The relationship between clinical subtypes, prognosis, and treatment in ICU patients with acute cholangitis using unsupervised machine learning methods

Abstract

Background: Acute cholangitis (AC) presents with significant clinical heterogeneity, and existing severity classifications have limited prognostic value in critically ill patients. Subtypes of AC in critically ill patients have not been investigated.

Objective: The study aimed to offer a novel approach to identify clinical subtypes and improve individualized risk assessment and treatment strategies using an unsupervised analysis.

Methods: We conducted a retrospective analysis of ICU patients with AC from the Medical Information Mart for Intensive Care-IV (MIMIC-IV) database. K-means clustering was applied to 24 routinely available clinical variables from the first 24 h of ICU admission to identify clinical subtypes. The primary outcome was 28-day all-cause mortality. Multivariable Cox regression was used to assess associations between subtypes, mortality, and biliary drainage strategies. Furthermore, a simplified model using the top five SHapley Additive exPlanations (SHAP) ranked variables (hemoglobin, RDW, albumin, CCI, and sodium) was developed and subsequently validated in an external cohort from MIMIC-III.

Results: Two distinct subtypes were identified: a mild-inflammatory subtype (n = 451) and an inflammatory-dysfunctional subtype (n = 348). The inflammatory-dysfunctional subtype was characterized by higher levels of RDW, ALP, bilirubin, creatinine, and coagulopathy markers, and exhibited significantly higher 28-day mortality (30.17% vs. 5.32%, p < 0.001). This subtype remained an independent predictor of mortality after multivariable adjustment (HR = 2.38,95%CI:1.42-3.9, p = 0.001). Within the inflammatory-dysfunctional subtype, ERCP was associated with lower mortality (HR = 0.56, p = 0.01), whereas PTCD was associated with higher mortality (HR = 1.63, p = 0.031), potentially reflecting underlying disease severity. The simplified model retained strong discriminative performance (AUC = 0.87) and was successfully validated in the external cohort, confirming the reproducibility and prognostic relevance of the subtypes.

Conclusions: We identified and externally validated two clinically meaningful AC subtypes with distinct prognoses. A simplified model using five readily available variables facilitates clinical application and can support more individualized treatment approaches. These data-driven subtypes offer additional prognostic discrimination beyond the Tokyo Guidelines, serving as a valuable complement to existing severity classifications for guiding precision management of ICU patients with AC.

Keywords: Acute cholangitis; K-means clustering; Prognosis; Subtype.

Fig. 1

The study flowchart

Fig. 2

Using PCA, t-SNE, and UMAP methods to visualize the subtypes of AC patients. A PCA, B t-SNE, C UMAP. Circles represent individual patients (blue represents Mild-inflammatory subtype, yellow represents Inflammatory-dysfunctional subtype)

Fig. 3

28-day survival curves for each clinical subtype in MIMIC-IV patients

Fig. 4

Prognostic value of Tokyo Guidelines severity and machine learning subtypes in AC. A Kaplan-Meier curves showing 28-day survival by Tokyo Guidelines 2018 severity grade. B 28-day mortality comparison between the two subtypes within each severity grade. ***p < 0.001

Fig. 5

ROC curve for the establishment of predictive model (A) and The SHAP summary plot for the inflammatory-dysfunctional subtypes of critically ill patients with AC (B)

Fig. 6

Performance and validation of the simplified model using five key variables. A ROC curve showing the discriminative ability of the simplified clustering model; B SHAP summary plot for the inflammatory-dysfunctional subtype; C 28-day Survival Curves for Each simplified clustering Subtype in MIMIC-III patients. D Forest plot of multivariable Cox regression analysis 28-day mortality in MIMIC-III patients