Researchers tasked an AI computer with analysing 1.77 million electrocardiogram (ECG) – which monitor electrical activity of the heart – results from 400,000 people to see who had a higher risk of dying within the next 12 months. The team from US healthcare provider Geisinger in Pennsylvania trained two sets of AI algorithms. One of the algorithms was only given the ECG data, while the other was given the ECG data and the patient’s sex and age.
The team then measured the AI’s success using AUC – which examines how an algorithm distinguishes between two groups.
The two groups were divided into those who survived, and those who died within a year. Both algorithms consistently scored above 0.85 – with 1.0 being a perfect score.
Doctors, however, only scored between 0.65 and 0.8 and the AI was able to determine who was most at risk even when the healthcare professionals could not do so.
What is bizarre however is that the researchers are unsure how the AI is managing to do it.
Study lead author Brandon Fornwalt told New Scientist: “No matter what, the voltage-based model was always better than any model you could build out of things that we already measure from an ECG.
“That finding suggests that the model is seeing things that humans probably can’t see, or at least that we just ignore and think are normal.”
In August this year, a study from the Max Planck Institute for Biology of Ageing, Germany, showed that doctors were able to measure 14 metabolic substances in blood which can give a clear indication of whether you will live for more than another decade.
The study was able to predict the results with astonishing accuracy after analysing 44,168 people, aged between 18 and 109.
This figure dropped to around 72 percent when the individual was older than 60-years-old, however.
The team wrote in their paper published in Nature Communications: “We subsequently show that the prediction accuracy of 5- and 10-year mortality based on a model containing the identified biomarkers and sex is better than that of a model containing conventional risk factors for mortality.
“In combination, these biomarkers clearly improve risk prediction of 5- and 10-year mortality as compared to conventional risk factors across all ages.
“These results suggest that metabolic biomarker profiling could potentially be used to guide patient care, if further validated in relevant clinical settings.”