This study employed a prospective cohort design, conducting a 12-month longitudinal follow-up of all enrolled T2DM patients, with follow-up time points at baseline, 3 months, 6 months, and 12 months. Data were collected through telephone or outpatient follow-up visits, with a dropout rate of 4.2% (21/500), primarily due to patient migration or refusal to continue participation. Treatment plans for all patients (oral hypoglycemic agents or insulin injection) were standardized according to the Chinese Guidelines for the Prevention and Treatment of Type 2 Diabetes (2017 Edition) (19) to minimize the impact of treatment heterogeneity on the results. The detailed study flowchart is shown in Figure 1 .

Patients diagnosed with T2DM admitted to the Ninth People’s Hospital of Suzhou from May 1, 2023, to April 30, 2024 were recruited, with specific inclusion and exclusion criteria as follows:

Inclusion criteria: i) diagnosis of T2DM regarding the diagnostic criteria outlined in the Chinese Guideline for the Prevention and Treatment of Type II Diabetes Mellitus (2017 Edition); ii) patients with FT were diagnosed according to the diagnostic criteria outlined in the Chinese Guideline for the Diagnosis and Treatment of Thyroid Diseases-Hyperthyroidism (20), confirmed by thyroid color Doppler ultrasound, and excluded cases of acute or subacute thyroiditis and Hashimoto’s thyroiditis; iii) patients aged between 18 and 65 years old; iv) patients confirmed with FT had not received treatment with anti-thyroid medications prior to study participation.

Exclusion criteria: i) patients with severe cardiovascular, cerebrovascular, or vascular diseases; ii) patients with severe organ dysfunction, including liver or kidney impairment; iii) patients who received treatment for relevant diseases within the three months prior to study participation, including treatment with iodine preparations for FT; iv) patients who experienced acute metabolic disorders such as diabetic ketoacidosis within the three months prior to study participation; v) patients with malignant tumors, immune system disorders, or other endocrine disorders; vi) patients with psychiatric disorders; vii) patients with infectious diseases.

Based on the aforementioned criteria, 500 T2DM patients were included. Of which, 60 patients developed FT and were categorized into T2DM+FT group, while the remaining 440 patients did not develop FT and were categorized into T2DM group. Additionally, 200 healthy individuals undergoing routine health examinations during the same period were recruited as controls. Sample size calculation was based on previous studies (21), setting a significance level of α = 0.05 and a power of 1-β = 0.8. The expected incidence of FT in T2DM patients was 12%, and the estimated total sample size required was 480 cases. A total of 500 T2DM patients (including 60 with T2DM+FT) were actually enrolled, meeting the statistical requirements. Although the sample size in the T2DM+FT group was relatively small, the risk of overfitting was reduced through LASSO regression and cross-validation. All participants provided informed consent and agreed to participate in this work. This research adhered to the ethical standards outlined in the Helsinki Declaration and was approved by the Ethics Committee of the Ninth People’s Hospital of Suzhou (Approval Number: KY2023-029-01).

Basic information (including age, gender, height, and weight), disease characteristics (including duration of diabetes), and details of diabetes treatment (including types of medication) were collected through clinical records and other data sources for all patients.

Fasting venous blood samples of approximately 5 mL were collected from each patient. Following blood collection, samples were centrifuged at 4,000 rpm for approximately 3 minutes using an Eppendorf 5418R centrifuge (Eppendorf AG, Germany, Catalog No: 5401000516). After centrifugation, serum portion was separated using pipettes and stored at -80°C in a Thermo Scientific Forma 900 freezer (Thermo Fisher, USA, Catalog No: 900-005) for subsequent research.

The serum samples prepared in Section 2.4 were analyzed employing an automated biochemical analyzer with corresponding reagents (Beckman Coulter AU480, Beckman Coulter, USA, Catalog No: A11857) to measure fasting blood glucose (FBG), 2-hour postprandial blood glucose (2hPBG), and hemoglobin A1c (HbA1c) levels for all patients. The normal value for FBG was defined as 5.6 mmol/L, with a diagnostic threshold for diabetes set at 7.0 mmol/L. The normal value for 2hPBG was below 7.8 mmol/L, while a value above 11.1 mmol/L was used as the criterion for diagnosing diabetes. Additionally, fasting insulin levels were measured using an automated electrochemiluminescence immunoassay (Roche Cobas e411, Roche Diagnostics, Switzerland, Catalog No: 07027742190) to calculate the homeostasis model assessment of insulin resistance (HOMA-IR).

