Abstracts

Rationale: The apparent unpredictability of seizures leads to a sensation of loss of control and worsens the quality of life of patients and families. However, seizures actually tend to follow complex but repetitive rhythms. The objective of this study was to compare the performance of machine learning models for predicting the occurrence and the duration of future events captured with wearable sensors. Methods: We used the SmartMonitor database of deidentified events automatically detected by SmartWatch (https://smart-monitor.com/). The database consists of time and duration for each event, and raw motion data from the watch’s inertial measurement unit (IMU). We aimed to predict the time of occurrence and the duration of future events based only on the time of occurrence and duration of prior events; the raw motion data from the IMU was not included in our analysis. The goal was to describe the predictive power of common supervised machine learning techniques for predicting the time of event occurrence for event N and the duration of the event for event N based on time since first event for events up to event N-1 and event duration for events up to event N-1. Models were developed in the training set. The performance of the model was then evaluated by comparing the predicted with the actual outcomes for the test set. Results: We analyzed the first 20 events for each of the 525 patients in the database who met our inclusion and exclusion criteria. As gender was a voluntary field, many patients did not enter this information. Among those who entered gender, approximately two thirds were females. The median (p25-p75) interval between events was 1.04 (0.02-6.68) days. The median (p25-p75) duration of the events was 21 (11-57) seconds. For time of event occurrence and for event duration and for each model, the absolute difference between the model-predicted and the real value was calculated for each patient and each event resulting in a matrix of results with 156 rows (patients in the test set) and 11 columns (event number from 10 to 20). To summarize these results, we calculated the median, 25th and 75th percentiles for each column (each event from 10 to 20) and presented them in the Figures 1 and 2.  The best model for time of seizure occurrence was robust regression with a median absolute difference between model-predicted and real time of occurrence of 3.24 days followed by MARS with two-way interaction, MARS with three-way interaction, MARS with no interaction, boosted trees, linear regression, and random forest. The best model for seizure duration was robust regression followed by MARS with three-way interactions, boosted tree, MARS with two-way interaction, MARS with no interactions, random forest, and linear regression. Conclusions: In this proof of concept study we compare different machine learning algorithms for predicting the times and duration of future events, using only times and durations of past events. Our data show that, even in the absence of motion data, the time of occurrence and the duration of future events may be to a certain extent predictable. Funding: BCH team funded by the Epilepsy Research Fund.