Multi-Task Deep Neural Networks for Irregularly Sampled Multivariate Clinical Time Series.

Area of Science:

• Biomedical Informatics
• Machine Learning
• Clinical Data Science

Background:

• Clinical time series data from Electronic Health Records (EHR) are often irregular with missing values and varying time intervals.
• Current deep learning methods use recurrent neural networks and time decay for imputation and prediction separately, limiting overall performance.

Purpose of the Study:

• To develop a multi-task deep neural network that simultaneously performs imputation and risk prediction on clinical time series data.
• To enhance accuracy in both imputation of missing EHR data and prediction of in-hospital mortality.

Main Methods:

• A novel multi-task deep neural network architecture was designed, integrating imputation as an auxiliary task for risk prediction.
• The model incorporates time decay mechanisms to effectively handle variable time intervals in clinical data.
• Validation was performed on two public EHR databases for clinical time series imputation and in-hospital mortality prediction.

Main Results:

• The proposed deep imputation-prediction network significantly outperformed state-of-the-art methods in both imputation and prediction tasks.
• Experimental results demonstrated the critical role of time decay mechanisms in achieving superior performance.
• The simultaneous approach yielded more accurate imputation and prediction results compared to sequential methods.

Conclusions:

• The novel deep imputation-prediction network offers improved accuracy for EHR data analysis.
• Simultaneous multi-task learning, enhanced by time decay mechanisms, is a promising direction for clinical time series modeling.
• Future research should explore advanced time decay mechanisms to further boost multi-task learning performance.