Story has been updated to more accurately state the company's planned FDA submission timeline.
NEW YORK – Clinical diagnostic startup Prenosis is applying machine learning to big data to better predict and diagnose sepsis and enable earlier treatment.
Prenosis, a University of Illinois at Urbana-Champaign spinout currently based in Chicago, uses "holistic data of many different dimensions," including biomarkers, vital signs, demographic information, and past medical histories to provide digital diagnostics for sepsis, CEO Bobby Reddy said.
While the firm’s proprietary software platform isn’t currently commercially available, the company plans to submit it to the US Food and Drug Administration for de novo clearance by the end of this year and hopefully receive clearance in 2022.
Reddy noted that a key point about sepsis that many people don't understand is that it is "a syndrome that is a collection of a lot of different diseases that kind of present under one roof." It indicates a response to a bacterial infection, rather than being an infection itself. That makes it difficult to diagnose using a specific test, he said.
Right now, clinicians often use systemic inflammatory response syndrome, or SIRS, criteria to identify potential sepsis patients. Those criteria are based on elevated heart and respiratory rate, body temperature, and abnormalities in the number of white blood cells in the blood stream.
However, those abnormalities can also be seen in patients who do not have sepsis, which can result in sepsis patients being missed.
Following a mantra of "collect everything," including multiple biomarkers and medical record data, the company applied machine learning methods to narrow down which features were most predictive of certain kinds of outcomes, as well as which were best for separating different subtypes of sepsis, Reddy said.
Prenosis' Sepsis ImmunoScore algorithm, which runs on its Immunix platform, looks at parameters that can signify sepsis such as a combination of vital signs, past medical history, and protein biomarkers and uses predictive modeling to score patients based on risk of developing sepsis and provide a diagnosis, he said.
"We're really about being agnostic to the type of data,” he said, adding that “it's about the overall holistic impact of this data" on the patient.
The algorithm can also provide a measure of objectivity compared to a doctor's interpretation of a patient's vitals or past medical history, Reddy continued.
Data collected from a patient in the emergency room is entered into the Prenosis system, and if there are signs of sepsis — based on the data from past patients who deteriorated into sepsis — a warning is triggered and the system requests additional information, Reddy said.
Once further tests for relevant biomarkers have been performed, that information is sent to the system, which combines it with the original data to provide a comprehensive score determining whether a patient will develop sepsis.
Prenosis doesn't do the testing itself, he noted — though it does have a testing product in development — but instead receives the information from the hospital laboratory after tests are done. Several of the biomarkers the algorithm looks at aren't routinely tested, and one has not yet been approved by the FDA for sepsis, although Prenosis is working with an undisclosed partner to get it approved, Reddy said.
He also noted that "almost every hospital in the US at least either has the capability or can easily adapt the capability" to do these biomarker tests.
The algorithm was trained on a dataset Prenosis built itself with more than 10,000 patients and more than 50,000 blood samples, and the company plans to double that number in the next year, Reddy noted.
The company prospectively collects blood samples from hospitals and measures 40 different proteins, including the specific proteins for the ImmunoScore — procalcitonin, interleukin-6, and C-reactive protein — in the samples itself, joining this information with other previously collected data, such as vital signs, to build its NOSIS dataset. Prenosis believes the set is "one of the largest if not the largest hybrid biomarker clinical datasets for sepsis," Reddy said.
The performance of the algorithm has been evaluated at six different hospital sites, with plans for an additional two, and Prenosis is currently processing its data now — it's having three separate doctors review each patient chart to compare results, he said.
Right now, 40 percent of patients aren't diagnosed correctly with sepsis after two hours in the hospital. Prenosis' algorithm brings that down to 7 percent, he said. In a paper published earlier this year in Clinical Translational Science, researchers at the University of Illinois at Urbana-Champaign found the test had "good diagnostic and prognostic capability at the time of initial blood culture."
"If you look at the sensitivity, specificity, and positive predictive value of what doctors are doing 12 hours after the patient comes in [to the hospital], that performance is similar to what the Sepsis ImmunoScore is doing one hour after," Reddy said.
Kevin Klauer, CEO of the American Osteopathic Association and an emergency medicine physician in Chicago who has reviewed the technology prior to investing in it, said that a key problem with sepsis diagnosis is that "when complex patient presentations such as sepsis … blend in with so many other illnesses, opportunities are missed to diagnose sepsis earlier when treatment is most effective."
"In other words, sepsis in its late stages is easy to diagnose but very difficult to treat," he said. "In contrast, early sepsis is much easier to treat, but much more difficult to diagnose."
He added that ER physicians need something better than SIRS criteria, such as the ImmunoScore, which he described as "clinical decision support that can assist the clinician in a real-time fashion using a scoring system built from former actual sepsis cases, combined with state-of-the-art laboratory testing strategies."
Klauer also noted that while sepsis is the main focus for Prenosis, its technology could be used for any difficult-to-diagnose clinical entity lacking diagnostic support.
Prenosis is bringing its algorithm through the FDA and expects de novo clearance next year, Reddy said. The company has already met with the FDA about its clearance plans, he added.
When and if that clearance is obtained, Reddy said the firm plans to commercialize the algorithm, potentially starting at the six hospitals that it built its NOSIS dataset with. It also has implementation studies planned at four new sites over the next six months, Reddy added.
Prenosis also intends to seek European regulatory approval, but decided to begin with the FDA because it's "hardest to get approval" from and requires the most extensive validation, he said.
As for the cost, Prenosis will charge hospitals likely less than $100 to use the algorithm and the impact on a hospital will be "tens of thousands of dollars," Reddy said.
Adoption from the clinical community may be a hurdle for the company, Klauer said. The firm will need to figure out how to introduce its tech into the "natural workflow of clinicians" so that it is "useful, additive to the clinical decision-making process, but not intrusive," he said.
Prenosis has thus far completed a Series A financing round for an undisclosed amount with a strategic investor and has also received government contracts from the National Institutes of Health, the US Centers for Disease Control and Prevention, and the US Department of Health and Human Services, raising "a little less than $20 million” in total, Reddy said.
The firm is also planning an investment round for later this year, he said.
Addressing the problems with sepsis diagnostics is important to Prenosis, and Reddy, because it's "a travesty that it's still such a huge problem in 2021 when we discovered the cure to sepsis" more than 100 years ago, Reddy said.
"It's just about getting the right treatment to the right patient at the right time," he said.
"It's more of an information and logistical problem than it is a fundamentally biological problem."