
Contact the VUNO team anytime.
APICS 2024
Objectives:
Perioperative stroke (PS), often leading to death or significant brain injury, is a devasting complication
of cardiothoracic and vascular surgery (CTVS). In this study, we newly defined surgery-related major
adverse cerebral events (sMACE) as PS and death. We then developed a deep-learning-based model
for early detection of sMACE in ICU-admitted patients following CTVS.
Methods:
Data were collected from CTVS patients admitted to the ICU between March 3, 2010, and March 4,
2021, at Pusan National University Yangsan Hospital (PNUYH) in Korea. We split the dataset into a
development set (2010-2017) and a validation set (2018-2021). Our interest is to detect sMACE, a
primary outcome defined as death and major PS, within 24 hours after ICU admission. Major PS
includes extensive strokes or hypoxic-ischemic brain damage with or without status epilepticus. Cases
with PS detected during presurgical brain imaging were excluded. Seven vital signs including the
Glasgow Coma Scale (GCS) score and 16 routinely done blood tests were utilized as input features for
early sMACE detection.
Results:
Out of 4,455 surgery cases, 218 (4.89%) experienced sMACE. Conventional scoring systems like the
modified early warning score (MEWS) and national early warning score (NEWS) achieved AUROC
scores of 0.729 (95% CI, 0.677-0.780). Machine learning algorithms such as Random Forest and Light
Gradient Boosting Machine (LightGBM) attained AUROC scores of 0.773 (95% CI, 0.726-0.818) and
0.773 (0.723-0.823), respectively. However, our model (Figure 1) significantly outperformed (p < 0.05)
these methods, yielding an AUROC of 0.826 (95% CI, 0.781-0.871).
Conclusions:
Early detection of surgical complications is challenging due to consistent management practices, e.g.,
ventilation and sedation, during the initial 24 hours of ICU admission. Our deep-learning-based early
detection system for sMACE demonstrated superior performance compared to conventional scoring
systems and other machine learning models, providing a promising approach to mitigate adverse
outcomes following CTVS.