Endocrine Abstracts

JOINT819

Background: Hypothyroidism is a common endocrinological disorder affecting approximately 10% of the general population, with a 10-fold higher prevalence in women. Despite levothyroxine being one of the most prescribed treatments worldwide, 30–50% of patients are over- or under-treated, with many reporting dissatisfaction with their treatment and quality of life. Hypothyroidism patients present with significant heterogeneity in symptoms, co-morbidities, and treatment requirements, yet large and comprehensive studies addressing this complexity remain lacking. Our study aims to identify clustering patterns and varying trajectories of hypothyroidism patients by leveraging longitudinal health and demographic data. We seek to characterise patient clustering patterns by defining respective co-morbidities and risk/protective factors, while also exploring the overlapping disease mechanisms involved in these co-morbidities.

Methods: We identified 21,790 (17,028 women and 4,761 men) hypothyroidism patients from the Copenhagen Hospital Biobank based on hypothyroidism diagnosis in hospital records and levothyroxine prescription data, ensuring a comprehensive dataset reflecting diverse patient profiles. We constructed longitudinal health sequences capturing each patient’s history of hospital diagnoses, prescriptions, procedures, and laboratory Results To address the complexity and high dimensionality of the data, we applied a machine learning approach: the Deep Generative Decoder (DGD), which learns low-dimensional representations even for small datasets, based on a transformer architecture to identify long-term dependencies and intricate patterns in sequential health data. The DGD integrates a Gaussian Mixture Model to capture latent space distributions and uncover data substructures. Patient representations generated through this pipeline were clustered, enabling the identification of unique patterns. These clusters were further characterized to highlight key patterns.

Results: Preliminary findings indicate that hypothyroidism patients are clustered with distinct and varying longitudinal patterns, notably differentiating between reproductive co-morbidities such as infertility and miscarriage (observed exclusively in women) and cardiovascular co-morbidities such as essential hypertension and atrial fibrillation (present in both sexes). Ongoing model refinement with additional data will improve these findings. Further detailed analyses of these clustering patterns and their associated health outcomes will be discussed, providing insights into the underlying disease mechanisms and potential ways for providing more personalised prevention, diagnosis and treatment options.

Conclusion: This study highlights the importance of understanding the heterogeneous nature of hypothyroidism patients in a sex-specific manner. By identifying patient clustering patterns, we aim to enhance clinical management and therapeutic approaches, ultimately improving patient outcomes in this diverse population. Further results and detailed characterisations will be presented at the conference.