characteristic | 30-day mortality status | p-value | |||
|---|---|---|---|---|---|
| Overall (N=3158) | Survived (N=2423, 76.7%) | Dead (N=735, 23.3%) | |||
| Gender No.(%) | |||||
| Male | 1326 (42.0) | 983 (40.6) | 343 (46.7) | 0.004 | |
| Female | 1832 (58.0) | 1440 (59.4) | 392 (53.3) | ||
| Length of hospital M[IQR] | 9.23 [5.39, 15.72] | 10.30 [6.51, 17.11] | 4.97 [1.71, 10.64] | <0.001 | |
| Age M[IQR] | 68.10 [57.03, 78.57] | 67.06 [55.80, 77.08] | 72.07 [61.00, 82.41] | <0.001 | |
| Length of ICU M[IQR] | 3.98 [2.02, 7.90] | 4.05 [2.16, 7.89] | 3.72 [1.43, 7.93] | <0.001 | |
| Weight M[IQR] | 80.00 [67.90, 95.53] | 81.00 [68.50, 96.03] | 77.30 [65.00, 93.93] | <0.001 | |
| Heart rate M[IQR] | 72.00 [62.00, 83.00] | 71.00 [62.00, 81.00] | 76.00 [62.00, 89.00] | <0.001 | |
| Respiration rate M[IQR] | 12.00 [10.00, 15.00] | 12.00 [10.00, 14.50] | 14.00 [10.00, 17.00] | <0.001 | |
| MAP M[IQR] | 56.67 [50.67, 62.00] | 57.50 [52.33, 63.00] | 52.33 [44.33, 59.00] | <0.001 | |
| SpO2 M[IQR] | 92.00 [89.00, 94.00] | 92.00 [90.00, 95.00] | 90.00 [82.00, 93.00] | <0.001 | |
| PH M[IQR] | 7.30 [7.22, 7.36] | 7.31 [7.24, 7.36] | 7.24 [7.13, 7.34] | <0.001 | |
| PaO2 M[IQR] | 81.00 [68.00, 100.00] | 83.00 [70.00, 103.00] | 73.00 [61.00, 90.00] | <0.001 | |
| PaCO2 M[IQR] | 35.00 [31.00, 40.00] | 35.00 [31.00, 40.00] | 34.00 [29.00, 40.00] | 0.007 | |
| FiO2 M[IQR] | 50.00 [40.00, 50.00] | 50.00 [40.00, 50.00] | 50.00 [40.00, 70.00] | <0.001 | |
| Ratio of PaO2:FiO2 M[IQR] | 156.00 [101.49, 212.00] | 166.00 [111.00, 220.00] | 120.00 [78.17, 180.00] | <0.001 | |
| BE M[IQR] | -4.00 [-8.00, 0.00] | -3.00 [-6.00, 0.00] | -8.00 [-13.00, -1.00] | <0.001 | |
| Lactate M[IQR] | 1.40 [1.00, 2.10] | 1.30 [1.00, 1.70] | 2.10 [1.40, 4.05] | <0.001 | |
| White blood cell M[IQR] | 10.40 [7.20, 14.40] | 10.20 [7.30, 13.70] | 11.70 [6.60, 17.10] | <0.001 | |
| Hemoglobin M[IQR] | 9.60 [8.30, 11.10] | 9.60 [8.30, 11.20] | 9.40 [8.10, 11.00] | 0.014 | |
| Platelet M[IQR] | 153.00 [103.00, 219.00] | 156.00 [109.25, 218.00] | 140.00 [74.00, 224.00] | <0.001 | |
| Red blood cell M[IQR] | 3.18 [2.74, 3.72] | 3.19 [2.77, 3.74] | 3.13 [2.66, 3.64] | 0.001 | |
| Creatinine M[IQR] | 0.90 [0.70, 1.40] | 0.90 [0.70, 1.20] | 1.40 [0.90, 2.30] | <0.001 | |
| BUN M[IQR] | 19.00 [13.00, 31.00] | 17.00 [12.00, 26.00] | 31.00 [20.00, 46.00] | <0.001 | |
| SOFA M[IQR] | 2.00 [1.00, 4.00] | 2.00 [0.00, 4.00] | 3.00 [1.00, 5.50] | <0.001 | |
| Intubation at first day of ICU No.(%) | |||||
| No | 250 (7.9) | 194 (8.0) | 56 (7.6) | 0.815 | |
| Yes | 2908 (92.1) | 2229 (92.0) | 679 (92.4) | ||
| GCS No.(%) | |||||
| <=12 | 581 (18.4) | 434 (17.9) | 147 (20.1) | 0.192 | |
| 13-15 | 2574 (81.6) | 1988 (82.1) | 586 (79.9) | ||
ICU = intensive care unit, MAP = mean arterial pressure, pH= pondus hydrogenii, BE = base excess, BUN = blood urea nitrogen, SOFA = sequential organ failure assessment, GCS = Glasgow Coma Scale.
