Reproduce “Clinical Risk Prediction with Temporal Probabilistic Asymmetric Multi-Task Learning” on OpenResearch

About this paper

A. Tuan Nguyen, Hyewon Jeong, Eunho Yang, Sung Ju Hwang

Researchers from KAIST and University of Oxford developed Temporal Probabilistic Asymmetric Multi-Task Learning (TP-AMTL), a framework for clinical risk prediction that uses feature-level uncertainty to guide dynamic, asymmetric knowledge transfer across tasks and timesteps. The model consistently achieved higher AUROC scores on clinical datasets and effectively mitigated negative transfer, providing interpretable insights into task interactions.

Key insights

• Feature-level uncertainty serves as a more reliable and dynamic criterion for knowledge transfer than task-level loss, effectively preventing negative transfer in complex multi-task scenarios.
• Explicitly modeling and enabling knowledge transfer across both tasks and their respective timesteps is crucial for accurately capturing the dynamic interdependencies in clinical time-series data.
• The probabilistic framework, incorporating both epistemic and aleatoric uncertainties, is vital for the robustness of the knowledge transfer mechanism, preventing overfitting and ensuring reliable sharing.

Summary by alphaXiv.