Big data in healthcare – big errors lie ahead

A colleague recently said that the problem of non-adherence to prescribed medication will not be solved by getting individuals to take their medication, but will be solved by “big data”. “It is a big data solution” he said. I was so puzzled by what he said, and so shocked by the meaninglessness of the statement that the opportunity passed for me to ask him just what he had meant. I still have no idea what that statement means.

Data is simply a tool. Often a very powerful tool, but simply a tool none-the-less. The value of “big data” is often overstated and its ability to solve problems or help businesses prosper is rapidly becoming the stuff of legend and folklore. It turns out that Google cannot predict Influenza outbreaks by big data gleaned from search terms. Data allows the development of insights, understandings, and new knowledge. This is only useful if it is applied, in healthcare (and I daresay in other sectors as well), back to the benefit of the individual. Back to the level of “little data” if you like.

Fabulous insights around drug resistance in Tuberculosis, or medication adherence patterns stratified by all sorts of data from the quantified self are useless unless they impact on the treatment of individuals with drug resistant Tuberculosis. Or, contribute to individuals being better adherent to their medication regimens.

In looking for the insights and knowledge that may lurk in big data, a very deep understanding of statistical methods is required. Simply applying a statistical method to large sets of data and looking for correlations is likely, no, it will certainly lead to meaningless insights, fallacious understanding and serious errors. Consider this: tests for statistical significance accept a p value of 0.05 as being significant in healthcare. In summary, this means that there is a 5% probability that the result obtained was due to randomness, and not a real correlation.

Now apply this to big data, imagine someone trawling through massive data sets looking for correlations. They will certainly find some, and 5% will be random, simply chance findings. Which 5%? Your guess is as good as mine. Now if we take the “knowledge” thus gleaned and apply it to individual patients … well, that is when big data will result in big errors.