Supervised deep learning with vision transformer predicts delirium using limited lead EEG.

As many as 80% of critically ill patients develop delirium increasing the need for institutionalization and higher morbidity and mortality. Clinicians detect less than 40% of delirium when using a validated screening tool. EEG is the criterion standard but is resource intensive thus not feasible for widespread delirium monitoring. This study evaluated the use of limited-lead rapid-response EEG and supervised deep learning methods with vision transformer to predict delirium. This proof-of-concept study used a prospective design to evaluate use of supervised deep learning with vision transformer and a rapid-response EEG device for predicting delirium in mechanically ventilated critically ill older adults. Fifteen different models were analyzed. Using all available data, the vision transformer models provided 99.9%+ training and 97% testing accuracy across models. Vision transformer with rapid-response EEG is capable of predicting delirium. Such monitoring is feasible in critically ill older adults. Therefore, this method has strong potential for improving the accuracy of delirium detection, providing greater opportunity for individualized interventions. Such an approach may shorten hospital length of stay, increase discharge to home, decrease mortality, and reduce the financial burden associated with delirium.

Feature Selection for Unsupervised Machine Learning.

Compared to supervised machine learning (ML), the development of feature selection for unsupervised ML is far behind. To address this issue, the current research proposes a stepwise feature selection approach for clustering methods with a specification to the Gaussian mixture model (GMM) and the k-means. Rather than the existing GMM and k-means which are carried out based on all the features, the proposed method selects a subset of features to implement the two methods, respectively. The research finds that a better result can be obtained if the existing GMM and k-means methods are modified by nice initializations. Experiments based on Monte Carlo simulations show that the proposed method is more computationally efficient and the result is more accurate than the existing GMM and k-means methods based on all the features. The experiment based on a real-world dataset confirms this finding.