The importance of clinical context in evaluating algorithmic fairness: insights from a medication adherence...

Area of Science:

• Health equity research
• Clinical informatics
• Artificial intelligence in healthcare

Background:

• Artificial intelligence (AI) algorithms risk exacerbating health disparities by misallocating resources.
• An algorithm predicting medication nonadherence in heart failure (HF) patients was evaluated for potential bias.

Purpose of the Study:

• To assess if an AI algorithm predicting HF medication nonadherence could worsen existing health disparities.
• To examine algorithm performance across socioeconomic and racial subgroups.

Main Methods:

• Retrospective study of 34,697 HF patients.
• Compared algorithm performance (accuracy, false positive/negative rates) between low/high neighborhood socioeconomic status (nSES) and Black/White patients.
• Applied results to hypothetical clinical use cases.

Main Results:

• Algorithm accuracy was similar across nSES and racial subgroups.
• Higher false positive rates for nonadherence in low nSES and Black patients.
• Lower false negative rates for nonadherence in Black patients compared to White patients.
• Worsening disparities were noted in use cases where false positives could negatively impact care prioritization.

Conclusions:

• Despite similar overall accuracy, differential false positive and negative rates can lead to AI algorithms worsening health inequities.
• Contextual evaluation of AI predictions within specific clinical use cases is crucial to prevent unintended harm and promote equitable care.