Improvements in clinical prediction research

Created by W.Langdon from gp-bibliography.bib Revision:1.4420

  author =       "Kristel Josephina Matthea Janssen",
  title =        "Improvements in clinical prediction research",
  school =       "Utrecht, Universiteit Utrecht, Faculteit Geneeskunde",
  year =         "2007",
  address =      "Holland",
  keywords =     "genetic algorithms, genetic programming, clinical
                 prediction research, prediction models, derivation,
                 (external) validation, updating, logistic regression,
                 penalised maximum likelihood estimation, genetic
                 programming, missing values, multiple imputation",
  URL =          "",
  URL =          "",
  isbn13 =       "978-90-393-4668-6",
  size =         "160 pages",
  abstract =     "This thesis aims to improve methods of clinical
                 prediction research. In clinical prediction research,
                 patient characteristics, test results and disease
                 characteristics are often combined in so-called
                 prediction models to estimate the risk that a disease
                 or outcome is present (diagnosis) or will occur
                 (prognosis). This thesis focuses on the derivation,
                 validation, updating, and application of prediction
                 models. Dealing with missing values is an under
                 appreciated aspect in medical research. Three methods
                 were compared that can handle missing predictor values
                 when a prediction model is derived (complete case
                 analysis, dropping the predictor with missing values
                 and multiple imputation). Multiple imputation
                 outperformed both other methods in terms of bias,
                 coverage of the 90percent confidence interval, and the
                 discriminative ability. Similarly, six methods were
                 compared that can handle missing predictor values when
                 a physician applies a prediction model for an
                 individual patient with missing predictor values.
                 Multiple imputation proved to be best capable of
                 improving the predictive performance of the prediction
                 model, compared to imputation of the value zero, mean
                 imputation, subgroup mean imputation, and applying a
                 submodel consisting of only the observed predictors.
                 Many prediction models are derived with dichotomous
                 logistic regression analysis. Alternative methods are
                 logistic regression with inherent shrinkage by
                 penalised maximum likelihood estimation (PMLE) and
                 genetic programming (a novel and promising search
                 method that may improve the selection of predictors).
                 The effect of four derivation methods was compared,
                 namely logistic regression, logistic regression with a
                 single shrinkage factor, logistic regression with
                 inherent shrinkage by PMLE, and genetic programming.
                 The performance measures of the four models were only
                 slightly different, and the 95percent confidence
                 intervals of the areas mostly overlapped. The choice
                 between these derivation methods should be based on the
                 characteristics of the data and situation at hand. The
                 predictive performance of most derived prediction
                 models is decreased when tested in new patients.
                 Therefore, before a prediction model can be applied in
                 daily clinical practice, it needs to be tested (i.e.
                 externally validated) in new patients. However, when
                 the predictive performance is disappointing in the
                 validation data set, the original prediction model is
                 frequently rejected and the researchers simply pursue
                 to build their own (new) prediction model on the data
                 of their patients, thereby neglecting the prior
                 information that is captured in previous studies. The
                 alternative is to update existing prediction models.
                 The updated models combine the information that is
                 captured in the original model with the information of
                 the new patients. As a result, updated models are
                 adjusted to the new patients and thus based on data of
                 the original and new patients, potentially increasing
                 their generalisability. We show the effect of these
                 updating methods with empirical data, and give
                 recommendations for its application. This thesis ends
                 with an overview of the promises and pitfalls of using
                 electronic patient records (EPR) as a basis for
                 prediction research to enhance patient care, and vice
                 versa. The EPR are medical records in digital format
                 that facilitate storage and retrieval of data on
                 patient care. Though the primary aim of the EPR is to
                 aid patient care it creates highly attractive
                 opportunities for prediction research.",

Genetic Programming entries for Kristel J M Janssen