Fair Machine Learning Model for Bladder Cancer Survival Prediction
Author Information
Author(s): Focsa Mircea, Bozzo Anthony, Hong Pengyu, Carbunaru Samuel, Neshatvar Yassamin, Do Hyungrok, Murray Katie, Ranganath Rajesh, Nayan Madhur
Primary Institution: New York University School of Medicine
Hypothesis
Can a machine learning model predict survival after radical cystectomy for bladder cancer while ensuring fairness across sex and racial subgroups?
Conclusion
The study found that the machine learning model exhibited bias across sex and racial subgroups, but fairness was improved through algorithm unfairness mitigation techniques.
Supporting Evidence
- The best naive model was extreme gradient boosting (XGBoost) with an F1-score of 0.860.
- All unfairness mitigation techniques increased the equalized odds ratio (eOR).
- The final mitigated model achieved an eOR of 0.750.
Takeaway
Researchers created a computer program to help predict how long bladder cancer patients might live after surgery, and they made sure it works fairly for everyone, no matter their gender or race.
Methodology
The study used data from the National Cancer Database to train and compare various machine learning algorithms to predict 5-year survival after radical cystectomy.
Potential Biases
The model showed bias in performance across different sex and racial subgroups.
Limitations
The final model did not achieve perfect fairness, indicating residual unfairness, and the study focused only on sex and race without considering other potential biases.
Participant Demographics
23.1% female, 91.5% White, 5% Black, 2.3% Hispanic, 1.2% Asian.
Statistical Information
Confidence Interval
95% CI 0.849-0.869
Digital Object Identifier (DOI)
Want to read the original?
Access the complete publication on the publisher's website