Hidden Markov model using Dirichlet process for de-identification.

Area of Science:

• Natural Language Processing
• Machine Learning
• Biomedical Informatics

Background:

• Clinical text de-identification is crucial for privacy protection.
• Traditional methods often require extensive feature engineering.
• Generalizing models to new datasets remains a challenge.

Purpose of the Study:

• To introduce a novel non-parametric Bayesian hidden Markov model with a Dirichlet process (HMM-DP).
• To reduce task-specific feature engineering and improve generalization for de-identification.
• To enhance performance by incorporating long-range and cross-document context.

Main Methods:

• Developed a variational method for model learning and an efficient approximation algorithm for prediction.
• Designed feature functions to handle out-of-vocabulary words.
• Employed a skip-chain conditional random field model to integrate HMM-DP outputs.

Main Results:

• The HMM-DP model achieved state-of-the-art accuracy without manual feature engineering.
• The model demonstrated an understanding of local context cues for accurate predictions.
• Integrating long-range context further improved de-identification performance.

Conclusions:

• The HMM-DP offers a robust and generalizable approach to clinical text de-identification.
• Reduced feature engineering requirements make the model more adaptable.
• Combined models show promise for advanced context-aware de-identification.