Gender-sensitive word embeddings for healthcare

Area of Science:

• Biomedical Informatics
• Natural Language Processing (NLP)
• Clinical Trial Analysis

Background:

• Women are historically underrepresented in clinical trials, leading to potential biases in AI models trained on this data.
• Existing NLP systems trained on clinical trial literature may exhibit deficits in knowledge pertaining to female patients.
• Representational biases in training data can negatively impact the performance of AI models for underrepresented populations.

Purpose of the Study:

• To analyze gender bias within clinical trial data.
• To develop an algorithm mitigating gender representation biases in NLP systems.
• To evaluate the performance of the developed gender-sensitive algorithm.

Main Methods:

• Analysis of 16,772 PubMed abstracts from clinical trials (2008-2018) to identify gender bias.
• Augmentation of word embeddings by weighting abstracts based on the number of female participants.
• Evaluation of gender-sensitive embeddings on clinical prediction tasks: comorbidity classification, hospital length of stay, and ICU readmission.

Main Results:

• Gender-sensitive models demonstrated improved performance across all evaluated clinical prediction tasks.
• Statistically significant improvements were observed for female patients, including higher AUROC for comorbidity classification (0.86 vs. 0.81) and ICU readmission (0.69 vs. 0.67).
• Enhanced accuracy was noted in length of stay prediction (MAE 4.59 vs. 4.66).

Conclusions:

• The proposed method enables gender-sensitive utilization of clinical trial publications for training NLP word embeddings.
• Gender-sensitive embeddings outperform baseline embeddings in clinical prediction tasks, demonstrating the algorithm's efficacy.
• Addressing representational biases in training data is crucial for improving AI performance for underrepresented groups in healthcare.