EGG’s fundamental problem is the problem of extracting signal from noise. Confounding this further we don’t actually know the signal we are looking for. Perhaps you’re looking for a spike in electrical potential such as an Event Related Potential (ERP). Or perhaps it’s a pattern of electrical potentials of arbitrary length instead of a single spike. What paths could you take or tools could you use in an attempt to make use of this noisy data?
Historically, progress was made leveraging the techniques and algorithms known in the mathematical and engineering realms of signal processing and information theory. Finding spiky behavior was where a lot of initial progress was made producing many of the popular ERPs used in research today. This evolved to more complex evaluation of EEG signals with techniques like Filter Bank Common Spatial Pattern (FBCSP). Largely the realm of signal processing of this kind ends up being the work of manually tuning filter parameters to produce a meaningfully accurate result.
Enter Artificial Intelligence (AI). AI has a storied history including what’s know as AI Winter that started in the 80s; however, in the 2010s that all changed. The level of research and investment (both corporate and academic) increased exponentially. New techniques were discovered and published. Because of the types of industries investing in AI, and use cases important to those industries, a lot of progress in image processing/classification and natural language processing (NLP) was made. As part of this progress there was inspiration in other fields of study that leveraged the findings in many other fields. Since EEG is, by definition, time-series data it wasn’t long before these techniques and tools were applied there as well.
As mentioned earlier: EEG’s fundamental problem is extracting signal from noise combined with knowing what signal you’re looking for. AI is uniquely good at understanding these patterns. Let me explain with a basic example. Let’s build a relatively shallow Neural Networks to mimic the pipeline of FBCSP (Frequency Filtering, Spatial Filtering, Feature Selection, and Classification). Generally speaking each stage of FBCSP would have a corresponding layer in the neural network. As an example we can create a stage of our network that extracts the frequency bands of interest in an EEG signal. The difference between the FBCSP example and our AI driven example is that what frequencies to extract, what spacial filters to apply, and what features to use in classification are all learned through the training process. FBCSP and other non-AI signal processing techniques require you to tune these filters and select these features in a mostly manual fashion.
What this implies is that we can build even lager and more complex networks that learn to recognize even more obscure or long-lived patterns in our brain’s electrical potentials. Allowing us to accurately classify, predict, or observe very complex neurological activity. The success in using AI models to classify emotional states is a good example of the potential here.