Diagnostic and prognostic models: applications and methods

Prediction modelling, both diagnostic and prognostic, has become a major topic in clinical research and practice. Traditionally, clinicians intuitively combine and judge the documented patient information, on e.g. risk factors and test results, to implicitly assess the probability or risk of having (in diagnostic estimations) or developing (in prognostic estimations) for certain diseases or outcomes. In this thesis, various clinical prediction models, both diagnostic and prognostic, were developed using empirical data. First, a prediction model of major depressive disorder in primary care patients was proposed. This model allows primary care physicians to select patients at high risk for this disorder. Second, prediction modelling was used to identify the best prognostic predictors for classifying the herpes zoster patients at highest risk for the development of postherpetic neuralgia. Third, prognostic predictors for short term and long term complications in patients with a pacemaker were studied. Even though we identified several effective predictors for the last two applications, these models have to be evaluated further before the results can be applied in practice. In recent years there has been an increasing interest in the methodology of prediction research. This research resulted in methodological recommendations for the design and analysis for prediction studies. The reporting of design and analysis issues and to what extent recommendations were followed in recently published prediction research was studied in a systematic review. Positive aspects of prediction research were the presence of relatively many prospective designs, adequate description of predictors and good reporting of the selection of predictors. Improvement is notably needed in blinded assessment of predicted outcomes, handling of continuous predictors, the investigation predictor interactions, reporting on statistical power, the reporting of the amount of missing data, the presentation of the results of multivariable analysis, and the methods used to quantify and validate the predictive performance of prediction models. Alternative methods for the development of prediction models with a limited effective sample size were also evaluated. The first method was the reduction of the number of predictors with principal components analysis. This method showed similar performance as other advanced methods for model development. If a strong predictor dominates the predictive performance, however, principal components analysis may not be a good option for variable reduction. The second method was aimed at dichotomous outcomes that were derived from a continuous variable. Here, the development of prediction models for the dichotomous versus the continuous outcome was evaluated. The two types of models showed a very similar performance in data with a large effective sample size. In smaller samples, the analysis of continuous outcomes is recommended to prevent development of too optimistic and overfitted prediction models.