Author Affiliations: Department of Hygiene and Epidemiology, University of Ioannina School of Medicine, Ioannina, Greece (Drs Siontis and Tzoulaki); Department of Epidemiology and Biostatistics, Imperial College of Medicine, London, England (Dr Tzoulaki); and Stanford Prevention Research Center, Departments of Medicine and Health Research and Policy, Stanford University School of Medicine (Dr Ioannidis), and Department of Statistics, Stanford University School of Humanities and Sciences, (Dr Ioannidis), Stanford, Connecticut.
Van Dijk et al raise several important points regarding prognostic models of chronic obstructive pulmonary disease, which would probably generalize to many other diseases. Prognostic tools should be evaluated in several sequential stages: initial model performance (model development), prospective validation in independent cohorts (external validation of a model), impact on patient management and outcome, and cost-effectiveness.1,2 However, even for established and widely used prognostic tools, many of these steps suffer from methodological limitations and, in many cases, are missing. Van Dijk et al are correct to highlight the paucity of evidence around their impact on patient management and clinical outcomes. Such important evidence would ideally come from randomized control trials, which compare the outcomes of patients whose treatment is guided by the proposed prognostic tool with the outcomes of patients who are treated without it. However, there are so many prognostic tools, that it is impossible to evaluate all of them in randomized control trials. Efforts should focus around those with most promising results.3 In selecting which models to test in randomized trials, one may wish to consider not only satisfactory, validated discriminating ability, but also what is the respective change in disease management that can be anticipated; how effective are the available preventive or treatment interventions for the disease and how much room exists for improvement; what is the expected cost to get the information required for building the model and to implement it in practice; and how likely it is that the model can be used widely by nonexpert health practitioners. Going through such a checklist is likely to eliminate most of the proposed prognostic models.
Siontis GCM, Tzoulaki I, Ioannidis JPA. The Clinical Utility of Prognostic Indices: The Proof of the Pudding Is in the Eating—Reply. Arch Intern Med. 2012;172(2):194-195. doi:10.1001/archinte.172.2.195-a