Predicting high-cost healthcare users: development and external validation of multivariable models using the HUNT and Tromsø studies linked to Norwegian health registries

Rising healthcare costs and demand call for better identification of individuals at risk of high-cost healthcare use. Few prediction models use detailed survey data or address persistent high-cost use in the general population. This study aimed to develop and externally validate prediction models for all-cause single-year and persistent high-cost healthcare use, and to assess whether adding survey data to administrative registry data improved performance. This was a prognostic study based on two population-based cohorts, the Trøndelag Health Study (HUNT4; model development) and the Tromsø Study (Tromsø7; external validation), linked to prospectively collected health registry data from primary and secondary care. Outcomes were (1) single-year high-cost use, defined as being in the top 25% of total healthcare costs in year one after survey completion, and (2) persistent high-cost use, defined as being in the top 25% in both years one and two. Predictors included self-reported sociodemographic and health-related variables and health registry data (prior-year costs and a morbidity index). Logistic regression models were developed for each outcome and internally validated via five-fold cross-validation. Model performance was assessed through discrimination and calibration. XGBoost models were trained and tested for benchmarking. External validation applied the developed models without refitting. We also developed and validated registry-only and survey-only models to compare performance against the full model. The development cohort included 42,049 individuals, and the external validation cohort included 20,942. In internal validation, the full logistic regression model achieved C-statistics of 0.79 (95% CI 0.78–0.79) for single-year high-cost use and 0.83 (95% CI 0.83–0.84) for persistent high-cost use. Corresponding C-statistics in external validation were 0.78 (95% CI 0.77–0.78) and 0.82 (95% CI 0.81–0.83). The models appeared well-calibrated on calibration plots. Full models showed significantly higher C-statistics than registry-only models (p < 0.001). Prediction models for identifying all-cause single-year high-cost and persistent high-cost healthcare use in the general adult population were developed and validated, showing good discrimination and calibration. The models can inform targeted preventive strategies and population health management. Incorporating self-reported survey data improved predictive performance, supporting the use of combining data sources for risk stratification.