Development and validation of an exercise adherence prediction model 107 backward elimination procedure was repeated to increase the likelihood of selecting variables that are genuinely related to the outcome. Variables that remained in the model in more than half of the bootstrap samples were included in the final prediction model. Model evaluation Model performance was assessed through discrimination (how well predictions differentiated participants who experienced the outcome from those who did not, quantified as the area under the receiver operating characteristic curve (AUROC), calibration (agreement between predicted and observed risk, assessed using calibration slopes, calibration-in-the-large, and calibration plots), and clinical utility (assessed using decision curve analysis and quantified as net benefit) [29]. An ideal calibration slope is 1, while calibration-in-the-large should be 0 if the number of observed outcome events matches the number predicted [29]. Decision curve analysis was used to calculate the clinical “net benefit” for the prediction model in comparison to default strategies of “treating” all or no patients [30]. In this study the benefit of the model is that it correctly identifies which patients are adherent and who are non-adherent. Preference refers to how healthcare providers value different outcomes for a given patient, a decision that is often influenced by a discussion between the healthcare provider and that patient [30]. Validity was assessed via bootstrapping (n = 500) to quantify any optimism in model performance. Adjustment for overoptimism of the overall performance enabled to better approximate the expected model performance in novel samples. Bootstrapping also estimated a uniform shrinkage factor to enable adjustment of the estimated regression coefficients for over-fitting [31]. When poorly calibrated predictions at validation were found, algorithm updating was considered to provide more accurate predictions for new patients [32]. An intercept adjustment was protocoled if the calibration intercept was not close to 0. Finally, the “optimal” cut-off value for the prediction model was calculated. The study was approved by the Ethical Committee Psychology of the University of Groningen (PSY-1920-S-0504). Results Participants From January 2021 until August 2022, patients from 53 different physiotherapy practices participated in the study. Out of 199 patients who gave informed consent, data from 196 patients were analyzed. The percentage of missing values across all 83 variables throughout the main study varied between 13.7% and 22.9%. In total 151-169 out of 196 patients had a complete data set. There was no association
RkJQdWJsaXNoZXIy MjY0ODMw