Development and external validation of a prediction model for patient-relevant outcomes in patients with chronic widespread pain and fibromyalgia

Abstract

Background The objective of this study was to develop prediction models and explore the external validity of the models in a large sample of patients with chronic widespread pain (CWP) and fibromyalgia (FM).

Methods Patients with CWP and FM referred to rehabilitation services in Norway (n = 986) self-reported data on potential predictors prior to entering rehabilitation, and self-reported outcomes at one-year follow-up. Logistic regression models of improvement, worsening and work status, and a linear regression model of health-related quality of life (HRQoL), were developed using lasso regression. Externally validated estimates of model performance were obtained from the validation set.

Results The number of participants in the development and the validation sets was 771 and 215 respectively; only participants with outcome data (n = 519–532 and 185, respectively) were included in the analyses. On average, HRQoL and work status changed little over one year. The prediction models included 10–11 predictors. Discrimination (AUC statistic) for prediction of outcome at follow-up was 0.71 for improvement, 0.67 for worsening, and 0.87 for working. The median absolute error of predictions of HRQoL was 0.36 (0.22–0.51). Reasonably good predictions of working at follow-up and HRQoL could be obtained using only the baseline scores as predictors.

Conclusions Moderately complex prediction models (10–11 predictors) generated poor to excellent predictions of patient-relevant outcomes. Simple prediction models of working and HRQoL at follow-up may be nearly as accurate and more practical.

Significance Prediction modelling of outcome in rehabilitation has been sparsely explored. Such models may guide clinical decision-making. This study developed and externally validated prediction models for outcomes of people with chronic widespread pain and fibromyalgia in a rehabilitation setting. Multivariable prediction models generated poor to excellent predictions of patient-relevant outcomes, but the complexity of these models may reduce their clinical utility. Simple univariable prediction models were nearly as accurate and may have more potential for use in clinical practice.