Dear Editor,
We are grateful to Dr. Biswas and Dr. Roy for their comments regarding our original article titled “Predictive machine learning model in intensive care unit patients with acute-on-chronic liver failure and two or more organ failures” [
1] and for further elaborating on the important findings [
2]. Their insightful critiques not only highlight the strengths of our work but also thoughtfully underscore its limitations, many of which we fully acknowledge and are actively addressing in the future research.
Below, we respond to the key concerns raised:
Firstly, the data used in our study was sourced from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database, a publicly available database that documented de-identified electronic health records from intensive care units (ICUs) at the Beth Israel Deaconess Medical Center (BIDMC) between 2008–2019 [
3-
5]. We collected patient data from the ICU environment, and it is true that there are regional and institutional differences in admission criteria for ICUs in different regions. Our training cohort cannot represent all acute-on-chronic liver failure (ACLF) patients, and there is a certain selection bias. However, when we conducted external validation using a cohort of patients with ACLF from a tertiary university clinic in Germany with regional and institutional differences to the training cohort. Nevertheless, the model performed well in the external validation cohort, demonstrating its wide applicability and stability.
Secondly, our study explored the European Association for the Study of the Liver-Chronic Liver Failure (EASL-CLIF) definition [
6] and then North American Consortium for the Study of End-stage Liver Disease (NACSELD) definition [
7] cohorts of ACLF, but did not investigate the Asian Pacific Association for the Study of the Liver (APASL) definition [
8] cohort, which may prevent the results of the study from being generalizable to Southeast Asia or East Asia. This is because the MIMIC database we used in our study is a retrospective cohort from the United States where ICU-patients were included, and therefore the APASL criteria applying to less severe patients would inherit an unavoidable selection bias. Therefore, when establishing the model, we decided to use NACSELD guideline and definition. Given the significant similarities between the EASL-CLIF definition and the NACSELD definition, we also established an EASL-CLIF definition cohort for mortality risk prediction. In contrast, the APASL definition differs markedly from NACSELD definition, so we believe that using Asian patients, and especially not restricted to the ICU, to establish an APASL definition prediction model would be more appropriate. Promoting the standardization of ACLF definition is a long-term goal in hepatology, which is being currently followed by the major international societies of hepatology [
9]. In the future research, we will also introduce Asian cohorts and adopt the future harmonized definition of ACLF.
Thirdly, ACLF has a high short-term mortality rate and is a rapidly changing disease [
2,
10]. Clinical decisions are made based on changes in vital signs, liver function, lactate, consciousness status, and other parameters. The prognostic model of the study is based on baseline data from the ICU environment and can only predict the static mortality risk at each moment, but cannot reflect the dynamic nature of the disease. In the future, we will use longitudinal monitoring data and more complex prediction models such as deep learning to explore the causal relationship between disease changes and characteristic parameters, in order to capture the time-varying characteristics of ACLF disease and improve the effectiveness of disease prognosis management.
Fourthly, we excluded patients with previous liver transplant (LT) before the occurrence of ACLF, because the pathophysiology of ACLF between those with a native liver and a liver graft is substantially different. By excluding these patients, we were able to ensure that our study should not be affected by the confounding factor. Inclusion of patients who had undergone LT for ACLF, along with controls who did not, would help in identifying optimal transplant candidates as well as help develop timelines for transplant evaluation and referral. However, due to the small proportion of patients with LT in the MIMIC database ACLF used in the study, there are not enough cases to establish a control cohort. In the future studies, we will introduce specialized transplant cohorts to develop independent prediction tools, considering the impact of LT on patient prognosis and helping clinical doctors evaluate transplantation.
In summary, our research attempts to use machine learning methods for prognosis prediction and management of ACLF patients, while providing a prediction website that can determine the current risk of death in ACLF patients with only 12 parameters, demonstrating strong clinical applicability. In the future research, we will conduct in-depth analysis and exploration in dynamic prediction.
FOOTNOTES
-
Authors’ contribution
Study conception: M.Z., Y.H.Y., J.Z., J.T., F.J. Drafting of the manuscript: M.Z. Data interpretation and critical review of the manuscript: M.Z., Y.H.Y., J.Z., J.T., F.J.
-
Conflicts of Interest
Jonel Trebicka: Speaker and/or consulting fees: Versantis, Gore, Boehringer-Ingelheim, Falk, Grifols, Genfit and CSL Behring. Fanpu Ji: Speaker: Gilead Sciences, MSD and Ascletis. Consulting/advisory board: Gilead, MSD. All other authors do not have conflict of interest.
Abbreviations
acute-on-chronic liver failure
Asian Pacific Association for the Study of the Liver
Beth Israel Deaconess Medical Center
European Association for the Study of the Liver-Chronic Liver Failure
Medical Information Mart for Intensive Care-IV
North American Consortium for the Study of End-stage Liver Disease
REFERENCES
- 1. Yeo YH, Zhang M, McCoy MS, Zu J, He Y, Liu Y, et al. Predictive machine learning model in intensive care unit patients with acute-on-chronic liver failure and two or more organ failures. Clin Mol Hepatol 2025;31:1355-1371.
- 2. Biswas S, Roy A. Harmonized ACLF prognostication with explainable machine learning models: Traversing from counts to composition: Editorial on “Predictive machine learning model in intensive care unit patients with acute-on-chronic liver failure and two or more organ failures”. Clin Mol Hepatol 2026;32:1386-1389.
- 3. Johnson A, Bulgarelli L, Pollard T, Horng S, Celi L A, Mark R. MIMIC-IV (version 2.2). PhysioNet web site, <https://doi.org/10.13026/6mm1-ek67>. Accessed 1 Sep 2023.
- 4. Johnson AEW, Bulgarelli L, Shen L, Gayles A, Shammout A, Horng S, et al. MIMIC-IV, a freely accessible electronic health record dataset. Sci Data 2023;10:1.
- 5. Goldberger AL, Amaral LA, Glass L, Hausdorff JM, Ivanov PC, Mark RG, et al. PhysioBank, PhysioToolkit, and PhysioNet: components of a new research resource for complex physiologic signals. Circulation 2000;101:E215-E220.
- 6. European Association for the Study of the Liver. EASL Clinical Practice Guidelines on acute-on-chronic liver failure. J Hepatol 2023;79:461-491.
- 7. Bajaj JS, O’Leary JG, Reddy KR, Wong F, Biggins SW, Patton H, et al. Survival in infection-related acute-on-chronic liver failure is defined by extrahepatic organ failures. Hepatology 2014;60:250-256.
- 8. Sarin SK, Choudhury A, Sharma MK, Maiwall R, Al Mahtab M, Rahman S, et al. Acute-on-chronic liver failure: consensus recommendations of the Asian Pacific association for the study of the liver (APASL): an update. Hepatol Int 2019;13:353-390.
- 9. Luo J, Hu M, Feng T, Zhang L, Huang Y, Huang Y, et al. Performance of the China-CLIF framework in acute-on-chronic liver failure: a multicohort study across all aetiologies. Gut 2025;75:131-146.
- 10. Kim JH. Rethinking ACLF prognosis: Can machine learning outperform conventional scores?: Editorial on “Predictive machine learning model in intensive care unit patients with acute-on-chronic liver failure and two or more organ failures”. Clin Mol Hepatol 2026;32:1383-1385.
Citations
Citations to this article as recorded by
