An AI system was recently trained to evaluate a decade of generational health data submitted by over a half million people in the UK, then the AI system was tasked to predict if individuals were at risk of dying prematurely. Predictions were made by algorithms which were significantly more accurate than predictions delivered by a model that did not use deep machine learning, according to Dr. Stephen Weng.
To evaluate likelihood of mortality two types of AI were tested: 1) deep learning in which layered information processing networks help the computer to learn for examples; and 2) random forest which is a simpler type of AI that combines tree like models to consider possible outcomes. The two computer model conclusions were compared to results from a standard Cox model algorithm.
Using data from the UK Biobank containing health data of over 500,000 people between 2006 to 2016, during which time close to 14,500 participants died, the 3 models analyzed information to determine factors such as gender, age, smoking, and prior cancer diagnosis as being the top variables for assessing likelihood of a person’s early death, however the diverged over other key factors.
- A) Cox modeling learned towards physical activity and ethnicity while the other machine learning models did not. B) Random forest modeling emphasized body fat percentage, waist circumference, amounts of produce people consumed, and skin tone. C) Deep learning modeling had top factors including exposure to job related hazards, exposure to air pollution, alcohol intake, and use of certain medications.
When the algorithms had finished their modeling the deep learning algorithm was found to have delivered the most accurate prediction which correctly identified 76% of the participants who died during the timeframe; random forest model correctly predicted 64% of premature deaths; and the Cox model identified 44%.
This is not the first time the predictive power of AI has been harnessed, in 2017 experts demonstrated that AI could learn to spot early signs of Alzheimer’s disease by evaluating brain scans to predict likelihood of a person developing the disease; the algorithm correctly predicted with 84% accuracy.
A different study also found AI could predict the onset of autism in 6 month old babies who were at high risk of developing the condition. AI was found to be able to predict signs of encroaching diabetes through analysis of retina scans in another study. By using data derived from retinal scans artificial intelligence also predicted the likelihood of a patient experiencing a stroke or heart attack.
It appears as if with careful tuning AI deep learning algorithms can be used to predict a range of outcomes. Using artificial intelligence in this manner may be unfamiliar to many in the healthcare field, methods such as those used in this study may help with scientific verification and future development of this field, explains Professor Joe Kai.
