Fairness in machine learning-based hand load estimation: A case study on load carriage tasks.

Area of Science:

• Occupational Health and Safety
• Biomechanical Engineering
• Machine Learning in Ergonomics

Background:

• Predicting external hand load is crucial for workplace ergonomic assessments.
• Current methods often require direct observation or supplementary data.
• Existing machine learning models show systematic bias related to biological sex, particularly with imbalanced datasets.

Purpose of the Study:

• To develop a fair predictive model for external hand load that mitigates sex-based bias.
• To improve the accuracy and fairness of hand load predictions in ergonomic assessments.
• To address health and safety disparities for specific worker groups.

Main Methods:

• Developed a fair predictive model using a Variational Autoencoder with feature disentanglement.
• Separated sex-agnostic from sex-specific motion features for unbiased prediction.
• Compared the proposed algorithm against conventional machine learning models (k-NN, SVM, Random Forest).

Main Results:

• The proposed algorithm achieved a mean absolute error of 3.42.
• Demonstrated improved fairness metrics, including statistical parity and residual differences.
• Outperformed conventional models, especially when trained on imbalanced sex datasets.

Conclusions:

• Fairness-aware algorithms are essential to prevent health and safety disadvantages in the workplace.
• Feature disentanglement in Variational Autoencoders can create unbiased predictive models.
• The developed model offers a more equitable approach to ergonomic exposure assessments.