Machine Learning and Public Health: Identifying and Mitigating Algorithmic Bias through a Systematic Review

Most public health ML studies fail to explicitly discuss fairness: The review of 35 Dutch public health studies found that although researchers documented data sampling and missing data practices, "most studies omit explicit fairness framing, subgroup analyses, and transparent discussion of potential harms"—meaning potential biases affecting different populations often go unexamined.

Algorithmic bias can worsen existing health disparities: The paper warns that without systematic attention to bias, machine learning systems "may inadvertently reinforce existing health disparities," suggesting that algorithms trained on biased data could perpetuate or amplify inequities in healthcare access and outcomes for vulnerable groups.

A structured framework can improve bias detection: The authors developed ACAR (Awareness, Conceptualization, Application, Reporting)—a four-stage framework with specific guiding questions—to help researchers systematically address fairness throughout the entire machine learning development process, from initial design through final reporting.

Transparency and subgroup testing are critical gaps: The systematic review identified that studies rarely conduct subgroup analyses (testing how algorithms perform across different demographic groups) or transparently discuss potential harms, which are essential practices for identifying whether an algorithm works equally well for all populations.

Actionable standards are needed for public health ML: The paper concludes with recommendations for practitioners to "consistently consider algorithmic bias and foster transparency," emphasizing that without deliberate attention to fairness, algorithmic innovations risk undermining rather than advancing health equity.