Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Learning to Select the Best Forecasting Tasks for Clinical Outcome Prediction

Yuan Xue, Nan Du, Anne Mottram, Martin Seneviratne, Andrew M. Dai

TL;DR

The paper tackles data-scarce clinical outcome prediction from high-dimensional EMR time-series and introduces AutoSelect, a meta-learning framework that automatically weights a broad set of self-supervised trajectory forecast tasks to pretrain an encoder-based patient representation. It uses a bi-level optimization with an inner self-supervised pretraining loop and an outer meta-objective that updates task weights via gradient-based updates, implemented efficiently with a first-order approximation. Empirical results on the MIMIC-III dataset show AutoSelect yields superior predictive performance, especially when labeled data are limited, and ablations reveal that top-selected tasks align with clinically meaningful signals rather than the full task set. The work highlights the value of targeted self-supervised pretraining and automated task selection for robust, data-efficient clinical risk prediction, while acknowledging broader ethical considerations in EMR-based modeling.

Abstract

We propose to meta-learn an a self-supervised patient trajectory forecast learning rule by meta-training on a meta-objective that directly optimizes the utility of the patient representation over the subsequent clinical outcome prediction. This meta-objective directly targets the usefulness of a representation generated from unlabeled clinical measurement forecast for later supervised tasks. The meta-learned can then be directly used in target risk prediction, and the limited available samples can be used for further fine-tuning the model performance. The effectiveness of our approach is tested on a real open source patient EHR dataset MIMIC-III. We are able to demonstrate that our attention-based patient state representation approach can achieve much better performance for predicting target risk with low resources comparing with both direct supervised learning and pretraining with all-observation trajectory forecast.

Learning to Select the Best Forecasting Tasks for Clinical Outcome Prediction

TL;DR

The paper tackles data-scarce clinical outcome prediction from high-dimensional EMR time-series and introduces AutoSelect, a meta-learning framework that automatically weights a broad set of self-supervised trajectory forecast tasks to pretrain an encoder-based patient representation. It uses a bi-level optimization with an inner self-supervised pretraining loop and an outer meta-objective that updates task weights via gradient-based updates, implemented efficiently with a first-order approximation. Empirical results on the MIMIC-III dataset show AutoSelect yields superior predictive performance, especially when labeled data are limited, and ablations reveal that top-selected tasks align with clinically meaningful signals rather than the full task set. The work highlights the value of targeted self-supervised pretraining and automated task selection for robust, data-efficient clinical risk prediction, while acknowledging broader ethical considerations in EMR-based modeling.

Abstract

We propose to meta-learn an a self-supervised patient trajectory forecast learning rule by meta-training on a meta-objective that directly optimizes the utility of the patient representation over the subsequent clinical outcome prediction. This meta-objective directly targets the usefulness of a representation generated from unlabeled clinical measurement forecast for later supervised tasks. The meta-learned can then be directly used in target risk prediction, and the limited available samples can be used for further fine-tuning the model performance. The effectiveness of our approach is tested on a real open source patient EHR dataset MIMIC-III. We are able to demonstrate that our attention-based patient state representation approach can achieve much better performance for predicting target risk with low resources comparing with both direct supervised learning and pretraining with all-observation trajectory forecast.
Paper Structure (19 sections, 11 equations, 2 figures, 9 tables, 2 algorithms)

This paper contains 19 sections, 11 equations, 2 figures, 9 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Learning Tasks
  3. Automatic Task Selection
  4. Problem Formulation
  5. Bi-level Optimization
  6. Efficient Gradient-based Learning Algorithm
  7. Experiments
  8. Baselines and Experiment Setting
  9. Performance Comparison of Clinical Outcome Prediction
  10. Ablation Study
  11. Related Work
  12. Conclusion
  13. Broader Impact
  14. Appendix
  15. Full Algorithm
  16. ...and 4 more sections

Figures (2)

  • Figure 1: Schematic for guiding pretraining by supervised learning on a primary outcome via a nested-loop meta-learning process. The inner loop learning uses a sequence-to-sequence architecture for trajectory prediction. The utility of the learned representation from the encoder is measured by the supervised learning in the outer loop, and then used to update the weight of each trajectory prediction task in the inner loop via gradient descent.
  • Figure 2: (a) Pretraining and finetuning curves of competing methods. AUC-ROC on the validation set is reported. (b) The learned weights of the 96 auxiliary tasks for mortality prediction as the primary task. (c) Predictive performance with respect to different meta-learning processes of AutoSelect.