AI models are applied to every dataset imaginable, but produce inconsistent results. This is as true in the medical world as it is in any other field, but a startup called Piramidal believes it is on the safe side with a basic model for analyzing brain scan data.
Co-founders Dimitris Sakellariou and Kris Pahuja found that while electroencephalography (EEG) technology is used in virtually every hospital, it is spread across many different types of equipment and requires specialized knowledge to interpret. Software that can continuously detect worrisome patterns regardless of time, place or equipment type could improve treatment outcomes for people with brain disorders while easing the burden on overworked nurses and doctors.
“In the neuralgia intensive care unit, there are nurses who monitor patients and look for signs on the EEG. But sometimes they have to leave the room, and these are acute conditions,” Pahuja said. An abnormal reading or alarm could mean an epileptic seizure, a stroke or something else – nurses don’t have that training and even specialists may recognize one but not the other.
The pair founded the company after spending years working on the feasibility of computer tools in neurology. They found that there was certainly a way to automate the analysis of EEG data that would be beneficial for care, but there was no easy way to deploy this technology where it was needed.
“I have experience with this, and I mean, I’ve sat next to neurologists in the operating room to understand exactly why these brain waves are useful and how we can build computer systems to identify them,” Sakellariou said. “They are useful in many contexts, but every time you use an EEG machine, you have to rebuild the whole system for the problem at hand. You have to get new data, you have to have people annotating the data from scratch.”
This would be difficult enough if all EEG systems, hospital IT facilities and data formats were the same. But there are big differences in the most basic elements, such as the number of electrodes on the machine and their placement.
Piramidal’s founders believe – and claim to know, even though the culmination of their work has not yet been published – that a fundamental model for EEG measurements could make life-saving detection of brainwave patterns work immediately, rather than after months of study.
To be clear, it’s not a one-size-fits-all medical platform – a better analogue might be Meta’s Llama series of (relatively) open models that bear the initial cost of developing the basic ability to understand language. Whether you’re building a customer service chatbot or a digital friend is up to you, but without the basic ability to understand human language, none of that will work.
But AI models aren’t limited to language—they can also be trained for fluid dynamics, music, chemistry, and more. For Piramidal, the “language” is brain activity as read by EEGs, and the resulting model would theoretically be able to understand and interpret signals from any setup, any number of electrodes or machine model, and any patient.
So far, no one has built one – at least not publicly.
While careful not to exaggerate their current progress, Sakellariou and Pahuja said, “We’ve built the basic model, we’ve done our experiments with it, and we’re now in the process of making the code base production-ready so it can scale to billions of parameters. It’s not about research – from day one, it’s been about building the model.”
The first production version of this model will be deployed in hospitals early next year, Pahuja said. “We are working on four pilot projects that will begin in the first quarter; all four will be tested in the ICU, and all four are keen to work with us.” This will be valuable proof that the model works in the different circumstances of each care unit. (Of course, PIramidal’s technology will go beyond any monitoring that would normally be provided to patients.)
The base model still needs to be fine-tuned for specific applications, a job that Pahuja will initially do itself. Unlike many other AI companies, they don’t plan to develop a base model and then collect fees for API usage. However, they made it clear that it is incredibly valuable as is.
“There is no world where a model trained from scratch will perform better than a pre-trained model like ours; a good start can only make things better,” Sakellariou said. “It’s still the largest EEG model ever, infinitely larger than anything else out there.”
To move forward, Piramidal needs the two things that are essential for any AI company: money and data. They’ve already started with the first things, with a $6 million seed round co-led by Adverb Ventures and Lionheart Ventures, with participation from Y Combinator and angel investors. The money will be used for compute costs (which are huge for training models) and staffing.
As for the data, they have enough to train their first production model. “It turns out there is a lot of open source data – but also a lot of open source in silos Data. We are in the process of aggregating and harmonizing these in a large integrated data store.”
However, the partnerships with hospitals are likely to provide valuable and extensive training data – thousands of hours. These and other sources could help push the next version of the model beyond human capabilities.
For now, Sakellariou said, “we can safely deal with this set of defined patterns that doctors look for. But a larger model will allow us to detect patterns that are smaller than the human eye can consistently and empirically detect.”
That’s still a long way off, but superhuman abilities are not a prerequisite for improving the quality of care. The ICU pilot projects should allow for a much more thorough evaluation and documentation of the technology, both in the scientific literature and probably in investors’ conference rooms.
