Delirium, a serious yet often underrecognized condition, continues to exert a significant burden on patients and the healthcare system. Characterized by a sudden onset of confusion, disorientation, and changes in cognition, delirium can arise in various medical settings, including hospitals, long-term care facilities, and even at home. Its impact on individuals' well-being, coupled with the challenges it presents to healthcare providers, necessitates a deeper understanding and increased attention to this debilitating condition.
It is a widespread problem, with as many as 80% of critically ill patients developing the condition. Yet, it has been estimated that only about 40% of delirium cases are detected using the traditional screening methods. It is important that all cases be detected so that treatment can be initiated, as delirium increases the need for institutionalization and results in higher morbidity and mortality rates.
Despite the fact that there are dozens of validated screening methods, less than 10% of clinicians report regularly screening their patients for delirium because the process tends to be resource intensive. Moreover, not all patients can participate in the screening procedures, due to being in a comatose or deeply sedated state, or due to other medical issues.
The electroencephalogram (EEG) has been proven to be a useful tool in diagnosing delirium, but due to the expertise required for making a diagnosis from this data, it is rarely used. A team led by a researcher at the University of South Carolina realized that the measurement equipment is a valuable diagnostic tool, but the interpretation of the results needs to be automated for it to be applied widely. They showed that this was possible by developing a supervised deep learning algorithm that can diagnose delirium with a very high degree of accuracy.
A vision transformer model with a Transformer architecture was designed to read in 10-electrode rapid response EEG measurements and predict the likelihood that they correspond with a diagnosis of delirium. A publicly available dataset was leveraged to train the model to make the association between the data and the prediction. On average, the model was found to make the correct diagnosis in 86.33% of cases.
These results were validated in a clinical study involving thirteen participants, seven of whom had delirium. In these real world tests, the method yielded a 99.9% average training accuracy, and a 97% accuracy level with the test dataset. While these results are very impressive, it is important to note that the tests were conducted on a very small cohort. Further validation will be need on larger groups of patients in the future.
The rapid response EEG machine used in this pilot study is easy to use, and accessible to many members of a hospital’s staff with minimal training. The specialized knowledge needed to operate a traditional, large EEG machine, and to place the electrodes properly, is no longer a limiting factor with this new technique. The system also eliminates the need for highly skilled interpreters, which makes it feasible to monitor critically ill older adults across medical, surgical, and cardiac ICUs.