Clinical Phenotyping Prediction via Auxiliary Task Selection and Adaptive Shared-Space Correction
Published in CACAI, CAAI A, 2023
Recommended citation: Xiao Yang, Ning Liu*, Jianbo Qiao, Haitao Yuan, Teng Ma, Yonghui Xu, Lizhen Cui. (2022). Clinical Phenotyping Prediction via Auxiliary Task Selection and Adaptive Shared-Space Correction. In: Fang, L., Povey, D., Zhai, G., Mei, T., Wang, R. (eds) Artificial Intelligence. CICAI 2022. Lecture Notes in Computer Science(), vol 13605. Springer, Cham. https://doi.org/10.1007/978-3-031-20500-2_36 https://doi.org/10.1007/978-3-031-20500-2_36
Clinical Phenotyping is a fundamental task in clinical services, which assessments whether a patient suffers a medical condition of interest. Existing works focus on learning better patients’ representations. Recently, multi-task learning has been proposed to transfer knowledge from different tasks and achieved promising performance. However, the existing multi-task models still suffer from the serious negative transfer and slow convergence problem when multiple phenotype tasks are trained together. Meanwhile, phenotype relatedness is ignored, limiting to boost the performance of the multi-task learning for the phenotype prediction. To address these issues, we propose a private-shared multitask framework with auxiliary task selection and adaptive shared-space correction for phenotype prediction (MTL_AC). To start with, we design an auxiliary task selection method to find the most compatible phenotype task against one task by using phenotype relatedness. And then, a novel adaptive shared-space correction mechanism is proposed to address the negative transfer and slow convergence problem when two tasks are jointly trained under the private-shared multitask framework. The experimental results show that the proposed method performs better on various phenotype prediction tasks.