Prediction of thyroid cancer based on a multifactorial logistic model and the distribution of the linear predictor

Introduction. Altai region has high incidence of thyroid cancer (TC) which served as motivation for development of a multivariable logistic prognostic model based on clinical and amnestic patient data with determination of low, intermediate, and high risks of the disease. The use of the model in clinical practice can promote optimization of examination volume and formation of priority groups for in-depth diagnostics.

Aim. To evaluate diagnostic value of a multivariable logistic model for TC prognosis and analyze distribution of a linear predictor in various clinical groups for identification of areas of low, intermediate and high risks.

Materials and methods. Single-center retrospective study including 1463 patients with diseases of the thyroid was conducted. The main group consisted of 505 sequentially examined and surgically treated patients with morphologically confirmed TC; the control group included 958 patients without oncological pathology of the thyroid (operated due to suspicion of a tumor with subsequent morphological exclusion and selected during prophylactic examinations). The set of clinical-anamnestic, anthropometric, and phenotypic characteristics was formed based on the results of previously published study on TC risk factors in the Altai population. Using survey patient data, a set of characteristics Х₁–X₂₆ was composed. Statistical analysis included descriptive statistics, univariate logistic analysis with categorization of quantitative variables per ROC criteria, as well as development of a multivariate logistic model with stepwise predictor selection. For the final model, χ² test, area under the ROC curve (AUC), sensitivity and specificity at optimal threshold were assessed; internal validation was performed using the bootsrap method. For clinical interpretation, distribution of a linear predictor z was analyzed in the TC group, non-TC group, as well as in patients with diffuse and nodular diseases of the thyroid and subgroups with stage T1a–4 tumors.

Results. The final multivariate logistic model included a limited set of independent demographic, anthropometric, and anamnestic predictors and demonstrated satisfactory discriminating ability with acceptable AUC values per internal validation. Distribution of the linear predictor showed increase in z values in patients with TC. Small range of values of this parameter was predominantly observed in patients without TC, presence of marked "grey area" – in patients with diffuse and nodular thyroid diseases. Comparison of T categories did not show clear correlation between T stage and the value of the linear predictor.

Conclusion. The multivariate logistic model developed based in clinical and anamnestic data of the Altai patients allows to identify 3 areas of linear predictor z values corresponding to low, intermediate, and high TC risks. Its application in clinical practice can promote rationalization of examination volume and identification of priority groups for in-depth diagnostics in regions with high TC morbidity.