AI sucks at an awful lot of things. That includes but is no way limited to predicting sport wins, designing Halloween masks, and cracking jokes. But there is at least one thing it is remarkably good at and that is predicting when you are going to die.
Indeed, researchers at the University of Nottingham, UK, have built a machine-learning algorithm that can work out who will die prematurely with a 76 percent accuracy, making it better than current approaches, its creators say.
The study, published in the journal PLOS One, is based on previous research that found four AI algorithms ('random forest', 'logistic regression', 'gradient boosting', and 'neural networks') were better cardiovascular disease detectors than those used in hospitals today.
For their latest trick, they trained an artificially intelligent algorithm on medical data submitted to the UK Biobank between 2006 and 2010. This included demographic, biometric, clinical, and lifestyle information on more than 500,000 citizens aged 40 to 69.
Once training was complete, the algorithm was programmed to predict who from this group would die prematurely – and, rather impressively, it correctly identified 76 percent of the 14,500 participants who did by the time of the follow-up in 2016.
Next, the researchers compared its performance to those of two other models. One was a standard algorithm, the Cox model, and the other a simpler AI program that uses several tree-like models – hence its name, 'random forest'.
While all three took factors like age, gender, smoking history, and previous cancer diagnosis into consideration, the Cox model relied heavily on ethnicity and exercise data, which the other two did not. The 'random forest' model focused more on waist circumference, body fat percentage, diet, and skin tone, whereas the new model emphasized air pollution exposure, job-related hazards, alcohol consumption, and the risk of taking certain medications.
The new machine-learning algorithm came out on top, followed by the 'random forest' model at 64 percent, and the Cox model at 44 percent.
Though this all sounds a bit Bran Stark (aka the three-eyed raven), it's not all doom and gloom. The researchers hope that by better predicting those who are at risk of premature death, medics will be able to take preventative action.
"Preventative healthcare is a growing priority in the fight against serious diseases so we have been working for a number of years to improve the accuracy of computerized health risk assessment in the general population," lead author Stephen Weng, assistant professor of epidemiology and data science, said in a statement.
"Most applications focus on a single disease area but predicting death due to several different disease outcomes is highly complex, especially given environmental and individual factors that may affect them."