Is there such a thing as a machine learning "approach"?
Science has its rules, its guidelines for how to do things the right way. The goal is validity – logical and conceptual soundness that permit us to rely on the conclusions drawn from the data and its analysis. Does data science follow those same rules or, as part of a new paradigm in which data reign, are there new rules to be written?
The question is timely. Every area of our life and our economy is now affected by the analysis of data using machine learning. Health and medical research is no exception. Looking at PubMed, the database of the US National Library of Medicine, we see a rise from a handful of papers in PubMed that had “machine learning approach” in their title in 2000 to 15 in 2010 and 53 in 2016. Updating the figures to 2018, we see exponential increases.
These studies appear to explicitly claim to be using a new approach – a machine learning approach. Studies in medicine can quite simply be matters of life and death, where validity matters more than anywhere else. Is there a new way of doing science that is the "machine learning approach"? If so, we should be very clear on what the guidelines are for this new approach. Or do we simply have a new set of tools (new kinds of data and new algorithms) that are deployed in the practice of science?
For a fine exemplar of a good machine-learning-based study, see "Identifying Distinct Subgroups of ICU Patients: A Machine Learning Approach” [see link below], which offers the possibility of more efficient and tailored ICU care.
A careful examination of "machine learning approach" studies in medicine is perhaps overdue. Are good scientific practices being adhered to or does the allure of new tools and the appeal of a catchy title override? These are important questions to answer.
POSTSCRIPT:
For a really fun short discussion of "good science," see the UC Berkeley's Museum of Paleontology discussion at the link below.
For a really serious, extended discussion of the scientific enterprise, see Stanford University's Encyclopedia of Philosophy, starting with the entry below (and take note of the intriguing suggestion that "computer methods" are in fact a "third way" of doing science).
© Colleen Crangle, 2017