TLDR: A systematic review of Dutch public health machine learning research from 2021-2025 found significant gaps in how algorithmic bias is identified, discussed, and reported. To address this, researchers developed the Risk of Algorithmic Bias Assessment Tool (RABAT) to evaluate 35 studies, revealing a lack of explicit fairness framing, subgroup analyses, and transparent harm discussions. In response, they propose the ACAR (Awareness, Conceptualization, Application, Reporting) framework, a four-stage guide with actionable questions to help researchers integrate fairness considerations throughout the machine learning lifecycle in public health, aiming to advance health equity.
Machine learning (ML) holds immense promise for transforming public health, offering advancements in disease surveillance, risk assessment, and resource allocation. However, without careful attention to algorithmic bias (AB), these powerful tools could inadvertently worsen existing health disparities. A recent systematic review delves into this critical issue, examining how AB is identified, discussed, and reported in Dutch public health ML research.
The paper, titled “Machine Learning and Public Health: Identifying and Mitigating Algorithmic Bias through a Systematic Review,” was authored by Sara Altamirano, Arjan Vreeken, and Sennay Ghebreab from the Informatics Institute, University of Amsterdam. Their work highlights a pressing need for more systematic approaches to fairness in public health machine learning.
To conduct their review, the researchers developed a specialized tool called the Risk of Algorithmic Bias Assessment Tool (RABAT). This tool integrates elements from established frameworks like the Cochrane Risk of Bias, PROBAST, and the Microsoft Responsible AI checklist, tailoring them specifically for public health ML. They applied RABAT to 35 peer-reviewed studies published between 2021 and 2025, focusing on research conducted within the Netherlands, a country with advanced health infrastructure and a diverse population where health disparities persist.
Key Findings: Significant Gaps in Bias Reporting
The analysis using RABAT revealed pervasive gaps in how algorithmic bias is addressed. While practices related to data sampling and handling missing data were generally well-documented, most studies fell short in several crucial areas:
- Explicitly framing discussions around ML fairness.
- Conducting subgroup analyses to understand how models perform across different population groups.
- Transparently discussing potential harms or negative societal impacts of their algorithms.
- Identifying specific at-risk subgroups.
- Reporting on sensitive attributes (like race, ethnicity, or disability) and their influence on model outcomes.
- Articulating bias risks in a structured and comprehensive manner.
- Reporting on how fairness-related harms were identified, assessed, or mitigated.
In essence, the review found a clear asymmetry: while methodological rigor in data handling was often present, considerations for fairness and equity were largely underdeveloped or absent.
Also Read:
- Assessing LLM Fairness: A New Framework Prioritizes Real-World Harm
- Uncovering Hidden Biases in Large Language Models with Adaptive Question Generation
Introducing the ACAR Framework: A Path Towards Fairness
In response to these identified gaps, the researchers introduce a four-stage, fairness-oriented framework called ACAR, which stands for Awareness, Conceptualization, Application, and Reporting. This framework is designed to guide public health ML researchers in addressing fairness throughout the entire machine learning lifecycle:
- Awareness: Recognizing that bias can emerge from data, model design, or social context, and reflecting on who might be affected and how societal impacts could arise.
- Conceptualization: Defining algorithmic bias, fairness, and subgroup risks in relation to research objectives and methodology early in the process.
- Application: Implementing strategies to address bias in data and modeling workflows, including careful sampling, subgroup testing, and bias mitigation techniques.
- Reporting: Clearly communicating how bias risks and fairness considerations were addressed, including structured discussions, subgroup findings, and transparency about limitations and ethical elements like consent.
The ACAR framework provides guiding questions for each stage, aiming to translate the insights from the systematic review into actionable steps for researchers. It emphasizes that addressing ethical concerns in public health ML requires more than just technical fixes; it demands a rethinking of how algorithms are evaluated, designed, and reported for the public good.
The study concludes with actionable recommendations for public health ML practitioners, urging them to consistently consider algorithmic bias and foster transparency. By adopting ACAR-informed practices and leveraging national transparency mechanisms, the goal is to ensure that algorithmic innovations truly advance health equity rather than undermine it. You can read the full research paper here.
