Health-Related Quality of Life Among US Adults with Long COVID: Development and External Validation of a Prediction Model

Long COVID is associated with substantial impairment in health-related quality of life (HRQoL), and social determinants of health may improve the identification of individuals with greater HRQoL burden. However, few prediction models have been externally validated, and the apparent contribution of social factors may be inflated when models include predictors that are conceptually close to the outcome. We aimed to develop and externally validate prediction models for poor HRQoL among adults with long COVID and to quantify the incremental contribution of social determinants after removal of predictors with potential construct overlap. Models were developed using the 2022 Behavioral Risk Factor Surveillance System (BRFSS) and applied without refitting to the 2023 cycle, in which long COVID ascertainment changed from a history of prolonged symptoms to current symptoms. Poor HRQoL was defined using the CDC Healthy Days measures. Logistic regression and 3 tree-based ensemble methods were evaluated across prespecified nested predictor sets, and incremental discrimination was estimated using paired bootstrap resampling among the same validation participants. Extreme gradient boosting had the highest discrimination, with an area under the receiver operating characteristic curve of 0.851 (95% CI, 0.838–0.865) in the internal test set and 0.844 (95% CI, 0.838–0.851) in the external validation cohort; the external calibration slope was 0.974. Adding the full set of social determinants increased external discrimination by 0.025 (95% CI, 0.021–0.029). After removal of predictors prespecified as overlapping conceptually with HRQoL, the incremental gain was reduced to 0.002 (95% CI, −0.002 to 0.007). Reintroducing perceived stress and life dissatisfaction alone restored an increase of 0.024 (95% CI, 0.020–0.028). This pattern was consistent across all 4 algorithms. In models that excluded functional-status measures, social connection contributed more incremental discrimination than material hardship (0.016 vs. 0.003). This study shows that poor HRQoL among adults with long COVID can be identified with good discrimination using routinely collected clinical, functional, and social characteristics, with performance maintained in an independent temporal cohort. Much of the apparent predictive contribution of social determinants was attributable to psychosocial and functional measures closely related to HRQoL, whereas social connection provided more independent predictive information than material hardship.