more_vert
COMPUTER analysis of routine blood tests can predict which hospital patients are likely to experience a critical event the next day, according to preliminary Australian research.
Melbourne researchers used a computer program to examine six million blood tests taken from patients in the wards and emergency department of the Austin Hospital in the past 5 years, and combined this data with hospital records of critical events.
“We wanted to see if the blood tests contained patterns of abnormalities … to predict whether a patient would, the next day, die, go to intensive care, or receive a MET [medical emergency team] call”, said Professor Rinaldo Bellomo, director of intensive care research at Austin Health.
On average, the computer analysis was able to predict a critical event 10.2 hours before it occurred in a ward patient, and 11.9 hours beforehand for an emergency patient. The analysis included blood tests for urea and electrolytes, liver function, full blood count and arterial blood gasses.
Professor Bellomo said the analysis was 85% accurate at predicting which patients would die the next day.
“It’s kind of cool, kind of creepy … but I think it’s potentially useful”, Professor Bellomo said.
He acknowledged that knowing which patients were at risk may not improve their outcomes. However, improving knowledge of the timing and severity of risks could also assist decision making, such as in end-of-life discussions with families, he said.
The research was presented recently at the Australian and New Zealand College of Anaesthetists Trials Group workshop in Queensland.
Professor Bellomo said the next step was to assess whether having the predictive information available would change practice.
“Whether doctors and nurses in the wards already knew that these people were at risk — we don’t know that yet. What we do know is that we can identify people at high risk and we can do that in a way that hasn’t been done before”, he said.
Ultimately, the data could be incorporated into an electronic decision-assistance tool.
Dr Ross Kerridge, associate professor of anaesthesia and perioperative medicine at the University of Newcastle, said the study was an interesting example of the move towards using more “engineered” processes at hospitals.
“It’s a really exciting result … it’s a tantalising glimpse of the potential of cleverly designed information systems.”
Dr Kerridge said the computerised predictions could be used as an adjunct to clinical decision making.
“It can’t replace clinical decision making, but if the computer says someone is at very high risk then it’s a prompt for doctors to look at them again”, he said.
Using such sophisticated computer systems would make the hospitals of the future “safer, more efficient, more effective and better places to work”.
However, there would also be new dangers, such as if doctors relied on computerised systems at the expense of their clinical judgement, Dr Kerridge said.
Professor Bellomo said it was difficult to predict how doctors would react to this sort of predictive information. “They may rebel, burn the beeper, call the police, I don’t know.”
“But this has established that the information obtainable from the lab every day can be harnessed to identify those at risk, and now we have a case to take that a step further”, he said.
– Sophie McNamara
Posted 22 August 2011
Classic ICU staff blinded by data, and using computers to try and second guess the bedside doc.
Would it improve outcome? How many of these patients were terminal?
Most emergency docs can recognise seriously ill data, and the red flags from FBE U&E CRP LFT
I am flabbergasted – an article worthy of headlining April 1 news; a candidate for an Ig Nobel Prize.
It’s “85% accurate at predicting which patients would die the next day” – is it really? Or did it simply match (sort of) 85% of deaths in the sample with the blood test results. And out of this 85%, what proportion could Dr Blind Freddy have picked as having less than 24 hours to live? And what of the other 15%? – why didn’t their blood tests ring alarm bells? – or did they (with humans , that is).
And what is this “On average, the computer analysis was able to predict a critical event 10.2 hours before . . . .” ? Average? what sort of average? 10.2 hours – laughably precise – what is the standard deviation of the set of predicted times?
What a waste of resources – I am flabbergasted.
Prof Bellomo and his team would do better to use our scarce financial resources by asking a few thousand patients for their real life stories of accidents.
I have the same objection to this result being used as I do to electronic medical records. How many times have I sat opposite a doctor who hasn’t – and isn’t about to – read my previous history. And I’m sure this doesn’t only happen to me, tho I have a particularly long and tortuous history of multiple illnesses.
Objections: 1. Without the whole picture more mistakes could be made than without this info eg, I may have ended up with a stent which would have only made matters worse when I actually had takotsubo, but the noted results seemed to indicate a normal heart attack – and most practitioners have no idea what takotsubo is, tho it too is immediately life threatening.
2. Drs already complain all the time about how busy they are: They’re not going to attend to a third party notification.
3. I couldn’t count the number of times I’ve found out one Dr has written a referral giving completely incorrect information – they were only discovered by my questioning — and the possibility of this type of mistake (x by machine errors) will be more prevalent because the patient will not be privy to what is written, so will be unable to ask questions based on their own knowledge of their health history.
This depends on the medical team being aware of the results of tests ordered. I know of two patients who died as a result of the medical staff ordering a test and then not seeking the result. 1. A persistently falling haemoglobin not noticed till the patient had a massive and fatal bleed from a DU. 2. A blood pressure patient sent home after having a test for phaeochomocytoma. The positive result was not found till he came back in with a coronary occlusion.
He died in surgery.
Interesting.. The headline “Computers may predict deaths” is wrong and misleading. “Computers” are programmed by mere humans! Pooled data from populations has variable relevance to any given particular, individual patient. Sure this machine may be useful as an “adjunct to clinical decision [making]”, however machines have lots of fallicies & failures.
Since when has ‘man’s outcome’ been predicted by ‘machine’
Proverbs 3:5