TY - GEN
T1 - Fairness in Survival Outcome Prediction for Medical Treatments
AU - Li, Yunqi
AU - Chen, Hanxiong
AU - Zhang, Lanjing
AU - Zhang, Yongfeng
N1 - Publisher Copyright: © 2024 IEEE.
PY - 2024
Y1 - 2024
N2 - Cancer is a life-threatening disease and affects millions of people globally. Predicting the survival outcome of patients after receiving a cancer diagnosis is of great importance for personalized treatment planning. While predictive machine learning models have shown promising results in predicting survival outcomes, they may result in unfair treatments against certain groups, such as those based on age, gender, race and socioeconomic status. In this paper, we analyze and identify what kind of fairness criterion should be developed in cancer survival predictions. Considering the high stakes involved, we believe the basis for developing fairness in this field should be assisting to improve the model performance of diverse patients, rather than hurting the performance of certain patients in exchange for better performance of other patients. We suggest min-max fairness as an appropriate concept that seeks to minimize the group error of the worst-off group, instead of pursuing a strict equal error rate across various patient groups which would reduce the utility of the advantaged groups. We investigate the application of min-max fairness through active sampling in predicting the survival outcome of cancer patients using machine learning models, and evaluate the fairness-aware method on several publicly available cancer datasets. The experimental results demonstrate that the fairness-aware predictive method can help improve the model performance of the disadvantaged patient group without hurting the advantaged patient group, and thus increase the overall model performance for all patients.
AB - Cancer is a life-threatening disease and affects millions of people globally. Predicting the survival outcome of patients after receiving a cancer diagnosis is of great importance for personalized treatment planning. While predictive machine learning models have shown promising results in predicting survival outcomes, they may result in unfair treatments against certain groups, such as those based on age, gender, race and socioeconomic status. In this paper, we analyze and identify what kind of fairness criterion should be developed in cancer survival predictions. Considering the high stakes involved, we believe the basis for developing fairness in this field should be assisting to improve the model performance of diverse patients, rather than hurting the performance of certain patients in exchange for better performance of other patients. We suggest min-max fairness as an appropriate concept that seeks to minimize the group error of the worst-off group, instead of pursuing a strict equal error rate across various patient groups which would reduce the utility of the advantaged groups. We investigate the application of min-max fairness through active sampling in predicting the survival outcome of cancer patients using machine learning models, and evaluate the fairness-aware method on several publicly available cancer datasets. The experimental results demonstrate that the fairness-aware predictive method can help improve the model performance of the disadvantaged patient group without hurting the advantaged patient group, and thus increase the overall model performance for all patients.
KW - AI for Health
KW - Algorithmic Bias
KW - Trustworthy AI
UR - http://www.scopus.com/inward/record.url?scp=85190627140&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85190627140&partnerID=8YFLogxK
U2 - 10.1109/CISS59072.2024.10480160
DO - 10.1109/CISS59072.2024.10480160
M3 - Conference contribution
T3 - 2024 58th Annual Conference on Information Sciences and Systems, CISS 2024
BT - 2024 58th Annual Conference on Information Sciences and Systems, CISS 2024
PB - Institute of Electrical and Electronics Engineers Inc.
T2 - 58th Annual Conference on Information Sciences and Systems, CISS 2024
Y2 - 13 March 2024 through 15 March 2024
ER -