Plasma ammonia levels predict hospitalisation with liver-related complications and mortality in clinically stable outpatients with cirrhosis

Plasma ammonia levels predict hospitalisation with liver-related complications and mortality in clinically stable outpatients with cirrhosis Graphical abstract 3 tertiary liver units 754 outpatients with stable cirrhosis Complete demographic, clinical and biochemical profiling Baseline ammonia sampling Follow-up until liver-related complications, transplant or death Training and test dataset 1.4 times AMM-ULN was the best cut-off to predict time to hospitalisation due to liver-related complications in the training cohort Ammonia is an independent predictor of hospitalisation with liver-related complications and mortality in clinically stable outpatients with cirrhosis and performs better than traditional severity scores in predicting liver-related complications. Test cohort Validation dataset Independent cohort of 130 patients 0.00 0.25 0.50 0.75 1.00 True positive rate (sensitivity) Time-dependent ROC curve (AUC) 1-year complications 0.00 0.25 0.50 0.75 1.00 False positive rate (1-specificity) AMM-ULN (AUC = 77.9) MELD (AUC = 66.1)Child-Pugh (AUC = 72.2) ++ + + + + + + + + ++ ++ + + ++++++++ + + ++ + + ++++ + +++ + + ++++++ + ++++++++ ++ +++++++++++ +++ ++++++ ++++++++ +++ ++++++ + + ++++ + ++++ +++ + +++ ++ +++++ 0.00 0.25 0.50 0.75 1.00 0 500 1,000 1,500 2,000 2,500 Complications probability 7983000 106 33 19 11 5 0Low High p <0.0001 Days N° at risk Liver-related complications AMM-ULN +High +Low ++++++++++ +++++ + ++ ++++++++++ +++ ++++ ++ ++ +++ ++ ++ ++ +++ + + + + ++++++++++++++++++++++++++ ++++++++++ +++++++++++++ ++++ +++++ +++ +++++++++++ +++ + + +++++ + ++ +++ p <0.0001 0.00 0.25 0.50 0.75 1.00 OS probability 12 5 2 1 39 24 15 7 0 500 1,000 1,500 2,000 2,500 79 0 106 0 Days 3 9 2 4 1 5 7 106 0 Low N° at risk High Low N° at risk High Survival AMM-ULN +High +Low AMM-ULN +High +Low + ++ ++ + +++++ + + + + + + ++ + +++ ++ + + +++ + + ++++ + ++++++ ++ + ++++ +++++++++++++ +++ +++ ++ ++++ +++++++++++++++ ++++ ++++ ++++ p = 0.025 0.00 0.25 0.50 0.75 1.00 0 2,000 4,000 6,000 Complications prbability 33720 219 78 19 1 Days AMM-ULN vs. MELD (p <0.001) and Child-Pugh (p = 0.086) Highlights Ammonia is an independent predictor of both hospitalisation with liver-related complications and mortality in stable outpatients with cirrhosis. Ammonia performs better than traditional severity scores in predicting liver-related complications. A cut-off level of 1.4x the upper limit of normal ammonia defines the risk of both hospitalisation with liver-related complications and mortality. Ammonia is a key variable for the prediction of liver-related complications in a derivation cohort and upon external validation. Authors Thomas H. Tranah, María-Pilar Ballester, Juan Antonio Carbonell-Asins, ., Carmina Montoliu, Rajiv Jalan, Debbie L. Shawcross Correspondence r[email protected] (R. Jalan). Lay summary We conducted a prospective cohort study evaluating the association of blood ammonia levels with the risk of adverse outcomes in 754 patients with stable cirrhosis across 3 independent liver units. We found that ammonia is a key determinant that helps to predict which patients will be hospitalised, develop liverrelated complications and die; this was confirmed in an independent cohort of patients. https://doi.org/10.1016/j.jhep.2022.07.014 © 2022 The Authors. Published by Elsevier B.V. on behalf of European Association for the Study of the Liver. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/). J. Hepatol. 2022, 77, 1554–1563 Research Article Cirrhosis and Liver Failure Plasma ammonia levels predict hospitalisation with liver-related complications and mortality in clinically stable outpatients with cirrhosis Thomas H. Tranah 1,† , María-Pilar Ballester 2,3,† , Juan Antonio Carbonell-Asins 3,† , Javier Ampuero 4 , Gonçalo Alexandrino 1,5 , Andra Caracostea 1 , Yolanda Sánchez-Torrijos 4 , Karen L. Thomsen 6,7 , Annarein J.C. Kerbert 6 , María Capilla-Lozano 2 , Manuel Romero-Gómez 4 , Desamparados Escudero-García 2 , Carmina Montoliu 3,8 , Rajiv Jalan 6,9,‡, *, Debbie L. Shawcross 1,‡ 1 Institute of Liver Studies, Dept of Inflammation Biology, School of Immunology and Microbial Sciences, Faculty of Life Sciences and Medicine, King’s College London, London, United Kingdom; 2 Digestive Disease Department, Hospital Clínico Universitario de Valencia, Spain; 3 INCLIVA Biomedical Research Institute, Valencia, Spain; 4 Hospital Universitario Virgen del Rocio, Instituto de Biomedicina de Sevilla, Universidad de Sevilla, Ciberehd, Spain; 5 Gastroenterology and Hepatology Department, Hospital Prof. Doutor Fernando Fonseca, Amadora, Portugal; 6 Liver Failure Group, Institute for Liver and Digestive Health, University College London, Royal Free Campus, United Kingdom; 7 Department of Hepatology and Gastroenterology, Aarhus University Hospital, Palle Juul-Jensens Boulevard 99, 8200, Aarhus, Denmark; 8 Department of Pathology, Faculty of Medicine, University of Valencia, Spain; 9 European Foundation for the Study of Chronic Liver Failure (EF Clif), Spain Background & Aims: Hyperammonaemia is central in the pathogenesis of hepatic encephalopathy. It also has pleiotropic deleterious effects on several organ systems, such as immune function, sarcopenia, energy metabolism and portal hypertension. This study was performed to test the hypothesis that severity of hyperammonaemia