最优机器学习模型联合临床资料在桥本甲状腺炎诊断中的价值

Objective To explore the diagnostic value of different machine learning models and optimal machine learning model combined with clinical data for diagnosing Hashimoto′s thyroiditis (HT). Methods The thyroid gland images of 643 patients with 643 thyroid nodules who underwent preoperative ultrasound examination and had pathological results in the Second Affiliated Hospital of Xi′an Jiaotong University from December 2018 to March 2024 were retrospectively collected, and the images were divided into training set and test set according to a ratio of 7 to 3. Twenty ultrasound imaging omics models were constructed using pairwise combination of 5 feature screening components and 4 classifiers. The area under the curve (AUC) of each model in the test set was compared. Meanwhile, 3 basic network models were respectively used to construct deep learning models for diagnosing HT, and the diagnostic efficacies of the deep learning models and the ultrasound imaging omics models for HT were compared. The model with the greatest efficacy was selected as the optimal machine learning model. Further, the optimal machine learning model was combined with clinical data to construct a combined model. The ROC curves were plotted to compare the diagnostic efficacy of the optimal machine learning model and the combined model for HT. Results In the comparison of the efficacies of ultrasound imaging omics models and deep learning models in diagnosing HT, the efficacy of stable feature screening-logistic regression (LR) model was the greatest, and the accuracy, sensitivity and specificity of using the LR model in diagnosing HT in the test set were 78%, 75%, 74%, respectively, with an AUC of 0.82(95%CI=0.76-0.88). After combining the LR model with clinical data, the accuracy, sensitivity, and specificity of the combined model in the test set were 87%, 74%, and 95%, respectively, with an AUC of 0.91(95%CI=0.87-0.95), which was strongly consistent with pathology (Kappa value=0.708, P<0.001). Conclusions The optimal machine learning model (LR model) constructed in this study demonstrates a strong ability to diagnose HT and can accurately detect patients with atypical ultrasound manifestations of HT. The combination with clinical data can improve its diagnostic efficacy with higher accuracy and specificity.