Improved Statistical Methods are Needed to Advance Personalized Medicine

Abstract

Common methods of statistical analysis, e.g. Analysis of Variance and Discriminant Analysis, are not necessarily optimal in selecting therapy for an individual patient. These methods rely on group differences to identify markers for disease or successful interventions and ignore sub-group differences when the number of sub-groups is large. In these circumstances, they provide the same advice to an individual as the average patient. Personalized medicine needs new statistical methods that allow treatment efficacy to be tailored to a specific patient, based on a large number of patient characteristics. One such approach is the sequential k-nearest neighbor analysis (patients-like-me algorithm). In this approach, the k most similar patients are examined sequentially until a statistically significant conclusion about the efficacy of treatment for the patient-at-hand can be arrived at. For some patients, the algorithm stops before the entire set of data is examined and provides beneficial advice that may contradict recommendations made to the average patient. Many problems remain in creating statistical tools that can help individual patients but this is an important area in which progress in statistical thinking is helpful.