is a risk factor for liver-related complications in clinically stable outpatients with cirrhosis. Methods: We studied 754 clinically stable outpatients with cirrhosis from 3 independent liver units. Baseline ammonia levels were corrected to the upper limit of normal (AMM-ULN) for the reference laboratory. The primary endpoint was hospitalisation with liver-related complications (a composite endpoint of bacterial infection, variceal bleeding, overt hepatic encephalopathy, or new onset or worsening of ascites). Multivariable competing risk frailty analyses using fast unified random forests were performed to predict complications and mortality. External validation was carried out using prospective data from 130 patients with cirrhosis in an independent tertiary liver centre. Results: Overall, 260 (35%) patients were hospitalised with liverrelated complications. On multivariable analysis, AMM-ULN was an independent predictor of both liver-related complications (hazard ratio 2.13; 95% CI 1.89–2.40; p<0.001) and mortality (hazard ratio 1.45; 95% CI 1.20–1.76; p<0.001). The AUROC of AMM-ULN was 77.9% for 1-year liver-related complications, which is higher than traditional severity scores. Statistical differences in survival were found between high and low levels of AMM-ULN both for complications and mortality (p<0.001) using 1.4 as the optimal cut-off from the training set. AMM-ULN remained a key variable for the prediction of complications within the random forests model in the derivation cohort and upon external validation. Conclusion: Ammonia is an independent predictor of hospitalisation with liver-related complications and mortality in clinically stable outpatients with cirrhosis and performs better than traditional prognostic scores in predicting complications. Lay summary: We conducted a prospective cohort study evaluating the association of blood ammonia levels with the risk of adverse outcomes in 754 patients with stable cirrhosis across 3 independent liver units. We found that ammonia is a key determinant that helps to predict which patients will be hospitalised, develop liver-related complications and die; this was confirmed in an independent cohort of patients. © 2022 The Authors. Published by Elsevier B.V. on behalf of European Association for the Study of the Liver. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4. 0/). Introduction The Child-Pugh (CP) score and the model for end-stage liver disease (MELD) score are the most utilised non-invasive tools for prediction of survival in patients with cirrhosis but are limited by interobserver subjectivity and their initial derivations in predicting survival after surgery and transjugular intrahepatic portosystemic shunt, respectively. 1,2 Furthermore, the composite features of the MELD score (bilirubin, albumin, international normalised ratio [INR] and creatinine) reflect incomplete facets of the pathophysiology of cirrhotic portal hypertension that are restricted to liver synthetic dysfunction and renal insufficiency. Predictive prognostic models for the development of liverrelated complications in patients with stable compensated cirrhosis are also limited. Endoscopic surveillance for varices is useful in determining the risk of variceal bleeding, neuropsychiatric tests provide insight into the risk of overt hepatic Keywords: Cirrhosis; Ammonia; Liver-related complications; Hepatic encephalopathy; Variceal bleeding; Ascites; Bacterial infection. Received 14 March 2022; received in revised form 7 July 2022; accepted 8 July 2022; available online 22 July 2022 *Corresponding author. Address: Liver Failure Group, Institute for Liver and Disease Health, University College London, Royal Free Campus. Rowland Hill Street, London, NW3 2PF, United Kingdom; Tel.: +44 02074332795. E-mail address: r[email protected] (R. Jalan). † Joint first authors. ‡ Joint senior authors. https://doi.org/10.1016/j.jhep.2022.07.014 Journal of Hepatology 2022 vol. 77 j1554–1563 Research Article Cirrhosis and Liver Failure encephalopathy (HE) and detection of clinically significant portal hypertension using hepatic venous pressure gradient (HVPG) provides insights into the risk of development of liver-related complications. 3–5 However, measurement of HVPG is expensive, invasive, requires specialist equipment, and is not routinely available. A simple, cost-effective, and widely available tool to define the risk of future liver-related complications remains a key area of unmet need. Ammonia has been long established as a gut-derived neurotoxin whose impaired metabolism in chronic liver disease plays a pivotal role in the development of HE. 6 Whilst hyperammonaemia is directly associated with cerebral oedema and raised intracranial pressure in acute liver failure, which is a critical and often fatal phenomenon, 7 the role of hyperammonaemia in HE complicating cirrhosis is less well defined and its relevance in routine clinical practice remains controversial. Ammonia has been implicated in the pathogenesis of other liver-related complications such as liver cell injury, immune dysfunction, sarcopenia and portal hypertension. 8 Recently, ammonia levels have been shown to be an independent predictor of mortality in patients with acute decompensation or acute-on-chronic liver failure irrespective of the severity of HE, suggesting that it may be a useful biomarker for predicting other liver-related complications. 