Table-II.
Baseline characteristics of patients stratified by one-year mortality status.
| Characteristic | one-year mortality status | p-value | ||
|---|---|---|---|---|
| Overall (N=3158) | Survived (N=2068, 65.5%) | Dead (N=1090, 34.5%) | ||
| Gender No.(%) | ||||
| Male | 1326 (42.0) | 823 (39.8) | 503 (46.1) | 0.001 |
| Female | 1832 (58.0) | 1245 (60.2) | 587 (53.9) | |
| Length of hospital M[IQR] | 9.23 [5.39, 15.72] | 9.88 [6.25, 15.89] | 7.97 [2.99, 14.97] | <0.001 |
| Age M[IQR] | 68.10 [57.03, 78.57] | 65.79 [54.63, 75.76] | 73.78 [61.41, 82.38] | <0.001 |
| Length of ICU M[IQR] | 3.98 [2.02, 7.90] | 3.89 [2.10, 7.31] | 4.25 [1.85, 9.00] | 0.196 |
| Weight M[IQR] | 80.00 [67.90, 95.53] | 82.00 [70.00, 97.15] | 76.10 [63.80, 92.40] | <0.001 |
| Heart rate M[IQR] | 72.00 [62.00, 83.00] | 71.00 [62.00, 81.00] | 74.00 [62.00, 86.00] | <0.001 |
| Respiration rate M[IQR] | 12.00 [10.00, 15.00] | 12.00 [10.00, 14.00] | 13.00 [10.00, 16.00] | <0.001 |
| MAP M[IQR] | 56.67 [50.67, 62.00] | 57.83 [52.67, 63.33] | 54.00 [46.42, 59.33] | <0.001 |
| SpO2 M[IQR] | 92.00 [89.00, 94.00] | 92.00 [90.00, 95.00] | 91.00 [85.75, 94.00] | <0.001 |
| PH M[IQR] | 7.30 [7.22, 7.36] | 7.31 [7.24, 7.36] | 7.26 [7.16, 7.35] | <0.001 |
| PaO2 M[IQR] | 81.00 [68.00, 100.00] | 84.00 [70.00, 103.00] | 75.00 [63.00, 92.00] | <0.001 |
| PaCO2 M[IQR] | 35.00 [31.00, 40.00] | 35.00 [31.00, 39.00] | 35.00 [30.00, 41.00] | 0.317 |
| FiO2 M[IQR] | 50.00 [40.00, 50.00] | 50.00 [40.00, 50.00] | 50.00 [40.00, 60.00] | <0.001 |
| Ratio of PaO2:FiO2 M[IQR] | 156.00 [101.49, 212.00] | 166.00 [112.86, 218.00] | 134.64 [85.71, 196.92] | <0.001 |
| BE M[IQR] | -4.00 [-8.00, 0.00] | -3.00 [-6.00, 0.00] | -6.00 [-12.00, -0.25] | <0.001 |
| Lactate M[IQR] | 1.40 [1.00, 2.10] | 1.30 [1.00, 1.70] | 1.80 [1.20, 3.30] | <0.001 |
| White blood cell M[IQR] | 10.40 [7.20, 14.40] | 10.20 [7.30, 13.60] | 11.20 [6.82, 16.20] | 0.001 |
| Hemoglobin M[IQR] | 9.60 [8.30, 11.10] | 9.60 [8.30, 11.20] | 9.40 [8.10, 11.00] | 0.005 |
| Platelet M[IQR] | 153.00 [103.00, 219.00] | 156.00 [111.00, 217.00] | 147.00 [83.00, 227.00] | <0.001 |
| Red blood cell M[IQR] | 3.18 [2.74, 3.72] | 3.20 [2.77, 3.74] | 3.14 [2.67, 3.69] | 0.001 |
| Creatinine M[IQR] | 0.90 [0.70, 1.40] | 0.90 [0.70, 1.20] | 1.30 [0.80, 2.10] | <0.001 |
| BUN M[IQR] | 19.00 [13.00, 31.00] | 16.00 [12.00, 24.00] | 29.00 [18.00, 43.00] | <0.001 |
| SOFA M[IQR] | 2.00 [1.00, 4.00] | 2.00 [0.00, 4.00] | 3.00 [1.00, 5.00] | <0.001 |
