Neuroscience and Neurology

Using artificial intelligence to predict chances of recovery from a coma

Dr. Florence Aellen

University of Bern
Inselspital, Bern University Hospital

Prof. Dr. Athina Tzovara

University of Bern
Inselspital, Bern University Hospital

Most survivors of cardiac arrest (CA) are initially in a coma. Currently used outcome prediction techniques mainly rely on expert assessment of clinical variables and physiological measurements such as electroencephalography (EEG). Assessing neural functions of coma patients and predicting chances for recovery still remain challenging. Notably, the existing markers for prognostication leave up to one third of patients with uncertain prognosis, in a ‘grey zone’.

Therefore, Florence Aellen, Athina Tzovara, and their colleagues challenged the traditional approach for predicting coma outcome and used artificial intelligence to assess the integrity of neural functions in coma and chances of recovery. They utilized state-of-the-art deep learning algorithms applied to coma patients’ EEG responses to sound stimuli.

During the first 24 hours of coma following cardiac arrest, 134 comatose patients in the intensive care units of four different Swiss hospitals were presented with sound stimuli via headphones. The researchers then trained deep learning algorithms to predict whether a given patient would survive the coma three months later based on EEG responses to the sounds. 

The analysis showed that neural responses to sounds in combination with deep neural networks can indeed be used to predict a patient’s chances of awakening from coma. Crucially, outcome prediction was at similar levels in a cohort of 48 ‘grey zone’ patients, whose outcome would be indeterminate based on existing clinical tests. Moreover, the confidence of the neural network in predicting patients’ outcome was reflecting interpretable properties of EEG signals. 

The researchers show, for the first time, systematic evidence that a deep learning framework can disentangle auditory processing in coma patients and assist in prognosticating their chances to recover. This work not only provides novel insights on neural functions that are preserved without consciousness but may further have implications for the field of neuro-critical care and outcome prognostication.

Auditory stimulation and deep learning predict awakening from coma after cardiac arrest. Florence M. Aellen, Sigurd L Alnes, Fabian Loosli, Andrea O. Rossetti, Frédéric Zubler, Marzia De Lucia, Athina Tzovara. Brain. 2023 Feb 13;146(2):778-788