Cerebellum-inspired memtransistor chip detects irregular heartbeat in real time

Researchers have created a minute electronic chip that copies the brain’s knack for tuning out the routine and responding only when something unusual happens.

Designed around the cerebellum - the reflex centre located at the base of the skull - the cerebellum-inspired device identified abnormal heartbeats almost as soon as they emerged.

It achieved this while requiring only about a ten-thousandth of the computing steps that current artificial intelligence typically needs to perform the same task.

That kind of efficiency could make always-on health tracking, self-driving vehicles, and other systems that must react immediately far more practical, without exhausting batteries or depending on faraway data centres.

The brain’s reflex centre

Most attempts to build brain-like computers take their cues from the cerebrum, the folded outer region responsible for thinking, memory, and language.

Researchers at Northwestern University opted to focus elsewhere. The cerebellum manages balance and reflex actions with little to no conscious input - and it does so with striking energy economy.

The study was co-led by Mark C. Hersam, the Walter P. Murphy Professor of Materials Science and Engineering at Northwestern. He characterises the cerebellum as an expert in holding back.

“The cerebellum is excellent at ignoring the expected and reserving its resources for reacting to the unexpected,” said Hersam. Ignoring things costs the hardware almost nothing.

Neuromorphic computing - the field of chips that emulate the nervous system - usually pursues the brain’s sheer computational muscle.

This approach targets something less common: the brain’s frugality, reflected in a dramatic reduction in energy demand.

Cutting the energy waste

Today’s computer systems tend to operate without pause, unlike biological brains.

“Today’s AI is remarkably good at recognizing patterns, but it often spends enormous amounts of computing power to continuously analyze streams of data,” said Hersam.

A camera or heart monitor might process every frame or every beat with the same intensity even when conditions remain unchanged.

Some inefficiency comes from how conventional hardware is connected. Standard computers move information back and forth between separate memory and processing components.

Instead, Hersam’s team combines storage and computation within a single element known as a memtransistor, so data can be kept and processed in the same place.

In earlier work, the lab showed that just two of these devices could carry out tasks that would typically require more than 100 transistors.

Inside the new device

The chip is made using molybdenum disulphide, a semiconductor only a few atoms in thickness.

To build it, the team positioned the electrodes asymmetrically, with one electrode partially overlapping the semiconductor across a thin insulating layer. That subtle imbalance altered the way electrical current flows through the component.

Changing the polarity of the applied voltage switches how the chip behaves. In one configuration, it behaves like an excitatory neural connection.

In that state, the response strengthens the longer the input persists. When the voltage is reversed, it becomes inhibitory - it reacts powerfully at first and then rapidly diminishes.

The brain’s balance

Within the cerebellum, excitatory and inhibitory activity generally offsets itself, remaining stable until an unexpected event disrupts the equilibrium. The chip reproduces this push-and-pull dynamic.

As a familiar signal repeats, the two operating modes remain aligned and the device shows very little activity; when a novel signal appears, the balance shifts and the device responds.

Brain-imaging studies have shown the human cerebellum activating in response to unforeseen sensory events - precisely the kind of signals this chip is intended to detect.

Spotting the odd beat

To see whether the concept worked in practice, the researchers supplied the chip with real electrocardiogram recordings - electrical traces from the beating heart - combining normal rhythms with hazardous arrhythmias.

The device largely ignored the regular pattern. When an irregular beat occurred, it identified it within one-fifth of a single heartbeat, before the beat had even completed.

It delivered more than 98% accuracy while using roughly 10,000 times fewer operations than a conventional silicon system performing the same task.

“Our cerebellum-inspired memtransistor detected an irregular heartbeat within a fraction of a second, before the heartbeat even ended,” said Hersam.

He said it was operating at more than double the speed of standard AI. And because the analysis is performed directly on the chip, there is no requirement to transmit data to the cloud and wait for a response.

Because dangerous heart rhythms can become fatal within seconds, detection speed is crucial.

A monitor that can raise an alert within a single beat - rather than after uploading data and waiting for a server to respond - could enable faster intervention.

Where this could lead

Energy demand is a major part of the promise. One projection suggests global data centres are heading towards roughly doubling their electricity consumption by 2030, largely due to AI.

Shifting more of that work onto low-power chips inside devices could reduce pressure on the electricity grid.

Until now, many brain-inspired chips have mainly aimed to classify data faster or at lower cost. This one demonstrates that a device can isolate the unexpected on its own, in real time, using a tiny fraction of the energy.

That points to sensors that can watch continuously and only signal when something is wrong.

Teaching the chip to learn

So far, Hersam’s team has reproduced only a single portion of the cerebellum’s function.

Their next aim is to add learning. If the chip experiences the same event repeatedly, it would stop treating it as surprising - similar to how the brain learns to disregard familiar noises.

“We have demonstrated one part of the cerebellum neural circuit, but there is more that we have not yet emulated,” said Hersam.

For the moment, the group has demonstrated that computing hardware can borrow the brain’s ability to notice when the world stops behaving as expected.