CHICAGO (360Dx) – Data exists for hospitals to address the problems of sepsis and septic shock, but as with so many other areas of healthcare, that information has long been held in silos.
A four-year-old Israeli company named Clew Medical is trying to solve this dilemma by collecting data from multiple sources within hospitals, then applying machine learning and predictive analytics to catch potentially deadly events before they happen, allowing for earlier intervention and more appropriate treatment. The technology also has shown promise as way to rule out sepsis.
Clew, known as Intensix until late 2017, concluded an $8.3 million Series A financing round in February 2017. It has raised $10.5 million to date, including a 2015 seed round worth $2.2 million.
In a preliminary study of the efficacy of its technology, Clew collected 10 billion data points from 200,000 high-acuity patients at five hospitals, including Mayo Clinic's flagship facility in Rochester, Minnesota. The technology predicted the forthcoming presentation of sepsis 56 percent of the time and successfully deduced that sepsis was not happening in 97 percent of cases where it ruled out the condition, according to non-peer-reviewed data supplied by the company.
Mayo intensivists presented these initial results at the Society of Critical Care Medicine's 2018 annual congress. A peer-reviewed paper is in the works, though it has not yet been submitted for publication, according to a Clew spokesperson.
"We have a complex method that makes sure that the data that we feed the model, mainly in terms of diagnosis, is improved … by looking into actions that were taken by the physicians that might imply that the patients were having septic shock or any respiratory failure, and we try to fine-tune the exact time that this deterioration happened," explained Uri Keler, Clew's vice president of clinical development.
The system next looks through retrospective data patterns that might indicate the duration of sepsis onset and presentation. Clew scientists and users then validate these models before running the technology on live patient data.
The patterns are validated before the duration occurs and to ensure that they aren’t appearing in the general population, Keler said. "By doing that, we are able to develop those models that we test internally on retrospective data … making sure that they are viable in different hospital settings with different monitoring protocols," he explained.
"We're trying to look at the mathematical patterns in data that comes in from the patients in high-acuity areas and look for mathematical abnormalities that detect that patients are deteriorating before the appearance of any physiological signs," added Izik Itzhakov, Clew’s vice president of business development.
The Clew system, which is a form of advanced clinical decision support, can be tied into electronic medical records, medical devices, and telemetry equipment.
"We are trying to connect to all the different feeds," Itzhakov said. "In critical care, we are getting all the data from the EMR, including the [admission, discharge, and triage] data, labs, microbiology, diagnostics, documentation done by the clinician, medication, fluids. On top of that, we're looking at the data that comes in from all medical devices — basically, everything that's connected transmits in high resolution once a minute from the monitors, the ventilators, dialysis machines, pumps."
Clew takes that information and looks at multiple parameters simultaneously in order to detect subtle changes that might indicate a patient’s condition is changing.
While Clew has mainly tested the technology in intensive care units and other high-acuity hospital departments, the technology could be used in other clinical areas, as well as for administrative purposes, according to Itzhakov.
"For clinical purposes, we're looking at deterioration. We're looking at different types of complications," Itzhakov said. He added that this approach could be used to predict a patient’s length of stay, the possibility that he or she may be discharged, the possibility of readmission, or when a patient could be taken off certain drugs or ventilators.
Clew’s platform “could be used for basically everybody that touches or gets involved with a patient, he said.
The firm is taking patient-specific data from a large cohort of patients, Keler explained, and, added Itzhakov, "Right now, we are in the process of trying to look at everything that we've done retrospectively and implement this in real-time environments for hospitals and clinicians in critical care so we could actually make decisions based on our predictions."
In the case of sepsis, the analytic technology is looking for ways to define a specific time frame in which a specific patient is diagnosed with the potentially deadly complication, as well as other crucial information.
