Combining Statistical and Deep Learning Models for Insomnia Detection.

Area of Science:

• Medical Informatics
• Artificial Intelligence in Healthcare
• Clinical Natural Language Processing

Background:

• Insomnia is frequently underdiagnosed in clinical settings due to information being hidden in unstructured text.
• Automated tools are needed to detect insomnia and supporting evidence for improved diagnosis and research.
• Current models often lack interpretability, hindering clinical adoption.

Purpose of the Study:

• To develop and evaluate a hybrid AI model that balances accuracy and interpretability for insomnia detection in clinical text.
• To improve the identification of insomnia and its supporting evidence within unstructured electronic health records.
• To facilitate clinical adoption through transparent and explainable AI methods.

Main Methods:

• A hybrid approach combining Finite Context Models (FCMs) for insomnia classification and BERT-based models for evidence extraction.
• Utilized structured annotations from the MIMIC-III dataset for model training and validation.
• Employed character-level classification for insomnia status and token classification for evidence identification.

Main Results:

• The hybrid model achieved accurate prediction of insomnia status.
• The system successfully extracted textual evidence supporting insomnia diagnoses.
• The approach demonstrated a balance between predictive performance and model explainability.

Conclusions:

• The developed hybrid AI method offers a promising solution for identifying insomnia in clinical text.
• This approach enhances diagnostic capabilities and supports transparent decision-making in healthcare.
• The balance of accuracy and interpretability is crucial for the clinical utility of AI tools in medical text analysis.