How Piramidal is using AI to decode the human brain

The human brain is ultimately one of the last frontiers — a paradoxical black box that we can’t even begin to understand ourselves.But what if, just as paradoxically, AI could interpret the complexities of the brain to help identify and diagnose some of our most serious diseases? That’s exactly what Y Combinator-backed startup Piramidal has set out to do. The company is building a first-of-its-kind foundation model that can detect and understand complex “brain language” or brainwaves. It can be fine-tuned to a range of electroencephalography (EEG) use cases and has implications in other areas of medicine, as well as in pharmacology and even consumer products. 

The startup is also announcing today a $6 million fundraise from Y Combinator, Adverb Ventures, Lionheart Ventures and angels including founders of Intercom, Plangrid and Guilded. 

“We’re training an AI model on brainwave data the same way ChatGPT is trained on text,” Kris Pahuja, Piramidal co-founder, told VentureBeat. “It is the largest model ever trained on EEG data.”

EEG data too much for one person to interpret

Today, when patients with brain-related conditions seek medical treatment, their EEG brain waves are mapped, and then inspected by neurologists. But this can be highly time-consuming and error-prone, with a margin of error up to 30%, according to Pahuja. 

Compounding this is the fact that there is an “extreme shortage” of neurologists — particularly those who can interpret EEGs — in the U.S. Pahuja pointed out that patients’ brain waves are recorded for several days or weeks when they are in the intensive care unit (ICU) — and no human could possibly go through all that. Instead, physicians take random samples and perform quick pattern recognition, but this can miss out on a lot of diagnosis.

EEG data is also incredibly complex, difficult to interpret and has significant signal variability. Pahuja pointed out that when someone is looking at an MRI image, for instance, they are looking at an image in one distinct period of time. 

But an EEG, by contrast, is “very difficult to read, it changes thousands of times a second across 10 to 20 channels,” said Pahuja. He noted that even specialized doctors can miss many details, and some may only be trained in certain areas such as epilepsy or brain injury, so they don’t know all the markers to look for. 

Another challenge lies in scarce labels/annotations for EEG recordings, which can inhibit the training of more large-scale, generalized models, he noted. Further, narrow models aimed at specific tasks can’t be repurposed for new use cases. 

“We want to train our model to be at the level of an expert neurologist, but also not miss anything while an EEG is going on,” said Pahuja. 

Trained on every EEG use case

Advancements in time-series models trained on diverse, unlabeled data to evolve to a variety of tasks is allowing Piramidal — named for pyramidal neurons found in areas of the brain — to overcome these significant challenges, according to the startup. 

The company is first fine-tuning its model for the neuro ICU; that product will be able to ingest EEG data and interpret in near-real time, providing outputs to medical staff on the occurrence and diagnosis of disorders such as seizures, traumatic brain bleeding, inflammations and other brain dysfunctions. 

“It is truly an assistant to the doctor,” said Pahuja, noting that the model can ideally help provide quicker and more accurate diagnoses that can save doctors’ time and get patients the care they need much more quickly (which can also help reduce overall healthcare costs). 

“Brainwaves are central to neurology diagnosis,” Piramidal co-founder and CEO Dimitris Sakellariou, who holds a PhD in neuroscience, told VentureBeat.

By automating analysis and enhancing understanding through large models, personalized treatment can be revolutionized and diseases can be predicted earlier in their progression, he noted. And, as wireless EEG sensors become more mainstream, models like Piramidal’s can enable the creation of personalized agents that “continuously measure and monitor brain health.”

“These agents will offer real-time insights into how patients respond to new treatments and how their conditions may evolve,” said Sakellariou.

The company’s model has seen every EEG use case from both proprietary and open-source datasets, said Pahuja. It can tackle certain biomarkers that exist on certain disorders right away (such as seizures, brain bleeding or low blood flow) and can find other biomarkers that don’t yet exist (such as for diseases such as Parkinson’s or Alzheimer’s. 

Piramidal is currently piloting in two hospitals in England, at King’s College and Saint Thomas. 

“No one else is building an EEG model like ours,” said Sakellariou. 

He pointed out that it requires significant time and money to ensure “generalisability and reliability” from the start. 

“AI has the potential to transform healthcare, especially neurological diagnostics,” said Sakellariou. “Piramidal aims to be at the forefront of this transformation.”

Inspired by psychedelic and sleep research

The revolutionary model was initially inspired by Sakellariou’s experiences in various EEG studies, ranging from psychedelics to sleep research — both as a subject and an observer. In these studies, he explained, a technician attaches electrodes to the scalp and the system records brainwaves.

“Surprisingly, the process of capturing brain activity through scalp and hair is straightforward — you simply attach some wires to your head, and you can monitor what’s happening in your cortex,” said Sakellariou.

However, researchers and clinicians then have to visually analyze these “wavy lines,” which could represent hours, days or even weeks of brainwave data, to extract useful information for the subject or patient.

He explained that this process is error-prone and subject to misinterpretation for a couple of reasons. Firstly, acquiring the necessary training to interpret brainwaves is “highly empirical”; secondly, the extensive duration of the recordings does not allow for “meticulous inspection,” especially since brain changes reflected in EEG data can occur in milliseconds.

Beyond the ICU

But for Piramidal, the ICU is just the start, according to its founders: Their model has significant potential beyond that niche area of medicine. 

For instance, Pahuja noted, it could be implemented into general neurology, epilepsy units, longer-term monitoring situations and in neuropsychiatry (which uses EEGs to study mental health disorders and cognitive decline). Further down the line, it could be used in every physician clinic to help with different types of patient screenings. 

It could also be “huge for pharmacy,” providing real-time efficacy, said Pahuja, as well as in consumer products that rely on EEG data (such as Ray Ban Meta or the multitude of health monitoring devices on the market). 

“As technology evolves, you can get through the noise,” he said. 

In the near future, it’s possible that humans will have the opportunity for “quantified introspection” through everyday devices such as earphones equipped with neural sensors, Sakellariou pointed out. For example, we could measure how stress levels decrease after reducing screen time, train ourselves to enhance meditation by monitoring relaxation levels in a closed loop or boost memory during periods of “intense learning” through targeted auditory stimuli during specific sleep stages.

“All of this will be possible via personalized agents powered by large-scale models like ours,” said Sakellariou.

Pahuja said he has always been fascinated by the brain, describing “neurotech as the next frontier.” 

As he put it: “The most complex thing we have is our brain, but that is completely not understood at the moment. Can we find a way to decode the brain?”