Medical risk stratification has received considerable attention in recent years, particularly in the US. This interest has been driven primarily by competition between provider organizations and the development of national performance standards. The authors of this book describe risk stratification as nothing more than a way of formalising the often heard claim that "our results are not as good as those of institution X down the street because we treat sicker patients". Given their focus I probably would have entitled the book Clinical risk stratification rather than Risk stratification which has a much broader meaning in epidemiology.
This book is a very practical guide to the process of clinical risk stratification and appropriately has been written by two cardiac surgeons and their statistical advisor. It is aimed at clinicians like the two clinical authors who want to understand their own data, and at other health professionals responsible for undertaking risk stratification for health care organizations. It also claims to be a reference text for a component of clinical research methods (i.e. clinical risk stratification) that is not well covered elsewhere.
Many of the individual chapters are excellent, particularly those focused on practical issues like data collection (Chapter 2) and the more technical aspects of analysis (Applying published risk estimates to local dataChapter 4, and Interpreting risk modelsChapter 6). However, I found the more methodological chapters, particularly Risk and published studies (Chapter 3), to be a little simplistic and naive from an epidemiological perspective. For example, the authors do not appear to understand the major similarities (both strengths and weaknesses) between cohort and case-control studies and that many, if not most, case-control studies are actually incidence studies.
The authors correctly point out that the only real difference between (clinical) risk stratification studies and efficacy-orientated clinical research is that the former focus on differences in outcomes between populations treated the same while the latter focus on differences in outcomes between groups treated differently. However, they go on to describe differences between classical clinical research and risk-stratification studies, that I believe are more apparent than real. For example, they state that classical clinical studies tend to focus on the effect of single factors while risk-stratification studies don't. However, any good quality clinical study will usually have considered the potential impact of multiple confounders or predictors simultaneously.
They discuss, in depth, the rationale for randomization in classical clinical studies of efficacy in order to deal with confounding. The potential for confounding in clinical risk stratification studies and need to be very cautious in the interpretation of all non-randomized research on interventions is also emphasized. However, I am concerned that most readers will focus on the how to do risk stratification components of the book and ignore the major potential pitfalls of non-randomized studies of therapies. In their enthusiasm for the subject I think the authors may have unintentionally overstated the ability of multivariate modelling to deal with confounding or to adjust for case mix, which is really just another way of describing the same problem. Indeed confounding may be a bigger problem in risk-stratification research than in classical clinical research given the wide variation in most clinical practice and the resulting wide variation in case mix in different settings. Moreover as most clinical risk stratification is done post hoc using data of limited quality that is generally collected for other purposes, the ability to assess even those confounders we know about is often poor.
Many of us who work in clinical research are becoming increasingly cautious about the results of non-randomized studies of therapy. The magnitude of the differences in findings between recently published randomized trials examining the effects of hormone replacement therapy on coronary disease risk and the previous best evidence from cohort studies was surprising and highlights our inability to fully adjust for potential confounders in non-randomized studies.
Whilst it is essential to attempt to adjust risk estimates for case-mix before making comparisons between the performance of different health care providers, I would want to attach a large health warning to any clinical risk stratification guideBEWARE OF RESIDUAL CONFOUNDING.