Abstract

Background: Heterogeneity among patients with acute-on-chronic liver failure (ACLF) confer variable outcomes (mortality-range: 0-100%). While prognostic scores capture the known associations, machine learning (ML) can identify the intricate hidden patterns between patient characteristics without any explicit hypothesis or labelling that remain unexplored in ACLF. We employed ML to explore, describe, and predict unknown clusters in ACLF patients.

Methods: We applied unsupervised ML on the data of 1568 ACLF patients defined by APASL or EASL criteria, recruited ambispectively over 2015-2023 with a 90-day follow-up at a tertiary care centre. After initial processing and evaluation of cluster tendency, the data, including clinical details, investigations, and organ failures at day-0, 3, and 7 of admission, were subjected to distance, density, and model-based clustering. We interpreted a final model with least BIC and identified clusters through inferential statistics. Then, we explored the cluster associations with disease evolution and mortality. Finally, we employed supervised ML in 70% of data to train the models and 30% of remaining data to predict and interpret the clusters.

Results: We enrolled ACLF patients aged 44.3 (11.3) years, 87% males, 62.9% with alcoholic hepatitis, 29.6% with APASL, 15.5% with EASL, 54.9% with both APASL and EASL criteria, and a MELD of 29(7) with survival of 50.5%. Nonsurvivors were likely to have both EASL+APASL or EASL-ACLF, >1 acute precipitant, tense ascites, grade III-IV HE, infection, organ failures, and poor severity scores (p<0.001, each).

Of nine algorithms, the latent class model identified four distinct clusters. Cluster 1 vs. 2 (HR: 1.94), 1 vs.3 (HR: 2.23), and 1 vs.4 (HR: 4.12) were associated with mortality (Fig.1A), that remained significant after adjustment for CLIF-C ACLF or MELD or AARC score (p<0.001, each). Patient’s stratification into clusters could increase the discriminative ability of CLIF-C ACLF for mortality prediction by 11%.

Distinct clinical profiles, organ failures, and severity scores were attributed to the clusters. APASL-ACLFs were frequently noted (70%) in cluster1. Females, acute or chronic viral hepatitis was common and alcoholic hepatitis, infections, and tense ascites were less prevalent in cluster2. Clusters 1 to 4 had an increasing gradient of severity, organ failures, mortality, EASL+APASL ACLF, infections, >1 acute precipitant, alcoholic hepatitis, tense ascites, and grade III-IV HE (p<0.001).

While extreme gradient boost was best (AUC: 0.989), an interpretable decision tree could also predict clusters with very good AUC of 0.875 (Fig.1B).

Conclusion: ML for the first time could identify and precisely predict hidden clusters with distinct phenotypes and outcomes in ACLF. Stratification of patients into clusters can guide their prognosis, resource allocation, intensive care management, and liver transplant discussions in ACLF.