Development of a machine learning model for the diagnosis of atypical primary hyperparathyroidism

Background: Atypical primary hyperparathyroidism (PHPT), which includes normocalcemic and normohormonal variants, can be a diagnostic challenge. We sought to create a machine learning model to predict the probability of a patient having atypical presentations of PHPT. Methods: A model was constructed using logistic regression of PHPT patients and were compared to controls. Variables included sex, body mass index (BMI), calcium, PTH, 25-hydroxyvitamin D, phosphorus, chloride, sodium, alkaline phosphatase, and creatinine. The performance of the model was evaluated using the area under the curve (AUC). Results: The study included 4987 controls and 433 patients with atypical PHPT. Calcium, PTH, vitamin D, phosphorus, BMI, and sex were found to significantly contribute to the performance of the model, achieving an AUC of 0.999. The sensitivity, specificity, positive and negative predictive values were 92.9 %, 99.7 %, 96.3 % and 99.4 %, respectively. Conclusion: Machine learning can reliably aid in the recognition of PHPT in patients with atypical variants. Clinical relevance: When evaluating patients with atypical variants of primary hyperparathyroidism, the clinician needs to be able to identify subtle relationships in the patient laboratory test to make the diagnosis. These relationships can be found with machine learning and incorporated to predictive models which can ease and improve the diagnosis.