Combining Statistical and Deep Learning Models for Insomnia Detection

Insomnia is a common but often underdiagnosed condition in clinical settings, where relevant information is typically buried in unstructured free-text notes. Automated tools that can identify both the presence of insomnia and the supporting evidence are essential to improve diagnosis and enable large-scale studies. However, existing models often prioritize accuracy at the cost of interpretability, which is critical for clinical adoption. To address this, we explore a hybrid approach that balances performance with explainability. Our method combines Finite Context Models (FCMs) for character-level classification of insomnia status with a BERT-based token classification model for extracting textual evidence, using structured annotations from the MIMIC-III dataset. This complementary setup enables both accurate prediction and transparent decision-making in clinical text analysis.