From Data Poverty to Data Sovereignty: Operationalizing Gold-Standard Biomedical Datasets in Low- and Middle-Income...

Area of Science:

• Biomedical informatics
• Artificial intelligence in healthcare
• Global health equity

Background:

• AI healthcare tools predominantly use data from high-income countries, leading to performance issues in LMIC populations.
• Underrepresentation of LMICs in AI datasets exacerbates health disparities.
• Current AI tools lack geographic and racial diversity, with fewer than 4% reporting such data.

Purpose of the Study:

• To address the disparity in AI healthcare data by proposing models for LMICs.
• To transition LMICs from "data poverty" to "data sovereignty."
• To enhance global AI equity through improved data representation.

Main Methods:

• Highlighting the Medical Imaging Datasets for India (MIDAS) initiative as a model.
• Utilizing a 4-domain Dataset Quality Matrix for representativeness, documentation, technical fidelity, and governance.
• Proposing a multilateral South-South Data Commons with a harmonized rubric, distributed governance, and outcome-linked incentives.

Main Results:

• The MIDAS initiative demonstrates the feasibility of creating high-quality, context-specific datasets.
• Initial MIDAS releases show value in developing robust, generalizable AI models.
• The proposed framework supports local stewardship and global collaboration for AI development.

Conclusions:

• LMICs can become autonomous data stewards, enhancing global AI equity.
• Institutionalizing quality control and accountability transforms LMICs into active contributors of biomedical data.
• The framework ensures the development of trustworthy and globally relevant AI datasets.