"We look at the specific interventions that this patient underwent and then we understand what was the point in time where the clinicians bedside understood that this patient is actually septic or [has] any other iterations that we are looking at," Keler said. From that, Clew develops models to predict patient deterioration, as well as the reasons behind this expectation of a downturn.
The model, he added, can be trained to predict sepsis hours before the presentation of symptoms.
Sometimes, a patient’s clinical indications may be suggestive of sepsis, but, in fact, it could be something else, he noted. "We mark this specific spot and then we highlight the time frame before that, spot and we train the model to detect those patterns that we see in the data,” Keler said. “We gather all the patterns that we come across and we train the model to alert when we notice those patterns in real time."
Historically, hospitals have attempted to detect sepsis by looking at vital signs and laboratory values to spot symptoms, but such parameters “change frequently,” which could result in false alarms, he said. For example, a patient might have a single lab result or blood pressure measurement that seems abnormal. Current, rules-based engines like an in-house system at Mayo called Sniffer tend to cause false alarms.
"They rely on a limited amount of parameters and those parameters are noisy by nature," Keler said. Clew is trying to reduce alert fatigue by providing more accurate predictions. To that end, criteria used by currently available models are used only as benchmarks for Clew’s algorithms, he said.
In doing so, Clew offers two advantages. "One is being able to alert ahead of time, and the other is being able to reduce the alarm fatigue by having a higher [positive predictive value]," he said.
Last year, Clew completed a proof-of-concept study that compared this predictive analytics system to Mayo's own Sniffer. The Clew model was able to detect 86 percent of the 275 events examined, with a PPV of 60 percent, the company reported, far superior to Sniffer's performance. There were false alarms, but far fewer than with older technology, including Mayo's.
"For every 10 alerts that you have to encounter, only one of them would be an actionable alert. We got to the point that for every three alerts that we fire, approximately two alerts would be actionable. It gives you much less alarm fatigue," Keler said.
This, he said, is because the Clew model learns patterns.
"In the cases where we provide alerts and the patient wasn't septic, it's not that the patient wasn't doing well and there was some kind of artifact in the monitoring data. It means that the patient was having some kind of a deterioration that showed a pattern similar to the pattern that the model learned, but it wasn't sepsis," Keler explained.
Recently, through Mayo, Clew was able to land the University of Massachusetts' UMass Memorial Health Care, as well as WakeMed Health and Hospitals in North Carolina, as customers. The firm has started a project with UMass to retrospectively analyze archived data on deterioration from respiratory failure and hemodynamic instability.
"At UMass, we tried to break it down into the different physiological systems that are deteriorating," Keler said. CLEW aims “to alert the physician that this patient is having a respiratory failure, a hemodynamic failure, maybe the combination of both. We are trying to highlight the patients that the model believes will require significant interventions in the near future," meaning the next four to eight hours, he explained.
Clew, thus, is training its analytical model to detect which patients might need mechanical ventilation, which could get by with noninvasive positive-pressure ventilation like BiPAP and CPAP, and which could be treated with vasoactive medications, Keler said.
In addition to alerting which patients are deteriorating, Clew’s technology tells physicians what kinds of interventions may be needed for the patients in the near future, he said, adding that this would not only allow for early treatment of serious cases, but possibly keep some patients out of the emergency department and the ICU.
Clew also plans on developing data models based on patient outside of the ICU setting. The challenge, Keler said, is to create datasets that can provide similar insights with data that is “not as rich as the ICU data.”
Clew is also working with other, undisclosed companies that offer and are developing continuous, noninvasive monitoring systems for patients in standard hospital wards as well as in ambulatory settings. "Those are new data feeds that are less explored," Keler said, though he did not specify any indications beyond the Mayo and UMass work.
He hopes to have the analytic model ready by the time noninvasive patient monitoring technology is in wide use.
"Because the technology is evolving and like all machine learning, the more data you have, the better your models are, so our customers will have the advantage that our models will constantly be evolving and improving compared to current, rule-based systems that are stationary in that area," Keler said.