Fairness and bias correction in machine learning for depression prediction across four study populations

Area of Science:

• Mental Health Research
• Artificial Intelligence in Healthcare
• Computational Social Science

Background:

• Mental healthcare faces significant stigma and inequality, particularly in underserved populations.
• Existing data inequalities can be amplified by machine learning (ML) models, reinforcing structural biases.
• Biased ML models pose a risk in clinical applications like depression prediction.

Purpose of the Study:

• To systematically investigate bias in ML models for depression prediction across diverse populations and countries.
• To evaluate the effectiveness of standard and novel mitigation techniques in reducing unfair bias.
• To highlight the importance of fairness analysis and transparent reporting in ML for mental health.

Main Methods:

• Conducted a systematic study of ML models for depression prediction using four distinct case studies.
• Applied standard ML approaches and evaluated their inherent biases.
• Implemented and assessed both established and a novel post-hoc bias mitigation technique.

Main Results:

• Standard ML approaches consistently exhibited biased behavior in depression prediction.
• Bias mitigation techniques, including the proposed post-hoc method, effectively reduced unfair bias.
• No single ML model achieved equal outcomes across all populations, underscoring the complexity of fairness.

Conclusions:

• Fairness must be a key consideration during ML model selection for depression prediction.
• Transparent reporting on the impact of debiasing interventions is essential.
• Practitioners should adopt positive habits and address open challenges to enhance fairness in mental health ML models.