Clinical prediction rules are increasingly being published in the medical literature for a variety of clinical problems. Such rules aim to support physicians in decision making on diagnostics (e.g. ordering a test) or therapy (e.g. a risky surgical intervention) in individual patients. The rules contain prognostic factors that predict a clinical outcome (e.g. presence of disease, mortality). In current clinical practice, published prediction rules are not yet widely used. One of the problems of clinical prediction rules may be that clinicians doubt the validity and usefulness of the rule for their patient population. Time will have passed between treating the patients, data analysis, publication, and considering application of the rule. Treatment may have changed meanwhile, especially when the time delay is long. Further, predictive relationships may differ according to place, e.g. due to definitions of risk factors and differences in the patient population. In this project, we aim to study the problems that may be encountered when trying to apply a prediction rule to a patient in a specific hospital. Further, we aim to develop and apply methods for time and place adjustment of a prediction rule for such a specific hospital.
A prediction rule should at least be internally valid, which implies that it gives accurate (i.e. well-calibrated and discriminative) predictions for the patient population underlying the sample considered for the development of the rule. Internal validity is determined by epidemiological aspects (biases in study design, selection of patients, etc.) and statistical aspects (prediction method used, correction for overfitting). Internal validity will only imply external validity (validity in another patient population than that used to construct the rule), when the influence of time and place is negligible. To achieve external validity, adjustment may be required. In this project, methods of adjustment will be studied for a number of prediction problems. Individual outcome data will be analyzed from well-defined cohorts of patients from different centers, using advanced statistical techniques.
The project will consist of a methodological part and an applied part. In the methodological part we will address issues in validation and adjustment from an epidemiological and biostatistical perspective. After a literature review of methodological and clinical studies, a statistical approach will be developed in close collaboration with experts in this field. Simulations will be performed with the GUSTO-I data set, where 30-day mortality is registered for over 40,000 patients with an acute myocardial infarction. The participating centers in this trial have a wide geographical distribution, which enables study of the effect of place. For practical realisation, existing collaborations will be intensified, especially with Prof. F.E. Harrell Jr (Dept of Health Evaluation Sciences, University of Virginia, Charlottesville,VA, USA). Three working visits of 2 weeks each are planned for year 1, 2, and 3 of the project.
The results from the methodological part will he applied in actual clinical problems, which are selected on the basis of ongoing research and collaborations with national and international researchers. The applicant is involved in the projects of four recently appointed researchers (three PhD students, one scientific researcher) who concentrate on the development, application and validation of prognostic models. The proposed approaches to adjustment of prediction rules will be applied in the research topics in these projects. The topics include 1) prediction of histology in testicular cancer; 2) abdominal aneurysm surgical mortality; 3) syndrome approach to management of sexually transmitted diseases; 4) prediction of pregnancy chances in subfertile couples; 5) diagnosis of renal artery stenosis; 6) outcome after head injury.
Time will especially be relevant in topic 2), place in topics 3), 4), 5), and both time and place in topics 1) and 6). |
