Contributors

• Kelly Lee, cl129
• Praneetha Bhogi, pbhogi

Type of Contribution

Standalone Task

Original Paper

Zhang et al. "Hurtful Words: Quantifying Biases in Clinical Contextual
Word Embeddings." ACM CHIL 2020.
https://arxiv.org/abs/2003.11515

Description

Implements MortalityTextTaskMIMIC3, a standalone PyHealth task that
reproduces the in-hospital mortality prediction task from the Hurtful
Words paper. The task generates synthetic clinical note templates
populated with real patient demographics (gender, age) from MIMIC-III
PATIENTS and ADMISSIONS tables, enabling fairness evaluation across
gender, ethnicity, insurance, and language subgroups.

Files to Review

• pyhealth/tasks/mortality_text_task.py — main task implementation
• tests/test_mortality_text_task.py — unit tests
• examples/mimic3_mortality_text_clinicalbert.py — ablation study
• docs/api/tasks/pyhealth.tasks.mortality_text_task.rst — documentation