Automatic phenotyping of electronical health record: PheVis algorithm.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Electronic Health Records (EHRs) often lack comprehensive and reliable patient medical condition annotations.
• Automated methods are needed to accurately identify patient conditions from complex EHR data.
• Existing unsupervised algorithms like Phenorm provide a foundation for condition identification.

Purpose of the Study:

• To extend the Phenorm algorithm for visit-level medical condition identification.
• To develop an interpretable machine learning model (PheVis) for predicting medical condition occurrence probability at each patient visit.
• To evaluate PheVis performance on both chronic and acute conditions using real-world clinical data.

Main Methods:

• PheVis integrates diagnosis codes with medical concepts extracted from clinical notes.
• A machine learning approach incorporates patient's past medical history.
• The model was applied to rheumatoid arthritis (chronic) and tuberculosis (acute) using data from the University Hospital of Bordeaux.
• Performance was evaluated using cross-validated AUROC and AUPRC metrics.

Main Results:

• PheVis achieved high performance for rheumatoid arthritis (chronic condition) with AUROC of 0.943 and AUPRC of 0.754.
• For tuberculosis (acute condition), PheVis achieved AUROC of 0.987 and AUPRC of 0.299.
• Performance for acute conditions was limited in French due to natural language processing challenges in excluding past medical history.
• PheVis demonstrated significantly better performance than state-of-the-art unsupervised methods, especially for chronic diseases.

Conclusions:

• PheVis effectively predicts medical condition occurrence at the visit level, outperforming existing unsupervised methods for chronic diseases.
• The model's interpretability and integration of diverse data sources enhance its clinical utility.
• Further refinement of natural language processing is needed to optimize performance for acute conditions in specific language contexts.