J. Imbus1, R. W. Randle1, S. C. Pitt1, R. S. Sippel1, D. F. Schneider1 1University Of Wisconsin,Department Of Surgery,Madison, WI, USA
Introduction: Most patients with primary hyperparathyroidism (PHPT) have a single adenoma (SA), but 20-25% of cases will have multigland disease (MGD). Preoperatative localization of SAs allows for a minimally invasive surgical approach, but these studies are less accurate or unnecessary in MGD. Therefore, pre-operative identification of MGD could direct the need for imaging as well as the operative approach, and potentially referral to experienced surgeons. Machine learning (ML) uses computer algorithms to build predictive models from labeled datasets. The purpose of this study is to use ML methods to predict MGD.
Methods: We reviewed a prospectively managed database of patients undergoing parathyroidectomy from 2001 to 2016. Patients (age ≥ 18 years) with PHPT who underwent initial, curative resection were included. MGD was defined as > 1 gland removed. Patients with genetic syndromes, a history of lithium use, prior neck surgery, or parathyroid carcinoma were excluded. The ML platform WEKA was utilized to compare different classifiers for predicting SA vs MGD from demographic, clinical, and laboratory features. The meta-algorithm, bagging, which reduces variance by averaging probability estimates, was applied. We selected the model with the best overall accuracy and separately used cost-sensitive classifier to maximize sensitivity for MGD. 10-fold cross validation was used to evaluate accuracy.
Results: 2035 patients met inclusion criteria: 1522 patients had SA (75%) and 513 had MGD (25%). After testing many algorithms, we selected the rule-based algorithm, PART, for its accuracy and potential integration in a clinical decision-support tool. Sample rules are shown in the figure. Using PART with bagging achieved 78% accuracy; 78% recall (sensitivity), 45% specificity, 76% precision (PPV), 0.710 Area Under the Receiver Operating Characteristics curve (AUC). To maximize sensitivity of detecting MGD, the cost-sensitive classifier achieved 89% sensitivity, 0.697 AUC for MGD. To validate the algorithm’s impact on practice, we reviewed imaging from a separate test set of 50 patients with MGD. The algorithm correctly identified 49 of these 50 patients (98%). Among these, 43 sestamibi scans and 14 ultrasounds were performed. However, only 14 sestamibi scans and 4 ultrasounds were correct. Eliminating the incorrect or non-localizing studies would have provided a potential cost savings of over $1200/patient.
Conclusion: Rule based ML methods can help distinguish SA from MGD early in the clinical evaluation to guide further workup including localization studies. ML can potentially save money spent on unnecessary imaging studies or guide referral to high volume surgeons who are comfortable with bilateral exploration for MGD.
