Diagnostic Performance of Routine Biochemical Markers and Internally Validated Multivariable Models in Primary and...

The aims of this study were to compare the biochemical characteristics of primary hyperparathyroidism and normocalcemic primary hyperparathyroidism and to evaluate the diagnostic performance of routinely available biochemical markers using validated patient-level multivariable statistical modeling. This retrospective multicenter study screened 1,832 individuals between January 2018 and December 2023. Baseline descriptive characteristics were obtained from an initial retrospective clinical cohort comprising 1,575 patients. Following the verification of the original database, duplicate encounters were consolidated, and all inferential statistical analyses were reconstructed at the patient level using the first eligible encounter for each patient. The final analytical cohort consisted of 375 patients (177 primary hyperparathyroidism and 198 normocalcemic primary hyperparathyroidism). Multivariable logistic regression with patient-level stratified cross-validation, bootstrap resampling and permutation testing, along with Spearman's correlation analyses, was performed. The effect of the synthetic minority over-sampling technique was evaluated using in-fold resampling. In the verified patient-level cohort, the multivariable logistic regression model demonstrated modest discrimination between primary hyperparathyroidism and normocalcemic primary hyperparathyroidism area under the curve: 0.578; 95% confidence interval: 0.516-0.639). Permutation testing indicated a statistically significant but limited diagnostic signal (p=0.02). Application of the synthetic minority over-sampling technique did not materially improve model performance (area under the curve: 0.577). No significant association was observed between the parathyroid hormone and 25-hydroxyvitamin D in the overall cohort (Spearman's ρ=- 0.079, p=0.185) or within the primary hyperparathyroidism and normocalcemic primary hyperparathyroidism subgroups. An exploratory visit-level screening model based solely on laboratory test-ordering patterns achieved higher discrimination (area under the curve: 0.733), suggesting that laboratory test-ordering patterns contributed to the observed classification signal. Primary hyperparathyroidism and normocalcemic primary hyperparathyroidism exhibit overlapping biochemical characteristics that provide only modest patient-level discrimination when evaluated using routinely available laboratory variables. Multivariable predictive modeling demonstrated limited diagnostic performance after rigorous patient-level validation, and the synthetic minority over-sampling technique did not improve model performance. These findings underscore the importance of patient-level validation and standardized analytical pipelines when developing predictive models using retrospective clinical data sets. Prospective studies with standardized biochemical assessment and external validation are required before such predictive models can be considered for routine clinical applications.