A machine learning-based diagnosis modeling of IgG4 Hashimoto’s thyroiditis

• DOI: 10.1007/s12020-024-03889-y
• Publication Date: 2024-05-29T15:02:25Z
• Authors: Chenxu Zhao; Zhiming Sun; Yang Yu; Yiwei Lou; Liyuan Liu; Ge Li; Jumei Liu; Lei Chen; Sainan Zhu; Yu Huang; Yang Zhang; Ying Gao

This study aims to develop a non-invasive diagnosis model using machine learning (ML) for identifying high-risk IgG4 Hashimoto's thyroiditis (HT) patients.A retrospective cohort of 93 HT patients and a prospective cohort of 179 HT patients were collected. According to the immunohistochemical and pathological results, the patients were divided into IgG4 HT group and non-IgG4 HT group. Serum TgAb IgG4 and TPOAb IgG4 were detected by ELISAs. A logistic regression model, support vector machine (SVM) and random forest (RF) were used to establish a clinical diagnosis model for IgG4 HT.Among these 272 patients, 40 (14.7%) were diagnosed with IgG4 HT. Patients with IgG4 HT were younger than those with non-IgG4 HT (P < 0.05). Serum levels of TgAb IgG4 and TPOAb IgG4 in IgG4 HT group were significantly higher than those in non-IgG4 HT group (P < 0.05). There were no significant differences in gender, disease duration, goiter, preoperative thyroid function status, preoperative TgAb or TPOAb levels, and thyroid ultrasound characteristics between the two groups (all P > 0.05). The accuracy, sensitivity, and specificity were 57%, 78%, and 79% for logistic regression model of IgG4 HT, 80 ± 7%, 84.7% ± 2.6%, and 75.4% ± 9.6% for the RF model and 78 ± 5%, 89.8% ± 5.7%, and 64.7% ± 5.7% for the SVM model. The RF model works better than SVM. The area under the ROC curve of RF ranged 0.87 to 0.92.A clinical diagnosis model for IgG4 HT established by RF model might help the early recognition of the high-risk patients of IgG4 HT.