Society: AASLD

Background: Due to development of an immune dysregulated phenotype, advanced liver disease in all its forms predisposes to sepsis acquisition and related mortality. Immunodeficiency develops despite immune activation, predisposing patients to opportunistic infections, including invasive fungal infections. Despite high mortality, little data exists on fungal infection within a medical liver intensive unit (MILU). Specifically, data is lacking describing such infection in acute on chronic liver failure.
We aimed to characterize epidemiology, baseline factors and outcomes of fungal infections in our MILU.

Methods: From our prospective registry of patients admitted to our MILU, we identified 17 fungal and 183 bacterial culture-positive infections between August 2018-September 2022. Data on baseline comorbidities, acute characteristics of infection and clinical outcomes was manually extracted. Univariate analysis was used to compare fungal infections to bacterial counterparts.

Results: Patients with fungal infections were younger than bacterial counterparts (mean age 51.5±14.2 vs 58.4±11.6, p=0.024, table 1). Etiology of liver disease in fungal cases included cirrhosis, acute liver failure (11.6%) and Byler's disease post-transplant (5.8%); five patients had acute surgical illnesses including pancreatitis, cholecystitis and perforated viscus. Patients with fungal infections more frequently had hepatorenal syndrome, hepatic encephalopathy, thrombocytopenia, and malnutrition as complications of underlying liver disease (p=0.012-0.048, table 1). Fungal cases were associated with higher bilirubin (16.8±10.5 vs 9.4±10.1, p=0.05) and presence of central lines (75.5% vs 39.7%, p=0.003). Most patients with fungal infections had acute on chronic liver failure (ACLF) grade 3 compared to bacterial (70.6% vs 34.8%, p=0.029). MELD-Na and Child Pugh scores at infection were higher in case of fungal vs bacterial infection. Mean admission MELD-Na for fungal cases was 29.6±7.24, and medial survival time was 1 week relative to 4 weeks for bacterial infections (p<0.0001, fig 1). More patients with fungal infections required pressors (100% vs 72%, p=0.011) and intubation (95.2% vs 66.3%, p=0.013); 12 (71%) of patients had received at least 5 days of antibiotics prior to positive fungal culture.

Conclusion:
All patients with fungal infections passed away within 2 weeks, reflecting 100% mortality. Majority of cases of fungal infection also had severe ACLF, likely accounting for the significantly worse survival relative to bacterial counterparts. Further prospective studies examining timing of early antifungal therapy and empiric antifungal coverage are critical. Our findings suggest antifungal coverage should be considered in patients with ACLF-3 who fail to improve on antibiotic therapy.

Table 1: Comparison of Baseline Variables and Acute Characteristics Between Fungal and Bacterial Infection.

Table 1: Comparison of Baseline Variables and Acute Characteristics Between Fungal and Bacterial Infection.

Figure 1. Comparison in survival from time of ICU admission between patients with fungal vs bacterial illness, as demonstrated in this Kaplan Meier curve.

Figure 1. Comparison in survival from time of ICU admission between patients with fungal vs bacterial illness, as demonstrated in this Kaplan Meier curve.

Introduction:
Studies show that fungal infections in cirrhotics have increased mortality, but early detection and treatment has shown mortality benefit. Cirrhotic patients with candidemia are more likely to develop acute on chronic liver failure(ACLF) and experience mortality. In studies of candidemia, the sensitivity of cultures was 21%-71%. Given the increase in mortality of fungal infections, it is vital to have a reliable test in those with cirrhosis. Our study evaluates mortality of early detection, MELD score, and organ dysfunction in cirrhotic patients positive for T2 panel.

Methods:
We performed a retrospective study and included cirrhotic patients admitted from 2017-2021. Data collected includes baseline characteristics, type of cirrhosis, MELD score, mortality outcomes with early T2 detection, INR, bilirubin(Bili), creatinine(Cr), hepatic encephalopathy(HE), mean arterial pressure(MAP), and PaO2/FiO2(P/F)ratio. Factors such as HE, MAPs, and P/F ratio were graded based on CLIF-C ACLF criteria(Table 1). Univariate and multivariate regression analysis was performed for these outcomes.

Results:
Our cohort(n=734) had a mean age 56.13 ± 11.81, BMI 29.77 ± 8.91, and male gender of 52%. Majority of the cohort had NASH (29%) or alcoholic(39%) cirrhosis. Out of 734 cohort, 94 (12.8%) had a positive T2 with the majority being NASH(25%) and alcohol(37%) cirrhosis. Mean number of days between admission and positive T2 in the expired group was 13.20 ± 17.45 vs 8.53 ± 10.24 in the alive group(p-value of 0.1152)(Table 2). Admission MELD of 23.47 ± 9.65 changed to MELD of 26.52 ± 9.81 when T2 turned positive(p-value 0.0001)(Table 3). Positive T2 expired vs positive T2 alive cohort shows an increase in the following: Bili by 3.32 ± 8.70 vs 0.39 ± 2.17(p-value 0.0323), INR by 0.71 ± 1.17 vs -0.10 ± 0.52(p-value 0.0001), Cr by 0.31 ± 1.13 vs -0.62 ± 2.62 (p-value of 0.0374), HE grade by 1.04 ± 1.10 vs 0.62 ± 0.91 (p-value 0.0461), MAP rating by 1.23 ± 0.88 vs 0.57 ± 0.77 (p-value 0.0002), and P/F ratio rating by 1.08 ± 0.97 vs 0.67 ± 0.82(p-value 0.0282)(Table 4).

