Identifying Transitional High Cost Users from Unstructured Patient Profiles Written by Primary Care Physicians

Area of Science:

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

Background:

• Identifying High Cost Users (HCUs) enables proactive interventions, improving patient outcomes and reducing healthcare expenditures.
• Electronic medical records contain valuable unstructured text data crucial for predicting patient risk.
• Traditional methods often struggle with the nuances of clinical notes, including domain-specific language.

Purpose of the Study:

• To predict the 2016 High Cost User (HCU) status of patients using 2015 electronic medical record data.
• To evaluate the effectiveness of context-specific word embeddings versus pre-trained embeddings for this prediction task.
• To assess the performance improvement of an aggregated word embedding model (EmbEncode) over traditional bag-of-words approaches.

Main Methods:

• Utilized free-form text data from 2015 cumulative patient profiles in Ontario family care practices.
• Developed context-specific word embeddings from unstructured clinical notes, accounting for domain-specific abbreviations and spellings.
• Compared a novel EmbEncode model against a bag-of-words model using held-out AUROC metrics.

Main Results:

• Context-specific word embeddings significantly outperformed pre-trained embeddings (Wikipedia, MIMIC, Pubmed).
• The EmbEncode model achieved a higher held-out AUROC (82.48±0.35%) compared to bag-of-words (81.85±0.36%, p = 3.2 × 10-4).
• EmbEncode used substantially fewer features and non-zero coefficients, indicating greater efficiency.
• Removing recurrent HCUs from the training set enhanced the prediction of new HCUs with minimal impact on predicting recurrent ones.

Conclusions:

• Aggregated word embeddings derived from clinical text (EmbEncode) offer a more effective and efficient method for predicting High Cost Users.
• Tailoring feature extraction to the specific clinical context is crucial for improving predictive model performance.
• A refined approach focusing on transitional HCUs can optimize intervention strategies within healthcare systems.