Natural Language Processing Methods to Empirically Explore Social Contexts and Needs in Cancer Patient Notes

Area of Science:

• Computational linguistics in healthcare
• Oncology and health disparities research
• Social determinants of health analysis

Background:

• Cancer disparities are driven by social determinants of health (SDOH).
• Empirical extraction of SDOH from clinical notes is needed.
• Natural language processing (NLP) offers a method for this extraction.

Purpose of the Study:

• To explore NLP methods for extracting social contexts and needs from clinical documentation.
• To empirically investigate drivers of cancer disparities using linguistic analysis of patient notes.

Main Methods:

• Retrospective analysis of 230,325 clinical notes from 5,285 radiotherapy patients (2007-2019).
• Comparison of linguistic features across racial/ethnic, insurance, and sex demographics.
• Application of variational autoencoder topic modeling and machine learning classifiers to identify demographic biases.

Main Results:

• Over-representation of terms related to housing/transportation for non-White and low-income patients.
• Over-representation of physical activity terms for White and higher-income patients.
• Significant variation in social history topic probability across demographic groups; poor classifier performance for minority groups.

Conclusions:

• Linguistic analysis of clinical notes reveals differences in social contexts and needs among cancer patients.
• These linguistic differences may contribute to cancer disparities.
• Further research is required to validate the role of these findings in cancer disparities.