| Intubation at first day of ICU No.(%) | ||||
| No | 250 (7.9) | 152 (7.4) | 98 (9.0) | 0.111 |
| Yes | 2908 (92.1) | 1916 (92.6) | 992 (91.0) | |
| GCS No.(%) | ||||
| <=12 | 581 (18.4) | 354 (17.1) | 227 (20.9) | 0.010 |
| 13-15 | 2574 (81.6) | 1713 (82.9) | 861 (79.1) | |
ICU = intensive care unit, MAP = mean arterial pressure, pH= pondus hydrogenii, BE = base excess, BUN =blood urea nitrogen, SOFA = sequential organ failure assessment, GCS = Glasgow Coma Scale
Supplementary Table-I.
The relationship between Glasgow Coma Scores and respiratory parameters.
| Variables | Overall(N=3158) | GCS<=12(N=581, 18.5%) | 12<GCS<=15(N=2574,81.5%) | p-value |
|---|---|---|---|---|
| Respiratory rate M(IQR) | 12.00 [10.00, 15.00] | 12.00 [10.00, 15.00] | 12.00 [10.00, 15.00] | 0.224 |
| SpO2 M(IQR) | 92.00 [89.00, 94.00] | 92.00 [88.00, 94.00] | 92.00 [89.00, 94.00] | 0.184 |
| pH Mean (SD) | 7.28 (0.12) | 7.27 (0.12) | 7.28 (0.12) | 0.094 |
| PaO2 M(IQR) | 81.00 [68.00, 100.00] | 79.00 [66.00, 97.00] | 81.00 [68.00, 101.00] | 0.069 |
| PaCO2 M(IQR) | 35.00 [31.00, 40.00] | 35.00 [31.00, 40.00] | 35.00 [31.00, 40.00] | 0.497 |
| FiO2 M(IQR) | 50.00 [40.00, 50.00] | 50.00 [40.00, 50.00] | 50.00 [40.00, 50.00] | 0.325 |
| PaO2/FiO2 ratio M(IQR) | 156.00 [101.49, 212.00] | 151.67 [100.00, 206.25] | 158.00 [102.00, 214.00] | 0.177 |
| BE Mean (SD) | -4.63 (6.43) | -4.63 (6.33) | -4.63 (6.45) | 0.992 |
| Lactate Mean (SD) | 1.96 (1.92) | 2.00 (1.77) | 1.95 (1.93) | 0.607 |
| PEEP M(IQR) | 5.00 [5.00, 5.00] | 5.00 [5.00, 5.00] | 5.00 [5.00, 5.00] | 0.713 |
| Ventilator respiratory rate total M(IQR) | 16.00 [14.00, 19.00] | 16.00 [14.00, 19.00] | 16.00 [14.00, 18.00] | 0.855 |
| Ventilator respiratory rate spontaneous M(IQR) | 0.00 [0.00, 0.00] | 0.00 [0.00, 0.00] | 0.00 [0.00, 0.00] | 0.408 |
| Ventilator minute volume M(IQR) | 7.39 [6.20, 8.69] | 7.50 [6.20, 8.78] | 7.30 [6.20, 8.60] | 0.295 |
| Ventilator tidal volume observed M(IQR) | 413.00 [340.00, 480.00] | 420.50 [330.00, 490.00] | 411.00 [340.84, 479.00] | 0.443 |
| Ventilator FiO2 M(IQR) | 40.00 [40.00, 50.00] | 40.00 [40.00, 50.00] | 40.00 [40.00, 50.00] | 0.210 |
pH, pondus hydrogenii; BE, base excess; PEEP, positive end-expiratory pressure. The results indicate there was no relationships between Glasgow Coma Scores and respiratory parameters including blood gas analysis, respiratory parameters and ventilator-related parameters.
