Literature

Reviews of the existing literature around predictive modelling of Thyroid cancer are detailed below.

Improving the diagnosis of thyroid cancer by machine learning and clinical data

Used a range of models (logistic regression, gradient boosting, linear discriminant analysis, support vector machine and random forest) to predict malignancy based on 18 predictor variables. Assessed via accuracy, precision, sensitivity and specificity and Area Under the Recevier Operating Characteristic (ROC).

Methods

10-fold cross validation (splitting data into ten subsets)

Performance

Model	Accuracy	AUROC	Sensitivity	Specificity	Precision
GBM	0.7741	0.8497	0.8750	0.5741	0.8029
Logistic	0.7834	0.8422	0.8352	0.6806	0.8384
LDA	0.7790	0.8394	0.8452	0.6477	0.8263
SVM (Radial)	0.7688	0.8237	0.8435	0.6206	0.8149
SVM (Linear)	0.7661	0.8200	0.8322	0.6349	0.8186
Random Forest	0.7931	0.8541	0.8629	0.6547	0.8321

Table 2

Figure 3

Variable Importance

Interesting approach using permutation prediction importance.

Six most important variables are shown in table 6 and are all features of the nodules, no other biological features such as gender or age.

Figure 4