What makes a Clinical Decision Support tool fair?

Abstract: Healthcare systems increasingly embed “Clinical Decision Support (CDS)” tools into the electronic healthcare record. CDS tools use AI-generated algorithms to identify patients at high risk for important clinical outcomes and prompt clinicians to take specific treatment actions. Ensuring that CDS tools lead to fair treatment for patients belonging to disadvantaged groups is of critical bioethical and legal importance.

Panelist 1, a pulmonary and critical care physician, clinical ethicist, and data scientist, will open the AI “black box” and non-technically explain the relevant statistical details to the audience using a case study of a CDS designed to trigger palliative care consults for hospitalized patients at high risk of death. They will introduce and apply three mutually exclusive definitions of algorithmic fairness (independence, separation, and sufficiency) to the CDS case.

Panelist 2, a general internist, bioethicist, and health equity scholar, will react to these formal definitions of algorithmic fairness. They will argue that a deep understanding of structural inequity in US healthcare is essential when selecting an algorithmic fairness definition for a specific clinical situation in which a CDS is deployed to avoid biased treatment against disadvantaged groups.

Panelist 3, a health law scholar, will analyze how US anti-discrimination law applies to CDS tools. They will specifically discuss which algorithmic fairness definition best aligns with the law. They will also discuss the potential legal implications for healthcare systems that ignore anti-discrimination laws when implementing CDS tools.