Improving medical code classification for death certificates using ontology-adapted contrastive loss in BERT models

Accurate identification of the Underlying Cause of Death (UCOD) is crucial for informed healthcare policy and planning. The World Health Organization supports the use of the ICD-10 system to standardize the coding of death certificates, a task increasingly supported by automated systems built on top of language models. This study advances the effectiveness of state-of-the-art BERT-based models for UCOD identification by incorporating a novel ontology-adapted contrastive loss function. Extensive experimentation on a dataset from the U.S. National Center for Health Statistics show that BERT models equipped with this specialized contrastive loss function outperform traditional state-of-the-art models.