Advancing Education, Research, and Quality of Care for the Head and Neck oncology patient.
Background: Primary Hyperparathyroidism (pHPT) is the leading cause of hypercalcemia and up to 75% of hypercalcemic patients go undiagnosed. Predicting primary hyperparathyroidism in data may facilitate earlier diagnosis and treatment. The purpose of this study was to examine the use of predictive modeling using a large clinical database to predict pHPT in patients with benign thyroid nodules.
Methods: This study was a retrospective analysis and predictive modeling of pHPT using a large discharge database. Data were from the Vizient Clinical Database (CDB), which is a large hospital discharge database from over 1,000 hospitals including academic health centers. We identified 2,541,901 patients with benign thyroid nodules between 2020 and 2023 in the Vizient CDB, of whom 83,555 (3.29%) had pHPT. The primary outcome measure was presence of pHPT, which was identified using ICD-10 codes. A predictive model of pHPT was created using logistic regression and compared to three machine learning algorithms: a Guassian naive Bayes classifier, a stochastic gradient descent classifier, and a histogrammic gradient boosting classifier. Model performance was compared using the area under the receiver operating characteristics (ROC) curve.
Results: In the baseline predictive model, several demographic characteristics were significantly associated with pHPT, including including age> 80 (odds ratio [OR]=2.8, p<0.0001), male sex (OR=0.66, p<0.0001), and Asian race (OR=0.48, p<0.0001). Hypotension was associated with a 30% lower odds of having pHPT (OR=0.71, p<0.0001). Chronic kidney disease was associated with twice the odds of pHPT (OR=2.13, p<0.0001). While use of bisphosphonates was associated with greater risk of pHPT (OR=2.90, p<0.0001), use of PPIs (OR=0.92, p<0.0001) and tobacco (OR=0.81, p< 0.0001) were associated with lower risk of pHPT. The logistic regression model had an area under the ROC curve of 68.1%, which was only slightly lower than that of the histogrammic gradient boosting model (68.7%) but equivalent to the gradient descent classifier (68.1%). Furthermore, the logistic regression model correctly classified 80.4% of pHPT cases, compared to 80.5% for both the histogrammic gradient boosting classifier and the gradient descent classifier. A threshold of 5% yielded a sensitivity of 38.5% and specificity of 81.8% for logistic regression.
Conclusions: Predictive modeling of pHPT among patients with benign thyroid nodules is possible using a large clinical database. This predictive model could be built into decision support systems to alert clinicians to potentially undiagnosed pHPT and aid in timely diagnosis and treatment of pHPT.
| Model | ROC-AUC | True Positive | False Positive | True Negative | False Negative | Accuracy |
|---|---|---|---|---|---|---|
| Logistic Regression | 68.1% | 1.3% | 17.6% | 79.1% | 2.0% | 80.4% |
| Gaussian Naive Bayes | 63.9% | 2.6% | 62.0% | 34.7% | 0.7% | 37.3% |
| Stochastic Gradient Descent Classifier | 68.1% | 1.3% | 17.4% | 79.3% | 2.0% | 80.5% |
| Histogramic Gradient Boosting Classifier | 68.7% | 1.3% | 17.5% | 79.2% | 2.0% | 80.5% |