Predictive Analytics in the Clinic
The ability to acquire and interpret vast amounts of health data has expanded dramatically in the past decades, leading to improved insights into a variety of medical conditions, ranging from cancer to autoimmune disease. Each breakthrough in scientific understanding, however, increases the medical knowledge clinicians are expected to absorb and apply to their patients. A powerful tool that helps clinicians manage this increased knowledge comes in the form of predictive analytics, which uses patient demographic and risk factor information to calculate the likelihood an individual will experience a particular health outcome or benefit from a certain intervention. Predictive analytics has been used, for example, to determine whether a woman is likely to be diagnosed with breast cancer in her lifetime. Despite their utility, as a recent JAMA Viewpoint article points out, predictive analytics models have an important limitation that clinicians must navigate to provide optimal care.
Using adverse cardiovascular events as an example, the article describes a weakness central to predictive analytics models: the risk calculated by these models applies to groups of people with similar characteristics, not to individual patients. As a result, even if a model can accurately predict that, as a hypothetical example, 10% of 60- to 75-year-old men with a family history of cardiovascular disease will suffer from a myocardial infarction within a year versus 1% of men aged 30- to 45, it will not be able to determine if a 65-year-old patient at hand will be one of those 10%. Because all of the information necessary to accurately determine whether an individual will experience an event is unknown, prediction models cannot reliably determine who in a group of similar individuals will ultimately suffer from an event, raising the question of how the medical field should proceed in incorporating predictive analytics into the management of patient care.
The authors suggest two approaches: the first involves increasing the amount of data used in formulating predictions, which can improve model accuracy, though a level of uncertainty will remain until all relevant information about the patient is known. Because of this residual uncertainty, the second approach authors requires clinicians to take into account patient characteristics not included in predictive analytics models to devise appropriate treatment strategies. Using both of these strategies in tandem will reduce the burden of new knowledge clinicians are expected to assimilate while ensuring patients receive evidence-based care.