Conclusion:
Fungal infections play a large role in mortality in cirrhotic patients, and early detection and early treatment have mortality benefits. Positive T2 panels were more common in those with NASH and alcoholic cirrhosis. Positive T2 patients had an elevation of MELD from admission to positive T2. In expired T2 positive patients, there was an elevation of bilirubin, INR, creatinine, HE grade, P/F ratio rating, and MAP rating. This shows that once a cirrhotic patient had candidemia there was a high chance of mortality and development of ACLF. This study shows the diagnostic value of obtaining T2 in cirrhotic patients especially those with worsening organ dysfunction. Other factors such as earlier detection of fungemia and detection time of T2 positivity mortality were not statistically significant.

Background: The Neutrophil-to-lymphocyte ratio (NLR) is a simplified tool to assess status of bacterial infection in hepatitis B virus (HBV)-related acute-on-chronic liver failure (ACLF) patients. Serum lactate is one of key predictors of mortality in ACLF patients despite it is not routinely tested. We aimed to evaluate the prognostic value of NLR in predicting the short-term prognosis of patients with any cause of acute decompensation (AD) and ACLF particularly compared to serum lactate.
Methods: A prospective cohort of adult ACLF Thailand multicenter (9 tertiary hospitals) enrolled the acute decompensation and ACLF patients who were diagnosed using EASL-CLIF criteria between January 2015 to June 2020. The primary outcome was death during follow-up. The area under the receiver operating characteristic (AUROC) curves was used to evaluate the accuracy of models. The survivals of patients were classified as NLR grade compared using the Log Rank test.
Results: Of 602 patients, 301 (50.0%) patients were ACLF, 396 (65.8%) were male, with mean age of 58.5±14.2 years. The majority etiology of chronic liver disease was alcohol (242 patients, 40.2%), followed by hepatitis C virus (119, 19.8%), HBV (112, 18.6%), and non-alcoholic steatohepatitis (55, 9.1%). Of those, 300 (49.8%) patients had a bacterial infection at the date of enrollment. Patients with bacterial infection had significantly higher NLR than those without bacterial infection (13.4±21.8 vs 8.3±11.2; p<0.001). At the endpoint of 30 days and 90 days after diagnosis, 250 (41.5%) and 316 (52.5%) patients deceased, respectively. The AUROC (95% CI) of NLR and lactate in predicting 90-day mortality were 0.75 (0.69-0.80); p<0.001 and 0.57 (0.51-0.63); p=0.031, respectively (Figure). NLR was an independent associated with 90-day mortality, and was categorized into three risk grades (grade 1; <3.10, grade 2; 3.10–4.78, and grade 3;>4.78) with 90-day cumulative mortalities of 22.8%, 44.2%, and 66.4%, respectively; p<0.001. Patients with NLR grade 1 and 2 had significantly longer survival than those with NLR grade 3 (>90 vs >90 vs 29.0 (26.0-33.9) days; p<0.001). In multivariable analysis, NLR grade 2 and 3 were independently associated with 90-day mortality with adjusted HRs (95%CI) of 2.25 (1.41-3.60), and 3.38 (2.26-5.07); p=0.001 (both groups), respectively. When combination of NLR >4.78 and lactate ≥4 mmol/L, HR provided additively increased to 3.10 (2.46-3.90), compared to HRs of 2.58 (2.00-3.32) and 1.33 (1.02-1.73) for NLR >4.78 and lactate ≥4 mmol/L; p<0.001, respectively (Table).
Conclusion: Our study demonstrated that NLR reflects the presence of bacterial infections in AD/ACLF patients. For short-term mortality prediction, NLR has a better predictive value than serum lactate. NLR could be a facilitating tool in primary hospitals to manage and provide prognostic guidance for AD/ACLF patients.

<b>Figure.</b> The AUROCs curve of NLR and lactate for 90-day mortality prediction.

Figure. The AUROCs curve of NLR and lactate for 90-day mortality prediction.

<b>Table</b>. Univariate and multivariate Cox regression analysis of factors associated with 90-day mortality in AD and ACLF patients.

Table. Univariate and multivariate Cox regression analysis of factors associated with 90-day mortality in AD and ACLF patients.

Patients with cirrhosis have high comorbidity burdens and complex medical regimens, placing them at risk for polypharmacy. Polypharmacy can cause unintended drug interactions and adverse events, and vulnerable patients with cirrhosis may be susceptible. This study aims to examine polypharmacy in patients with cirrhosis and the association between polypharmacy and overt hepatic encephalopathy (HE).

