Early prediction of low cardiac output syndrome in post-cardiotomy patients using supervised machine learning
J J Garate Togo, R Gopar-Nieto, V N Ramirez Frankiz, F Alvarez Ovando, A Arias-Mendoza, S Leon Delgado, D Manzur SandovalAbstract
Background
Post-cardiotomy low cardiac output syndrome (PC-LCOS) complicates ~10% of cardiac surgeries and is associated with increased mortality, prolonged ICU stay and greater need for mechanical circulatory support. It is characterised by refractory haemodynamic instability requiring inotropes, vasopressors and/or mechanical support, although diagnostic criteria remain heterogeneous. Current risk stratification relies on clinical assessment and perioperative scores with limited predictive performance. Supervised machine learning (ML) may improve prediction by integrating complex perioperative and early postoperative data. Although high discrimination (AUC >0.89) has been reported, most models are retrospective and single-centre, and externally validated tools for early PC-LCOS prediction are lacking.
Purpose
To develop and validate supervised ML models for early prediction of PC-LCOS in adults admitted to a postoperative cardiac ICU using routinely available perioperative and early ICU data.
Methods
This retrospective study included adults undergoing cardiac surgery at a tertiary centre and admitted to the postoperative cardiac ICU. PC-LCOS during ICU stay was the binary primary outcome, defined as clinician-diagnosed low cardiac output syndrome and/or escalation of inotropic or mechanical circulatory support beyond routine care. Demographic, perioperative, haemodynamic and laboratory variables available within an early postoperative window were analysed. Logistic regression, random forest and gradient boosting models were trained using random training/testing split with cross-validation. Performance was assessed by AUROC, sensitivity and specificity.
Results
Among 1,041 patients, 173 (16.6%) developed PC-LCOS. Affected patients had higher EuroSCORE II, longer cross-clamp and cardiopulmonary bypass times (all p < 0.001), and more frequent previous cardiac surgery, heart failure and atrial fibrillation (all p ≤ 0.016). Logistic regression achieved AUC 0.83 (95% CI 0.77–0.89), sensitivity 75% and specificity 81% (accuracy 80%, NPV 94%). Random forest showed similar discrimination (AUC 0.83, 95% CI 0.76–0.89) with higher sensitivity (81%), specificity (83%) and NPV (96%) (accuracy 82%). Gradient boosting yielded AUC 0.76 (95% CI 0.69–0.83). Key predictors were operative complexity (bypass and cross-clamp duration), EuroSCORE II and early haemodynamic markers (cardiac index, lactate, central venous pressure).
Conclusions
Supervised ML demonstrated good discrimination for early PC-LCOS prediction using routinely collected perioperative and ICU data. Random forest achieved comparable AUROC to logistic regression with improved sensitivity and NPV. Operative complexity and early haemodynamic instability were dominant predictors. ML-based tools may support early risk stratification in postoperative cardiac ICU populations.ROC Curves Comparison