Related Experiment Video
Updated: Nov 1, 2025
05:01
A Detailed Protocol for Physiological Parameters Acquisition and Analysis in Neurosurgical Critical Patients
Published on: October 17, 2017
7.2K
Multitask prediction of organ dysfunction in the intensive care unit using sequential subnetwork routing.
Subhrajit Roy1, Diana Mincu1, Eric Loreaux1
1Google Health, London, United Kingdom.
Summary
Sequential subnetwork routing (SeqSNR) improves multitask learning (MTL) performance on electronic health records. This novel architecture enhances label efficiency, outperforming single-task and naive multitask models, especially when data is limited.
Area of Science:
- Machine Learning
- Biomedical Informatics
- Artificial Intelligence
Background:
- Multitask learning (MTL) with electronic health records (EHRs) enables simultaneous prediction of multiple clinical outcomes.
- While MTL improves model performance and efficiency, it can suffer from negative transfer when tasks are poorly selected.
- A new architecture is needed to effectively route related tasks and mitigate negative transfer in EHR-based MTL.
Purpose of the Study:
- Introduce a novel sequential subnetwork routing (SeqSNR) architecture for EHR-based multitask learning.
- Evaluate SeqSNR's performance and label efficiency against single-task (ST) and naive multitask (MTL) models.
- Demonstrate SeqSNR's utility in scenarios with limited labeled data.
Main Methods:
- Developed a deep neural network utilizing a sequential subnetwork routing (SeqSNR) architecture with soft parameter sharing.
- Trained and compared models on the MIMIC-III dataset to predict six clinical endpoints: acute kidney injury, continuous renal replacement therapy, mechanical ventilation, vasoactive medications, mortality, and length of stay.
- Assessed discriminative performance and label efficiency across ST, naive MTL, and SeqSNR models.
Main Results:
- SeqSNR achieved a statistically significant performance improvement on 4 out of 6 tasks compared to ST and naive MTL.
- SeqSNR demonstrated superior label efficiency, outperforming ST models across all tested dataset sizes (1%, 5%, and 10% of labels).
- Average area under the precision-recall curve boosts for SeqSNR were 2.1% (1% labels), 2.9% (5% labels), and 2.1% (10% labels).
Conclusions:
- The SeqSNR architecture offers superior label efficiency compared to single-task and naive multitask learning approaches.
- SeqSNR is particularly valuable in clinical settings where obtaining endpoint labels for electronic health records is challenging.
- This approach facilitates more effective cross-learning between related tasks in multitask learning models.
