Objective:

To provide a visual explanation of the mathematical basis of machine learning models used in seizure detection devices.

Background:

Automating seizure detection is an area of active research. Recent technologies have employed machine learning models to develop seizure detection software. Development of such systems requires effective collaboration between epileptologists and data scientists. A visual explanation of the mathematical principles of machine learning could better equip physicians to effectively communicate their needs to data scientists.

Design/Methods:

We developed a pattern recognition machine learning model on simulated data. The simulated data consisted of five patterns. One of the patterns was arbitrarily declared pathological and we developed a software for detecting the “pathological pattern”. The flow of data through this model was then visually demonstrated with a parallel explanation of the underlying mathematics.

Results:

The software achieved a sensitivity of >97% and positive predictive value of >96 % with a 100% event detection rate. The results were validated across 20 runs of repeat experimentation with random re-generation of simulated dataset(s). As the data flowed through the model, the pathological pattern elicited responses starkly different from the responses to the normal patterns. This can be seen numerically by conducting statistical tests and visually by plotting graphs.

Conclusions:

Our work fills the need for demonstrations of machine learning principles developed for physicians, especially epileptologists.