Machine learning and therapeutics 2.0: Avoiding hype, realizing potential

Clarifying the elements of an algorithm--and their distinctive impact--will be increasingly important if machine learning is to overcome skepticism among healthcare stakeholders. Machine learning algorithms must offer insights that are credible and aligned with the scientific or clinical consensus. An algorithm that fails to replicate established findings or counters the established body of evidence is more likely an indication of a methodological oversight or a data artifact than a truly novel insight. A pharmaceutical manufacturer recently described a scenario in which a machine learning algorithm concluded that reducing low-density lipoprotein cholesterol after a heart attack was not associated with cardiac outcomes. This finding does not change 20-plus years of established clinical science, but rather speaks to nuances in the data and analytic structure. Without such a context, machine learning could conclude that cigarette lighters cause lung cancer. This context is provided by domain-specific expertise. Its absence results in decisions that, while analytically sound, produce algorithms that are not likely to be adopted. For instance, a recent machine learning algorithm to predict cardiovascular events included "lack of data" as a key risk factor.12