2.1. Setting

This study was a retrospective cohort study based on data from the Medical Information Mart for Intensive Care IV (MIMIC IV 0.4) from 2008 to 2019 [13]. MIMIC IV is a publicly available database of patients admitted to the Beth Israel Deaconess Medical Center in Boston, MA, USA. The use of data from the MIMIC IV database was approved by the Institutional Review Boards of the Massachusetts Institute of Technology and Beth Israel Deaconess Medical Center [13]. The requirement for individual patient consent was waived because the study did not impact clinical care, and all protected health information was deidentified. We included data from 69,619 ICU patients, which included 19,658 patients with sepsis, among whom 1,377 patients with sepsis had their serum ammonia levels evaluated.

2.2. Patients

First, the database was searched for patients with sepsis. Sepsis was defined as a suspected infection combined with an acute increase in the Sequential Organ Failure Assessment (SOFA) score of ≥2 [1]. Patients with an infection site or prescriptions for antibiotics and those who underwent sampling of bodily fluids for a microbiological culture were considered to have a suspected infection. When the antibiotic was administered before sampling, we assumed the microbiological sample must have been collected within 24 h; when microbiological sampling was conducted before administering antibiotics, we assumed the antibiotic must have been administered within 72 h [14]. Second, we searched for patients with sepsis who underwent serum ammonia level testing. We excluded patients diagnosed with acute and chronic liver disease and diseases that affect consciousness. We divided the patients into an SAE group and a non-SAE group. SAE in the study was defined as a Glasgow Coma Scale (GCS) score of <15, diagnosed delirium, or usage of haloperidol in patients with sepsis [4, 5, 15].

The inclusion criteria were as follows: (1) patients who fulfilled the definition of sepsis 3.0, (2) patients >18 years, and (3) an ICU stay of >24 h.

The exclusion criteria included the following [4, 5]: (1) patients with brain injuries, such as traumatic brain injury, meningitis, encephalitis intracerebral hemorrhage, cerebral embolism, ischemic stroke, epilepsy, or intracranial infection and other cerebrovascular diseases (Supplementary materials 1–Supplementary materials 5); (2) mental disorders and neurological diseases (Supplementary materials 6); (3) chronic alcohol or drug abuse (Supplementary materials 7); (4) metabolic encephalopathy, hepatic encephalopathy, hypertensive encephalopathy, hypoglycemic coma, and other liver or kidney diseases affecting consciousness (Supplementary materials 8); (5) severe electrolyte imbalances or glycemic disturbances, including hyponatremia (<120 mmol/l), hyperglycemia (>180 mg/dl), or hypoglycemia (<54 mg/dl); (6) a partial pressure of carbon dioxide (PaCO₂) ≥80 mmHg; (7) patients with acute and chronic liver diseases [16, 17], including hepatitis, hepatic cirrhosis, hepatic encephalopathy, hepatorenal syndrome, hepatic injury, or other chronic liver diseases (Supplementary materials 9); and (8) missing serum ammonia values.

2.3. Data Collection

Patients’ baseline variables, including age, sex, type of admission, ethnicity, coexisting illness (Supplementary materials 10–14), site of infection, microbiology type, length of hospital stays, and hospital mortality were extracted from the database. The patient’s vital signs, including heart rate, systolic blood pressure, diastolic blood pressure, mean arterial pressure, respiratory rate, temperature, and arterial oxygen saturation (SpO₂), were extracted. The patients’ laboratory parameters, including alanine aminotransferase, aspartate aminotransferase, albumin, ammonia, lactate, bilirubin, creatinine, blood urea nitrogen, glucose, and hemoglobin levels along with platelet count, partial thrombin time, international normalized ratio, prothrombin time, white blood cell count, neutrophil count, lymphocyte count, partial pressure of carbon dioxide (PaCO₂), and partial pressure of oxygen (PaO₂), were extracted. The patient’s disease severity score, including Sequential Organ Failure Assessment (SOFA) score, Simplified Acute Physiology Score II (SAPS), and Glasgow Coma Scale (GCS) score in the first 24 h of the patient’s admission to the ICU, was extracted. The analgesic and sedative drugs used, including midazolam, propofol, and/or opioids, were extracted. The use of vasopressors, mechanical ventilation, and renal replacement therapy was extracted. The worst laboratory parameters and the mean value of vital signs during the first 24 h of ICU admission were used in the analysis in this study. R statistical software (R Foundation for Statistical Computing, Vienna, Austria) was used to extract the patient’s indexes.

2.4. Statistical Analysis

All continuous and nonnormally distributed variables were assessed using the Shapiro–Wilk test. Age, Charlson’s comorbidity index, vital signs, laboratory parameters, SAPS II score, SOFA score, GCS score, and length of hospital stays were expressed as the medians (P25, P75) (interquartile range, IQR). Categorical variables, including sex, type of admission, ethnicity, coexisting illness, site of infection, microbiology type, length of hospital stays, hospital mortality, use of analgesic and sedative drugs, vasopressors, and mechanical ventilation, were expressed as numbers and proportions. Continuous variables were examined using the Mann–Whitney U test in the SAE group and the non-SAE group; categorical variables were compared using Pearson’s exact test.

