Identify and mitigate bias in electronic phenotyping: A comprehensive study from computational perspective.

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Health Equity

Background:

• Electronic phenotyping is crucial for precision medicine and real-world evidence generation.
• Machine learning has advanced electronic phenotyping using electronic health records.
• Current focus is on accuracy, neglecting fairness, leading to patient subgroup underrepresentation.

Purpose of the Study:

• To comprehensively study bias in electronic phenotyping models.
• To evaluate the effectiveness of debiasing methods for phenotyping.
• To address the gap in fairness considerations for patient subgroup identification.

Main Methods:

• Benchmarked 9 electronic phenotyping methods (rule-based to data-driven).
• Evaluated 5 bias mitigation strategies (pre-processing, in-processing, post-processing).
• Focused on pneumonia and sepsis as target diseases for bias identification.

Main Results:

• Identified specific biases within electronic phenotyping models.
• Assessed the performance of various debiasing techniques across different phenotyping methods.
• Provided insights into the effectiveness of bias mitigation strategies in healthcare.

Conclusions:

• Electronic phenotyping models can exhibit significant bias, impacting healthcare equity.
• Bias mitigation strategies show potential but require careful selection based on the phenotyping task.
• Further research is needed to ensure fair and equitable patient subgroup identification in digital health.