9–12 We hypothesised that hyperammonaemia is a risk factor for the development of not only HE but also other liver-related complications and consequent mortality in clinically stable outpatients with cirrhosis. In this study, our primary aim was to determine whether ammonia levels define the risk of subsequent hospitalisation with liver-related complications such as bacterial infection, variceal bleeding, overt HE and ascites. The secondary aims were to determine whether ammonia levels were associated with mortality and the individual liver-related complications. We also sought to determine a threshold value of ammonia that defines the risk of complications and mortality and, to develop a prognostic model for the prediction of these events in stable outpatients with cirrhosis. Patients and methods Study design and patient selection The AMMON consortium was created to determine the role of ammonia in the pathogenesis and treatment of complications of cirrhosis. This analysis is part of the ongoing studies within the Consortium. This is the first study of this programme and evaluates data from 4 independent liver units in Europe: King’s College Hospital (KCH), London, United Kingdom (UK), Hospital Clínico Universitario de Valencia (HCUV), Spain, Royal Free Hospital (RFH), London, UK and Virgen del Rocio University Hospital (VRUH), Seville, Spain. A prospective observational cohort study of clinically stable outpatients with cirrhosis was conducted across these sites. Inclusion criteria were patients aged > −18 years with cirrhosis based on histological criteria, characteristic radiological findings and/or typical clinical presentation. Patients with both compensated and decompensated cirrhosis were included in the study, however hospital inpatients at the time of assessment and patients hospitalised with an acute decompensation within the previous 4 weeks were excluded. Baseline ammonia levels were measured in all patients at the beginning of the study. Specific characteristics of patients comprising the training and test and validation cohorts were: - Training and test cohort BCohort 1 –KCH, n = 447: patients sequentially assessed for liver transplantation. Arterial ammonia levels were measured. BCohort 2 - HCUV, n = 156: consecutive patients reviewed without previous episodes of overt HE or hepatocellular carcinoma (HCC). Venous ammonia levels were measured. BCohort 3 - RFH, n = 151: consecutive patients without previous episodes of overt HE or HCC, not candidates for liver transplantation. Venous ammonia levels were measured. - Validation cohort: consecutive outpatients from VRUH (n = 130). Venous ammonia levels were measured. Patients were followed up until liver transplantation, death, or study closure. Clinical datasets came from studies approved by ethical review boards at each study site. Variables Ammonia was measured in each hospital either for routine clinical purposes or as part of other studies addressing the role of ammonia in the pathogenesis of complications of cirrhosis at the time of outpatient clinical evaluation. All centres used standard operating procedures for ammonia measurement that involved collection of the sample in cooled EDTA tubes, rapid sample transport to the laboratory on ice and spectrophotometric assays. There was a requirement to standardise ammonia levels to correct for differences in phlebotomy and laboratory handling protocols between centres; we transformed the crude ammonia measurement to a calibrated ammonia level (AMM-ULN) using the formula: AMM-ULN = serum ammonia ( l mol/L)/reference laboratory upper limit of normal for ammonia ( l mol/L). In this manner, we were able to express the ratio of the patient ammonia level corrected to a normal population measured with the same local test system and test conditions. The following clinical and demographic information were collected at baseline: age, sex, anthropometric data, aetiology of cirrhosis, comorbidities, co-prescribed medications, previous liver-related complications suggesting decompensated disease, laboratory parameters and disease severity assessed by MELD, MELD-Na and CP scores. The presence of portal hypertension was determined using its surrogates; this included presence of varices, evidence of portosystemic collaterals, splenomegaly and/or use of beta-blockers. The primary endpoint of the study was hospitalisation due to liver-related complications which represented a composite endpoint of bacterial infection, variceal bleeding, overt HE and new onset, or worsening ascites; more details on the definitions of liver-related complications are available in the supplementary materials. Secondary endpoints included overall survival (OS), type of complication and both liver-related complications and survival at 3, 6, and 12 months and 5 years. Statistical analyses Continuous demographic, clinical and laboratory variables were analysed for normality using the Shapiro-Wilk test. Normally distributed data were reported as mean (SD) and non-normally distributed data were reported as median (IQR). Comparisons between hospitals were performed using ANOVA or KruskalWallis tests for normally and non-normally distributed data, Journal of Hepatology 2022 vol. 77 j1554–1563 1555 respectively. Categorical data were reported as n (%) and comparisons analysed by Chi-squared (s 2 ) test. All correlations were performed using Pearson’s correlation coefficient except for ordinal variables where Kendall’s tau was used. Multivariable competing risk frailty analysis was performed, considering liver transplantation as a competing risk, to identify factors independently associated with complications and mortality using Fine-Gray subdistribution hazard modelling. Only original variables (not CP or MELD score) were included in the multivariable model to avoid multicollinearity. A competing risk cause-specific Cox regression model was applied to study the effect of original variables in each liver-related complication. Time-dependent receiver-operating characteristic (ROC) curves were constructed and compared considering the competing risk within the model. 