Slashdot Mirror
AI Can Predict When Patients Will Die From Heart Failure 'With 80% Accuracy' (ibtimes.co.uk)
New submitter drunkdrone quotes a report from International Business Times: Scientists say they have developed an artificial intelligence (AI) program that is capable of predicting when patients with a serious heart disorder will die with an 80% accuracy rate. Researchers from the MRC London Institute of Medical Sciences (LMS) believe the software will allow doctors to better treat patients with pulmonary hypertension by determining how aggressive their treatment needs to be. The researchers' program assessed the outlook of 250 patients based on blood test results and MRI scans of their hearts. It then used the data to create a virtual 3D heart of each patient which, combined with the health records of "hundreds" of previous patients, allowed it to learn which characteristics indicated fatal heart failure within five years. The LMS scientists claim that the software was able to accurately predict patients who would still be alive after a year around 80% of the time. The computer was able to analyze patients "in seconds," promising to dramatically reduce the time it takes doctors to identify the most at-risk individuals and ensure they "give the right treatment to the right patients, at the right time." Dr Declan O'Regan, one the lead researchers from LMS, said: "This is the first time computers have interpreted heart scans to accurately predict how long patients will live. It could transform the way doctors treat heart patients. The researchers now hope to field-test the technology in hospitals in London in order to verify the data obtained from their trials, which have been published in the medical journal Radiology.
The point, as in TFA, is to suggest which patients should receive more intensive treatment. If there is a higher chance a patient is going to die, you would be more willing to apply riskier treatments with more potential to cause harmful side effects. If the chances are low, you then are more careful with treatments, because sometimes the treatments can do more damage if there is too low of a risk from the actual condition.
This is a common problem with most AI announcements. Is 80% accurate better than a simple statistical model? Often not. Does it scale up from a small sample size? Remember the recent face recognition thing that managed with only a hundred or so pixels? Sounded impressive, until you realise that the training set and the testing set were the same and that they only included around 1,000 faces, so simple information theory tells you that you only need 10 bits of information to identify each one and 800 bits doesn't sound quite so impressive.
I am TheRaven on Soylent News
Sounds like you need to get a better doctor.
Yeah. Generic Lipitor is $9 at Walmart. Generic Crestor is $15. KMart has generic Zocor for $3. That $.10-.$50 a day for the 3 dominate statins. That's not making any pharmaceutical companies rich.