Extracting Medical Information From Free-Text and Unstructured Patient-Generated Health Data Using Natural Language...

Area of Science:

• Computational linguistics
• Health informatics
• Clinical data science

Background:

• Patient-generated health data (PGHD) from digital tools offers insights into patient health journeys outside clinical settings.
• Unstructured PGHD, such as notes and diaries, provides a comprehensive view of health conditions.
• Natural Language Processing (NLP) shows potential for analyzing unstructured PGHD to derive meaningful insights.

Purpose of the Study:

• To demonstrate the feasibility of an NLP pipeline for extracting medication and symptom information from real-world patient and caregiver data.
• To evaluate the effectiveness of NLP techniques in processing unstructured PGHD.

Main Methods:

• A secondary data analysis utilized free-text patient notes from parents of children with special health care needs (CSHCN).
• An NLP pipeline employing a zero-shot approach, Named Entity Recognition (NER), and medical ontologies (RXNorm, SNOMED CT) was developed.
• Sentence-level dependency parse trees and part-of-speech tags were used for detailed information extraction.

Main Results:

• The study analyzed 87 patient notes, identifying medication and symptom-related entries.
• The NLP pipeline achieved satisfactory performance with precision >0.65, recall >0.77, and F1-score >0.72 for medication and symptom extraction.
• Results highlight the potential of NLP with NER and dependency parsing for extracting information from unstructured PGHD.

Conclusions:

• The developed NLP pipeline is feasible for extracting medication and symptom information from real-world unstructured PGHD.
• Leveraging unstructured PGHD through NLP can enhance clinical decision-making, remote monitoring, and self-care.
• NLP models offer a flexible approach for extracting clinical information from unstructured PGHD, even in low-resource settings.