13 Data was randomly split into training (75%) and test (25%) sets and the optimal cut-off for AMM-ULN was calculated using maximally selected rank statistics for time to development of liver-related complications (Fig. S1). This value was then used in the test sample to construct Kaplan Meier curves for time to development of liver-related complications and OS. Differences in survival were assessed using log-rank tests. AMM-ULN was subsequently evaluated in an independent external validation set. Fast unified random forests with 500 trees, using log-rank splitting criteria with bootstrapping and cross-validation, were Table 1. Description of patient characteristics in the 3 prospective hospital cohorts. Parameter Total (n = 754) RFH (n = 151) HCUV (n = 156) KCH (n = 447) pvalue Demographic Age (mean, SD) 56 (11) 57 (11) 63 (9) 54 (11) <0.001 Sex (no. male, %) 498 (66) 80 (53) 125 (80) 293 (66) <0.001 Disease aetiology, n (%) ALD 300 (40) 73 (48) 69 (44) 158 (35) <0.001 NAFLD 106 (14) 3 (2) 14 (9) 89 (20) Viral hepatitis 182 (24) 60 (40) 58 (37) 64 (14) Autoimmune liver disease 110 (15) 9 (6) 4 (3) 97 (22) Other 56 (7) 6 (4) 11 (7) 39 (9) Comorbidities, n (%) Diabetes mellitus 263 (35) 55 (36) 55 (35) 153 (34) 0.875 Co-prescribed medications, n (%) Non-selective b-blockers 355 (47) 104 (69) 58 (37) 193 (43) <0.001 Lactulose 266 (35) —55 (35) 211 (47) <0.001 Rifaximin 176 (23) —1 (0.6) 175 (39) <0.001 Laboratory parameters mean (SD) Albumin (g/dl) 3.3 (0.7) 2.9 (0.6) 3.9 (0.6) 3.3 (0.6) <0.001 Bilirubin (mg/dl) 3.4 (4.4) 4.9 (5.7) 1.3 (1.2) 3.5 (4.4) <0.001 Creatinine (mg/dl) 0.9 (0.5) 1.0 (0.6) 0.9 (0.3) 0.9 (0.4) 0.001 INR 1.5 (0.4) 1.4 (0.3) 1.2 (0.3) 1.6 (0.5) <0.001 WBC (x10 9 /L) 5.3 (2.7) 6.6 (3.8) 5.9 (2.2) 4.7 (2.3) <0.001 Platelets (x10 9 /L) 115 (71) —123 (60) 112 (75) 0.126 Sodium (mmol/L) 136 (5) 135 (5) 139 (3) 136 (5) <0.001 Ammonia level ( l mol/L) 63 (40) 61 (17) 29 (29) 75 (42) <0.001 AMM-ULN 1.4 (0.8) 1.5 (0.4) 0.9 (0.9) 1.5 (0.8) <0.001 Disease severity, n (%) Portal hypertension 708 (94) 148 (98) 127 (81) 433 (97) <0.001 Hepatocellular carcinoma 79 (10) 0 0 79 (18) <0.001 Decompensated cirrhosis 484 (64) 127 (84) 25 (16) 332 (74) <0.001 Child-Pugh score, mean (SD) 8 (2) 9 (2) 6 (1) 9 (2) <0.001 Child-Pugh group A 197 (26) 16 (11) 106 (69) 75 (17) <0.001 B 346 (46) 83 (55) 42(28) 221 (50) C 207 (28) 52 (34) 5 (3) 150 (34) MELD Score (mean, SD) 14 (6) 14 (7) 10 (4) 15 (6) <0.001 MELD-Na Score (mean, SD) 15 (7) 17 (7) 9 (3) 17 (6) Outcome, n (%) Transplant 359 (48) 0 (0) 3 (2) 355 (79) <0.001 Development of liver-related complication 260 (35) 73 (48) 53 (34) 134 (30) <0.001 Type of complication, n (%) Ascites 56 (7) 16 (11) 16 (10) 24 (5) <0.001 Variceal bleeding 28 (4) 9 (6) 10 (7) 9 (2) Infection 123 (16) 30 (20) 15 (10) 78 (17) Hepatic encephalopathy 52 (7) 16 (11) 12 (8) 24 (5) Mortality, n (%) 120 (16) 29 (19) 40 (26) 51 (11) <0.001 Follow-up (days; median, range) 223 (2,451) 232 (422) 1,385 (2,418) 138 (1,341) <0.001 Continuous variables are presented as mean (SD), categorical variables are presented as n (%). Comparisons were performed between patient cohorts using ANOVA and Kruskal Wallis tests for normally and non-normally distributed data, respectively and Chi-squared for categorical data. p<0.05 was considered statistically significant. ALD, alcohol-related liver disease; AMM-ULN, ammonia level corrected to the upper limit of normal; HCUV, Hospital Clínico Universitario de Valencia; INR, international normalized ratio; KCH, King’s College Hospital; MELD, model for end-stage liver disease score; NAFLD, non-alcoholic fatty liver disease; RFH, Royal Free Hospital; WBC, white blood cell count. 1556 Journal of Hepatology 2022 vol. 77 j1554–1563 Research Article Cirrhosis and Liver Failure used to predict mortality and complications in the combined dataset. Fourteen different models were included using different subsets of variables; the external validation set was used to test model performance of all 14 models. Performance in risk prediction for each model was evaluated using Brier score comparisons. Variable importance (VIMP) was extracted using subsampling to calculate confidence intervals and standard errors. All analyses were performed with R (version 4.0.2, R Core, 2021) with the cut-off for statistical significance set at 0.05. Maxstat package 14 was used for maximally selected rank statistics; Survival 15 and Survminer 16 for Kaplan Meier curves and logrank statistic calculation; timeROC 13 package for timedependent ROC curves and AUC comparison; randomForestSRC package 17 for the random forest analyses; and pec package 18 for Brier score model comparison. Results Patient characteristics A total of 754 patients were included (66% males; mean 56 years) in the training and test cohort. Patient characteristics from KCH, HCUV, RFH, and the combined dataset are