Using non-invasive wearable technology to forecast seizures in epilepsy is both feasible and imminent, according to an Australian-led research team.
The feasibility study, comprising 11 patients with refractory epilepsy, used machine learning and a neural network model to estimate seizure risk from smartphone seizure diaries and smartwatch data such as sleep, heart rate cycles and physical activity.
The study found hourly and daily forecasts performed better than chance when predicting seizures.
“Individual forecasters performed better than chance with all people when an hourly prediction horizon was used, and with 10 of 11 people when a daily prediction horizon was used,” the study said.
“These results indicate that non-invasive seizure forecasting is possible for people with epilepsy with seizure warning periods of up to 24 h.”
The study, published in Frontiers of Neurology, found the hourly forecaster resulted in more accurate predictions than the daily forecaster.
“The superior performance in the hourly forecasts may be attributed to a number of factors, such as the inclusion of circadian heart rate cycles, hourly step count and rate of change in heart rate.”
“The resolution of the daily forecaster would also have played a role in the loss of information. For example, high frequency seizure days (>1 seizure occurred on a day) were weighted equally to low seizure frequency days (1 seizure on a day).”
Cyclic features (heart rate cycles and previous seizure timing) were the strongest predictors of seizures in most cases while sleep features appeared to be useful predictors of seizure likelihood for some people using the daily forecaster but weak predictors in the hourly forecaster.
“Physical activity features were also predictive of seizures in some people, namely in the hourly forecaster.”
The researchers said smartwatches were not yet suitable substitutes for electrocardiography and polysomnography. As well, the number and frequency of seizures were limiting factors in any seizure forecasting.
However the investigators said they had shown that forecasting was possible.
“Prospective analysis and clinical trials should also be undertaken on longitudinal datasets in the future.”