Model construction:
We evaluated multiple models for predicting 30-day and one-year mortality using different variable-selection approaches. Fig.1 presents the ROC curves and corresponding AUC values for the training (Fig.1 a & b) and test sets (Fig.1 c & d); panels a and c depict 30 days mortality, whereas panels b and d depict one-year mortality. The models were assessed on the basis of the AIC, BIC, and AUC (results represented in Table-III).
Fig.1.
ROC curves and AUC values for different models in predicting 30-day (Panels a and c) and one-year (Panels b and d) mortality. a) and b): ROC curves and AUC values for different models in the training set. c) and d): ROC curves and AUC values for different models in the test set.
Table-III.
The corresponding evaluation indices (AIC, BIC, etc.) are constructed for building models via different filtering variable methods to reflect the goodness of fit and complexity.
| Target | Method | Index | |||||
|---|---|---|---|---|---|---|---|
| AIC1 | BIC1 | AUC1 | Cut-off | Sensitivity | Specificity | ||
| 30 days mortality | Best subset | 2341.53 | 2404.24* | 0.816 | 0.477 | 0.733 | 0.744 |
| LASSO2 | 2335.9 | 2404.32 | 0.818 | 0.494 | 0.758* | 0.737 | |
| Boruta | 2328.94* | 2408.76 | 0.820* | 0.478 | 0.742 | 0.749 | |
| RF2 | 2422.40 | 2462.31 | 0.797 | 0.463 | 0.703 | 0.759* | |
| One-year mortality | Best subset | 2480.33 | 2543.05 | 0.787 | 0.471 | 0.706 | 0.733* |
| LASSO2 | 2469.37* | 2549.19 | 0.790* | 0.479 | 0.716 | 0.729 | |
| Boruta | 2474.31 | 2542.73* | 0.788 | 0.481 | 0.722* | 0.723 | |
| RF2 | 2535.41 | 2586.73 | 0.772 | 0.473 | 0.684 | 0.715 | |
AIC: Akaike information criterion; BIC: Bayesian information criterion; AUC: area under the curve.
LASSO: least absolute shrinkage and selection operator; RF: random forest.
Denotes the best evaluation index of each model.
We selected the Boruta model for 30-day mortality, which achieved an AUC of 0.820, a sensitivity of 0.742, and a specificity of 0.749. The LASSO model for one-year mortality had an AUC of 0.790, a sensitivity of 0.716, and a specificity of 0.729. Fig.2 illustrates the variables selected for 30-day mortality prediction (Fig.2a) and one-year mortality prediction (Fig.2b). Notable odds ratios (ORs) for 30-day mortality included lactate (OR: 2.63) and Cre (OR: 2.66). For one-year mortality, significant ORs included pH (OR: 0.17), lactate (OR: 1.69), Cre (OR: 3.35), and GCS (OR: 0.74).
