Introducing Task-Adaptive Loss to Multitask Learning for Electronic Healthcare Prediction

Document Type

Article

Abstract

Multi-Task Learning (MTL) is a machine learning paradigm that allows for a model to simultaneously train and predict multiple related tasks/targets using a shared core network and task-specific branching layers. The related nature of the tasks can be exploited to regularize the model and has been shown to increase performance in all tasks. Medical healthcare records (MHRs) taken on events such as pregnancy often track multiple related targets such as maternal morbidity factors and long-term health outcomes which serve as an excellent testbed for MTL models. Imbalance in tasks leads to MTL models being dominated by a few tasks as well as struggling to progress during training due to conflicting loss function parameters, resulting in generalized but unspecific models. This study introduces Task-Adaptive Loss weighting, a novel method to address imbalanced tasks and conflicting loss and demonstrates iton a prenatal MHR dataset to create a prediction tool for adverse maternal outcomes.

Digital Object Identifier (DOI)

https://doi.org/10.1109/ICHI57859.2023.00133

APA Citation

Tsien, E., Wu, D., Tong, Y., Fede, A. L.-D., & Gareau, S. (2023). Introducing Task-Adaptive Loss to Multitask Learning for Electronic Healthcare Prediction. 2022 IEEE 10th International Conference on Healthcare Informatics (ICHI), 727–730. https://doi.org/10.1109/ICHI57859.2023.00133

Rights

© Copyright 2025 IEEE - All rights reserved, including rights for text and data mining and training of artificial intelligence and similar technologies.

Link to Record

Share

COinS