Prostate cancer risk prediction (PCRP) is crucial in guiding clinical decision-making and ensuring accurate diagnoses. The area under the receiver operating characteristic curve (AUC) is typically used for the evaluation of PCRP models. However, AUC considers regions with high false positive rates (FPRs), which are not applicable in clinical practice. To address this concern, we propose to use partial AUC (pAUC) as a more clinically meaningful metric which evaluates PCRP models with restricted FPR. Moreover, we propose a new PCRP framework named pAUCP, which optimizes pAUC to train PCRP models and adopts model ensemble to further enhance its usability. We construct clinical datasets obtained from two medical centers over an extended period to evaluate the proposed pAUCP framework. Extensive experiments demonstrate the rationality and superiority of the pAUCP framework, especially the cross-time and cross-center transferability of the obtained PCRP model.
Citation:
@inproceedings{zhu2024improving,
title = {Improving Prostate Cancer Risk Prediction through Partial AUC Optimization},
author = {Zhu, Xinyuan and Ren, Xiaohan and Shi, Wentao and Wang, Changming and Liu, Xuehan and Liu, Yuqing and Tao, Tao and Feng, Fuli},
booktitle = {Companion Proceedings of the ACM Web Conference 2024},
series = {WWW ’24 Companion},
location = {Singapore, Singapore.},
url = {https://doi.org/10.1145/3589335.3651458},
doi = {0.1145/3589335.3651458},
numpages = {4},
year = {2024}
}