Machine learning methods help you take diagnostic tests out of the lab

What if you could detect cancer and other diseases at the same speed and ease as a pregnancy test or blood glucose meter? Researchers at the Carl R. Woese Institute for Genomic Biology believe this is near a step towards achieving this goal by integrating machine learning-based analytics into point-of-care biosensing technology.

A new method called Loca-Pram was reported in the journal Biosensors and Bioelectronics Improves biomarker detection accessibility by eliminating the need for technical experts to perform image analysis. The paper is titled “Localization and quantification of physically grounded, deep learning-enabled gold nanoparticles in photonic resonator absorption microscopy for digital resolution molecular molecular diagnostics.”

Traditional medical diagnostic techniques require a physician to send a patient's blood or tissue sample to a clinical laboratory. There, expert scientists will carry out testing procedures and data analysis.

“With current technology, patients need to visit hospitals to get a diagnosis. This takes time. Many people also have barriers where more appointments are not financially or spatially feasible,” says Hanley, the first author of the study and a graduate student at the Nanosensor Research Group.

“I think we can make a difference by developing more point-of-care technologies that people can use.”

Point-of-care diagnosis is performed and results are achieved at the patient care site, at home, at the clinic or anywhere in between. This allows for low-cost, easy-to-use, and quick testing that will help you inform your next step.

Some examples already employed in daily life include urine pregnancy tests, Covid-19 antigen test kits, and glucose meters that allow diabetics to respond to lower levels of blood glucose and spikes throughout the day.

In the field of point-of-care, researchers are investigating new ways to integrate these types of techniques into patient care settings, including appointments with professionals such as oncologists and oral surgeons. This helps reduce patient time and financial burdens while improving physicians' real-time decision-making.

“Doctors say they want something similar when you get a bacterial infection. They will test you right away and send you home from an appropriate antibiotic appointment to treat the specific bacteria you have,” said Brian Cunningham, a professor of electrical and computer engineering, CGD leader.

“So why not try doing the same thing to choose the right anti-cancer drug or determine if the medication you've been taking for a few weeks is starting to work?”

Previously, the group reported a novel biosensing method called photonic resonator absorption microscopy (PRAM) to detect molecular biomarkers.

PRAM allows for the detection of single biomarker molecules, including nucleic acids, antigens and antibodies. Instead, common biosensing techniques detect cumulative signals of hundreds to thousands of molecules.

Cunningham said, “Essentially, what we do is make a red LED light shine at the bottom of the sensor. Then, at the top of the sensor, it's detected whenever a molecule lands and there are small particles made of gold.

Images generated using Pram depict a red background with small black spots scattered around them. However, these images themselves look relatively simple, but to get accurate counts you need a trained eye that can decipher the spots that really correspond to the AUNP-tagged biomarker molecule.

“There are many types of artifacts, such as dust particles and nanoparticle aggregates. Without a lot of experience, it's difficult to distinguish between them,” Lee said. “The traditional counting algorithm we use requires a lot of tweaking parameters to remove these artifacts.”

We proposed moving this process out of the laboratory and integrating machine learning into the image analysis process as it is suitable for a point-of-care environment.

“Han really was himself, and after taking classes in college, he became interested in machine learning to learn about it,” Cunningham said. “He said he thought he could come to me one day and create a machine learning algorithm.

Compared to other biosensing techniques, Pram is suitable for incorporating deep learning algorithms as it generates microscopic images rather than detecting optical signals. However, these algorithms are as good as the data they trained it, so Lee decided to image the same sample using both PRAM and a scan electron microscope.

The AuNP is 1000 times smaller than human hair, and is only shown as small black spots in the pram image, allowing for more clearly visualization by electron microscopes. In the time-intensive process, Lee used electron microscope images to cross-reference all the spots in the pram image to obtain highly accurate data from the machine learning training set.

“It was very difficult to find a suitable place to compare because it's like finding needles in the desert. One way I came up with was to create a reference point like a sea lighthouse, where you can find the exact same place as the registration,” Lee said.

The resulting deep learning-based method, called localization, with integrated context recognition with PRAM, allows for real-time, high-precision detection of molecular biomarkers without the need for the eyes and experience of technical experts.

On testing, the team found that Loca-Pram surpasses traditional technology with accuracy, detecting lower levels of biomarkers and minimizing false positives and negative rates.

“Every trip for my PhD started because I wanted to change the Point of Care Field,” Lee said. “I just want to do everything on my own to develop more advanced technologies that could have an impact in the future.”