Fig.2.
Nomograms for Predicting 30-day and one-year Mortality. Panel a) depicts the nomogram for predicting 30-day mortality, while Panel b) illustrates the nomogram for predicting one-year mortality.
Model evaluation and validation:
The calibration curves showed good fit (Fig.3, a & b). The Hosmer–Lemeshow test yielded p-values of 0.517 and 0.385 for the 30-day and one-year models, respectively. Decision curve analysis (DCA) plots (Fig.3, c & d) indicated substantial net benefit across a relevant range of threshold probabilities, highlighting the predictive capabilities of the models. To validate the predictive performance of our models, additional analyses were conducted on different subpopulation datasets and unbalanced data (Fig.4a). Patients were stratified by their initial SOFA score upon ICU admission (SOFA score <2 vs. ≥2), and models were validated within these subgroups. Results are presented in Supplementary Table-II, Fig.4b (SOFA<2), and Fig.4c (SOFA≥2). These analyses highlight the robustness and generalizability of our models.
Fig.3.
Calibration curves and DCA curves for predicting 30-day and one-year mortality. Panels a) and c) present the results for predicting 30-day mortality, whereas Panels b) and d) illustrate the results for predicting one-year mortality. The numbers on the curves indicate the corresponding risk threshold values.
Fig.4.
Model Performance in Predicting 30-day and One-year Mortality. These figures represent the ROC curves, AUC results, calibration curves, and DCA curves for predicting 30-day and one-year mortality across different datasets and patient subgroups. Specifically, a) illustrates model performance in an unbalanced dataset without SMOTE, b) focuses on patients with SOFA scores less than 2, and c) examines patients with SOFA scores of 2 or higher.
Supplementary Table-II.
Model performance evaluation on imbalanced data across different populations.
| Data | Target | Index | ||||
|---|---|---|---|---|---|---|
| AUC1 | Cut-off | Sensitivity | Specificity | Hosmer-Lemeshow test | ||
| No-SMOTE2 | 30-day | 0.804 | 0.244 | 0.766 | 0.699 | 0.357 |
| One-year | 0.778 | 0.370 | 0.771 | 0.649 | 0.917 | |
| SOFA2< 2 | 30-day | 0.793 | 0.163 | 0.640 | 0.812 | 0.137 |
| One-year | 0.786 | 0.260 | 0.645 | 0.783 | 0.862 | |
| SOFA2 ≥ 2 | 30-day | 0.813 | 0.250 | 0.756 | 0.737 | 0.453 |
| One-year | 0.786 | 0.260 | 0.645 | 0.783 | 0.699 | |
AUC, area under curve.
SMOTE, synthetic minority over-sampling technique; SOFA, sequential organ failure assessment.
DISCUSSION
Sepsis is a severe multisystem disorder with high mortality. Our previous work elucidated the link between sepsis-associated AKI and mortality5, and the present study further explores factors related to SAE and establishes a relevant model. Our study revealed that the 30-day and one-year mortality rates for sepsis patients with ARDS were 23.27% and 34.52%, respectively. Our study also showed that 30 days mortality was associated with a higher respiratory rate and lower oxygen saturation, pH, PaO2, PaCO2, oxygenation index. The pulmonary-brain axis has received increasing attention in recent years.6 In 64% of all sepsis cases, the infection leading to sepsis originates in the lungs.1 Inflammatory factors from chronic inflammation in asthma can impact brain regions, leading to neurological symptoms.7 Animal experiments also confirmed that lung infections can damage brain nerves.8 For now, the only reliable clinical measurement of brain function was the GCS score.9-11 Interestingly, we found no correlations between the GCS score and respiration parameters, possibly because of indirect pathways such as those involving inflammatory factors.12
More intriguingly, our findings demonstrated that several factors differed significantly between patients who died within one year of ICU admission and those who died within 30 days. Specifically, PaCO2 differed markedly between the 30-day non-survivors and longer-term survivors. Nevertheless, PaCO2 was not retained in the final predictive model for one-year mortality, suggesting that its prognostic relevance may be confined to short-term outcomes. In contrast, the severity of neurologic injury emerged as a pivotal determinant of long-term survival. Lu et al. found that mechanical ventilation parameters influence patients’ prognosis, which is not supported by the findings of our study.13 Notably, their research was a single-center study with a small cohort that employed regression analysis. These difference might account for variations in medical standards across different regions. This discrepancy between the two studies warrants further verification through future randomized controlled trials.