Serum ammonia in the SAE group was significantly higher than that in the non-SAE group as shown in Table 1. We applied propensity score matching to match the covariates with in-between differences as shown in Supplementary Material 15. We considered that it may be related to the occurrence of SAE. Our data were all nonnormally distributed. We used generalized additive models [18] to further determine the relationship between serum ammonia levels and SAE. Our cohort study found that serum ammonia levels of ≥45 μmol/L can increase the incidence of SAE. For further verification, we used the multivariate logistic regression analysis. After adjusting for other risk factors, serum ammonia levels of ≥45 mmol/l were identified as an independent risk factor for SAE. We used the survive package to observe the trend of changes between serum ammonia levels and the hospital mortality of SAE. Data were analyzed using R software.

3.1. Baseline Characteristics

As shown in Figure 1, 1,377 patients with sepsis had data on their serum ammonia level in the MIMIC IV database. In total, 393 patients were diagnosed with acute and chronic liver diseases, 264 patients were diagnosed with traumatic brain injury, and other diseases were excluded. Finally, 720 patients were included in the final cohort.

Table 1 describes the characteristics and outcomes of patients with SAE in the original cohort. Compared with patients in the non-SAE group, patients in the SAE group were younger (64 (55–72) vs. 68 (58–75), p = 0.002) and had lower platelet levels (140 (82–206) vs. 173 (98–242), p = 0.001), higher serum ammonia levels (41 (31–63) vs. 21 (16–31), p < 0.001), and higher SOFA scores (6 (4–10) vs. 5 (3–8), p < 0.001). In addition, compared to the non-SAE group, more patients in the SAE group used opioids (7.5% vs. 3.1%, p = 0.017), underwent mechanical ventilation (53.8% vs. 41.2%, p = 0.017), had a longer hospital stay (4.6 (1.9–11.7) vs. 2.6 (1.3–6.2), p < 0.001), and had higher hospital mortality (20.6% vs. 1.2%, p < 0.001). To verify the accuracy of our results, propensity score matching analysis was performed on the original cohort; patients were matched for age, urinary infection, lung infection, catheter infection, abdominal cavity infection, Klebsiella infection, Escherichia coli infection, Pseudomonas aeruginosa infection, Staphylococcus aureus infection, platelet counts, and use of propofol and morphine drugs in the original cohort in a 1 : 1 match. The standardized mean differences (SMD) of the original cohort when compared with those of the IPW cohorts showed that almost all variables had SMD values of <10%, thereby indicating a good matching effect (Supplementary material 15). After propensity score matching, our cohort still showed that serum ammonia levels were significantly higher in the SAE group than in the non-SAE group (p < 0.001).

3.2. Relationship between Serum Ammonia Levels and SAE Incidence

A generalized linear model was used to explore the relationship between serum ammonia levels and SAE incidence (Figure 2(a)). Serum ammonia ≥45 μmol/L can increase the incidence of SAE. Furthermore, Figure 2(b) shows that as serum ammonia levels increase, the incidence of SAE increases.

3.3. Multivariate Logistic Analysis of Risk Factors for SAE

After adjusting for other risk factors, serum ammonia levels of ≥45 μmol/L (odds ratio (OR): 3.501, 95% confidence interval (CI): 2.326–5.269, p < 0.001), use of opioids (OR: 3.433, 95% CI: 1.360–8.669, p = 0.009), and SOFA scores (OR: 1.126, 95% CI: 1.062–1.194, p < 0.001) were risk factors for the occurrence of SAE (Table 2).

3.4. Relationship between Serum Ammonia Levels and SAE Hospital Mortality

The survival package of R Studio was used to observe the relationship between serum ammonia levels and the hospital mortality of patients with SAE. Figure 3 demonstrates that, when serum ammonia levels were <150 μmol/L, an increase in the serum ammonia level was associated with a gradual increase in the hospital mortality of patients with SAE.

3.5. The Effect of Serum Ammonia Level ≥45 μmol/L on Prognosis of SAE Patients

Patients with SAE who have serum ammonia levels of ≥45 μmol/L had higher SOFA scores (7 (5–10) vs. 4 (6–9), p = 0.022) than those with serum ammonia levels of <45 μmol/L; however, there was no difference in length of hospital stays and hospital mortality between the two groups (Table 3).

4.1. Limitations

This study had some limitations. First, our definition of SAE and nonhepatic serum ammonia were different from those used in previous studies. According to the reference literature retrieved from retrospective studies, the definition of SAE is based on the GCS score, the diagnosis of delirium, or the use of haloperidol drugs, and the exclusion of cerebral hemorrhage, brain trauma, and epilepsy, severe liver and kidney disease caused by consciousness disorder, or mental disorders. The definition of nonhepatic serum ammonia is based on the diagnostic code to exclude acute and chronic liver disease. They all lack imaging data, such as computed tomography, magnetic resonance imaging, electroencephalogram, and ultrasound data. Hence, there may be information bias in the SAE cohort. Second, the results of this study demonstrated that patients with sepsis who had serum ammonia levels of ≥45 μmol/L had a 3.508 times higher chance of developing SAE than those with serum ammonia levels of <45 μmol/L; however, a causal association with serum ammonia level ≥45 μmol/L and the incidence of SAE could not be proved due to the retrospective nature of our study. Its causality should be assessed in future studies using a prospective study design. The consciousness of patients with SAE could not be observed through our intervention of serum ammonia levels. Finally, critically ill patients often have various other comorbidities. Due to the interaction between the diseases, some confounding factors still cannot be ruled out and can cover up or exaggerate the relationship between the risk factors and SAE. The small sample size of this study may lead to bias in the study results and require larger sample sizes to validate in the future.