AI models have been applied to a variety of datasets with inconsistent results, as is true in the medical world and elsewhere, but a startup called Piramidal believes its basic model for analyzing brain-scan data holds promise.
Co-founders Dimitris Sakellariou and Chris Pahuja realized that while electroencephalography (EEG) technology is used in nearly every hospital, it's spread across many types of machines and requires expertise to interpret. Software that can consistently flag worrying patterns regardless of time, location, or type of equipment could improve outcomes for people with brain diseases while reducing the burden on overwhelmed nurses and doctors.
“In neuro-intensive care units, nurses actually monitor patients and look for EEG signs, but sometimes nurses have to leave the room if the patient is acutely ill,” Pahuja says. Abnormal readings or warnings could mean a seizure, stroke or other symptoms. Nurses aren't trained to do so, and even specialists may recognize one but not the other.
The duo founded the company after years of researching the feasibility of computational tools in neurology: They realized that while there were certainly ways to automate the analysis of EEG data to inform care, there was no easy way to deploy that technology where it was needed.
“I have experience with this – I've sat next to neurologists in the operating room and understood exactly why EEG is useful and how we can build computational systems to identify it,” Sakellariou said. “EEG is useful in many situations, but every time you use an EEG device, you have to re-build the whole system for that specific problem. You have to get new data and have humans annotate the data from scratch.”
It would be difficult enough if all the EEG systems, hospital IT setups, and data formats were the same, but the most basic elements, like how many electrodes the machines have and where they are placed, vary greatly.
Co-founders Dimitris Sakellariou (left) and Chris Pahuja. Image courtesy of Pyramidal.
Pyramidal's founders believe, and claim to know, that a basic model of EEG measurement will soon enable life-saving brainwave pattern detection without months of research, but the culmination of their research has yet to be published.
To be clear, this is not a one-size-fits-all healthcare platform. A closer analogy would be the (relatively) open model of Meta's Llama series, which covers the upfront cost of creating a foundational capability: language understanding. Whether you want to build a customer service chatbot or a digital friend, it's up to you, but neither will work without the foundational capability of understanding human language.
But AI models are not limited to language and can be trained to work in a variety of domains, including fluid mechanics, music, chemistry, etc. For Piramidal, “language” is brain activity as read by EEG, and the resulting model will conceptually be able to understand and interpret signals from any setting, with any number of electrodes or models of machines, and from any patient.
No one has built it yet – at least, publicly.
While careful not to overstate their current progress, Sakellariou and Pahuja said, “We built a basic model and ran experiments based on it. We are now in the process of productionizing the code base so that it can scale to billions of parameters. This has never been about research; it's been about building models from the beginning.”
The first production versions of the model are expected to be deployed in hospitals early next year, Pahuja said. “We've been working on four pilots since the first quarter, all four we'll be testing in the ICU, and all four they want to co-develop with us,” he said. This will provide a valuable proof of concept that the model can work in a variety of settings across any care unit. (Of course, PIramidal's technology will go beyond the monitoring typically provided to patients.)
The foundational models will need to be fine-tuned for specific applications, but Pahuja said the company is doing that work in-house to begin with. Unlike many other AI companies, the company doesn't plan to build foundational models and then earn royalties for using the API. But it's clear that it's very valuable as it stands.
“There's no way that a model trained from scratch can be better than a pre-trained model like ours. Having a warm start can only make things better,” Sakellariou says. “This is the largest EEG model that has ever existed, and it's infinitely larger than anything else out there.”
To move forward, Pyramidal needs two essential ingredients for any AI company: funding and data. First, they raised a $6 million seed round, co-led by Adverb Ventures and Lionheart Ventures, with participation from Y Combinator and angel investors. The funding will cover computing costs (training models is expensive) and headcount.
As for data, there's enough to train the first production models. “We found that there's a lot of open-source data, but also a lot of open-source siloed data, so we've been in the process of aggregating and harmonizing that into a big unified data store.”
But the hospital partnerships will provide a wealth of valuable training data — thousands of hours of data — that could help push the next version of the model beyond human capabilities.
“Right now we can confidently work with a set of defined patterns that doctors look at, but larger models will be able to pick out smaller patterns than the human eye can consistently and empirically determine exist,” Sakellariou said.
That's still a long way off, but superhuman abilities aren't a prerequisite for improving the quality of healthcare, and the ICU pilot should encourage the technology to be much more rigorously evaluated and documented, both in the scientific literature and perhaps in investor boardrooms.