summarised in Table 1. Alcohol and viral hepatitis were the main aetiologies in RFH and HCUV, while non-alcoholic fatty liver disease (NAFLD) was more prevalent in KCH. Severity of liver disease was higher (with higher CP and MELD scores) in the RFH and KCH cohorts than in the HCUV cohort. Median followup of the total cohort was 223 days (range: 2–2,453). Overall, 35% of patients developed a liver-related complication during follow-up, with infections (16%) representing the most frequent type of complication. Amongst 120 patients who died in the study, 36% (n = 43) died outside the tertiary centre with cause of death not known; amongst the remainder, 38% (n = 29) died of complications of cirrhosis or acute-on-chronic liver failure, 28% (n = 22) of infection, 10% (n = 8) of HCC, 12% (n = 9) of other malignancy and 12% (n = 9) died with other non-liverrelated conditions. AMM-ULN demonstrated differential association with disease aetiology (p= 0.002) and was most marked in patients with NAFLD cirrhosis (mean: 1.6, SD: 0.8). Furthermore, ammonia levels were higher in patients who had diabetes (1.5 vs. 1.3, p <0.001). AMM-ULN was predictably higher in patients with more advanced stages of cirrhosis (0.9, 1.5 and 1.6 in CP groups A, B and C, respectively and 1.0, 1.4 and 1.6 in MELD score groups < −9, 9–12 and > −12, respectively p<0.001) (Table S1). AMM-ULN also correlated with the MELD-Na score (r = 0.25, p<0.001), and individual markers of liver function –bilirubin (r = 0.083, p= 0.023), albumin (r = -0.322, p<0.001) and INR (r = 0.172, p <0.001) –as well as platelets (r = -0.209, p<0.001) and creatinine (r = 0.120, p= 0.001). Anthropometric data were not available at every centre, however within the KCH cohort no correlation was found between nutrition or muscle status and AMM-ULN (p>0.05). AMM-ULN as a prognostic biomarker for the prediction of complications and mortality AMM-ULN was an independent predictor of hospitalisation with liver-related complications on univariable analysis in each of the 3 cohorts (KCH: hazard ratio [HR] 2.43, 95% CI 2.17–2.72, p <0.001; HCUV: HR 1.65, 95% CI 1.39–1.97, p<0.001; and RFH: HR 5.13, 95% CI 3.06–8.59, p<0.001) and in the univariable (HR 2.21, 95% CI 1.96–2.49, p<0.001) and multivariable (HR = 2.13, 95% CI 1.89–2.40, p<0.001) analyses of the combined dataset. Other significant risk factors for hospitalisation with liver-related complications were diabetes (HR 1.53, 95% CI 1.12–1.94, p= 0.003) and INR (HR 1.77, 95% CI 1.39–2.27, p= 0.001) (Table 2). Subgroup analyses were performed within the KCH cohort to address the impact of nutrition and muscle status on the risk of hospitalisation with liver-related complications. A stepwise procedure including AMM-ULN and each anthropometric parameter showed the best model included only AMMULN (Table S2). AMM-ULN was also independently associated with the risk of hospitalisation with the individual liver-related complications; overt HE (HR 2.19, 95% CI 2.10–2.29, p<0.001), variceal bleeding (HR 1.93, 95% CI 1.73–2.16, p<0.001), ascites (HR 1.76, 95% CI 1.35–2.30, p<0.001) and bacterial infection (HR 2.35, 95% CI 1.93–2.87, p<0.001) (Table 3). The AUROC for AMM-ULN was 77.9% for 1-year complications; higher than the MELD score (66.1%, p<0.001) but not significantly higher than the CP score (72.2%, p= 0.062), Fig. 1A. The AUROCs of AMM-ULN for hospitalisation due to liver-related complications at 3 and 6 months and, at 5 years were 73.2%, 74.9% and 74.4%, respectively (Fig. S2A–C). AMM-ULN was an independent predictor of mortality in each independent hospital (KCH: HR 1.83, 95% CI 1.61–2.08, p<0.001; HCUV: HR 1.40, 95% CI 1.08–1.81, p= 0.011 and RFH: HR 4.89, 95% CI 2.28–10.48, p<0.001) and the combined cohort (HR 1.73, 95% CI 1.48–2.02, p<0.001). Furthermore, it was an independent risk factor for mortality in the multivariable frailty competing risk model (Table S3). The AUROC of AMM-ULN, 70.5%, was not significantly higher than CP (p= 0.803) and MELD scores (p= 0.073) for 1-year survival (Fig. 1B). Identifying the optimal cut-off of AMM-ULN to predict hospitalisation due to liver-related complications and mortality Maximally selected rank statistics determined AMM-ULN >1.4 as the best cut-off to predict time to hospitalisation due to liverrelated complications in the training cohort. Applying this to the test set, statistical differences in survival were found between high and low levels of AMM-ULN for liver-related complications (log-rank p<0.001) and mortality (log-rank p<0.001) (Fig. 2). Patients in the high-risk group were more likely to have diabetes (42% vs. 30%, p= 0.001), be receiving lactulose (42% vs. 30%, p<0.001) and rifaximin (33% vs. 16%, p<0.001) and have more advanced liver disease (MELD-Na 17 vs. 14, p<0.001; CP score 9 vs. 8, p<0.001) (Table S4). Patients in the high-risk group were more likely to require hospitalisation due to liver-related complications (58% vs. 17%, p<0.001), including bacterial infections (28% vs. 8%, p<0.001), and patients had a higher mortality (23% vs. 10%, p<0.001) over the course of the study follow-up. Prognostic model to predict hospitalisation due to liverrelated complications Cox regression and random forest models were developed and evaluated using bootstrap cross-validation. The integrated Brier score showed that the best model to predict liver-related complications was the random forest model with original variables; Journal of Hepatology 2022 vol. 77 j1554–1563 