target menu
... / ... / ... / Our (associate) professors / Ewout Steyerberg
 

Ewout Steyerberg

Prof.dr.E. SteyerbergProfessor of Medical Decision Making

Personal web Page

Prediction models to support Decision Making
Predictions are relevant in many domains, ranging from the weather ('will it rain today?') to the stock market ('what will the future value of this company stock be?'). In medicine, both diagnosis ('what disease is present?') and prognosis ('what will be the outcome of this disease?') involve predictions. Predictions for individuals can be made by statistical models in which patient and disease characteristics are combined. The hope is that better predictions lead to better clinical decisions regarding diagnostic and treatment strategies, and finally to better outcomes.

Prediction models: from development to implementation
A number of ingredients are necessary to make better prediction models. Obviously, we need strong predictors. Much research efforts are currently focused on discovering new risk factors for developing disease and new prognostic factors for outcome of disease, including biomarkers and new imaging techniques. Second, appropriate statistical methods are needed to combine factors into prediction models. When new data become available, specific challenges lie in defining sensible strategies for improving existing models. A dynamic updating strategy arguably is attractive to a de novo development of a model in each specific setting. Third, prediction models are useless if they are not implemented in clinical practice. The ongoing automatization offers many opportunities for such implementation. Input for a model that predicts whether or not a person is likely to benefit from a screening test for the early detection of a disease can, for example, automatically be gained from this person's disease history and through questionnaires. In the hospital, prediction models can be incorporated in electronic patient records, so that predictions of diagnoses and prognoses are immediately available to the treating physician.

Media attention
The media often cover seemingly spectacular but highly speculative issues in predictive modeling. The predictive power of genetic characteristics is, for example, often overestimated. The empirical underpinning of such predictions is only in its early stages. Further, single and recognizable predictors are more easily picked up by the media than powerful but more complex models. In a study in which we aimed to predict who dies and who survives surgery among American patients with esophageal cancer, it appeared that black patients were much less likely to undergo surgery than white patients, even after statistical adjustment for many medical and non-medical factors. Rightly, quite some media attention was drawn to this example of "undertreatment of blacks". In contrast, much less attention was given to the finally developed, more refined multivariable model to predict outcomes of surgery.

Multidisciplinary activities
Prediction research requires efforts from different disciplines. First of all, medical professionals have to identify relevant questions, indicate which predictors are relevant in a specific context, and eventually implement prediction models in their daily practice. Further, new predictors will undoubtedly emerge from the basic sciences, including genomics, proteomics, and research on biomarkers. Biostatisticians and epidemiologists should contribute to methodological improvements in prediction modeling. Successful implementation requires support from medical informatics and medical psychology. Financial costs need to be considered together with health economists. With these joint efforts from various fields, my prediction is that prediction research will often be in the spotlight in the coming years.

Ewout Steyerberg