Leveraging Temporal Trends for Training Contextual Word Embeddings to Address Bias in Biomedical Applications:...

Area of Science:

• Natural Language Processing
• Machine Learning
• Clinical Informatics

Background:

• Historically, women have been underrepresented in clinical trials, leading to potential biases in machine learning models trained on trial data.
• Word embeddings, fundamental to NLP, can inadvertently amplify gender-based performance disparities if trained on biased clinical trial abstracts.

Purpose of the Study:

• To capture temporal trends in clinical trials using temporal distribution matching on contextual word embeddings (BERT).
• To explore the impact of these temporal trends on bias in downstream clinical prediction tasks.

Main Methods:

• Introduced TeDi-BERT, a method leveraging the trend of increased female inclusion in clinical trials to train contextual word embeddings.
• Employed adversarial classification for temporal distribution matching, adapting models from older to more recent clinical trial data.
• Evaluated TeDi-BERT on predicting intensive care unit readmissions and hospital length of stay.

Main Results:

• TeDi-BERT improved prediction accuracy for female patients in readmission prediction (AUC 0.64 vs. 0.62) and length of stay (MAE 4.56 vs. 4.62).
• Performance also improved for male patients in readmission prediction (AUC 0.66 vs. 0.64) and length of stay (MAE 4.54 vs. 4.60).

Conclusions:

• TeDi-BERT enhanced clinical prediction performance for both female and male patients, demonstrating that improved accuracy for one gender does not necessitate compromising the other.
• Training contextual word embedding models on temporal trends effectively mitigates biases present in historical clinical trial data, leading to better outcomes for all patients.