Feature engineering from medical notes: A case study of dementia detection

Area of Science:

• Natural Language Processing (NLP) in Healthcare
• Machine Learning for Disease Prediction
• Clinical Informatics

Background:

• Medical notes offer rich, unstructured data for patient health assessment.
• Routine clinical notes can be leveraged for predicting chronic diseases like dementia.
• Evaluating methodologies for transforming free-text medical notes into predictive models is crucial.

Purpose of the Study:

• To investigate methods for converting routine medical notes into dementia risk classifiers.
• To assess the generalizability of these dementia risk classifiers across different patient populations and healthcare institutions.
• To compare different feature engineering and encoding techniques for medical note analysis.

Main Methods:

• Utilized TF-ICF (Term Frequency-Inverse Class Frequency) to identify keywords distinguishing dementia risk.
• Compared two encoding methods: BERT/Clinical BERT embeddings and aggregation of keywords into UMLS concepts.
• Developed classifiers using a neural network (for embeddings) and gradient boosted trees (for UMLS concepts).
• Considered misspellings of keywords to enhance predictive performance.

Main Results:

• Achieved an Area Under the Curve (AUC) of 75% for predicting dementia risk one year ahead using gradient boosted trees with UMLS concepts.
• Performance decreased when using embedded feature spaces or applying classifiers to external healthcare institutions.
• Analysis revealed that feature importance varied based on the inclusion of spelling variants.

Conclusions:

• Medical notes are a viable data source for developing dementia risk prediction models.
• Improving the generalizability of these models requires further research into note characteristics, data availability, and feature engineering.
• The study highlights the potential of NLP and machine learning in proactive healthcare.