Mitigating Bias in Machine Learning Models with Ethics-Based Initiatives: The Case of Sepsis.

Area of Science:

• Medical Informatics
• Machine Learning Ethics
• Health Equity

Background:

• Machine learning models are increasingly used for predicting clinical outcomes like sepsis onset.
• Bias in these models, stemming from social determinants of health (SDOHs) and model design, can reduce accuracy and lead to unfair treatment.
• Adverse SDOHs disproportionately affect socioeconomically disadvantaged and marginalized populations.

Purpose of the Study:

• To discuss ethics-based strategies for mitigating bias in machine learning models predicting sepsis onset.
• To analyze how various biases, particularly those linked to SDOHs, affect model predictive accuracy.
• To propose ethical approaches to prevent disparate or unfair treatment by these models.

Main Methods:

• Literature review and ethical analysis of bias in machine learning.
• Examination of bias sources, including SDOHs and model construction.
• Discussion of ethically-grounded mitigation strategies.

Main Results:

• Bias from SDOHs and model design can significantly reduce the predictive accuracy of sepsis onset models.
• Synergistic effects of different biases can exacerbate accuracy issues.
• Ethical strategies can be developed to address and mitigate bias.

Conclusions:

• Ethical considerations are crucial for developing fair and accurate machine learning models in healthcare.
• Mitigation strategies are necessary to ensure equitable outcomes for all populations, especially vulnerable groups.
• The findings are applicable beyond sepsis prediction to other conditions influenced by SDOHs.