Machine learning for differentiating shock types using inflammatory markers

Cardiogenic shock (CS) and septic shock (SS) involve systemic inflammatory activation associated with hemodynamic instability and high mortality. However, the specific contributions of individual inflammatory components remain unclear. This study aimed to identify blood biomarkers and hemodynamic parameters in CS and SS patients to distinguish between these shock types and assess their prognostic potential, including the possibility of a mixed shock phenotype.

Data were analysed from the MARS project, a prospective observational study conducted at two tertiary university hospital intensive care units. The cohort included adult patients with either CS or SS, admitted from January 2011 to September 2013 and matched 1:1 by age and sex. CS was defined by ST-elevation myocardial infarction, persistent hypotension, and the need for vasopressors or mechanical circulatory support (MCS). SS was defined by infection within 24 hours of ICU admission, with hypotension requiring norepinephrine or MCS. Daily samples of inflammatory biomarkers and physiological (including hemodynamic) parameters were collected during the first four days following admission. XGBoost, a gradient boosting algorithm optimized for complex data patterns, was used to classify shock types. Logistic regression was used to analyse associations between the identified groups and 30-day mortality.

The study included 111 matched patients with cardiogenic shock (CS) and septic shock (SS), with approximately equal proportions in each group (CS: 55 patients, SS: 56 patients). The 30-day mortality rates were similar (SS: 43%, CS: 41%, p = 0.56). XGBoost performed grouping using a set of pro- and anti-inflammatory markers (interleukin [IL]-6, IL-10, IL-13, IL-1β) and physiological parameters (blood pressure, heart rate, respiratory rate), resulting in three patient groups: Predominantly CS (38%), predominantly SS (46%), and a mixed phenotype (16%). The mixed phenotype group had significantly elevated IL-10 (median 150 pg/mL) and IL-1β (median 6 pg/mL) compared to available normal range values1 (IL-10: 0–10 pg/mL; IL-1β: 0–5 pg/mL). The median IL-13 level in the mixed group was 2.5 pg/mL, which is higher than the predominantly CS group (median IL-13: 1.0 pg/mL) and the predominantly SS group (median IL-13: 1.2 pg/mL) and has no clear established normal range. The mixed phenotype was associated with significantly lower 30-day mortality compared to the predominantly SS group (odds ratio [OR] 0.25, 95% confidence interval [CI] 0.06–0.84, p = 0.029). The odds ratio for mortality for the predominantly CS group was 0.40 (95% CI 0.15–1.07, p = 0.067). Within the mixed group, patients with the highest combined levels of IL-10, IL-1β and IL-13 had a lower risk of 30-day mortality (OR 0.21, 95% CI 0.08–0.54, p = 0.002).

Machine learning analysis suggested a mixed shock phenotype, which was linked to lower mortality compared to patients with CS and SS.