Serum samples collected in Section 2.4 were analyzed for serum NRG4 levels using the enzyme-linked immunosorbent assay (ELISA). The NRG4-ELISA kit was purchased from R&D Systems (USA Catalog No: EK713197), and experiments were conducted strictly regarding the manufacturer’s instructions. The concentration of NRG4 in samples was calculated using the standard curve method.

Serum samples prepared in Section 2.4 were analyzed employing automated electrochemiluminescence immunoassay to measure thyroid hormone levels, including free triiodothyronine (FT3), FT4, and thyroid-stimulating hormone (TSH) levels. The kit was Roche Diagnostics, Switzerland, Catalog No: 07027742190.

After the informed consent was signed by the patients, a preoperative assessment was conducted to ensure that the patients were suitable for thyroid puncture. Thereafter, thyroid tissue was obtained under ultrasound guidance using fine-needle aspiration biopsy (FNA), with an adequate amount of tissue samples (1–2 mL) being collected. An ultrasound-guided procedure was performed using a GE Logiq E9 ultrasound diagnostic device (GE Healthcare, USA).

The acquired thyroid tissue samples were immediately placed in a 10% formalin fixative (Fisher Scientific, USA, catalog number: 102-097-121) for a fixation period of 24 hours. After fixation, the samples underwent dehydration and clarification processes before being embedded in paraffin to prepare tissue sections (4-5μm) for hematoxylin-eosin (H&E) staining and immunohistochemistry (IHC) examination.

Paraffin-embedded tissue sections were retrieved and deparaffinized using a graded series of hydration. Subsequently, the sections were stained with HE staining, beginning with placement in hematoxylin solution (Sigma-Aldrich, USA, catalog number: H3136) for 10 minutes, followed by rinsing with deionized water. The sections were then differentiated in hydrochloric acid alcohol (Sigma-Aldrich, USA, catalog number: 320501) for 1–3 seconds, rinsed again with deionized water, and transferred to eosin staining solution (Sigma-Aldrich, USA, catalog number: E4009) for 1–3 minutes, followed by a gentle rinse with deionized water. Finally, the sections underwent dehydration, clarification (using xylene, Sigma-Aldrich, USA, catalog number: Xylene), and mounting. The stained sections were observed under a light microscope (Olympus BX53, Japan) to record the morphology, size, and arrangement of the observed tissue cells.

The sections were baked in an oven (Thermo Scientific, USA, model: Heratherm OGS) at 60°C for one hour, deparaffinized with xylene (Sigma-Aldrich, USA, catalog number: Xylene) and ethanol (Fisher Scientific, USA, catalog number: A956-1), and then rinsed with distilled water. After being placed into 0.01 M citrate buffer (Sigma-Aldrich, USA, catalog number: C9427, pH 6.0), the sections were heated in a pressure cooker until boiling and maintained for 15 minutes. They were then quickly cooled in ice water for 5–10 minutes to terminate the reaction. After antigen retrieval, the sections were washed three times with PBS (Gibco, USA, catalog number: 10010-031) to remove unbound substances. A specific NRG4 antibody (Cell Signaling Technology, USA, catalog number: 13167S) was selected and diluted with antibody dilution solution (Beyotime, China, catalog number: P0013) at a ratio of 1:500, incubation at room temperature for one hour. The sections were then incubated with the diluted antibody in a 4°C refrigerator overnight, washed three times with PBS, and incubated with horseradish peroxidase-labeled goat anti-rabbit IgG secondary antibody (Cell Signaling Technology, USA, catalog number: 7074S) at room temperature for one hour, followed by another three washes with PBS. According to the instructions of the IHC detection kit, the DAB substrate (Zhongshan, China, catalog number: ZLI-9018) was added for the color development reaction for 5–10 minutes at room temperature. The sections were rinsed with tap water and finally mounted. The sections were observed under an optical microscope, and the expression of NRG4 was quantitatively analyzed by recording the proportion of positive cells.

All patients underwent examination using a color Doppler ultrasound diagnostic device (GE Logiq E9, GE Healthcare, USA), with a high-frequency transducer (7.5–10 MHz) employed for both longitudinal and transverse scanning of the thyroid to assess its volume (calculated using the equation: Volume = Length × Width × Height × 0.52). Blood flow in the thyroid was evaluated, with blood flow velocity and resistance index (RI) recorded for various parts, and the blood supply to the thyroid was analyzed.

