EMIT: A New Approach for Irregular Time Series in Healthcare AI

Excited to share our paper “EMIT: Event-Based Masked Auto Encoding for Irregular Time Series” published at ICDM 2024. Together with A/Prof. Sen WANG, Dr Ruihong Qiu, A/Prof. Adam Irwin and Prof. Shazia Sadiq, we explore how irregular time series (like vital signs and lab results recorded at uneven intervals) challenge existing AI models and how our proposed framework, EMIT, improves clinical decision support through better representation learning. Special thanks to CIRES, Queensland Health and The University of Queensland for supporting this research.

Read full paper at https://arxiv.org/pdf/2409.16554

Our Approach
We introduce EMIT, a pretraining framework based on transformer architecture, tailored for irregular clinical time series data. EMIT learns by:

• Finding important points in irregular time series
• Pretraining by masking and predicting those points
• Use the pretrained model for any downstream task (e.g., outcome prediction)

Key Findings

Improved Representation Learning: EMIT captures important variations without losing timing information, outperforming generic pretext approaches for irregular time series.

Data Efficiency: On benchmark healthcare datasets (MIMIC-III & PhysioNet Challenge 2012), EMIT achieved strong results using only 50% of labeled data, reducing reliance on costly annotations.

Task Relevance: By designing pretext tasks specific to irregular time series, EMIT delivers more reliable clinical predictions compared to standard forecasting approaches.

How can we design AI that adapts to the messy, irregular reality of clinical data while still delivering trustworthy predictions?