Biomarker Discovery

The leading provider of In-Home Video EEG Testing in the United States hired Positronic in a consultative capacity to build a proprietary software application (unrelated to LIT) to perform feature extraction and training of deep learning models capable of detecting subtle differences between EEGs obtained in different clinical states for the creation of potential biomarkers of drugs and disease states.

 

DATA

The platform allows the user to select or upload a batch of EEG files. The platform provides controls for general file management functions to keep the ever-growing datasets manageable. Each sample is converted into a format readable by the pipeline.

CHANNELS

The platform provides a view that allows the user to specify which channels will be analyzed from each sample and what preprocessing should be applied to the incoming data; i.e. filtering, downsampling, normalization. The preprocessing options are modular & extensible.

FEATURES

The platform provides a view that allows the user to choose features to be extracted from the output of the Channels processing. The user may select one or more channels as input and apply one or more feature transformations to it; e.g. Power Spectral Analysis, Higuchi Fractal Dimension. Again, these transformations must be modular & extensible.

REDUCTION

Once the desired features are calculated, the platform performs correlation analysis, using the training set, to get a rough idea of what subset of the selected features best separates the two input states. The same view allows the user to select the features which will be given as inputs to the network.

MODEL

The platform provides a set of prebuilt deep learning models the user may select from. These models, like the rest of the platform are modular & extensible. They are simply python scripts that can be interrogated by the platform for name, icon, and parameters. Developers of these models have no restrictions with regards to authoring tool or platform dependencies.

OUTPUT

Finally, the platform provides the user with tools to analyze and measure the efficacy of the model. The output of the platform is an optimized deployable trained neural network.

 

TIMELINE

The Positronic team built this platform with a two-person team, a software engineer and a data scientist, in three months, start to finish.