Making glucose measurement in food easier with deep learning

Recently there has been excitement about deep learning technology, a branch of machine learning that relies on artificial neural networks. A collaborative research team has now developed a deep learning-based glucose-sensing method that is robust to variations in sample position. Their findings have been published in the journal Laser & Photonics Reviews.

Metamaterials are artificial materials with unique electromagnetic properties not found in nature, enabling them to manipulate electromagnetic waves such as light or microwaves. A common structure used in the design of metamaterials is the split ring resonator (SRR), which features a ring with a split in its centre. This design allows the SRR to absorb, penetrate or reflect electromagnetic fields at specific frequencies and amplify signals due to the interruption of smooth current flow, leading to electromagnetic resonance within the ring. While SRRs have been widely used in sensors, their effectiveness has been limited by inconsistent and unreliable measurements influenced by factors such as temperature, humidity and sample location.

In this study, the team aimed to address the issue of SRR-based sensor electrical signal fluctuations caused by changes in sample position. They began by optimising the sensor to amplify electrical signals in the 0.5 to 18 GHz frequency range using a photolithography process that creates patterns on semiconductors with light. The researchers then employed deep learning technology to enable the glucose sensors to learn from the electrical signals measured at various locations.

Building on this foundation, the team developed a one-dimensional convolutional neural network (1D CNN) and conducted experiments with it. The results demonstrated that the model effectively compensated for errors due to sample location variations, achieving a mean absolute error (MAE) of 0.695% and a mean squared error (MSE) of 0.876%.

In an experiment aimed at predicting the brix in real fruit juices including pineapple, Jeju citrus, and Shine Muscat, the glucose measurement sensor using the research team’s 1D CNN model demonstrated high accuracy, achieving an MAE of 0.45% and an MSE of 0.305%. By overcoming the inherent challenges of SRR, the researchers have developed a reliable glucose measurement sensor suitable for practical use.

Schematic illustrating glucose concentration sensing based on deep learning. Image credit: POSTECH.

One of the researchers, Professor Junsuk Rho of Pohang University of Science & Technology (POSTECH), stated, “We have successfully managed to control the electrical signal which is sensitive to changes in sample position, thus enhancing the consistency and reliability of the glucose measurement device.

“It is also noteworthy that this technology can be commercialised and mass-produced using a photolithography process which is already widely employed in the semiconductor industry,” he said.

Top image credit: iStock.com/Sedaeva