Uncovering distinct clinical phenotypes in disseminated intravascular coagulation through machine learning-enabled cluster analysis

TL;DR

This study uses machine learning to identify two distinct subtypes of disseminated intravascular coagulation (DIC) with different severity and outcomes.

Contribution

The novel use of unsupervised machine learning to stratify DIC patients into clinically meaningful subtypes.

Findings

Two distinct DIC subtypes were identified: mild and severe coagulation dysfunction.

The severe subtype was associated with significantly higher 7-day and 28-day mortality risks.

Subtypes showed differences in both model variables and other clinical parameters like heart rate and blood pressure.

Abstract

Disseminated intravascular coagulation (DIC) is a critical condition encountered in the intensive care unit (ICU), characterized by multiple etiologies and variable outcomes. Distinguishing between DIC phenotypes poses a significant challenge. This study aims to apply unsupervised machine learning (ML) algorithms to stratify DIC patients, thereby enabling more personalized treatment approaches.

We conducted a retrospective analysis of patients diagnosed with DIC upon admission to the ICU at a comprehensive teaching tertiary hospital in China, spanning from May 2015 to November 2022. We applied an unsupervised machine learning approach for consensus clustering using the R package Consensus Cluster Plus to identify clinical phenotypes in 134 patients with DIC. The analysis incorporated the key variables: Thrombin-Antithrombin Complex (TAT), Plasmin-α2-Plasmin Inhibitor Complex (PIC), tissue plasminogen activator-inhibitor complex (tPAIC), and thrombomodulin (TM). The elbow method, cumulative distribution function (CDF) plot, and consensus matrix were employed to ascertain the optimal number of clusters. Logistic regression (LR) analysis was used to investigate the association between the identified phenotypes and clinical endpoints.

The consensus cluster analysis delineated two distinct subtypes: a mild coagulation dysfunction subtype (n = 79) and a severe coagulation dysfunction subtype (n = 55). Notable differences were observed in both variables included in the analysis (e.g., thrombin-antithrombin complex [TAT], P < 0.05) and those not utilized for model training (e.g., heart rate [HR] P < 0.05 and systolic blood pressure [SBP] P < 0.05). Logistic regression revealed that the severe coagulation dysfunction subtype was significantly associated with increased odds of 7-day (OR 4.71; 95% CI 2.23–9.98; P < 0.001), 28-day (OR 2.29; 95% CI 1.11–4.72; P = 0.024).

The study identified two clusters with distinct laboratory profiles and mortality risk.

Figures