Can Race-sensitive Biomedical Embeddings Improve Healthcare Predictive Models?

Area of Science:

• Biomedical Informatics
• Computational Linguistics
• Health Services Research

Background:

• Biomedical embeddings are crucial for analyzing health data.
• Previous research suggested demographic data could enhance predictive model accuracy.
• The impact of race distribution weighting in embeddings remains under-explored.

Purpose of the Study:

• To develop and evaluate an algorithm for weighting race distribution in biomedical embeddings.
• To assess if race-sensitive embeddings improve prediction of length of hospital stay (LHS) and intensive care unit (ICU) readmission.
• To reproduce findings on the utility of demographic-sensitive embeddings in healthcare.

Main Methods:

• Extracted 12,864 PubMed abstracts (2000-2022) and linked to ClinicalTrials.gov for race distribution data.
• Trained Word2vec and BERT embeddings using race-weighted and neutral approaches.
• Evaluated embedding performance on predicting LHS and ICU readmission using MIMIC-IV electronic health record data.

Main Results:

• Race-sensitive embeddings showed comparable performance to neutral embeddings for LHS prediction (MAE 1.975 vs. 2.008).
• Accuracy (74.61% vs. 75.17%) and AUC (0.775) for ICU readmission prediction were similar between race-sensitive and neutral embeddings.
• No consistent significant improvement in predictive accuracy was observed with race-sensitive embeddings.

Conclusions:

• Weighting biomedical embeddings with race distribution data does not inherently enhance predictive accuracy for health outcomes like LHS or ICU readmission.
• Findings challenge previous reports suggesting significant benefits from demographic-sensitive embeddings in healthcare.
• Further research is needed to understand the complex interplay between demographic data and predictive model performance.