Artificial intelligence can predict risk of premature death

By Sally Robertson, B.Sc.Reviewed by Kate Anderton, B.Sc. (Editor)Mar 28 2019

Artificial intelligence may significantly improve the accuracy of premature death prediction, compared with standard models, report researchers at the University of Nottingham.

Min C. Chiu | Shutterstock

The team recently trained two artificial intelligence (AI) models to evaluate a decade’s worth of open-access health data and predict a person’s risk of dying prematurely from chronic disease.

The most advanced AI model accurately identified people at risk for premature death in 76% of cases, while a conventional model only predicted 44% of cases correctly.

Machine-learning significantly improved the accuracy of prediction of premature all-cause mortality in this middle-aged population, compared to standard methods.”

For the study, 502, 628 participants (aged 40 to 69 years) who submitted data to the UK Biobank between 2006 and 2010, were followed until 2016. Weng and team used two types of AI to assess the participants’ risk of premature mortality.

One type was “deep learning,” where layered information processing enables a computer to learn from examples. The other was a “random forest” algorithm that involved a combination of multiple branched models designed to account for various possible outcomes.

The results generated by the AI models were compared with those produced by a standard algorithm referred to as the Cox proportional hazards model, which links survival time to individual covariates.

All three models identified factors such as age, gender, previous cancer diagnosis and smoking history as important variables in assessing the risk of premature death. However, key differences were seen in how the models prioritized other factors.

The Cox model was heavily focused on ethnicity and physical activity, whereas the random forest model prioritized body fat percentage, waist circumference, fruit and vegetable intake and skin tone.

The deep-learning algorithm prioritized exposure to air pollution and job-related hazards, alcohol intake and the use of certain medications.

During the study period, almost 14,500 participants died, mainly as a result of heart disease, respiratory disease and cancer.

As reported in the journal PLoS ONE, the deep-learning model accurately identified the risk of premature death for 76% of patients who died during the study period, while the random forest algorithm identified 64% and the Cox model only 44%.

Weng says the study represents a major step forward in this field, providing a new and holistic approach to predicting premature death risk using machine learning.

Co-author Joe Kai says, although the methods may be unfamiliar to many healthcare professionals, they "could help with scientific verification and future development of this exciting field.”

This study illustrates the value of machine-learning for risk prediction within a traditional epidemiological study design, and how this approach might be reported to assist scientific verification.”