Patients ‘lumped’ into the syndrome acute-on-chronic liver failure (ACLF) differ on several observable levels: underlying aetiology of cirrhosis, precipitants, number and severity of organ failures (OFs). Despite phenotypic heterogeneity, consensus definitions of ACLF have changed how the condition is understood, recognised and studied.

A key unmet need is identifying which elements of ACLF heterogeneity matter. There have been attempts to address this. For example, patients with Hepatitis B virus (HBV)-associated ACLF have distinct OF patterns and worse outcomes compared with non-HBV-associated ACLF [1]. Similarly, extrahepatic precipitating insults are associated with more extrahepatic OFs and worse outcomes, whereas hepatic insults predominantly cause liver and coagulation failure [2].

Verma et al. take a different approach, suggesting that latent or unobservable heterogeneity in the clinical characteristics of patients with ACLF—ACLF sub-phenotypes or clusters—may account for differences in patient trajectories and outcomes [3]. If ACLF sub-phenotypes exist, are identifiable, and sub-phenotypes respond differently to treatments, this approach could be used to match the right treatment to patient sub-phenotype: personalised medicine.

Using clinical data from a single-centre, Indian cohort of patients with ACLF, the authors tested several clustering algorithms. The number of ACLF clusters identified varied depending on the algorithm. The latent class analysis model was deemed most robust and identified four ACLF clusters with distinct survival profiles. Adding cluster assignment to the Chronic liver failure (CLIF-C) OF ACLF score improved prognostic accuracy. In a validation cohort, cluster membership could be predicted using a limited set of variables. This is a tentative step towards personalised ACLF management, but several unanswered questions remain.

First, are these clusters reproducible? Verma et al. [3] did not take a clinically informed approach to variable selection and included composite severity scores (derived from clinical data). Informed selection of variables and use of raw data may alter clustering results. Also, region-specific differences in ACLF aetiology and outcomes are well recognised [4], and external validation in a heterogenous global cohort is required.

Second, will absence of a universal consensus definition for ACLF hinder progress? There are several overlapping ACLF definitions [5-8], with important differences: nature of precipitating insult, compulsory inclusion of liver failure in ACLF diagnostic criteria, definitions and thresholds for organ failure. In this study, patients who met the EASL and/or APASL ACLF definitions were included. Sub-phenotypes will likely differ depending on definition; sub-phenotyping ACLF in variably defined populations could lead to confusion rather than clarity.

Third, and most important, does clinical heterogeneity represent biological heterogeneity? To deliver personalised medicine, we need to identify subgroups of ACLF patients with shared biological pathways, that is, endotypes. Studies to sub-phenotype sepsis and acute respiratory distress syndrome used biological data (‘omic’ data and biomarkers) to identify possible endotypes [9] and then determined if bedside identification of these endotypes—using clinical data—is feasible [10]. We do not know if the identified ACLF clusters capture underlying biological pathways of ACLF.

ACLF is complex and multifaceted. Current management is largely supportive with no disease-modifying treatments. There is an urgent need for biological and clinical phenotyping to develop targeted therapies.

被“归类”为慢加急性肝衰竭 （ACLF） 综合征的患者在几个可观察到的水平上有所不同：肝硬化的潜在病因、诱因、器官衰竭 （OF） 的数量和严重程度。尽管表型异质性，但 ACLF 的共识定义改变了人们对该病的理解、识别和研究方式。

一个关键的未满足需求是确定 ACLF 异质性的哪些要素很重要。已经有人尝试解决这个问题。例如，与非 HBV 相关 ACLF 相比，乙型肝炎病毒 （HBV） 相关 ACLF 患者具有不同的 OF 模式和更差的结局 [1]。同样，肝外诱发性损伤与更多的肝外 OF 和更差的结局相关，而肝脏损伤主要导致肝脏和凝血功能衰竭 [2]。

Verma 等人采取了不同的方法，认为 ACLF 患者临床特征的潜在或不可观察的异质性（ACLF 亚表型或集群）可能是患者轨迹和结果差异的原因 [3]。如果 ACLF 亚表型存在、可识别，并且亚表型对治疗的反应不同，则此方法可用于将正确的治疗与患者亚表型相匹配：个性化医疗。

使用来自印度 ACLF 患者单中心队列的临床数据，作者测试了几种聚类算法。识别的 ACLF 集群数量因算法而异。潜在类别分析模型被认为是最稳健的，并确定了四个具有不同生存特征的 ACLF 集群。在 ACLF 评分的慢性肝衰竭 （CLIF-C） 中增加聚类分配提高了预后准确性。在验证队列中，可以使用一组有限的变量来预测集群成员资格。这是迈向个性化 ACLF 管理的试探性一步，但仍然存在几个未解之谜。

首先，这些集群是否可重现？Verma 等 [3] 没有采用临床知情的变量选择方法，并包括综合严重程度评分（来自临床数据）。明智的变量选择和原始数据的使用可能会改变聚类结果。此外，ACLF 病因和结局的地区特异性差异已得到充分认可 [4]，需要在异质性全球队列中进行外部验证。

其次，ACLF 缺乏普遍共识定义会阻碍进步吗？ACLF有几种重叠的定义[5-8]，但存在重要差异：诱发损伤的性质、ACLF诊断标准中强制包含肝衰竭、器官衰竭的定义和阈值。在本研究中，纳入了符合 EASL 和/或 APASL ACLF 定义的患者。亚表型可能会因定义而异;在定义不同的人群中对 ACLF 进行亚表型分析可能会导致混淆而不是明确。

第三，也是最重要的，临床异质性是否代表生物学异质性？为了提供个性化医疗，我们需要确定具有共同生物途径（即内型）的 ACLF 患者亚组。脓毒症和急性呼吸窘迫综合征亚表型的研究使用生物数据（“组学”数据和生物标志物）来识别可能的内型 [9]，然后确定使用临床数据在床旁识别这些内型是否可行 [10]。我们不知道已鉴定的 ACLF 簇是否捕获了 ACLF 的潜在生物途径。

ACLF 是复杂且多方面的。目前的管理主要是支持性的，没有疾病缓解治疗。迫切需要生物和临床表型来开发靶向治疗。