Perspective: Machine Learning for Health Should Consider Social Drivers of Health

Area of Science:

• Health Informatics
• Artificial Intelligence in Medicine
• Health Equity Research

Background:

• Clinical machine learning (ML) algorithms can perpetuate societal injustices.
• Disparities exist in the distribution of harms and benefits of health ML, with minoritized groups often bearing the brunt.
• Algorithmic harms in healthcare are frequently underemphasized and inadequately addressed.

Purpose of the Study:

• To analyze the relationship between algorithmic harms and social drivers of health (SDOH).
• To highlight the tension between ML development in health and the potential to exacerbate health inequities.
• To propose strategies for mitigating algorithmic harms and promoting health equity through ML.

Main Methods:

• Framework analysis correlating algorithmic harm typologies with social drivers of health (SDOH) concepts.
• Examination of how algorithmic harms impact human health via SDOH factors, particularly structural determinants.
• Literature review on existing approaches to algorithmic bias and health equity.

Main Results:

• Algorithmic harms impact health primarily through social drivers of health (SDOH), especially structural factors.
• The development of ML for health inherently risks worsening existing health inequities.
• A significant tension exists between the goals of ML in health and the imperative to reduce health disparities.

Conclusions:

• Integrating SDOH considerations throughout the ML development pipeline is essential for identifying and addressing algorithmic harms.
• Developing expertise in structural analysis and community-engaged methodologies is necessary for effectively considering SDOH.
• While SDOH can guide the creation of more equitable algorithms, interventions targeting the root causes of health inequities are paramount, as equitable algorithms may not always be feasible.