Weathering the Storm: Forecasting Patient Recovery after Coma
From checking the weather forecast for incoming rain, to checking traffic conditions to get to work on time, we base so many decisions in our life around our predictions for the future. Usually, more information leads to better accuracy when it comes to making daily decisions adapted to the forecasts. Uncertainty and poor predictions can leave you in an unexpected rainstorm without an umbrella. When confronting much more serious situations however, such as caring for a loved one with a life-threatening injury, accurate and reliable predictions become essential. Healthcare decisions are often made around the likelihood and quality of a patient’s recovery. Uncertainty in predicting that recovery risks far more severe consequences than simply getting caught in the rain. Unfortunately, for patients with severe brain injury, making predictions on their recovery is a very difficult task.
For a patient with severe brain injury, accurate prediction on their chance of recovery can make the difference between life and death. Critical brain injury can result in coma, a long-term state of unconsciousness. Coma patients are unresponsive and often need to be kept on life-support to survive. For many, making the decision to continue life-sustaining measures for a loved one depends on the predictions made about their expected recovery. But these predictions are often uncertain, and recovery can vary greatly between patients. Some coma patients recover exceptionally well, while others may never regain full consciousness. There are few known clinical signs to predict how well a patient will recover[1].
Since predicting recovery after coma is such a difficult task, researchers have been searching for better measures that could lead to more accurate predictions. At Western University’s Brain and Mind Institude, Owen Lab researchers Matthew Kolisnyk and Dr. Karnig Kazazian have been looking into neuroimaging. They suggest using resting-state brain activity to predict patient outcome[2]. Resting-state functional magnetic resonance imaging (fMRI) measures the activity of the resting brain, even in unconscious patients. It's a good choice of brain activity to measure in coma patients, as it doesn't require them to respond in any way. However, even the resting brain is amazingly complex. Although coma patients show different patterns of brain activity, it can be hard to tell which are related to a strong recovery and which are not. Trying to pick predictors of recovery out from all the other work the brain does can be like looking for a needle in a haystack. To solve this problem, Kolisnyk and Kazazian used the help of AI. They chose to use machine learning, a type of AI that excels at finding new patterns in large datasets – perfect for a study looking to predict coma recovery from brain activity.
The study was conducted at the London Health Sciences Centre in London, Ontario. Twenty-five coma patients completed a five-minute resting-state fMRI scan within one month of sustaining severe brain injury. A machine learning model then used the brain activity to create predictions of recovery. For each participant, the model ran several hundred simulations. Each simulation predicted the chance of good or poor recovery after a 6-month period. From all of these calculations, the machine learning model answered two important questions: whether participant recovery was expected to be good or poor, and how confident it was about its prediction. Confidence is extremely important when it comes to making reliable predictions. It makes the difference between being 95% certain a patient will have a good recovery and being only 60% certain.
Six months after patients had their brain activity measured, their actual recovery outcome was evaluated clinically. Once their true recovery progress was known, the researchers compared the predictions made by the machine learning algorithm to the actual outcomes of the patients. What they found was promising. The algorithm successfully predicted patient outcome from resting-state brain activity with a high degree of accuracy. It correctly predicted a good outcome for 8 out of the 10 patients who recovered well, and correctly predicted a poor outcome for 12 of the 15 patients who had not shown much progress.
Importantly, the machine learning model had a high level of confidence in its predictions as well. This means that brain activity measured only a few days after injury can already be used to create recovery predictions that are unlikely to change. The algorithm was not perfect, but it was far more accurate and confident than predictions made using traditional clinical measures. This is a very positive outcome to help better the treatment of patients with severe brain injury.
What brain activity did the algorithm use to make its predictions? As of right now, the research team isn’t exactly certain, but they did note that the machine learning model considered visual brain areas the most when making its predictions. This area of study is still new, and it will take much more work to fully understand how visual brain activity promotes the recovery of consciousness in coma patients. It also remains to be seen if machine learning models can create similarly accurate predictions using different types of brain activity, as some patients may be unfit for fMRI.
The work by Kolisnyk and Kazazian in The Owen Lab is a huge step forward which will help improve the care of patients of severe brain injury. Patients who are unconscious and unresponsive are also unable to make decisions regarding their own healthcare. They must rely on others making choices for their care based on the information they have. Machine learning and resting-state fMRI is shaping up to be a reliable tool for predicting coma recovery. With more accurate predictions to work from, health providers and caregivers can have more confidence they are making decisions that best suit the wishes of their patients.
Original Article [2]:
Kolisnyk, M., Kazazian, K., Rego, K., Novi, S. L., Wild, C. J., Gofton, T. E., Debicki, D. B., Owen, A. M., & Norton, L. (2023). Predicting neurologic recovery after severe acute brain injury using resting-state networks. Journal of Neurology, 270(12), 6071–6080. https://doi.org/10.1007/s00415-023-11941-6
Additional source [1]:
Weijer, C., Bruni, T., Gofton, T., Young, G. B., Norton, L., Peterson, A., & Owen, A. M. (2016). Ethical considerations in functional magnetic resonance imaging research in acutely comatose patients. Brain, 139(1), 292–299. https://doi.org/10.1093/brain/awv272