AI Model Eliminates Bias in Critical Decisions Across Health, Education, and Employment

Researchers at the University of Navarra have developed a novel methodology that enhances equity and reliability in critical decision-making models used across health, education, justice, and hiring sectors. The team, led by Alberto García Galindo, Marcos López De Castro, and Rubén Armañanzas Arnedillo from the University of Navarra’s Data Science and Artificial Intelligence (DATAI) Institute, has introduced a system that optimizes the parameters of reliable machine learning models.

These models, which are algorithmic AI tools ensuring transparent predictions with certain confidence levels, can potentially reduce disparities related to sensitive attributes such as race, gender, or socioeconomic status. The study, published in Machine Learning, combines advanced prediction techniques with evolutionary learning algorithms that offer rigorous confidence levels and equitable coverage across different social and demographic groups.

This new AI model ensures the same level of reliability regardless of individual characteristics, guaranteeing fair and unbiased results. The widespread use of AI in sensitive areas has raised ethical concerns due to potential algorithmic discrimination, explains Armañanzas Arnedillo, principal researcher at DATAI. Our approach allows businesses and policymakers to choose models that balance efficiency and equity according to their needs, responding to emerging regulations,” he adds.

The method was successfully tested on four real-world datasets related to economic income, criminal reincarceration, hospital readmissions, and school admissions. Results demonstrated that the new algorithms could significantly reduce disparities without compromising prediction accuracy.

“For instance, we found striking biases in predicting school admissions, indicating a significant lack of impartiality based on family financial status,” points out Alberto García Galindo, predoctoral researcher at DATAI and first author of the article. “Our methodology often achieves reducing such biases without compromising the model’s predictive capacity. In fact, we found solutions where discrimination was almost completely eliminated while maintaining prediction accuracy.”

Furthermore, the methodology offers a ‘Pareto frontier’, which allows visualizing the best available options according to priorities and understanding, for each case study, how algorithmic equity and precision relate. According to the researchers, this innovation has broad applicability in sectors where AI supports critical decision-making in a reliable and ethical manner. García Galindo notes that “our methodology not only contributes to equity but also provides a deeper understanding of how model configuration affects results, which could guide future investigations in AI regulation.” The researchers have made the study’s code and data publicly available to promote research and transparency in this emerging field.