Revolutionary Foundation Model for Brainwaves Could Transform EEG Technology
AI models are increasingly being applied to various datasets, yet their outcomes often lack consistency, particularly in the medical field. A startup called Piramidal aims to address this issue with a foundational model specifically designed for analyzing brain scan data.
Image Credits:Ā Piramidal
The cofounders, Dimitris Sakellariou and Kris Pahuja, have noticed that electroencephalography (EEG) technology, while commonly used in hospitals, is fragmented across different machines and requires specialized knowledge to interpret accurately. They believe that a software solution capable of consistently identifying critical patternsāregardless of the time, location, or equipment typeācould significantly enhance outcomes for patients with brain disorders while also reducing the burden on overstretched medical staff.
Cofounders Dimitris Sakellariou (left) and Kris Pahuja.
Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Ā Image Credits:Ā Piramidal
āIn the neural ICU, nurses monitor patients and look for signs on the EEG, but they can’t always be in the room. These are acute conditions,” Pahuja explained. “An abnormal reading or alarm could indicate an epileptic episode, a stroke, or something else. Nurses arenāt trained to recognize all of these, and even specialist doctors may only be familiar with some of them.ā
The duo founded Piramidal after years of research into the feasibility of computational tools in neurology. They concluded that automating the analysis of EEG data could be highly beneficial for patient care. However, deploying such technology consistently across different settings remains a challenge.
āIāve spent time in operating rooms with neurologists to understand why these brainwaves are important and how we can create computational systems to detect them,ā Sakellariou said. āEEG data is valuable in many contexts, but each time you use a new EEG device, you have to rebuild the system for that specific scenario. This involves collecting new data and having humans annotate it from scratch.ā
This challenge is further complicated by the wide variation in EEG systems, hospital IT setups, and data formats, even down to basic elements like the number of electrodes and their placement.
Piramidalās founders are confidentāthough their findings are not yet publishedāthat a foundational model for EEG readings could enable life-saving brainwave pattern detection to work seamlessly, without the need for months of studies. Their goal isn’t to create an all-encompassing medical platform; a more fitting comparison might be Metaās Llama series of relatively open models, which establish the fundamental capability of language understanding. Whether the application is a customer service chatbot or a digital companion, it all starts with the ability to understand human language.
AI models, however, are not limited to language; they can be trained in fields like fluid dynamics, music, chemistry, and more. For Piramidal, the ālanguageā is brain activity as recorded by EEGs. Their proposed model would ideally be capable of understanding and interpreting signals from any EEG setup, regardless of the number of electrodes, the machine model, or the patient.
So far, no one has built such a modelāat least, not publicly.
Although they were cautious not to overstate their progress, Sakellariou and Pahuja stated, āWe have built the foundational model, conducted our experiments, and are now in the process of productionizing the code base so it can scale to billions of parameters. This has never been just about researchāitās always been about building the model.ā
The first production version of this model is expected to be deployed in hospitals early next year. Pahuja mentioned that they are planning four pilot programs starting in Q1, all of which will test the model in ICUs. These pilots will co-develop the technology with Piramidal and will serve as a critical proof of concept, demonstrating that the model can operate effectively in the varied conditions of different care units. Piramidalās technology will complement, rather than replace, existing patient monitoring systems.
While the foundational model will still require fine-tuning for specific applicationsāa task Piramidal intends to handle internallyāit remains incredibly valuable in its current state.
āThereās no scenario where a model trained from scratch will outperform a pretrained model like ours; having a head start only improves the outcome,ā Sakellariou asserted. āOurs is the largest EEG model ever created, vastly surpassing anything else available.ā
To advance, Piramidal needs two critical resources: funding and data. They have already made progress on the funding front, securing a $6 million seed round co-led by Adverb Ventures and Lionheart Ventures, with support from Y Combinator and angel investors. This funding will be used for computing costsāsubstantial for training modelsāand expanding their team.
As for data, they have enough to train their first production model. āThereās a wealth of open-source data out there, but much of it is siloed. Weāve been working on aggregating and harmonizing it into a comprehensive data store.ā
The partnerships with hospitals will also provide valuable training dataāpotentially thousands of hoursā worthāthat could elevate future versions of the model beyond current human capabilities.
For now, Sakellariou said, āWe can confidently identify the specific patterns that doctors are trained to watch for. But a larger model will enable us to detect even subtler patterns that the human eye may miss.ā
While superhuman capability is still some way off, it isnāt necessary to significantly improve the quality of care. The upcoming ICU pilots will provide the opportunity for more rigorous evaluation and documentation of the technology, both in scientific literature and in meetings with potential investors.