Unsupervised machine learning to explore inflammation following cardiopulmonary bypass

Abstract

Introduction

Cardiac surgery with cardiopulmonary bypass (CPB) often induces systemic inflammatory reaction syndrome (SIRS), affecting postoperative outcome. We aimed to explore adaptive/maladaptive inflammation using unsupervised machine learning.

Methods

We conducted a post hoc analysis of 1908 adult patients who underwent elective cardiac surgery with CPB between June 2016 and June 2020 at a single institution. Patients were assessed for SIRS 12 hours post-surgery and clustered using the partitioning around medoids (PAM) algorithm based on Gower distance. The influence of SIRS on a composite outcome comprising death, stroke/TIA, renal replacement therapy, reoperation for bleeding, mechanical circulatory support, and ICU stay >96 hours was analyzed via multivariable logistic regression.

Results

SIRS occurred in 28.7% of patients (median age 69 years; 68.7% male). Clustering revealed two subgroups: maladaptive SIRS (52.9%) with higher preoperative risk and worse outcomes, and adaptive SIRS (47.1%) with favorable outcomes. Maladaptive SIRS patients had higher 30-day mortality (21.7% vs 1.6%, p < .001). Adaptive SIRS patients had outcomes similar to SIRS-negative controls. In selected clusters, SIRS was independently associated with a lower risk of the composite outcome (OR 0.44; 95% CI 0.26-0.74, p = .002).

Conclusion

Unsupervised machine learning effectively identifies adaptive and maladaptive SIRS in cardiac surgery patients, providing a basis for personalized postoperative care. Several clinical and procedural factors associated with maladaptive SIRS may be modifiable, supporting future precision strategies to reduce harmful inflammation after cardiac surgery.