Scientists Build an Artificial "Cerebellum" for AI That Spots Arrhythmia in a Split Second

A team from Northwestern University led by Mark C. Hersam published a paper on July 10, 2026, in the journal Nature Communications describing an electronic device that mimics the cerebellum, the part of the brain responsible for reflexes and reactions that occur without conscious thought. The device, called a memtransistor, is designed to detect anomalies faster and with far less energy use than conventional artificial intelligence systems.

How the Memtransistor Works

The device uses an asymmetric electrode structure, with the electrodes partially overlapping through a thin insulating layer. The direction of the applied voltage switches the device between an excitatory and an inhibitory mode, mirroring how the cerebellum maintains a balance between excitatory and inhibitory signals under normal conditions and shifts that balance when something unusual occurs.

In classic AI systems, every segment of a data stream, an EKG recording or a sensor signal, for example, is analyzed the same way regardless of whether anything has changed. The memtransistor works differently: it monitors the stream for novelty rather than analyzing each sample individually, so it stays in a resting state most of the time and only triggers full analysis when it detects a deviation from the norm.

Today's AI is remarkably good at recognizing patterns, but it often uses enormous amounts of computing power to continuously analyze data streams, even when nothing has changed - Mark C. Hersam, Northwestern University

Testing on Heart Arrhythmia

In proof-of-concept tests, the device analyzed electrocardiogram recordings and identified abnormal heart rhythms with accuracy exceeding 98 percent. Detection occurred within a time equivalent to one-fifth of a heartbeat, before the given beat had even fully completed.

Our cerebellum-inspired memtransistor detected an irregular heartbeat in a fraction of a second, before the heartbeat had even finished - Mark C. Hersam, Northwestern University

To build the device, the researchers used molybdenum disulfide, an atomically thin semiconductor that makes it possible to create transistors combining memory and computation in a single element. The project also involved Vinod K. Sangwan of the McCormick School of Engineering, neurobiologist Indira M. Raman, and Amit Trivedi of the University of Illinois at Chicago, giving the work an interdisciplinary character spanning materials science, electronics, and neurobiology.

Why It Matters for On-Device AI

Current AI models, including those used in wearable medical devices or autonomous vehicles, typically require a constant connection to computing servers or powerful batteries, since real-time data analysis consumes a lot of energy. The device designed at Northwestern is meant to run locally, on the device itself, and only switch into full operation when it detects something that requires attention, which could dramatically extend battery life in heart-monitoring patches, industrial sensors, or robots.

This approach fits into the broader trend of edge AI, moving artificial intelligence from the cloud directly onto edge devices. The edge AI market was worth about $30.9 billion in 2026, and analysts cited alongside the publication project growth to over $225 billion by 2035, driven precisely by demand for energy-efficient chips capable of operating without a constant connection to a data center.

Limitations and Next Steps

The researchers stress that the current device replicates only one element of the cerebellum's neural circuitry, not its full complexity. In humans and animals, the cerebellum is responsible not only for detecting novelty but also for precisely learning movements and gradually correcting them, something the memtransistor cannot yet do.

We've demonstrated one part of the cerebellum's neural circuit, but there's still more we haven't replicated - Mark C. Hersam, Northwestern University

The team says it plans further work to recreate the adaptive learning ability of the biological cerebellum, which would let the device improve its own performance over time without being reprogrammed. This builds on the same group's earlier research from 2023, in which a similar approach delivered a hundredfold reduction in energy use compared to conventional transistors.

Implications for Real-World Applications

The study's authors point to applications in wearable health devices, autonomous vehicles, robotics, and cybersecurity systems, where fast anomaly detection without continuous computing power use is critical. Unlike many high-profile language model launches, this research concerns the physical hardware layer of AI, an area where the US, China, and Europe are increasingly competing for an edge in energy efficiency and cost.

Sources: Cerebellum-inspired memtransistors enable emergent differentiation for hardware-efficient novelty detection (nature.com), AI Has A Cerebellum Now (science20.com), AI Chip Mimicking Brain's Reflex Center Developed (neurosciencenews.com), AI gets a 'cerebellum' (digitaljournal.com)