Machine Learning-Based Prediction of Postoperative Vasoplegia After Cardiac Surgery.

OBJECTIVE

To develop and internally validate a machine-learning model for the early prediction of postoperative vasoplegia after cardiac surgery.

DESIGN

A retrospective cohort study with model development and internal validation.

SETTING

Tertiary academic intensive care units (ICUs) using routinely collected perioperative and postoperative clinical data.

PARTICIPANTS

Adults admitted to intensive care after cardiac surgery.

INTERVENTIONS: None.

MEASUREMENTS AND MAIN

RESULTS

Incident postoperative vasoplegia (greater than 6 to 48 or less hours after ICU arrival) was defined as either (1) concurrent cardiac index greater than 2.2 L/min/m², systemic vascular resistance less than 800 dyne·s/cm⁵, and mean arterial pressure less than 65 mmHg or (2) cardiac index greater than 2.2 L/min/m² with combined norepinephrine plus epinephrine infusion greater than 0.2 µg/kg/min. Candidate predictors included demographics, comorbidities and perioperative factors, admission vital signs and laboratory tests, and vasoactive infusions within 0-6 hours of ICU arrival. On the held-out test set, the best-performing model achieved moderate discrimination (area under the receiver operating characteristic curve 0.74-0.75) with acceptable calibration. Influential features included underlying hypertension, congestive heart failure, diabetes, male sex, and preoperative angiotensin-converting enzyme inhibitor exposure.

CONCLUSIONS

A machine-learning model using routinely available perioperative and early postoperative data accurately predicted postoperative vasoplegia after cardiac surgery. These findings support prospective evaluation for risk stratification and may inform targeted prevention and management strategies.