1557 further models with a reduced number of variables had similar margins of prediction error (Fig. 3). The error rate for the selected model according to the number of trees and subsampled forests to estimate standard errors and confidence intervals for VIMP is described in Fig. 4. Variables that were retained in the model according to their VIMP to predict liver-related complications were AMM-ULN, sodium, creatinine, albumin, bilirubin, and INR with AMM-ULN identified as the single most important variable for predicting hospitalisation due to liver-related complications. The addition of a history of previous liver decompensation to the prognostic model did not increase its predictive ability (Fig. S3). External cohort validation of AMM-ULN as a predictor of hospitalisation due to liver-related complications AMM-ULN remained a risk factor for hospitalisation due to liver-related complications in an independent validation cohort of 130 patients (HR 1.56; 95% CI 1.18–2.05; p=0.002); comparative clinical data from the derivation and validation sets is summarised (Table S5). Kaplan-Meier analysis with logrank test using the AMM-ULN cut-off of 1.4 again demonstrated significant differences in predicting liver-related complications, p= 0.025 (Fig. 5). Application of the Cox regression and random forest models from our derivation sets to the validation cohort further demonstrated the optimal performance of random forest with original variables and the central importance of AMM-ULN as a prediction variable within the most accurate models (Fig. S4). Discussion In this study, we sought to consider hyperammonaemia outside of the context of HE and to define its role as a prognostic biomarker in patients with stable cirrhosis. For the first time, we have demonstrated that ammonia is an independent predictor of hospitalisation due to liver-related complications and mortality in clinically stable outpatients with cirrhosis. Multivariable analysis revealed that the component variables of traditional scoring systems, bilirubin, creatinine, albumin, INR, and sodium, remained significant in the prediction of liver-related complications. However, AMM-ULN was independently associated with adverse outcomes and carried greater weight than any other prognostic variable. Although the cohorts in the AMMON study were collected in 4 tertiary liver centres and encompassed patients followed up with different clinical characteristics, we observed concordance in the ability of AMM-ULN to independently predict both hospitalisation due to liver-related complications and survival. Furthermore, AMM-ULN performed better than traditional severity scores in predicting hospitalisation due to liver-related complications at 6 months, 12 months, and 5 Table 2. Univariable and multivariable model for liver-related complications. Parameter Univariable model Multivariable model HR 95% CI pvalue HR 95% CI pvalue Demographic Age 1.005 0.99–1.02 0.39 Sex 0.96 0.74–1.25 0.79 Disease aetiology ALD Ref. Ref. Ref. NAFLD 1.45 1.01–2.08 0.46 Viral hepatitis 0.67 0.48–0.93 0.02 Autoimmune liver disease 0.79 0.53–1.21 0.29 Other 1.17 0.73–1.88 0.5 Comorbidities Diabetes mellitus 1.75 1.37–2.24 <0.001 1.53 1.12–1.94 0.005 Co-prescribed medications Non-selective b-blockers 1.23 0.96–1.59 0.1 Lactulose 1.88 1.44–2.45 <0.001 Rifaximin 1.59 1.21–2.09 <0.001 Laboratory parameters Albumin 0.56 0.45–0.69 <0.001 Bilirubin 1.03 1.01–1.05 0.004 Creatinine 1.42 1.16–1.73 <0.001 INR 1.86 1.46–2.36 <0.001 1.77 1.39–2.27 <0.001 WBC 0.99 0.94–1.04 0.68 Platelets 0.995 0.992–0.997 <0.001 Sodium 0.97 0.95–0.99 0.015 0.97 0.94–1.00 0.070 AMM-ULN 2.21 1.96–2.49 <0.001 2.13 1.89–2.40 <0.001 Disease severity Child-Pugh class A Ref. Ref. Ref. B 2.46 1.65–3.67 <0.001 C 4.21 2.72–6.49 <0.001 MELD score 1.05 1.03–1.07 <0.001 MELD-Na 1.06 1.04–1.07 <0.001 Decompensated cirrhosis 3.01 1.88–5.09 <0.001 1.47 0.94–2.3 0.090 Competing risk frailty analysis was performed considering liver transplantation as a competing risk to identify factors independently associated with complications and mortality using Fine-Gray subdistribution hazard modelling. A backward-forward stepwise procedure was conducted for variable selection using Akaike’s information criterion. ALD, alcohol-related liver disease; AMM-ULN, ammonia level corrected to the upper limit of normal; HR, hazard ratio; INR, international normalized ratio; MELD, model for end-stage liver disease score; NAFLD, non-alcoholic fatty liver disease; WBC, white blood cell count. 1558 Journal of Hepatology 2022 vol. 77 j1554–1563 Research Article Cirrhosis and Liver Failure years. Our data demonstrate that AMM-ULN >1.4 correlates with a significantly higher risk of hospitalisation due to liver-related complications (58% vs. 17%, p<0.001) and mortality (23% vs. 10%, p<0.001). Contrary to expectation, prior decompensation did not predict hospitalisation with liver-related complications; moreover, addition of prior decompensation to the prognostic model did not increase its predictive ability –this could be because of its interaction with ammonia in defining outcomes. We derived a predictive model using random forest plots that demonstrated superiority to established predictive models such as the MELD and CP scores in addition to models derived from Cox regression within our dataset; AMM-ULN was the parameter that carried the highest VIMP in this model. On external validation we verified that ammonia was again associated with risk of hospitalisation due to liver-related complications and remains a highly important variable in the random forest models with the best predictive accuracy. Taken together, these data provide compelling evidence for the potential role of ammonia to risk stratify outpatients with stable cirrhosis. Hyperammonaemia is classically described in the context of HE. It crosses the blood-brain barrier and exerts several toxic effects leading to astrocytic swelling, neuroinflammation, cell signalling and alterations in neurotransmission. 