Prediction of transient hypocalcemia after total thyroidectomy using machine learning

Aim: We aimed to predict the occurrence of transient hypocalcemia in patients who underwent total thyroidectomy through machine learning. Materials and Methods: We examined 571 patients who underwent total thyroidectomy at the General Surgery Department of Çukurova University Faculty of Medicine between January 2015 and December 2021. The prediction scores of machine learning algorithms were examined using demographic, clinical, and laboratory data of patients. Results: Transient hypocalcemia was observed in 19.8% (n=113) of the patients. Incidental parathyroidectomy, tracheal injury, and recurrent laryngeal nerve injury significantly affected the occurrence of transient hypocalcemia (p<0.05). Preoperative calcium, ionized calcium, T3, postoperative calcium, and parathormone levels were significantly lower in the hypocalcemia group (p<0.05). The ROC-AUC score of logistic regression was found to be 0.610, and the accuracy score was observed to be 0.637 while the ROC-AUC score of the Linear Discriminant Analysis model was 0.602, and the accuracy score was 0.805. Discussion: Transient hypocalcemia after total thyroidectomy is a common complication of this surgery, and there are publications reporting that transient hypocalcemia can occur at a rate of 68%. In the literature, risk factors of developing postoperative hypocalcemia have been identified as weight, age, sex, lymph node dissection, hyperthyroidism, Hashimoto's thyroiditis, thyroid volume, incidental parathyroidectomy, postoperative calcium, and parathormone levels, and plunging goiter, whereas in our study, only incidental parathyroidectomy, postoperative calcium, and parathormone levels were found to be related to hypocalcemia. Linear Discriminant Analysis model has higher ROC-AUC and accuracy scores in predicting hypocalcemia compared to other models. Conclusion: Prediction of complications after surgery can be made through machine learning, while extensive databases will facilitate efficient integration of these methods.