Statistical analysis was performed using SPSS 22.0 (IBM Corporation, USA). Categorical variables (expressed as percentages) were compared using the chi-square test (χ² test). For continuous variables that followed a normal distribution, one-way analysis of variance (ANOVA) was used, and pairwise comparisons between groups were conducted using Tukey’s post hoc test. To account for the risk of Type I errors due to multiple comparisons, the Bonferroni correction was applied to adjust the significance level to α = 0.0045 (0.05/11 indicators). Pearson correlation analysis was employed to assess the association between serum NRG4 levels and clinical parameters (e.g., HOMA-IR, FT3, TSH), with the Holm-Bonferroni method used to adjust the significance threshold for multiple correlation analyses. The predictive performance of the SVM model and traditional NRG4 levels was evaluated through receiver operating characteristic (ROC) curve analysis, and the area under the curve (AUC) was calculated to quantify overall accuracy. Unless otherwise specified for threshold adjustments, a two-tailed test with P< 0.05 was considered statistically significant.

The baseline data were collected from Ctrl group, T2DM group, and T2DM+FT group, including age, gender, height, weight, duration of T2DM, and type of T2DM treatment medications ( Table 1 ). Upon comparison, no considerable differences were observed in the baseline characteristics among Ctrl group, T2DM group, and T2DM+FT group (P>0.05).

Table 2 summarizes the measurement results of FBG, 2hPBG, HbA1c, HOMA-IR, NRG4, FT3, FT4, and TSH levels in patients of the Ctrl group, T2DM group, and T2DM+FT group. The results showed that the levels of FBG, 2hPBG, HbA1c, and NRG4 in the T2DM and T2DM+FT groups were significantly higher than those in the Ctrl group (P<0.05), and the level of HOMA-IR in the T2DM+FT group was also significantly increased (P<0.05). Furthermore, the T2DM+FT group exhibited significantly higher levels of FT3 and FT4, as well as significantly lower levels of TSH (P<0.05).

All patients underwent neck ultrasound examinations to assess the structure and blood flow of the thyroid gland. The results showed that the volume of the thyroid gland in patients of the T2DM+FT group was significantly enlarged to (25.24± 6.14) mL, which was significantly different compared to the T2DM group (18.52 ± 4.33) mL and the Ctrl group (15.06 ± 3.8 mL) (P<0.05). Additionally, the blood flow velocity in the T2DM+FT group (20.75 ± 3.52) cm/s was also significantly higher than that in the T2DM group (15.55 ± 2.29) cm/s and the Ctrl group (13.21 ± 2.09) cm/s, while the RI in the T2DM+FT group (0.82 ± 0.08) was significantly higher than that in the T2DM group (0.65 ± 0.05) and the Ctrl group (0.55 ± 0.06) (P<0.05) ( Figure 4 ).

Based on the statistical results from Sections 3.2 and 3.3, Pearson’s correlation coefficient was employed to assess the relationship between NRG4 levels in patients with T2DM combined with FT and various indicators, including FBG, 2hPBG, HbA1c, HOMA-IR, FT3, FT4, TSH, the positive rate of NRG4 in thyroid tissue, thyroid volume, blood flow velocity, and RI ( Table 3 ). The levels of NRG4 were found to have a significant positive correlation with HOMA-IR, FT3, FT4, thyroid volume, blood flow velocity, and RI, and a significant negative correlation with the positive rate of NRG4 in thyroid tissue and TSH (P<0.05).

Through k-fold cross-validation analysis, different values of C parameters and kernel functions were evaluated based on accuracy, Recall, and F1 score values to determine the optimal parameters. The specific results are as follows:

For C parameters of 0.01, 0.1, 1, 10, 100, and 1,000, the 50 iterations of training yielded accuracy, Recall, and F1 score values as shown in Figures 5A–C . Observing the results, it was found that when C=1, the fluctuations in accuracy, Recall, and F1 scores across the 50 training iterations were the most stable. As depicted in Figure 5D , the average values for accuracy, Recall, and F1 scores were as follows: C = 0.01: accuracy = 0.736 ± 0.028, Recall = 0.667 ± 0.026, F1 = 0.701 ± 0.021; C = 0.1: accuracy = 0.766 ± 0.025, Recall = 0.709 ± 0.020, F1 = 0.738 ± 0.018; C = 1: accuracy = 0.798 ± 0.019, Recall = 0.736 ± 0.007, F1 = 0.767 ± 0.009; C = 10: accuracy = 0.774 ± 0.010, Recall = 0.719 ± 0.017, F1 = 0.747 ± 0.013; C = 100: accuracy = 0.759 ± 0.024, Recall = 0.712 ± 0.012, F1 = 0.736 ± 0.012; C = 1,000: accuracy = 0.764 ± 0.023, Recall = 0.714 ± 0.015, F1 = 0.739 ± 0.014. By comparison, when C = 1, the average values of accuracy, Recall, and F1 score were significantly higher than those at C = 0.01, 0.1, 10, 100, and 1,000 (P<0.05).

Based on the results obtained, the optimal value for the C parameter was determined to be 1, guiding the determination of the optimal kernel function for this work. The evaluation compared the performance metrics (accuracy, Recall, F1) across various kernel functions including linear, polynomial, RBF, and Sigmoid, over 50 training iterations as shown in Figures 6A–C . Observations revealed that the variations in accuracy, Recall, and F1 scores were most stable across 50 training iterations using the linear kernel function. As depicted in Figure 6D , the average values for the linear kernel were accuracy = 0.783 ± 0.004, Recall = 0.766 ± 0.004, F1 = 0.776 ± 0.003; for the polynomial kernel, accuracy = 0.742 ± 0.023, Recall = 0.709 ± 0.010, F1 = 0.728 ± 0.012; for the RBF kernel, accuracy = 0.709 ± 0.011, Recall = 0.683 ± 0.012, F1 = 0.671 ± 0.011; and for the Sigmoid kernel, accuracy = 0.681 ± 0.03, Recall = 0.662 ± 0.009, F1 = 0.643 ± 0.011. Comparison analysis revealed that the average accuracy, Recall, and F1 values of the linear kernel were greatly higher than those of the polynomial, RBF, and Sigmoid kernel functions (P<0.05).

The results of the ablation study showed that the CNN model achieved an accuracy of 78.5%, recall of 75.3%, and F1 score of 76.8%, while the LSTM model achieved an accuracy of 72.1%, recall of 70.5%, and F1 score of 71.3%. In contrast, the CNN+LSTM model significantly outperformed both individual models, with an accuracy of 85.2%, recall of 83.7%, and F1 score of 84.4%. These results demonstrate that the combination of CNN and LSTM better captures the spatial features and temporal sequence characteristics of ultrasound images ( Figure 7 ).

Based on the above, it was known that NRG4 showed a significant correlation with thyroid indicators such as FT3, FT4, TSH, HOMA-IR, the positive rate of NRG4 in thyroid tissue, thyroid volume, blood flow velocity, and RI. Therefore, this study obtained the NRG4 levels of patients with isolated T2DM and those with T2DM complicated by FT. An SVM model based on NRG4 levels combined with an ultrasound feature classification prediction model based on CNN+LSTM (SVM-CNN+LSTM) was applied to evaluate its predictive performance for T2DM complicated by FT and compared with traditional NRG4-based prediction methods. The ROC curve was plotted, and the results are shown in Figure 8 . The area under the ROC curve (AUC = 0.943) for the SVM-CNN+LSTM model was significantly higher than that of the traditional method (AUC = 0.732). The optimal cutoff value for this model corresponded to a sensitivity of 91.32% and a specificity of 94.18%.

According to the ROC curve, the specificity, specificity, accuracy, Youden index, and AUC of the SVM-CNN+LSTM classification model for prediction were 91.32%, 94.18%, 95.35%, 0.855, and 0.943, respectively. The prediction results of the SVM model based on NRG4 levels were 86.23%, 90.33%, 90.89%, 0.798, and 0.887, respectively. In comparison, the traditional NRG4 prediction method yielded specificity, specificity, Youden index, and AUC of 60.22%, 80.78%, 79.49%, 0.410, and 0.732 ( Figure 9 ). The SVM model based on NRG4 levels showed significant improvement over the traditional prediction method in terms of specificity, specificity, Youden index, and AUC, while the SVM-CNN+LSTM classification model further outperformed the NRG4-based SVM model in all prediction metrics (P<0.05).