We also found that higher GCS scores were significantly associated with lower long-term (one-year) mortality. As reported by Reynolds JC et al. and Malhotra AK et al., long-term mortality in patients with ARDS is associated with a lower GCS; a similar finding was observed in our cohort of septic patients with ARDS.14,15 COVID-19, flu, or bacterial pneumonia were shown to have a long-term impact on the cognitive function and concentration of patients and the impairments in cognitive function and concentration can lead to errors and delays in judgments related to hazard avoidance.16 Meanwhile, these defects affect multiple aspects of patients’ work and daily lives. Such consequences may cause to the increased long-term mortality rate among patients with ARDS.
Prior studies have demonstrated that acute respiratory distress syndrome (ARDS) occurs frequently in patients with severe acute brain injury (SABI) and portends worse neurological outcomes.17 However, three major conflicts exist between lung-protective ventilation recommended for ARDS and the cerebral-protective strategies required for SABI.1,18 These conflicts may further increase mortality in sepsis. In the current cohort of mechanically ventilated septic patients we observed that respiratory parameters were not correlated with the Glasgow Coma Scale (GCS). This finding indirectly suggests that the relationship between GCS and mortality is independent of ventilatory support, a result that appears to diverge from earlier reports.19 Consequently, reliance on GCS alone as a surrogate for cerebral function may be insufficient.
To predict 30 days and one-year mortality, we developed two robust models for clinical use (The detailed calculation method can be found in Supplementary Fig.3 and 4). Jiyeon Roh and colleagues previously constructed a nomogram to estimate long-term mortality in sepsis (including septic shock) using data from 446 patients (60.8% male; median age, 71 years) treated for ≥ 3 years at a university-affiliated tertiary care hospital.20 Multivariate analysis identified age ≥ 65 years, BMI < 18.5 kg/m², hematologic malignancy, and mechanical ventilation as predictors for both the 180- and 365-day models. The corresponding AUCs were 0.713 (95% CI, 0.668–0.758) and 0.697 (95% CI, 0.651–0.743), which are lower than those achieved by our current models.
Supplementary Fig.3.
Nomogram calculation for predicting 30 days mortality.
Supplementary Fig.4.
Nomogram calculation for predicting 1-year mortality.
Our study provides short-term and long-term prediction models as well as practical predictive tools for patients with ARDS caused by sepsis. Notably, the factors incorporated into the modeling are all actual predictors of ARDS patients’ prognosis, which offers important basis for future RCTs and mechanism-related studies. Meanwhile, with more studies suggesting that the brain is an easily overlooked yet critical and overloaded organ in sepsis, our study reveals that a low GCS score is associated with poor long-term prognosis in patients with sepsis complicated by ARDS—this provides important insights for future research.
Limitations
First, we excluded more than 80% of the cases with missing data during the data processing phase, which may have introduced potential bias. Second, we choose only the GCS score as a parameter for brain function, but there might many other potential parameters can reflect brain function. Third, despite the successful internal validation of our models, the lack of external data hinders additional validation.
