Integration of NLP2FHIR Representation with Deep Learning Models for EHR Phenotyping: A Pilot Study on Obesity...

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Clinical Data Analytics

Background:

• Health Level Seven Fast Healthcare Interoperability Resources (FHIR) is a standard for electronic healthcare information exchange.
• FHIR can model structured and unstructured electronic health record (EHR) data.
• The utility of FHIR for clinical data analytics, particularly with unstructured data, requires further investigation.

Purpose of the Study:

• To demonstrate the application of FHIR-represented unstructured EHR data in deep learning models for clinical phenotyping.
• To evaluate the effectiveness of FHIR data normalization pipelines in preparing data for text classification.
• To assess the portability and performance of deep learning models across different EHR systems.

Main Methods:

• Leveraged and extended the NLP2FHIR clinical data normalization pipeline.
• Conducted a case study using two obesity datasets.
• Tested deep learning text classifiers (CNNs, GRUs, Text GCNs) on raw text and NLP2FHIR-processed data.

Main Results:

• The combination of NLP2FHIR input and text graph convolutional networks (Text GCNs) achieved the highest F1 score.
• FHIR-based representation facilitated the use of unstructured EHR data for deep learning.
• Performance varied across different deep learning models and data types.

Conclusions:

• FHIR-based deep learning approaches show potential for enhancing EHR phenotyping.
• This methodology can improve the portability of phenotyping algorithms across diverse EHR systems and institutions.
• Standardizing EHR data with FHIR is crucial for advancing clinical data analytics and research.