An Early Identification of Central Precocious Puberty Using a Contrastive Learning Framework with Time Series Transformers

Central Precocious Puberty (CPP) is a pediatric endocrine disorder characterized by the early activation of
the hypothalamic–pituitary–gonadal axis, leading to premature development of secondary sexual
characteristics. Early and accurate diagnosis is critical for timely intervention but remains challenging due to
heterogeneous growth patterns and variability in clinical presentation. This paper proposes HormoCLR, a
novel framework that combines Time Series Transformers with Contrastive Representation Learning to model
longitudinal hormone and growth data. Each patient’s clinical history is structured as a multivariate timeseries, incorporating hormone levels (LH, FSH, Estradiol/Testosterone), anthropometric indicators (height,
weight, BMI), bone and chronological age, and Tanner staging. The model leverages self-supervised learning
through data augmentations such as jittering, time cropping, and masking to generate positive and negative
pairs for contrastive pretraining. A Transformer encoder captures temporal dependencies and outputs highdimensional embeddings that are later fine-tuned for downstream tasks, including CPP diagnosis, subtype
classification (idiopathic vs. organic), and time-to-onset prediction. Additionally, attention weights from the
Transformer provide interpretability, highlighting influential features and timepoints. Initial experimental
results indicate that HormoCLR achieves improved classification performance and robustness over traditional
recurrent models, especially in low-label settings. This approach offers a scalable, interpretable, and clinically
relevant tool for the early detection of CPP.