Machine learning models could save lives through personalized sepsis diagnostics

April 4, 2018
sepsis

Lawrence Livermore National Laboratory researchers Priyadip Ray (left) and Brenden Petersen and their teams, using machine learning algorithms, have developed computer models that can more accurately characterize a patient's progression through stages of sepsis and better predict mortality risk by integrating past medical history, real-time vital signs and other diagnostics. Photo by Charlie Hunts/LLNL (Download Image)

Machine learning models could save lives through personalized sepsis diagnostics

Jeremy Thomas, thomas244@llnl.gov, 925-422-5539

Researchers and clinicians may be able to track the progression of sepsis, a potentially life-threatening condition characterized by an extreme reaction to infection, with more precision and confidence using machine learning models developed at Lawrence Livermore National Laboratory (LLNL) in conjunction with health care provider Kaiser Permanente.

Typical diagnostic tools for sepsis used in clinics today rely on standardized threshold tests that take a single snapshot of a patient’s blood pressure, respiratory rate and other vital signs, which may not provide a full picture of risk or disease trajectory, researchers said. However, by using machine learning algorithms, scientists can more accurately characterize a patient's progression through stages of sepsis and better predict mortality risk by integrating past medical history, real-time vital signs and other diagnostics.

LLNL scientists presented work using machine learning algorithms called hidden Markov models at the recent Machine Learning for Health Workshop (ML4H) at Neural Information Processing (NIPS) in Long Beach. The paper was published online in January.

“It’s very important that patients are treated immediately in sepsis, because at later stages the mortality rate is very high,” said LLNL staff scientist Priyadip Ray, principal investigator on the project. “We came up with this model with memory, and we added one more thing: We said, ‘let’s not have a single set of model parameters for all patients -- it’s not a one size fits all.’ We bring in past information on the patient and allow for heterogeneity so you can have different model parameters for different patients.

“This model gives you a better idea of how the underlying disease states are changing,” he added.

To improve the model, the authors obtained and incorporated more than 20,000 sepsis patients’ vital signs and outcomes from Kaiser Permanente, adding multiple metrics to come up with individualized progression models. While the models haven’t been used clinically yet, researchers said the machine learning algorithms have the potential to substantially improve diagnostics and triaging, resulting in improved treatments for sepsis, which kills at least 250,000 Americans each year, according to the Centers for Disease Control and Prevention (CDC).

“Once a patient is admitted, vital signs will continually be recorded, and we can continuously run this model to provide a real-time analysis of risk state of the patient so that clinicians can take actions based on the patient’s health trajectory,” Ray said. “This kind of model is inherently more interpretable -- it can tell the doctor that a patient is trending a certain way based on this diagnostic data.”

The research, part of a Laboratory Directed Research and Development (LDRD) project to use data modeling to predict sepsis outcomes, was conducted by LLNL scientist Brenden Petersen (who designed and implemented the entire Python codebase) and included contributions from Lab summer intern and Cal State University Long Beach graduate student Kalvin Ogbuefi, former LLNL scientist and the project’s former principal investigator Michael Mayhew and research scientists Vincent Liu and John Greene of Kaiser Permanente, who received funding from the National Institutes of Health.

The project ended in September, but Ray said there is potential to include significantly larger datasets of patient records, not only including additional types of vital signs, but also laboratory test results such as white blood cell counts, to improve the model’s accuracy and utility.

“The idea would be to develop a framework to think about how to treat this patient so that he or she can quickly come to a healthy state,” Ray said. “Right now, it’s an open-loop model. Now we want to look at which drugs and treatments can bring patients to the desired state to close the loop.”

The LDRD project produced another paper utilizing the same patient data obtained from Kaiser Permanente, but with a different machine learning approach. Published in the Journal of Bioinformatics, LLNL researchers describe a model incorporating an array of different variables, including vital signs and static data such as age and sex, to look for patterns in sepsis mortality and indications that could preempt sudden patient health deterioration leading to death.

"We took an unsupervised approach to analyzing the Kaiser data, meaning that we tried not to make strong assumptions about what we expected the data to look like, and instead let the data speak for itself," Petersen said. "The model was predicated on the fact that patients exhibit considerable diversity: a 'normal' heart rate for you is not the same as a 'normal' heart rate for me. A good risk model needs to account for that."

The framework, called a composite mixture model (CMM), splits patients into clusters and identifies a variety of subgroups of patients with similar physiological characteristics upon hospital admission, and assessed mortality risk based on the patient’s similarity to others in that same subgroup. The researchers found that certain subgroups had a considerably higher risk of mortality, and that many patients even followed the same trajectory of subgroups.

“You can get sepsis from numerous starting points,” said LLNL researcher Ana Paula Sales. “It can set in for an elderly person with a general infection or a young, healthy person with a cut, so it’s highly variable. If you become septic, the mortality rate is really high, but if you can catch it early, the better the outcome. This model gets at being able to deal with all these types of variables.”

More than 1.5 million Americans are infected with sepsis every year, according to the CDC. The disease also impacts 19 million patients worldwide leading to as many as 6 million deaths, according to Kaiser Permanente’s Liu, who said several outstanding questions still need to be addressed to improve sepsis treatment.

“First, how early can sepsis be detected in patients? Earlier patient identification could enhance appropriate treatment and prevention. Second, how can we distinguish different subtypes of sepsis to allow for personalized treatment? Currently, our sepsis treatments are a one-size-fits-all approach,” Liu said. “The cutting-edge machine learning techniques used here are a step toward leveraging advanced computational approaches to combat sepsis. They maximize the use of highly complex and granular clinical data that can be developed to drive new insight into sepsis diagnosis and treatment.”

Sales and Petersen said they hope scientists will be able to use the models to create an “early warning system” that could immediately categorize patients into subgroups and alert physicians when a patient is at high risk for sepsis, thus enabling the physician to provide the best possible care. The researchers cautioned they are still “many steps away” from the machine learning framework being used in a clinical setting, It would require further funding and research into different types of data, including other vital signs, prescription drug use, previous health history and genetic data, to build a more robust model that could assist physicians in effectively triaging sepsis in an automated way.

“You can have high-risk patients that look very different from each other,” Sales said. “To say that a patient is high-risk could mean many things. I think this is a new way to look at sepsis prediction and really emphasizes the need to break down the sepsis population into smaller heterogeneous groups. It’s a good demonstration that this is the way to go… It’s high-stakes research that could have a real impact.”

LLNL researchers Sales and Petersen, Kaiser Permanente’s Liu and Greene and former Lab computer scientists Mayhew and Todd Wasson all contributed to the composite mixture model paper.

The two related papers, according to LLNL’s Director of Innovation Jason Paragas, represent part of a new toolset for biodefense, learning new lessons and strategies similar to the ones that helped in the fight against Ebola in 2014.

“AI and data science in general is quickly becoming our next medicines, and LLNL is leading the conversation by bringing the know-how,” Paragas said. “About 85 percent of sepsis patients can be diverted with early detection -- AI and machine learning with hospital records could be an important tool to achieve that goal.”