CONCLUSIONS
This study revealed that in sepsis patients with ARDS, higher GCS scores were significantly associated with lower mortality one year after admission to the ICU. Predictive models for both short-term and long-term mortality in patients were constructed and have demonstrated good predictive value.
Authors’ contributions:
ZR: Designed the study, participated in data collection, guided the research, and reviewed the manuscript. This author is also responsible and accountable for the accuracy and integrity of the work.
BR: Literature search, data analysis and finalized the manuscript and
MC: Literature search, Drafted the manuscript.
YL: Assisted in data collection and analysis and manuscript revision.
ZL: Participated in literature search, writing and revising the manuscript.
All authors have contributed equally to the manuscript and read and approved the final version.
Footnotes
Funding: This research was funded by a grant from the National Natural Science Foundation of China (No. 82102286) and Project of the 2024 “Technology Innovation Action Plan” of Shanghai Municipality (24SF1900703).
Conflicts of interest: None.
REFERENCES
- 1.Cecconi M, Evans L, Levy M, Rhodes A. Sepsis and septic shock. Lancet. 2018;392(10141):75–87. doi: 10.1016/S0140-6736(18)30696-2. doi:10.1016/S0140-6736(18)30696-2. [DOI] [PubMed] [Google Scholar]
- 2.Cochi SE, Kempker JA, Annangi S, Kramer MR, Martin GS. Mortality Trends of Acute Respiratory Distress Syndrome in the United States from 1999 to 2013. Ann Am Thorac Soc. 2016;13(10):1742–1751. doi: 10.1513/AnnalsATS.201512-841OC. doi:10.1513/AnnalsATS.201512-841OC. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Chaudhry N, Duggal AK. Sepsis Associated Encephalopathy. Adv Med. 2014. 2014:762320. doi: 10.1155/2014/762320. doi:10.1155/2014/762320. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Capuron L, Miller AH. Immune system to brain signaling:neuropsychopharmacological implications. Pharmacol Ther. 2011;130(2):226–238. doi: 10.1016/j.pharmthera.2011.01.014. doi:10.1016/j.pharmthera.2011.01.014. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Liu Y, Ren B, Cheng M, Du J, Ren R. Re-exploring the association between the central venous pressure and the risk of sepsis-associated acute kidney injury according to the latest definition:Analysis of the MIMIC-IV database. Pak J Med Sci. 2025;41(5):1393–1401. doi: 10.12669/pjms.41.5.12047. doi:10.12669/pjms.41.5.12047. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 6.Goldstein DS. Post-COVID dysautonomias:what we know and (mainly) what we don't know. Nat Rev Neurol. 2024;20(2):99–113. doi: 10.1038/s41582-023-00917-9. doi:10.1038/s41582-023-00917-9. [DOI] [PubMed] [Google Scholar]
- 7.Vafaee F, Shirzad S, Shamsi F, Boskabady MH. Neuroscience and treatment of asthma, new therapeutic strategies and future aspects. Life Sci. 2022;292:120175. doi: 10.1016/j.lfs.2021.120175. doi:10.1016/j.lfs.2021.120175. [DOI] [PubMed] [Google Scholar]
- 8.Yu X, Xiao H, Liu Y, Dong Z, Meng X, Wang F. The Lung-Brain Axis in Chronic Obstructive Pulmonary Disease-Associated Neurocognitive Dysfunction:Mechanistic Insights and Potential Therapeutic Options. Int J Biol Sci. 2025;21(8):3461–3477. doi: 10.7150/ijbs.109261. doi:10.7150/ijbs.109261. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Engelen T, Solcà M, Tallon-Baudry C. Interoceptive rhythms in the brain. Nat Neurosci. 2023;26(10):1670–1684. doi: 10.1038/s41593-023-01425-1. doi:10.1038/s41593-023-01425-1. [DOI] [PubMed] [Google Scholar]
