After cardiac arrest and resuscitation, part of the patients will be in a coma and treated at an intensive care unit. Their prospects are uncertain. What is needed to get an outcome prediction that is reliable? Researchers of the University of Twente and the 'Medisch Spectrum Twente' hospital, both in Enschede, The Netherlands, developed a learning network that is capable of interpreting EEG-patterns. It can make a reliable outcome prediction, and thus forms a valuable extra source of information. The researcher present their approach in Critical Care Medicine journal.
In The Netherlands, about one third of the people that had a cardiac arrest followed by resuscitation, will have to be treated at the ICU. These patients, about 7000 each year, are in a coma. More than half of them will not regain consciousness. The family will want to know what the prospects are and, if their relative regains consciousness, what will be the quality of life. The question 'does further treatment make sense?' can only be answered after careful analysis of the situation. One of the options, now, is the SEPP-test; if an electrical signal applied to the wrist does not reach the brain, this is no good news.
The electrical signals of the brain, the EEG pattern measured via electrodes on the head, give a lot of information as well. Analysis of EEG using artificial intelligence gives a very accurate outcome prediction, as the researchers show in their latest paper. Twelve hours after resuscitation, the learning network is capable of predicting a good outcome with 58 percent accuracy and a bad outcome with 48 percent. This is a better performance than the trained eye of a neurologist. Both computer and human, however, still have a category 'I don't know', in situations the EEG data are not specific enough.
The first author, Marleen Tjepkema, already made a plea for using EEG in the outcome prediction, in her PhD thesis in 2014 as a UT Technical Medicine graduate. She and her colleagues now takes this an important step ahead by introducing automated interpretation of the EEG scan. The learning network has been trained using 600 EEG patterns, it did not get any hints on what to look at. After that, it was fed with 300 EEG patterns to see how it performed in giving a prediction. Neurologists have to look at hundreds of EEG's as well, as part of their training. An experienced neurologist will point out specific characteristics. Still, the EEG-patterns are so information-rich that the computer outperforms the human eye.
Once trained, the network will be capable of judging the EEG very fast, well within a second. The researchers expect that this adds valuable information to human judgment. One of the other advantages is flexibility, a prediction can be made any time of the day. Using the new technology at ICU's will show if the 'intensivist' also sees as a valuable tool.
One of the next steps in this research is having a closer look at the learning strategy of the network, making it more transparent than a black box approach. For this, the neurophysiologists collaborate with computer scientists and mathematicians of the University of Twente. Other examples of the use of deep learning in medicine are interpreting X-ray images or classifying skin injuries.
University of Twente
Tjepkemna, M. et al. (2019) Outcome Prediction in Postanoxic Coma With Deep Learning. Critical Care Medicine. doi.org/10.1097/CCM.0000000000003854.