Sophisticated systems for detecting biomarkers – molecules such as DNA or proteins that indicate the presence of disease – are essential for real-time disease monitoring and diagnostic devices.
Holger Schmidt, Distinguished Professor of Electrical and Computer Engineering at UC Santa Cruz, and his group have long focused on developing unique, highly sensitive devices called photofluidic chips for detecting biomarkers.
Schmidt graduate student Vahid Ganjalizadeh led an effort to use machine learning to improve their systems by improving its ability to accurately classify biomarkers. The deep neural network he developed classifies particle signals with 99.8 percent accuracy in real time, on a relatively cheap and portable system for point-of-care applications, as described in a new paper in Scientific Nature Reports.
When vital sign detectors are taken out in the field or in a setting of care such as a health clinic, the signals received by the sensors may not be of as high quality as those in a laboratory or controlled environment. This may be due to a variety of factors, such as the need to use cheaper chips to reduce costs, or environmental characteristics such as temperature and humidity.
To address the challenges of a weak signal, Schmidt and his team developed a deep neural network that can identify the source of that weak signal with high confidence. The researchers trained the neural network with known training signals, teaching it to recognize potential differences that it could see, so that it could recognize patterns and identify new signals with very high accuracy.
First, the Parallel Cluster Wavelet Analysis (PCWA) approach designed in Schmidt’s lab detects the presence of a signal. Next, the neural network processes the potentially weak or noisy signal, and identifies its source. This system works in real time, so users can receive results in a split second.
“It’s all about making the most of low-quality signals, and doing it really quickly and efficiently,” said Schmitt.
A smaller version of the neural network model can run on mobile devices. In the paper, the researchers ran the system on a Google Coral Dev board, which is a relatively cheap edge hardware for rapid implementation of AI algorithms. This means that the system also requires less energy to perform processing compared to other technologies.
“Unlike some research that requires running on supercomputers to do the high-resolution detection, we’ve demonstrated that even a relatively small, portable, and cheap device can do the job for us,” Ganglizadeh said. “It makes it accessible, accessible, and portable for point-of-care applications.”
The entire system is designed to be used entirely locally, which means that data processing can happen without access to the Internet, unlike other cloud-based systems. This also provides the advantage of data security, because results can be produced without having to share the data with a cloud server provider.
It is also designed to be able to give results on a mobile device, eliminating the need to bring a laptop into the field.
“You can build a more robust system that you can use in under-resourced or less developed areas, and it still works,” Schmidt said.
This improved system will work with any other biomarkers that Schmidt’s lab systems have been used to detect in the past, such as COVID-19, Ebola, influenza, and cancer biomarkers. Although it is currently focused on medical applications, it is likely that the system could be adapted to detect any type of signal.
To push the technology even further, Schmidt and members of his lab plan to add more dynamic signal processing capabilities to their devices. This will simplify the system and combine processing techniques needed to detect signals at both low and high concentrations of molecules. The team is also working to fit discrete parts of the setup into the integrated photofluidic chip design.