Novel Insights on Induced Sparsity in Multi-Time Attention Networks.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Time Series Analysis

Background:

• Deep learning models often struggle with sparse, irregularly sampled time-series data common in electronic health records.
• Existing methods do not fully leverage the inherent sparsity of physiological data.

Purpose of the Study:

• To investigate the impact of controlled data sparsity on the predictive performance of Multi-Time Attention Networks (mTAN).
• To explore sub-sampling strategies for efficiently training deep learning models on sparse physiological time-series data.

Main Methods:

• Inducing sparsity by sub-sampling time-series data at rates from 10% to 90%.
• Evaluating the performance of the Multi-Time Attention Networks (mTAN) on sub-sampled data.
• Conducting experiments on the Human Activity dataset and the Physionet 2012 mortality prediction task.

Main Results:

• Sub-sampling coupled with mTAN improved performance by 2% on the Human Activity dataset, using 80% fewer training time-points.
• Comparable performance to baseline was achieved on the Physionet dataset with 30% fewer time-points.
• Demonstrated that time-series data can be coarsely acquired when using advanced sparse-data handling networks like mTAN.

Conclusions:

• Controlled time-point sub-sampling enhances deep learning performance on sparse physiological time-series data.
• This methodology significantly reduces data requirements, making it valuable for applications with data acquisition and labeling constraints.
• Suggests potential for more efficient data acquisition in healthcare settings using robust deep learning architectures.