19,20 Additionally, ammonia has multisystem deleterious effects that may contribute to the emergence of late-stage cirrhotic complications. 8 In this study, we have shown that hyperammonaemia is independently associated not only with the development of HE, but also other decompensating events such as bacterial infections, variceal bleeding, and ascites. Ammonia can directly induce hepatocyte cell death, thereby contributing to the progression of liver injury, fibrosis, and portal hypertension through the activation of hepatic stellate cells. 21,22 Hyperammonaemia also propagates circulating innate immune dysfunction in the context of cirrhosis. 23 Commensurate with this observation, and in the setting of significant portal hypertension, we report increased rates of bacterial infections in patients with AMM-ULN >1.4. Sarcopenia is associated with mortality in cirrhosis and is bidirectionally linked with hyperammonaemia; 24 progressive sarcopenia leads to loss of a compensatory route of extrahepatic ammonia detoxification through glutamine synthetase whilst ammonia has been shown to promote skeletal muscle loss via multiple mechanisms. 25 However, in our study, it was neither associated with AMM-ULN nor was it a significant variable Table 3. Results of multivariable cause-specific competing risk frailty models. Multivariable model HR 95% CI pvalue Ascites Diabetes 0.99 0.44–2.20 0.985 Bilirubin 1.03 0.99–1.06 0.106 Albumin 0.97 0.50–1.90 0.937 INR 1.77 1.57–1.99 <0.001 AMM-ULN 1.76 1.35–2.30 <0.001 Hepatic encephalopathy Diabetes 1.49 1.01–2.20 0.047 Bilirubin 1.05 1.03–1.08 <0.001 Albumin 0.77 0.30–2.01 0.600 INR 1.32 1.15–1.52 <0.001 AMM-ULN 2.19 2.10–2.29 <0.001 Variceal bleeding Diabetes 0.84 0.53–1.32 0.43 Bilirubin 1.12 1.05–1.19 <0.001 Albumin 0.34 0.25–0.46 <0.001 INR 0.28 0.14–0.55 <0.001 AMM-ULN 1.93 1.73–2.16 <0.001 Infection Diabetes 2.05 1.72–2.45 <0.001 Bilirubin 1.04 1.01–1.19 0.015 Albumin 0.56 0.38–0.83 0.004 INR 2.67 2.16–3.31 <0.001 AMM-ULN 2.35 1.93–2.87 <0.001 Results of multivariable cause-specific competing risk frailty models demonstrating variables independently associated with the risk of developing individual complications of liver disease. Competing risk frailty analysis was performed considering liver transplantation as a competing risk to identify factors independently associated with complications and mortality using Fine-Gray subdistribution hazard modelling. A backward-forward stepwise procedure was conducted for variable selection using Akaike’s information criterion. AMM-ULN was independently associated with each individual liver-related complication on multivariable analysis. p<0.05 was considered statistically significant. AMM-ULN, ammonia level corrected to the upper limit of normal; INR, international normalized ratio. 0.00 0.25 0.50 0.75 1.00 True positive rate (sensitivity) Time-dependent ROC curve (AUC) 1-year complications 0.00 0.25 0.50 0.75 1.00 0.00 0.25 0.50 0.75 1.00 False positive rate (1-specificity) 0.00 0.25 0.50 0.75 1.00 False positive rate (1-specificity) True positive rate (sensitivity) Time-dependent ROC curve (AUC) 1-year survival AMM-ULN (AUC = 70.6) MELD (AUC = 63.1)Child-Pugh (AUC = 70)AMM-ULN (AUC = 77.9) MELD (AUC = 66.1)Child-Pugh (AUC = 72.2) AB Fig. 1. Time-dependent AUROC curve. Comparison of AMM-ULN against Child-Pugh and MELD score for predicting (A) hospitalisation due to liver-related complications, and (B) mortality at 1 year. p<0.05 was considered statistically significant. AMM-ULN demonstrates improved predictive performance (77.9%) for the development of liver-related complications when compared against the MELD score (66.1%, p<0.001), and showed a tendency to perform better than the Child-Pugh score (72.2%, p= 0.086). AMM-ULN was not significantly better than the MELD (p= 0.073) and Child-Pugh (p= 0.803) scores for the prediction of 1year mortality. AMM-ULN, ammonia level corrected to the upper limit of normal; MELD, model for end-stage liver disease. Journal of Hepatology 2022 vol. 77 j1554–1563 1559 predicting the development of liver-related complications in the KCH cohort, where this relationship could be evaluated. Hyperammonaemia in cirrhosis may arise as a consequence of impaired liver synthetic function, portosystemic shunting, sarcopenia (with reduced extrahepatic ammonia clearance), renal dysfunction and intestinal dysbiosis in cirrhosis. 26 Thus, hyperammonaemia reflects complex, multiorgan interactions describing any combination of hepatic, renal, neurological, immunological and skeletal muscle functions that become dysregulated in cirrhosis and therefore represents an important biomarker of physiological reserve. 8 In this study, ammonia levels were higher in patients with NAFLD and patients with diabetes. The underlying mechanism of this association is unknown but may be due to the negative impact of liver steatosis on hepatic expression and activity of urea cycle enzymes. 