Endocrine Abstracts

JOINT3232

Introduction: Graves’ disease (GD) is one of the most common autoimmune thyroid disorders worldwide. Although GD is not life-threatening, complications significantly impact patients’ quality of life: 52% of patients relapse, 50% develop thyroid eye disease (TED), and thyroidectomy remains the last resort in challenging clinical cases. In recent years, data engineering and multivariate statistical methods have been increasingly applied in the medical field. They provide a broader analytical perspective, accelerate statistical analysis and contribute to the development of predictive algorithms. This study aimed to identify clinical and biochemical variables collected at diagnosis that are associated with thyroidectomy, relapse, and future TED onset in GD patients using these methods.

Methodology: We analyzed data from GD patients diagnosed at Hospital Universitario de La Princesa between 2010 and 2018, with follow-up until 2022. Data engineering techniques were employed to clean the dataset by removing low-quality variables. The analytical process was conducted in two steps: 1) a univariate analysis was performed to identify significant associations with the outcomes of interest (P <0.1), which were subsequently included in further models. 2) Next, supervised logistic regression (100-fold cross-validation-CV-) and time-to-event analysis (Cox models) were conducted.

Results: Multivariate analysis reported the following independent associations:

• Relapse: Significant associations were found with albumin (β=-1.782, P = 0.064) and CAS (β=1.238, P = 0.070). However, time-to-event analysis did not show a clear impact of these variables on relapse.

• TED: Independent associations were observed with basophils (β=-4.883, P = 0.016) and GPT(β=2.171, P = 0.084). Both variables had a weak, but significant correlation(r=-0.212, P = 0.004). Other correlations with thyroid hormones would suggest a link between them. Interestingly, basophils also showed significant independent associations in Cox models(β=-1.959, P = 0.076), inferring a potential relationship between their levels at diagnosis and future TED onset.

• Thyroidectomy: Associations were found with immature granulocytes (β=-22.272, P = 0.033), anti-TG antibodies (β=-23.643, P = 0.096), leukocytes (β=3.310, P = 0.066), eosinophils (β=2.969, P = 0.026), potassium (β=-3.801, P = 0.093) and goiter (β=2.308, P = 0.028). However, the Cox model did not report any significant relationship between surgery timing and these variables. The mean areas under the curve reflected poor quality models: relapse: 0.576±0.076; TED: 0.663±0.100 and thyroidectomy: 0.676±0.107. A class imbalance in the dataset, a limited sample size and the lack of biomarkers may have contributed to the reduced model performance.

Conclusions: 1. We obtained significant associations between relapse and albumin levels and CAS, already reported in the literature. 2. A novel relationship between TED onset and basophils. 3. Associations between thyroidectomy immune populations levels, goiter and potassium, which could contribute to its management.