Published: Nov. 6, 2019 By

Blood in an artery
Machine learning technology evaluates millions of images of blood samples to detect and identify pathogens.

New research using machine learning technology may help doctors identify pathogens in blood samples in a fraction of the time of current methods, leading to faster deployments of life-saving treatments in patients suffering from sepsis, especially newborns.

While researching injectable therapeutic preparations, Gillespie Professor Theodore Randolph, Assistant Professor Adjunct Christopher Calderon and graduate student Austin Daniels developed a machine learning-based technique for identifying and classifying types of particles found in their formulations. They soon realized that they could use this same analysis method to detect and identify invasive bacteria within blood samples.

In cases of sepsis — particularly in premature births — neonatologists often have to do a combination of guesswork and analysis to determine the nature of the illness. It is critically important to identify the right treatment for the right infection, or more harm can come from deploying the wrong antibiotics than from the disease itself.

“If doctors think there’s a blood infection, then they put the infant on antibiotics,” Randolph said. “Because they don’t always know exactly what kind of organism it is, they are basically guessing. If the infant is actually not infected and they are given antibiotics, that’s bad. The potential toxic effects of any antibiotic use are not good.”

After delivery, a newborn’s blood is sampled and analyzed for microbial infection. Currently, that analysis takes two days. If an infection is present, it can be another day or more before current analysis methods can identify the bacteria.

“Waiting three days to get the right treatment for the right infection may be too long,” Randolph said. “Unfortunately, it is a frequent occurrence that the kids die before the bacteria is identified. That’s the motivating factor: can we speed things up to save lives?”

“The combination of microscopy, machine learning and microfluidics enables us to find a needle in a haystack, where the haystack represents the large number of particles that can be resolved with an optical microscope — red and white blood cells, platelets, et cetera — in a blood sample,” Calderon said. “But we don’t just find the infection — we can also identify the bug in the blood sample.”

Daniels, Calderon, Randolph
 Reseachers Daniels, Calderon and Randolph.

This machine learning-based technique is similar to facial recognition technology, but it works at the microscopic level. A blood sample is sent through microfluidic channels under a microscope and photographs are taken of objects that are larger than one micron in size. This results in over a million photos taken of a relatively small blood sample. These images are then reviewed by the system, which can identify specific microbial organisms in a fraction of the time of traditional tests.

“We are designing the system so we only need a single drop of blood, which is advantageous, especially for premature infants,” Randolph said. “Twenty minutes after the sample is analyzed by the machine, the doctor gets an answer as to which antibiotic to use.”

To ensure their platform has the right impact for medical professionals, the team is working with James L. Wynn, Professor of Pediatrics, Pathology, Immunology, and Experimental Medicine at the University of Florida Medical School. Wynn is serving as a consultant for the project, helping to make the platform’s algorithm clinically relevant, including providing a prioritized list of bacteria for identification and conducting medical research literature reviews.

“A fast approach that can be deployed at a variety of hospitals worldwide for detecting and determining the root cause of sepsis from blood samples would address many issues facing sepsis detection and diagnosis,” Wynn said. “The implementation of this platform should have a major impact on antimicrobial treatment in all areas of the hospital.”

The team is preparing to test this technique with human blood samples in collaboration with physicians working in the field and anticipates two to three more years of FDA review before tests can begin on patients.

Randolph credits his brother David, a ChBE department graduate and medical doctor, with sparking the team’s conversations about the potential applications of their technology. David Randolph currently practices as a neonatologist.