Every year, one in five adults admitted to U.S. hospitals develop acute kidney injury (AKI), a serious condition in which the kidneys lose their ability to filter waste products from the blood. It’s even more common for patients in the ICU, more than half of whom develop the condition, putting them at great risk for permanent kidney damage and death.
Meanwhile, because of factors such as the aging of the population, increases in the number of surgical procedures and in conditions that increase susceptibility of AKI such as diabetes and hypertension, the incidence of AKI is increasing rapidly, doubling about every 10 years.
That’s the bad news. The good news is that AKI, which can occur in days or hours, can be halted or even cured for many patients—if it is discovered and treated soon enough. But the specific risk of conditions and exposures that lead to AKI are different not only between subpopulations of patients (e.g., cancer patients vs. cardiac surgery patients) but also between individual patients, making predicting AKI for these hospital patients complicated.
Mei Liu, Ph.D., associate professor of medical informatics in the Department of Internal Medicine at the University of Kansas Medical Center, is meeting that challenge head on. Liu and her colleagues have developed the first personalized artificial intelligence method to predict the risk of AKI in individual hospital patients. This new personalized approach, published in JAMA Network Open, outperformed existing one-size-fits-all methods for predicting AKI because it is designed to account for variations in different inpatient subgroups and between individual patients.
“AKI has many preventable risk factors, for example, nephrotoxic drugs,” said Liu. “So clinicians will be able to use the tool to identify high-risk patients and adjust their treatment strategies.”
The new method Liu and her team created with machine learning predicts the risk of AKI within 24 hours of onset, and can be adjusted to predict the risk within 48 hours.
Ultimately, the researchers hope their methodology could be implemented into hospital systems, such as electronic alert systems, to help clinicians protect patients from AKI.
A novel method with a jump start
To predict AKI, the new method uses machine learning, a form of artificial intelligence that enables computers to “learn” without being programmed. Large volumes of data — in this case data from the electronic health records (EHR) of hospital patients— are fed into a computer algorithm that identifies patterns in that data, and the algorithm then uses what it learned to make predictions about new data (new patients). Machine learning has been used to differentiate between images of healthy cells and tumors, and to predict diabetes and liver disease as well as AKI.
When Liu began working on developing a tool to predict AKI in inpatients in 2018, she initially expected to create a new and improved “global model.” Global models use predefined study cohorts; they capture knowledge that can be generalized to a population—such as hospital inpatients—but they ignore information specific to an individual or subpopulation.
“At first, we were doing what everybody else is doing—building a global model, but trying to achieve the best performance,” Liu said. “But then when we tested it on a subgroup [based on admission diagnoses], we found that its performance was not that equitable across different patient groups.”
There are some models to predict AKI that are based on subgroups, but the disease is so heterogeneous that exhaustive subgroup modeling is simply not possible.
The researchers decided to try a personalized modeling approach, which would utilize cohorts of clinically similar patients to assess the risk for AKI in a new patient. The problem—and the reason many personalized models never come to fruition—is that it can be difficult to find enough similar patients in one EHR system.
Liu and her team solved that problem by incorporating “transfer learning,” a new paradigm in machine learning that enabled them to transfer knowledge gained from a global model to “jump start,” as Liu put it, the personalized model and compensate for its smaller sample sizes.
A better predictor
The researchers used de-identified EHR data from The University of Kansas Health System for adult patients hospitalized for at least two days from November 2007 to December 2016, excluding those with moderate to severe kidney dysfunction at admission. The data variables included demographics, vital signs, medications, admission diagnoses and lab tests.
Their personalized model outperformed the comparison global model in predicting AKI by 2% in general hospital patients. It also bested the global model by 4% across 20 high-risk subgroups and by as much as 13% for certain high-risk patient subgroups such as patients admitted for liver transplant and cardiac surgery. Moreover, the personalized model also outperformed existing models based on patient subgroups.
Many other machine learning approaches function as ‘black boxes’ that would not, in this case, reveal the factors that contributed to an individual patient’s risk of AKI, noted Alan Yu, MB, BChir, director of the Division of Nephrology and Hypertension at KU Medical Center and an author on the study. When the personalized model generates a prediction for a particular patient, on the other hand, it is possible to identify every factor and the relative contribution of each to that patient’s AKI risk.
“This is potentially revolutionary,” said Yu. “Clinicians, armed with this information, could then tailor interventions to address only those risk factors specific for that patient, and thereby mitigate AKI risk or prevent it completely.”
Original source can be found here
Alerts Sign-up