Comparison of Machine Learning Models for Delirium Prediction in Patients Undergoing Aortocoronary Bypass Surgery

Background: Postoperative delirium is a prevalent and disabling mental disorder that occurs regularly in patients undergoing cardiac surgery. Delirium is associated with increased morbidity and mortality as well as a prolonged hospital stay. Identifying patients likely to develop delirium post-extubation at the earliest possible point in time may help to guide treatment decisions. This study compares machine learning models for the prediction of delirium in patients undergoing aorto coronary bypass surgery.

Method: In this prospective monocentric clinical trial, we included 256 patients who underwent coronary artery bypass grafting (CABG) between 2017 and 2018. Demographic, clinical, and hemodynamic data until extubation were used to develop different machine learning models. We modeled the postoperative occurrence of delirium in patients undergoing CABG surgery using multiple state-of-the-art machine learning models. The machine learning models used included decision trees, multilayer perceptrons, random forest, AdaBoost, support vector machine, logistic regression, and Bernoulli-naive Bayes. We refrained from oversampling. The training-to-test set ratio was 70:30. The test set area under the curve was used as metric to compare model performance. Patients were considered having delirium if either the 4AT or Confusion Assessment Method (CAM)-ICU test conducted within 7 days after extubation was positive and neuroleptic or sedative medication was required.

Results: A total of 256 patients were enrolled, of which 53 patients developed delirium as defined. Patients with delirium were significantly older (72.3 ± 7.9 vs. 66.4 ± 9.3 years, p = 0.002) and had a worse Logistic EuroSCORE (p < 0.001). In addition, the duration of ventilation was significantly longer (19.8 ± 14.4 vs. 13.0 ± 6.3 hours, p = 0.003) and these patients showed significant amount of fluid overload (p = 0.016). Our results indicate that the best performing models are based on random forest and Bernoulli-naive Bayes algorithm. The best classification performance was 0.83 for naive Bayes and 0.82 for random forest. The computing time of the trained models was no longer than 1.5 seconds per patient.

Conclusion: Our results indicate that models based on Bernoulli-naive Bayes and random forest are most suited for the prediction of postoperative delirium after CABG surgery. Due to the monocentric nature of this analysis, prospective validation in a separate cohort with large sample size is recommended.

Publikationsverlauf

Artikel online veröffentlicht:
03. Februar 2022