Relationship prediction between clinical subtypes and prognosis of critically ill patients with cirrhosis based on unsupervised learning methods: A study from two critical care databases

Abstract

Background

Our objective was to identify distinct clinical subtypes among critically ill patients with cirrhosis and analyze the clinical features and prognosis of each subtype.

Methods

We extracted routine clinical data within 24 h of ICU admission from the MIMIC-IV database. To determine the number of clinical subtypes, we employed the “elbow method,” “cumulative distribution function (CDF) plot,” and “consensus matrix.” Consensus k-means, k-means, and SOM methods were used to identify different clinical subtypes of critically ill cirrhosis. We validated our findings using patients from the eICU database. The SHapley Additive exPlanations (SHAP) method was used to explore the features of each clinical subtype, and 28-day Kaplan-Meier curves were generated. Survival differences among the clinical subtypes were assessed using the log-rank test.

Results

Our study included 2,586 patients from the MIMIC-IV database and 1,670 patients from the eICU database. Based on the clinical routine variables, we identified three clinical subtypes among patients in the MIMIC-IV database. Subtype A (N = 1424, 55.07 %) was labeled the “common subtype” and exhibited the lowest mortality. Subtype B (N = 703, 27.18 %) was classified as the “hyperinflammatory response subtype” and had a relatively high mortality. Subtype C (N = 459, 17.75 %) was identified as the “liver dysfunction subtype” and had the highest mortality. These findings were consistent with the results obtained from both the internal validation set (MIMIC-IV database) and the external validation set (eICU database).

Conclusions

Our study presents a novel and clinically applicable approach for subtyping critically ill cirrhosis.