Predictive medicine in multiple sclerosis: A systematic review

•A systematic review of developed and/or validated a predictive model for MS.•Despite finding more than 6000 studies, 15 articles were retained.•An over-interpretation of association in terms of prediction in the MS literature.•A need to integrate good standards in developing and validating predictive models.•Validated predictive tools for MS management are currently lacking. One of the main challenges in multiple sclerosis (MS) is to predict disease progression based on patient characteristics and therapeutic strategies. We therefore performed a systematic review to critically appraise the composite tools available for this purpose. We performed electronic database searches in MEDLINE, EMBASE, Web of Science and the Cochrane Library. We included studies in English or French that developed and/or validated a predictive model for MS patients. Two reviewers independently screened articles by title and abstract. Three teams of two reviewers assessed the full text of each relevant study. Database searches yielded 6,035 studies after deduplication. Among the 42 screened full texts, 15 articles satisfied the eligibility criteria. Of these, six articles examined the development of predictive tools, six articles aimed to validate existing tools and three articles proposed both development and validation. We identified numerous methodological pitfalls, especially the lack of adequate validations in terms of discrimination and calibration. Only two scoring systems were externally validated several times: the Rio and the modified Rio scores. Nevertheless, their accuracies were highly variable, ranging from 65% to 91%. Overall, there is a lack of validated predictive tools in MS, and further external validation of the existing ones are required. Demonstration of the clinical usefulness is also needed prior to being transferred into clinical practice. Finally, our study illustrates that the MS literature needs to integrate good standards in developing and validating predictive models.