Artificial neural network to predict post-operative hypocalcemia following total thyroidectomy

Area of Science:

• Endocrinology and Surgical Oncology
• Medical Artificial Intelligence
• Clinical Decision Support Systems

Background:

• Post-operative hypocalcemia is a common complication following total thyroidectomy.
• Accurate prediction of hypocalcemia and its severity is crucial for patient management.
• Current predictive methods may lack precision and comprehensive factor analysis.

Purpose of the Study:

• To develop and validate an artificial neural network (ANN) model for predicting post-operative hypocalcemia and its severity after total thyroidectomy.
• To determine the relative importance of various factors in predicting hypocalcemia using ANN.

Main Methods:

• A retrospective case series of 196 patients undergoing total thyroidectomy.
• Multilayer Perceptron (MLP) artificial neural network (ANN) for prediction and validation.
• Multivariate analysis to assess predictor weightage and construct validity.

Main Results:

• The ANN model achieved high accuracy, predicting hypocalcemia in 96.3% of test samples and severity in 92.8%.
• Mean incorrect predictions were low (3.18% training, 3.66% testing).
• Intact parathormone (iPTH) levels were identified as the most significant predictor (100% weightage), followed by the need for sternotomy (28.55%).

Conclusions:

• ANN models, particularly MLP, demonstrate significant potential for accurately predicting post-operative hypocalcemia and severity after total thyroidectomy.
• iPTH levels are a critical factor in predicting hypocalcemia post-thyroidectomy.
• While promising, the ANN model should complement, not replace, clinical expertise.