London School of Hygiene & Tropical Medicine

A study and a comment piece published in the Lancet journal have investigated predictors of mortality over a five year period in a UK sample of nearly 500,000 people. The authors of the study used UK Biobank data and questionnaires to predict the risk of death in middle-aged and elderly people.

Commenting on this, Professor Stephen Evans, Professor of Pharmacoepidemiology at the London School of Hygiene & Tropical Medicine, said: "If everyone read the commentary carefully and therefore modified the tendency to overstate the results from the paper by Ganna and Ingelsson, then everyone would benefit.

"This large study shows the hazards of 'big data' as well as their limited benefits. A large number (but not a random sample of the population) of men and women have been followed up for nearly five years on average, starting between age 40 and 70. 2.3% of the men have died and 1.2% of the women have died. If you are not currently seriously ill, then it is clear that age is the best predictor of your (quite low) chances of dying in the next five years. If you are currently seriously ill, then, not unexpectedly, your chances of dying are greater than those of someone who is not ill.

"None of these findings help in determining the reasons why you are ill or why you will die. It is almost a pity that the first analysis of data from Biobank does not use the data for which Biobank was set up, though it is not yet ready for good analysis.

"The authors' claims that their scoring system has major advantages is over-stated. They say "Clinicians might use this score to identify patients at high risk of mortality to target with specific interventions". In fact there is no evidence that their scoring system adds anything to clinicians' judgement, knowing the age, gender and health status of their patients, since they are already reasonably good at prediction of mortality in a group, though prediction for the individual is much more difficult. This scoring system is also limited in its ability to predict in individual cases.

"They also say "Finally, government and health organisations can use this information to prioritise public policy to decrease the burden of specific risk factors." This contradicts what they have said earlier, that what they have studied is associations, not causally-based risk factors. Reliance on studies of this type in 'big data' can mislead people into placing more reliance on the analysis than is warranted.

"The practical gains of the prediction are not stated clearly. The validation using data from Scotland is useful, but shows the results are not highly generalisable. The emphasis in the presentation of the 'C-index' is of technical interest, but does not communicate to a general reader the impact of that index. The vast majority of people in this age range do not die in the next five years, and it would seem, based on a calculation from their data, that even those with high scores on their index, 90% or more will survive the next five years. As the commentators Thompson and Willeit note, there is a real danger of 'Cyberchondria' and it seems likely that the publicity around this paper may feed that."