Explainable machine learning for preoperative prediction of tumor focality in thyroid cancer using lifestyle and clinical examination data

Abstract

Multifocal thyroid carcinoma is tougher to anticipate before surgery, especially in regions with low resources, even though it is more likely to recur and needs more competent surgical planning. In order to tackle this problem, the proposed research would construct an interpretable machine learning model that could predict tumor focality (unifocal vs. multifocal) using only non-invasive preoperative clinical and lifestyle data. Three hundred eighty-three patients were employed as a balanced sample set for training the classifiers XGBoost, Random Forest, and Logistic Regression. Logistic Regression was the best model when compared to other models, with the most excellent recall (70.4) and ROC-AUC (0.726), both of which were substantially different (p<0.05). The HAP-based analysis found that the most relevant predictors were age, smoking history, and physical examination. The concept may be applied in outpatient clinical settings as it is straightforward to learn, particularly in institutions without access to histology or imaging. This paper indicates that tumor focality is foreseeable using easily accessible data, enabling a trustworthy decision-support tool for early treatment planning in the treatment of thyroid cancer.