BiT-MAC: Mortality prediction by bidirectional time and multi-feature attention coupled network on multivariate...

Area of Science:

• Biomedical Informatics
• Machine Learning in Healthcare
• Clinical Data Analysis

Background:

• Mortality prediction is vital for patient care and prognosis assessment.
• Multivariate time series (MTSs) analysis of clinical data offers insights but is challenged by sparse, irregular, and incomplete data.
• Existing methods struggle with modeling inter-MTS couplings and lack interpretability.

Purpose of the Study:

• To develop a novel deep learning model for robust mortality prediction from incomplete clinical MTSs.
• To effectively capture both temporal dependencies (intra-MTS) and relationships between variables (inter-MTS).
• To enhance model interpretability in clinical time series analysis.

Main Methods:

• Proposed a bidirectional time and multi-feature attention coupled network (BiT-MAC).
• Utilized a bidirectional recurrent neural network for temporal dependencies (intra-MTS coupling).
• Employed multi-head attention to model relationships among variables (inter-MTS coupling).
• Fused intra- and inter-MTS representations for missing value estimation and prediction.

Main Results:

• BiT-MAC demonstrated superior performance in mortality prediction on PhysioNet'2012 and COVID-19 datasets, outperforming existing models.
• The model effectively handled missing data, improving the robustness of MTS-based predictions.
• Feature interpretability was validated through a case study, highlighting the model's clinical utility.

Conclusions:

• BiT-MAC offers a significant advancement in analyzing incomplete clinical MTSs for mortality prediction.
• Modeling both intra-MTS and inter-MTS couplings is crucial for accurate and interpretable clinical predictions.
• The approach holds promise for improving patient prognosis and clinical decision-making.