DI++: A deep learning system for patient condition identification in clinical notes.

Area of Science:

• Medical Informatics
• Natural Language Processing
• Artificial Intelligence

Background:

• Accurate patient condition recording in Electronic Health Records (EHRs) is crucial for clinical care, billing, and decision-making.
• Patient conditions are often documented in unstructured clinical notes, posing a challenge for data extraction.
• Distinguishing actual patient conditions from general disease mentions in notes is a key difficulty.

Purpose of the Study:

• To develop and evaluate a robust system for identifying patient conditions from unstructured clinical notes.
• To enhance the accuracy and efficiency of disease identification within EHR systems.
• To introduce and validate an advanced deep learning model for disease mention classification.

Main Methods:

• A two-step workflow was developed: disease mention extraction followed by disease mention classification.
• A prototype system, DI++ (Disease Identification), was implemented.
• An advanced CLSTM-Attention deep learning model was utilized for disease mention classification.

Main Results:

• The DI++ system demonstrated significant performance advantages over existing systems.
• Key metrics including F1 Score and Area Under the Curve showed substantial improvements.
• The CLSTM-Attention model outperformed other deep learning models in disease mention classification tasks.
• Evaluation was conducted on approximately one million pages of de-identified clinical notes, confirming efficiency and effectiveness.

Conclusions:

• The DI++ system effectively identifies patient conditions from clinical notes.
• The CLSTM-Attention model represents a state-of-the-art approach for disease mention classification.
• Accurate disease identification from clinical notes can enhance EHR data quality and support clinical decision-making.