Abstract
OBJECTIVE: To develop a machine learning tool to integrate clinical data for the prediction of non-benign thyroid cytology and histology. CONTEXT: Papillary thyroid carcinoma is the most common endocrine malignancy. Since most nodules are benign, the challenge for the clinician is to identify those most likely to harbor malignancy while limiting exposure to surgical risks among those with benign nodules. METHODS: Random forests (augmented to select features based on our clinical measure of interest), in conjunction with interpretable rule sets, were used on demographic, ultrasound, and biopsy data of thyroid nodules from children younger than 18 years at a tertiary pediatric hospital. Accuracy, false-positive rate (FPR), false-negative rate (FNR), and area under the receiver operator curve (AUROC) are reported. RESULTS: Our models predict nonbenign cytology and malignant histology better than historical outcomes. Specifically, we expect a 68.04% improvement in the FPR, 11.90% increase in accuracy, and 24.85% increase in AUROC for biopsy predictions in 67 patients (28 with benign and 39 with nonbenign histology). We expect a 23.22% decrease in FPR, 32.19% increase in accuracy, and 3.84% decrease in AUROC for surgery prediction in 53 patients (42 with benign and 11 with nonbenign histology). This improvement comes at the expense of the FNR, for which we expect 10.27% with malignancy would be discouraged from performing biopsy, and 11.67% from surgery. Given the small number of patients, these improvements are estimates and are not tested on an independent test set. CONCLUSION: This work presents a first attempt at developing an interpretable machine learning based clinical tool to aid clinicians. Future work will involve sourcing more data and developing probabilistic estimates for predictions.