One of the biggest challenges hospitals face is predicting when frail patients will decline into a life-threatening spiral. Subtle changes in health status get lost in a sea of data that is too vast for humans to effectively monitor. In a paper published Wednesday in the journal Nature, researchers at DeepMind describe a possible solution: A machine learning system capable of crunching hundreds of thousands of data points in electronic health records to alert physicians to an impending crisis long before it happens. To demonstrate the system’s potential, researchers at the London-based Alphabet artificial-intelligence subsidiary used it to predict the onset of acute kidney injury — a sudden decrease in kidney function — in hundreds of thousands of patients treated in Veterans Affairs hospitals across the U.S. They found the AI was able to predict 90% of these episodes that required subsequent administration of dialysis, with a lead time of 48 hours. The system was far from perfect: It reported two false positives for every accurate alert and its performance suffered when the injury was less severe. But the system outperformed by 20 percentage points an existing model used to assess the likelihood of kidney injury in hospitals, highlighting its ability to give clinicians more accurate warnings so they could intervene and possibly prevent harm. “The predictive nature of what they came up with is pretty impressive,” said Dr. Joe Bonventre, chief of the division of renal medicine at Brigham and Women’s Hospital in Boston who wasn’t involved in the study. He said getting 48-hour notice of acute kidney injury — when kidney function becomes compromised and waste products build up in the blood — could allow physicians to respond by regulating blood pressure and reducing the use of toxic medications that can contribute to a loss of organ function. Still, he said, “The question is always, will it apply to other populations in other settings outside Veterans Affairs hospitals?” The use of VA data limits the applicability of the findings in a couple of ways: The VA’s patients skew heavily male; only about 6% of the patients included in the training data were women. In addition, the VA, which operates its own electronic health record system, may also collect different data on patients than hospitals that use record-keeping systems supplied by private vendors. That might make it harder to replicate the AI’s performance in other facilities. The DeepMind researchers emphasized that their system still must be validated in clinical care so doctors can fully assess its impact on patient outcomes. But they said the early success of their approach points to a future where algorithms can be set loose on electronic health record data to help predict an array of crises in hospitalized patients, such as the onset of pneumonia, heart attack, or sepsis. “Giving doctors a head start on these major causes of patient deterioration that contribute to the deaths of literally hundreds of thousands of people every year could be absolutely transformative,” said Dr. Dominic King, clinical lead of DeepMind Health and a co-author of the paper. “The idea of proactive, preventive care that lots of us like talking about is more tangible when you look at this approach.” In the study, the AI system was trained on data from more than 700,000 VA patients treated in hospitals and outpatient clinics. It predicted the likelihood of kidney injury by looking for patterns in hundreds of thousands of data points contained in electronic health records. It is designed to track changes in patients over time, providing physicians with a continuously updated prediction of the likelihood a patient will suffer an acute kidney injury within the next 48 hours. The system can also pinpoint factors culled from the records that formed the basis for its conclusion. The DeepMind researchers said the next step is to validate the system’s performance on broader populations and in different clinical settings. It will likely be many months, if not years, before the AI could be sold commercially. Source
