the evolution of a malignancy risk prediction model for thyroid nodules using the artificial neural network

Background: clinically frank thyroid nodules are common and believed to be present in 4%to 10% of the adult population in the united states. in the current literature, fine needle aspirationbiopsies are considered to be the milestone of a model which helps the physician decide whether acertain thyroid nodule needs a surgical approach or not. a considerable fact is that sensitivity andspecificity of the fine needle aspiration varies significantly as it remains highly dependent on theoperator as well as the cytologist’s skills. practically, in the above group of patients, thyroidlobectomy/isthmusectomy becomes mandatory for attaining a definitive diagnosis where the majority(70%-80%) have a benign surgical pathology. the scattered nature of clinically gathered data andanalysis of their relevant variables need a compliant statistical method. the artificial neural networkis a branch of artificial intelligence. we have hypothesized that conduction of an artificial neural networkapplied to certain clinical attributes could develop a malignancy risk assessment tool to helpphysicians interpret the fine needle aspiration biopsy results of thyroid nodules in a context composedof patient’s clinical variables, known as malignancy related risk factors.methods: we designed and trained an artificial neural network on a prospectively formedcohort gathered over a four year period (2007-2011). the study population comprised 345 subjectswho underwent thyroid resection at nemazee and rajaee hospitals, tertiary care centers of shirazuniversity of medical sciences, and rajaee hospital as a level i trauma center in shiraz, iran afterhaving undergone thyroid fine needle aspiration. histopathological results of the fine needleaspirations and surgical specimens were analyzed and compared by experienced, board-certifiedpathologists who lacked knowledge of the fine needle aspiration results for thyroid malignancy.results: we compared the preoperative fine needle aspiration and surgical histopathologyresults. the results matched in 63.5% of subjects. on the other hand, fine needle aspiration biopsyresults falsely predicted malignant thyroid nodules in 16% of cases (false-negative). in 20.5% ofsubjects, fine needle aspiration was falsely positive for thyroid malignancy. the resilient backpropagation (rp) training algorithm lead to acceptable accuracy in prediction for the designed artificialneural network (64.66%) by the cross- validation method. under the cross-validation method, a backpropagation algorithm that used the resilient back propagation protocol - the accuracy in predictionfor the trained artificial neural network was 64.66%.conclusion: an extensive bio-statistically validated artificial neural network of certain clinical,paraclinical and individual given inputs (predictors) has the capability to stratify the malignancy riskof a thyroid nodule in order to individualize patient care. this risk assessment model (tool) can virtuallyminimize unnecessary diagnostic thyroid surgeries as well as fna misleading.