Time-sensitive clinical concept embeddings learned from large electronic health records

Area of Science:

• Artificial Intelligence
• Medical Informatics
• Computational Linguistics

Background:

• Deep learning for clinical concept representation is crucial in medicine.
• Existing methods often neglect temporal dynamics in patient electronic health records (EHRs).
• Ignoring temporal dependencies can lead to inaccurate contextual understanding.

Purpose of the Study:

• To enhance clinical concept embedding methods by incorporating time-sensitive information.
• To evaluate the impact of temporal considerations on concept embedding performance.
• To assess the utility of time-aware embeddings in downstream medical tasks.

Main Methods:

• Extended word2vec, PPMI, and FastText algorithms to include temporal features.
• Trained models on a large EHR database (approx. 50 million patients).
• Conducted intrinsic (concept similarity) and extrinsic (disease onset prediction) evaluations.

Main Results:

• Embeddings using only intra-visit information improved concept similarity.
• FastText generally outperformed word2vec and PPMI in intrinsic evaluations.
• Word2vec with a 30-day sliding window yielded optimal results for disease onset prediction.

Conclusions:

• Temporal constraints are vital for effective clinical concept embedding.
• Time-aware embeddings can enhance various downstream medical applications.
• This research provides a foundation for more accurate EHR data analysis.