27,28 The intestinal microbiome is highly dysregulated in both NAFLD and cirrhosis, which may contribute to excess ammonia production directly or indirectly by modulating the activity of intestinal glutaminase. 29 One area that has hampered widespread usage of ammonia measurement is the inter-laboratory differences in protocols of measurement and the reflected differences in the range and distribution of measurements. Here, AMM-ULN performed better than crude ammonia levels in predicting both complications and mortality (Akaike information criterion = 3,119 vs. 3,131 and 1,348 vs. 1,359, respectively). Furthermore, this study includes patients with both arterial (KCH) and venous (RFH, HCUV, VRUH) sampling of ammonia with the potential for arteriovenous differences in ammonia measurements secondary to muscle metabolism. 30 Despite this, AMM-ULN levels remained strongly predictive of hospitalisation due to liverrelated complications and mortality in stable cirrhosis. We propose that using AMM-ULN may harmonise ammonia levels being reported widely in the literature, which are often affected by local laboratory practices, in a manner akin to reporting of the international normalised ratio corrected to laboratory prothrombin time controls 31 and severity grading of ++ + + + + + + + + ++ ++ + + ++++++++ + + ++ + + +++++ +++ + + ++++++ + ++++++++ ++ +++++++++++ +++ ++++++ ++++++++ +++ ++++++ ++ ++++ + ++++ +++ + +++ ++ +++++ 0.00 0.25 0.50 0.75 1.00 0 500 1,000 1,500 2,000 2,500 Complications probability 7983000 106 33 19 11 5 0Low High ++++++++++ +++++ + ++ ++++++++++ +++ ++++ ++ ++ +++++ ++ ++ +++ + + + + ++++++++++++++++++++++++++ ++++++++++ +++++++++++++ ++++ +++++ +++ +++++++++++ +++ + + +++++ + ++ +++ p <0.0001 p <0.0001 0.00 0.25 0.50 0.75 1.00 OS probability AMM-ULN +High +Low AMM-ULN +High +Low 12 5 2 1 39 24 15 7 Days N° at risk 0 500 1,000 1,500 2,000 2,500 79 0 106 0Low High Days N° at risk A B Fig. 2. Kaplan Meier plots of hospitalisation due to liver-related complications and mortality. Kaplan Meier plots demonstrating cumulative probability of the development of (A) liver-related complications (composite endpoint of sepsis, variceal bleeding, overt hepatic encephalopathy, acute onset, or worsening ascites), and (B) OS during follow-up. Patients with AMM-ULN > −1.4 were allocated to the high-risk group. Differences in median overall survival were assessed by log-rank test. p<0.05 was considered statistically significant. AMM-ULN, ammonia level corrected to the upper limit of normal; OS, overall survival. 1560 Journal of Hepatology 2022 vol. 77 j1554–1563 Research Article Cirrhosis and Liver Failure drug-induced liver injury classified as a ratio of the ULN of alanine aminotransferase and aspartate aminotransferase. 32 The results of this study should be interpreted considering its limitations. First, the study patient cohorts were derived from 3 separate centres and included patients with different severities of cirrhosis. Despite this, concordant results were observed with respect to the independent association of ammonia with complications and mortality in cirrhosis across the centres that were 0.0 0.1 0.2 0 500 1,500 2,000 2,500 Time Brier scores (BS) Models Cox AMM-ULN Cox all variables Cox Child-Pugh class Cox MELD Cox MELD-Na Cox MELD-Na + AMM-ULN RF AMM-ULN RF all variables RF Child-Pugh class RF MELD RF MELD-Na RF MELD-Na + AMM-ULN RF Redux 1 RF Redux 2 1,000 Fig. 3. Integrated Brier score to evaluate model performance using bootstrap cross-validation to predict hospitalisation due to liver-related complications. The integrated Brier score calculates the predictive error over time with larger values of the integrated Brier score indicating worse performance of the predictive model. Fourteen models from our training set were compared; these were developed using Cox regression and RF modelling with input variables restricted to AMM-ULN, CP class, MELD score, MELD-Na score, or all variables (age, sex, aetiology of liver disease, diabetes, bilirubin, albumin, creatinine, INR, and sodium). Cox AMM-ULN and RF AMM-ULN performed better than their respective comparative models based on CP and MELD scores. The optimal predictive model was a RF model incorporating all variables. A further 2 models with reduced variables (RF Redux 1: AMM-ULN, INR, bilirubin, albumin, creatinine, sodium, and diabetes; RF Redux 2: AMM-ULN, INR, bilirubin, albumin, creatinine, and sodium) maintained an excellent predictive performance. AMM-ULN, ammonia level corrected to the upper limit of normal; CP, Child-Pugh; INR, international normalized ratio; MELD, model for end-stage liver disease; RF, random forest; ULN, upper limit of normal. (This figure appears in color on the web.) 0.26 0.27 0.28 0.29 0.30 Number of trees Error rate 100 x VIMP (time to complications) 010203040 AMM-ULN INR Bilirubin Albumin Creatinine Sodium 0 100 200 300 400 500 Fig. 4. Prognostic model using random forest to predict hospitalisation due to liver-related complications. Error rate for the selected random forest model according to number of trees (left panel) and subsampled forests to estimate standard errors and confidence intervals for VIMP (right panel). AMM-ULN represented the variable with the highest VIMP within the predictive model. AMM-ULN, ammonia level corrected to the upper limit of normal; INR, international normalized ratio; VIMP, variable importance. Journal of Hepatology 2022 vol. 77 j1554–1563 1561