Development of postoperative delirium prediction models in patients undergoing cardiovascular surgery using machine...

Area of Science:

• Cardiology
• Neurology
• Artificial Intelligence

Background:

• Postoperative delirium is a significant adverse event following cardiovascular surgery.
• Early identification of patients at high risk for delirium is crucial for implementing timely interventions.
• Current methods for predicting delirium risk in this population require enhancement.

Purpose of the Study:

• To develop and validate machine learning (ML) models for predicting delirium after cardiovascular surgery.
• To identify key preoperative predictors of post-cardiovascular surgery delirium.
• To compare the performance of various ML algorithms in delirium prediction.

Main Methods:

• Prospective enrollment of patients aged 40 years or older undergoing cardiovascular surgery.
• Assessment of preoperative and intraoperative factors, including demographics, cognitive function (Mini-Cog), functional status (Barthel Index), medical history (stroke, hemorrhage), and estimated glomerular filtration rate (eGFR).
• Development and validation of ML models (Bernoulli naive Bayes, SVM, Random Forest, Extra-trees, XGBoost) using stratified fivefold cross-validation.

Main Results:

• Out of 87 patients, 24 (27.6%) developed postoperative delirium.
• Key predictors identified: age, psychotropic drug use, low cognitive function, low activities of daily living index, history of stroke/hemorrhage, and reduced eGFR.
• The Extra-trees model achieved the highest Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.76 (SD 0.11), with 0.63 sensitivity and 0.78 specificity.
• XGBoost demonstrated the highest sensitivity (0.67) with an AUROC of 0.75 (SD 0.07) and 0.79 specificity.

Conclusions:

• Machine learning algorithms effectively predict the risk of postoperative delirium in cardiovascular surgery patients using preoperative data.
• These models offer a promising tool for identifying high-risk patients, enabling proactive management strategies.
• Further validation and implementation of these ML models can potentially reduce delirium incidence and improve surgical outcomes.