Active collaboration with Dr. Rhonda Szczesniak, Department of Biostatistics Epidemiology and the department of Pulmonary Medicine at Cincinnati Children’s Hospital Medical Center (CCHMC), Cincinnati, United States.
Objectives: To fit a univariate joint model (JM) of lung-function decline and pulmonary exacerbation (PE) onset and contrast its predictive performance with a class of multivariate JMs that include combinations of growth markers as additional submodels.
Study Design and Setting: Longitudinal cohort study on 17,100 patients aged 6-20 years (US CF Registry; 2003-2015). Primary outcomes included longitudinal lung function (FEV1), BMI, weight-for-age and height-for-age percentiles and onset of PE. Relevant demographic/clinical covariates were included in each submodel. We implemented a joint model of FEV1 and time-to-PE and four multivariate JMs including growth outcomes.
Results: All five JMs indicated a negative association between FEV1 and hazard of PE onset (HR from univariate joint model: 0.97, P < 0.0001), and all had reasonable predictive accuracy (AUC > 0.70 for cross-validations). Jointly modeling only FEV1 and PE onset yielded the most accurate (AUC = 0.75) and precise predictions (narrowest IQR from cross-validations). Dynamic predictions were accurate across forecast horizons (0.5-, 1-, and 2-years) but precision improved with age.
Conclusion: Including growth markers via multivariate JMs did not yield gains in prediction performance but joint modeling is a useful approach for monitoring CF disease progression.