- 10.Zhao S, Umpierre AD, Wu LJ. Tuning neural circuits and behaviors by microglia in the adult brain. Trends Neurosci. 2024;47(3):181–194. doi: 10.1016/j.tins.2023.12.003. doi:10.1016/j.tins.2023.12.003. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Huang YZ, Ma JX, Bian YJ, Bai QR, Gao YH, Di SK, et al. TRPV1 analgesics disturb core body temperature via a biased allosteric mechanism involving conformations distinct from that for nociception. Neuron. 2024;112(11):1815–1831. e4. doi: 10.1016/j.neuron.2024.02.016. doi:10.1016/j.neuron.2024.02.016. [DOI] [PubMed] [Google Scholar]
- 12.Rissel R, Schaefer M, Kamuf J, Ruemmler R, Riedel J, Mohnke K, et al. Lung-brain 'cross-talk':systemic propagation of cytokines in the ARDS via the bloodstream using a blood transfusion model does not influence cerebral inflammatory response in pigs. Peer J. 2022;10:e13024. doi: 10.7717/peerj.13024. doi:10.7717/peerj.13024. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 13.Lu W, Zhang J, Qiu Y, Fei N, Yin L. Correlations between APACHE-II score and pressure parameters of mechanical ventilation in patients with ARDS and their value in prognostic evaluation. Pak J Med Sci. 2023;39(6):1584–1588. doi: 10.12669/pjms.39.6.7190. doi:10.12669/pjms.39.6.7190. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 14.Reynolds JC, Michiels EA, Nasiri M, Reeves MJ, Quan L. Observed long-term mortality after 18,000 person-years among survivors in a large regional drowning registry. Resuscitation. 2017;110:18–25. doi: 10.1016/j.resuscitation.2016.10.005. doi:10.1016/j.resuscitation.2016.10.005. [DOI] [PubMed] [Google Scholar]
- 15.Malhotra AK, Shakil H, Smith CW, Huang YQ, Kwong JCC, Thorpe KE, et al. Predicting outcomes after moderate and severe traumatic brain injury using artificial intelligence:a systematic review. NPJ Digit Med. 2025;8(1):373. doi: 10.1038/s41746-025-01714-y. doi:10.1038/s41746-025-01714-y. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Blackburn KM, Wang C. Post-infectious neurological disorders. Ther Adv Neurol Disord. 2020;13 doi: 10.1177/1756286420952901. 17562⇂0952901. doi:10.1177/17562⇂0952901. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 17.Matin N, Sarhadi K, Crooks CP, Lele AV, Srinivasan V, Johnson NJ, et al. Brain-Lung Crosstalk:Management of Concomitant Severe Acute Brain Injury and Acute Respiratory Distress Syndrome. Curr Treat Options Neurol. 2022;24(9):383–408. doi: 10.1007/s11940-022-00726-3. doi:10.1007/s11940-022-00726-3. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Robba C, Poole D, McNett M, Asehnoune K, Bösel J, Bruder N, et al. Mechanical ventilation in patients with acute brain injury:recommendations of the European Society of Intensive Care Medicine consensus. Intensive Care Med. 2020;46(12):2397–2410. doi: 10.1007/s00134-020-06283-0. doi:10.1007/s00134-020-06283-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Ziaka M, Exadaktylos A. Brain-lung interactions and mechanical ventilation in patients with isolated brain injury. Crit Care. 2021;25(1):358. doi: 10.1186/s13054-021-03778-0. doi:10.1186/s13054-021-03778-0. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Roh J, Jo EJ, Eom JS, Mok J, Kim MH, Kim KU, et al. Factors predicting long-term survival of patients with sepsis on arrival at the emergency department:A single-center, observational astudy. Medicine (Baltimore) 2019;98(33):e16871. doi: 10.1097/MD.0000000000016871. doi:10.1097/MD.0000000000016871. [DOI] [PMC free article] [PubMed] [Google Scholar]