Real-time AI Classification Achieved by Nanoelectronic Devices, No Cloud Dependency Needed

Engineers at Northwestern University have developed a new nanoelectronic device that can perform precise machine learning classification tasks in the most energy-efficient way to date. Compared to current technology, this device reduces energy consumption by 100 times and can process large amounts of data and perform real-time artificial intelligence (AI) tasks without the need to transfer data to the cloud for analysis.

Due to its small size, ultra-low power consumption, and real-time analysis capabilities, this device is ideal for direct application in wearable electronic devices such as smartwatches and fitness trackers for real-time data processing and instant diagnostics.

To test this concept, engineers used the device to classify a large dataset of publicly available electrocardiogram (ECG) data. The device not only efficiently and accurately identified irregular heartbeats but also accurately determined subtypes of arrhythmias in six different categories, achieving an accuracy rate of nearly 95%.

This research, titled "Reconfigurable Heterogeneous Transistors for Personalized Support Vector Machine Classification," was published in the journal Nature Electronics on October 12th.

Mark Hersam, a professor at Northwestern University's McCormick School of Engineering and the Walter P. Murphy Professor of Materials Science and Engineering at the Feinberg School of Medicine, who is a senior author of the study, said, "Currently, most sensors collect data and send it to the cloud for analysis on high-energy-consuming servers, and then send the results back to the user. This approach is expensive, energy-intensive, and has time delays. Our device is so energy-efficient that it can be directly applied in wearable electronic devices for real-time monitoring and data processing, enabling faster intervention in health emergencies."

As a nanotechnology expert, Hersam is the Walter P. Murphy Professor of Materials Science and Engineering, director of the Materials Research Science and Engineering Center, and a member of the International Institute for Nanotechnology at Northwestern University's McCormick School of Engineering. Hersam co-led this research with Professor Wang Han from the University of Southern California and Assistant Professor Venkat Srinivasan from Northwestern University.

Before machine learning tools can analyze new data, they must first accurately and reliably classify training data into various categories. For example, if a tool is classifying photos based on color, it needs to identify which photos are red, yellow, or blue in order to classify them accurately. This is a simple task for humans but a complex and energy-consuming task for machines.

For current silicon-based technology, classifying large datasets such as electrocardiograms requires over 100 transistors, each consuming energy. However, Northwestern University's nanoelectronic device only requires two devices to perform the same machine learning classification. By reducing the number of devices, researchers significantly reduce energy consumption and develop a smaller device that can be integrated into standard wearable devices.

The secret of this new device lies in its unprecedented tunability, which is achieved through the use of hybrid materials. Instead of using traditional silicon materials, researchers construct microtransistors using two-dimensional molybdenum disulfide and one-dimensional carbon nanotubes. Therefore, compared to traditional techniques that require many silicon transistors (one for each data processing step), these reconfigurable transistors are flexible enough to switch between different steps.

"Integrating two different materials in one device allows us to control the flow of current by applying voltage, enabling dynamic reconfiguration. Having high tunability in a single device allows us to execute complex classification algorithms with a smaller size and lower energy consumption," said Hersam.

To test this device, researchers used publicly available medical datasets. They first trained the device to interpret electrocardiogram data, which typically requires trained medical professionals to spend a significant amount of time. Then, they asked the device to classify six types of heartbeats: normal heartbeat, atrial premature beat, ventricular premature beat, paced heartbeat, left bundle branch block heartbeat, and right bundle branch block heartbeat.

This nanoelectronic device accurately identified each type of arrhythmia in 10,000 electrocardiogram samples. By avoiding the need to send data to the cloud, this device not only saves critical time for patients but also protects privacy.

"Every data transfer increases the risk of data theft. If personal health data is processed locally, such as on a wristwatch, it reduces security risks. In this way, our device enhances privacy and reduces the risk of data breaches," said Hersam.

Hersam envisions that eventually, these nanoelectronic devices can be integrated into everyday wearable devices for personalized applications based on each user's health condition. They will enable people to make full use of the data they have already collected without consuming additional energy.

"Artificial intelligence tools are consuming an increasing portion of the power grid. If we continue to rely on traditional computer hardware, it is an unsustainable path," Hersam said.