Why is it so hard to catch a worsening patient in time?
In hospital, some patients get worse without anyone noticing early enough. The authors note that an estimated 11% of deaths in hospital follow a failure to promptly recognise and treat a deteriorating patient. Care teams are busy, and the warning signs are scattered across dozens of test results and notes.
Acute kidney injury is a good example. It is a sudden drop in how well the kidneys work, common and sometimes dangerous. The earlier it is spotted, the better the chance of limiting the damage, for instance by changing a medicine or giving fluids. But it does not always announce itself with clear symptoms.
A tool that watches every record around the clock, and calls attention to the few patients who need a closer look, could help busy teams decide where to look first.
How was the AI built?
The team used deep learning, a kind of AI that learns from a very large number of examples. The model was developed on a large set of electronic health records covering many different settings: 703,782 adult patients, 172 inpatient sites and 1,062 outpatient sites.
The idea is to update each patient's risk continuously, every time a new test result or measurement arrives. The model also gives a confidence level, the list of items in the record that weighed most in its judgement, and an estimate of how some blood tests may evolve.
In practice, this looks like a risk indicator that updates itself throughout the stay, a bit like a weather forecast for the patient. It does not replace examining the patient: it draws attention to those who may need it first.
What results did the AI achieve?
The AI flagged 55.8% of all acute kidney injury episodes that happened during a hospital stay, with a lead time of up to 48 hours. For the most severe forms, those that later needed dialysis (a treatment that takes over the work of the kidneys), it flagged 90.2%.
There is a price to pay: for every true alert, there were two false alerts. A tool like this therefore has to be designed so that it does not swamp clinicians with useless messages.
Key figures
- 703,782adult patients whose records were used to develop the AI
- 55.8%of inpatient acute kidney injury episodes flagged, up to 48 hours ahead
- 90.2%of cases that later needed dialysis flagged
What could this change for patients?
Having a few hours, or a day or two, of warning gives time to act: reviewing a medicine that strains the kidneys, monitoring more closely, or asking a specialist for advice. The authors describe a time window that could allow early treatment.
Beyond the kidneys, the study is an example: the same kind of continuous monitoring could one day be used for other complications in hospital.
Whether this helps in practice depends on how the alerts are delivered and how teams respond to them, which the study does not test.
What are the limits of this study?
The AI did not flag every episode: a large share went unspotted, so it cannot be treated as a guarantee. The number of false alerts is high. Above all, the evaluation used records from the past: the study does not show that using it at the bedside actually reduces complications or deaths.
Trials in real conditions are needed, with clinicians receiving the alerts, to find out whether it truly improves patients' health.
In hospital: keeping track of what you are told, with Mima
This AI is an early-warning tool for the kidneys, studied on hospital records. It is not yet routine, and it is the care team that decides to review a medicine or watch you more closely.
In hospital, everything moves fast. Mima helps a loved one note treatments, explanations and the questions that come up. A useful one: "Are my kidneys checked regularly during my stay, and with which tests?"
Frequently asked questions
What is acute kidney injury?
It is a sudden drop in the kidneys' ability to filter the blood. It is common in hospital and can be serious if it is not recognised and treated in time.
Can this AI prevent kidney injury?
It does not prevent it: it flags a risk up to 48 hours ahead. It is then up to the care team to decide what to do.
Why are there so many false alerts?
To catch many true cases, the model accepts also flagging patients who end up fine. Here there were 2 false alerts for every true one.
Is it already used in hospitals?
The study shows an approach that could work, tested on past records. It does not yet prove a benefit for patients in real-world care.
The original study
Tomašev N, Glorot X, Rae JW, et al. “A clinically applicable approach to continuous prediction of future acute kidney injury” Nature, 31 July 2019. DOI : 10.1038/s41586-019-1390-1
Read the full study in NatureThis article summarises a scientific study and is not medical advice.