One transistor mimics a synapse and sorts heartbeats, in lab

A single transistor, one device the size of a standard lab chip, is now doing two jobs that usually need separate parts: it remembers what came before, and it processes a signal in real time. At The University of Tokyo, a team led by Daisuke Kiriya, with Durgadevi Elamaran and colleagues, has shown that this combination works in the lab on electrocardiogram (ECG) recordings of the heart. The results appear in the journal ACS Nano.

The usual route to a memory device is to put a switching material throughout the channel that carries the current. The Tokyo team took another path. It worked on a planar MoS₂ memtransistor, a transistor made from molybdenum disulphide, a two-dimensional semiconductor, and treated only the points where the metal electrodes meet the material, using a solution-based AuCl₃ step. That treatment changes how charge gets trapped and released at the interface, which in turn alters the energy barrier, called the Schottky barrier, that controls how carriers are injected into the channel. The team's temperature-dependent measurements put the effective barrier at about 40.8 meV in untreated devices and about 196 meV after treatment. The authors describe the resulting switching as smooth and bipolar, with the channel itself left largely unmodified.

The device was then asked to behave like a synapse, the junction between two neurons whose strength changes with use. Under voltage pulses it showed gradual potentiation and depression, meaning its conductance rose or fell step by step, and its response depended on the amplitude and timing of the pulses. Its paired-pulse response had both fast and slow components, which the team links to more than one charge-relaxation process at the interface.

The heartbeat test is where the device had to use time itself. The team ran the transistor as a physical reservoir, a setup in which a dynamic system transforms an incoming signal and a simple readout layer interprets the result. Instead of an array of devices, one memtransistor was operated with 10 time-multiplexed virtual nodes, meaning one device was read at ten successive points in time to stand in for ten separate units. The team worked with 23,818 heartbeat segments from the MIT-BIH arrhythmia database, using 19,053 for training and 4,765 for testing. For telling normal beats from ventricular ectopic beats, the abnormal beats that start in the lower chambers of the heart, the analysis reached an area under the ROC curve of 0.965 and an overall accuracy of 93.9%. The researchers are explicit that this is not a diagnostic system: it demonstrates temporal signal processing with a single contact-engineered device, nothing more.

What changes, and for whom? The authors give no timeline, and their own list of open questions is long: large-scale integration, long-term reliability and manufacturing compatibility are all still to be evaluated. Still, the design points to a concrete route. The switching is localized at the metal–MoS₂ interface, and the treatment is a solution-processing step compatible with lithographic fabrication, the patterning method used to make chips. For engineers building low-power circuits that must analyse signals locally, such as wearable monitors, a memory-plus-processing element that fits in one device rather than an array would be a simpler building block, provided it survives those tests. Until the team or others reproduce these numbers on devices made outside the lab, the 93.9% figure stays a laboratory result.