We conducted a retrospective cohort study of patients with cirrhosis seen in the ambulatory hepatology clinics at Indiana University from 1/1/2019-12/31/2021. Demographic, liver disease characteristics, and medication exposure was collected at the time of index visit. Medication burden was examined as the number of medications with at least a 30-day exposure. They were followed until transplant, death, or administrative censoring. The primary outcome was HE hospitalizations, with death and liver transplant as competing risks. To examine the relationship between medication burden and HE, a cause-specific Cox proportional hazards model was used.

1,039 patients with cirrhosis were included, with a median follow up time of 840 days. 270 (26%) were taking <5 chronic medications, 444 (43%) were taking 5-9 medications, and 325 (31%) were taking ≥10 medications. The mean age of the cohort was 59 years old with 47% female. Non-alcoholic steatohepatitis (NASH) was the most common etiology at 37%, and 56% had Child-Pugh class B or C cirrhosis. 37% of the cohort had a history of HE at baseline. Further characteristics of the medication burden groups are demonstrated in the Table. Older patients and women were more likely to experience polypharmacy as were those with NASH. Polypharmacy was associated with greater MELD-Na, Child-Pugh score, and Charlson Comorbidly Index. The prevalence of HE at baseline was higher in those with polypharmacy. Those with polypharmacy were more likely to require hospitalization or experience death. Accounting for competing risks of liver transplant and death, the cumulative incidences of HE hospitalizations between the different medication burden groups is shown in the Figure, with greater medication burden associated with increased HE. Those taking ≥5 medications had an increase in incidence of hospitalizations for HE (HR 2.35, 95% CI 1.34-4.14). In a multivariable analysis adjusting for age, gender, and baseline MELD-Na, when compared no polypharmacy (0-4 meds) there was increasing risk for HE hospitalizations with increasing medication burden, with individuals on 5-9 medications HR 1.67 (95% CI 0.88-3.16) and on 10+ medications HR 2.831 (95% CI 1.50-5.34).

Patients with cirrhosis frequently experience polypharmacy. Greater medication burden and polypharmacy is associated with an increased risk for HE related hospitalization. Interventions focusing on reducing polypharmacy and medication burden in those with cirrhosis may improve outcomes.

Background:
ACLF is associated with up to 80% mortality at one-month. LT is a well-established treatment for ACLF but up to 40% mortality post-LT have been reported. Existing post-LT prediction models in ACLF have been limited by small derivation samples. We developed a novel machine learning (ML) model incorporating ACLF physician expert input to predict post-LT outcomes in ACLF.

Methods:
We identified all adults with ACLF who underwent LT from 2013-2015 in the University of California Health Data Warehouse, a de-identified electronic health record (EHR) database from five academic medical centers. We used the rtemis R package to train an EAML model to predict post-LT mortality at one-year. EAML applies RuleFit, an ML algorithm that extract rules from decision-trees, to ACLF patients’ EHR data to extract rules that are then evaluated on 5-point Likert scales by 15 physician experts who conduct clinical care and research in ACLF. We used AUROC statistics to compare the performance of EAML and RuleFit models versus other cross-sectional models used in ACLF care and research.

Results:
We identified 1,384 patients with ACLF and divided them into 60:20:20 train (N=841), validation (N=255), and test (N=288) sets. The three sets were similar in terms of age, race/ethnicity, liver disease etiologies, and comorbid conditions. Median MELDNa scores were 34 (IQR29-39), 34 (IQR30-38), and 34 (IQR30-38) in the train, validation, and test sets, respectively. 10% (87) in the train, 11% (28) in the validation, and 13% (34) in the test set died at one-year. We ultimately extracted 139 data features per patient.

After training the algorithm on the train set, tuning hyperparameters on the validation set, RuleFit generated 20 rules. These rules contained two data features on average, and largely consisted of rules governing laboratory values (e.g. serum glucose ≤ 137 AND serum sodium > 143). The greatest discrepancies between human experts and ML occurred in rankings of biomarkers more commonly utilized in clinical practice, such as age and MELDNa score. Across the test set, AUROCs were 0.707 (CI0.625-0.793) for EAML and 0.719 (CI0.640-0.800) for RuleFit models. For cross-sectional models, AUROC values were 0.496 (CI0.455-0.561) for MELDNa, 0.581 (CI0.507-0.664) for NACSELD, 0.407 (CI0.358-0.456) for CLIF-ACLF, and 0.388 (CI0.298-0.480) for TAM3 across all patients (see Figure).

Conclusions:
We found both ML models without (RuleFit) and with ACLF expert input (EAML) outperformed cross-sectional models used in ACLF care and research in assessing one-year mortality. Discrepant data features selected by human experts and ML may reflect current biases in ACLF care and indicate potential areas for further exploration.

Mark Pletcher

Appearances

PREDICTING POST-LIVER TRANSPLANT (LT) OUTCOMES IN ACUTE-ON-CHRONIC LIVER FAILURE (ACLF) PATIENTS USING EXPERT-AUGMENTED MACHINE LEARNING (EAML)

Presenter

Speakers