Moving From Clinical Trials to Precision Medicine: The Role for Predictive Modeling.

Evidence-based care has been defined as “the conscientious, explicit, and judicious use of current best evidence in making decisions about the care of individual patients.”1 However, until now, most treatments have been designed with a “one-size-fits-all” approach: useful for some patients but not helpful or even harmful for others.2 Analyses of clinical trials generally focus on summarizing overall average treatment effects without more deliberate investigation of which patients actually benefit. For example, if the number needed to treat using a new therapy is 50, then 50 individuals need to receive this treatment for 1 individual to benefit. But what characterizes that benefiting individual? Therapies can also both help and harm, successfully improving some outcomes while also placing patients at increased risk for other adverse events. The Dual Antiplatelet Therapy (DAPT) trial provides an example of such complexity. The DAPT trial concluded that, among patients who had undergone a coronary artery stent procedure in which a drug-eluting stent was placed, those who continued thienopyridine therapy beyond 12 months, on average, had a reduced risk of ischemic events but at a cost of increased risk of moderate or severe bleeding.3 Such mixed results leave clinicians and patients in a quandary, struggling to know how these overall benefit and risk estimates apply to their particular situation. More sophisticated approaches to analytics and decision tools are critically needed. Precision medicine calls for the customization of health care, with medical decisions tailored to the individual patient.3 Sometimes precision medicine can identify a single variable such as a gene or biomarker that can successfully differentiate individuals who benefit or are harmed by a given treatment. However, in many situations, the outcomes of intervention are associated with multiple variables. In these instances, statistical risk prediction models can estimate the likely implications of a therapeutic intervention and thereby assist medical decision making. Specifically, these models can simultaneously aggregate multiple patient characteristics into a simplified “risk prediction score” that can provide individualized probabilities of outcome with or without treatment. In this issue of JAMA, the study by Yeh et al4 describes an interesting application of predictive modeling as a means of better interpreting a clinical trial, in this case the DAPT study. Their analytic goal was to identify which patients who had received drug-eluting stents and were assigned to receive extended thienopyridine treatment (relative to without receiving extended treatment) had the most (or least) favorable absolute benefitrisk ratio. Specifically, the authors built predictive models estimating 4 distinct probabilities: the risk of ischemic events and the risk of bleeding if thienopyridine was extended, and similar ischemic and bleeding risk estimates if thienopyridine was not extended. If both ischemic and bleeding risks were reduced with extended treatment, then the obvious decision is to extend treatment with dual antiplatelet therapy. Conversely, if both risks were increased, stopping treatment would be the correct decision. However, many patients are in this category in which extending thienopyridine decreases the risk of ischemic events but also increases the risk of bleeding. Yeh and colleagues used linear regression to reduce this 4-dimensional prediction problem to a single dimension. Specifically, they developed a simplified prognostic tool that simultaneously identified 9 independent clinical factors that best maximized the absolute benefit-risk difference for each patient. These findings are important for several reasons. First, in the spirit of precision medicine, the study by Yeh et al moved the focus from a single overall conclusion for the DAPT trial to a decision analytic approach that recognizes patient diversity in response to therapy. Second, when investigating therapeutic heterogeneity, the authors employed an analytic framework that simultaneously considered more than a single factor at a time (eg, young vs old, or diabetes vs not diabetes). Third, the DAPT risk score (−2 to 10) seems to achieve its goal of differentiating patient subpopulations in whom benefits of treatment outweigh risks and vice versa. When applied in the original sample of 11 648 patients, of whom 348 developed ischemic events and 215 developed bleeding events, patients with a DAPT score of 2 or higher (n = 5917) who continued thienopyridine therapy vs placebo had a lower risk of ischemic events (2.7% for continued thienopyridine vs 5.7% for placebo) and similar risk of bleeding (1.8% for continued thienopyridine vs 1.4% for placebo). Among the 5731 patients with a low score (<2), those who continued thienopyridine therapy vs placebo had a similar risk of ischemic events (1.7% for continued thienopyridine vs 2.3% for placebo) and higher risk of bleeding (3.0% for continued thienopyridine vs 1.4% for placebo). Fourth, the authors took the additional important step of evaluating the generalizability of their risk score in an external sample. The study by Yeh et al also has several important limitations. First, any analysis is only as strong as the data on which it is derived. The DAPT score was derived from a population of patients selected to participate in a clinical trial and may differ from patients treated in routine clinical practice. Second, since the DAPT trial was originally designed, management practices in interventional cardiology have evolved. Related article page 1735 Opinion