Machine Learning-Derived Simple Score Predicts the Risk of Tertiary Hyperparathyroidism Requiring Surgical Treatment Among Kidney Transplant Recipients: The DPC score

Abstract

Tertiary hyperparathyroidism (THPT), a persistent elevation of parathyroid hormone (PTH) causing hypercalcemia in kidney transplant recipient, is associated with rapid bone loss and decline of kidney function. Give the central role of parathyroidectomy in THPT treatment, early and accurate prediction of surgical candidates for THPT may improve clinical outcomes after kidney transplant. We aimed to develop clinical models and simple-to-use integer-based score system to predict the risk of THPT requiring parathyroidectomy using machine learning algorithms. Models and integer-based score system were developed in derivation cohort (n=669, 75% training, 25% test set; kidney transplant recipient between 2009 and 2015, Severance Hospital, Seoul, Korea) and externally validated in KNOW-KT registry (n=542, multi-center registry between 2012 and 2016). THPT requiring surgical treatment was defined as individuals who underwent parathyroidectomy, those with calcium level higher than 11.5 mg/dL at any time point, or those who remained refractory (calcium level > 10.5 mg/dL) despite the use of cinacalcet after kidney transplant. Three key features (dialysis duration, intact PTH and calcium level prior to kidney transplant) were selected among 30 clinical features based on recursive feature elimination method to develop models using random forest, gradient boosted decision tree, support vector classifier, regularized logistic regression, and ensemble algorithm. Simple-to-use integer-based score system (DPC [Dialysis duration, PTH, and Calcium] score) was built based on beta coefficients from regularized logistic regression model, with weight-of-evidence method to allocate integer score for each category binned by tree-based method. THPT requiring surgical treatment occurred in 6.9% and 4.8% in derivation and external validation cohorts, respectively. Ensemble model showed best performance and good calibration among models to predict THPT requiring surgical treatment in both internal test set (AUROC 0.956, AUPRC 0.644, precision 0.82, recall 0.75, brier score 0.031) and external validation cohort (AUROC 0.894, AUPRC 0.509, precision 0.64, recall 0.54, brier score 0.032). DPC score ranged from 0 to 25 points (sum of each score, dialysis period [D] <5 months: 0, 5–34: 1, 35–84: 5, 85 or longer: 8; intact PTH [P] <120 pg/mL: 0; 120–159: 2; 160–519; 6; 520 or higher: 8; serum calcium [C] <8.0 mg/dL:0, 8.0–8.9: 3, 9.0–9.9: 5, 10.0 or higher: 9), which showed comparable performance with ensemble model in external validation (AUROC 0.905 vs. 0.894, p=0.579). One-point increment in DPC score was associated with 1.43-fold increased risk of THPT (95% CI 1.29–1.56), with precision 0.61 and recall 0.54 when high risk threshold was set at 21 or higher. Integer-based simple-to-use DPC score, developed using machine learning algorithms, predicted the risk of THPT in kidney transplant recipients.