This research reveals how patient medication non-adherence creates implicit bias in clinical machine learning models, potentially leading to flawed treatment recommendations.

• Study analyzed EHR data from 3,623 hypertension patients, extracting adherence information from clinical notes
• Non-adherence to prescribed treatments fundamentally distorts both causal inference and predictive modeling
• ML systems trained on incomplete adherence data may recommend inappropriate or potentially harmful treatments
• Demonstrates the critical need for adherence-aware ML model development in clinical settings

Why it matters: For healthcare organizations implementing AI, this work highlights a significant but often overlooked source of bias that could compromise patient outcomes and undermine trust in medical AI systems.

Treatment Non-Adherence Bias in Clinical Machine Learning: A Real-World Study on Hypertension Medication