Explainable Temporal Inference for Irregular Multivariate Time Series. A Case Study for Early Prediction of Multidrug...

Area of Science:

• Artificial Intelligence
• Healthcare Informatics
• Clinical Prediction Models

Background:

• Multivariate Time Series (MTS) data presents challenges in healthcare due to irregularity and temporal dependencies.
• Existing "MTS-to-TS" models often lack clinical explainability, hindering adoption.

Purpose of the Study:

• To introduce novel eXplainable Artificial Intelligence (XAI) methods for "MTS-to-TS" architectures.
• To enable tracking of patient evolution and identify key variables for adverse outcomes.
• To evaluate the framework on Multidrug Resistance (MDR) and circulatory failure prediction.

Main Methods:

• Developed Irregular Time SHapley Additive exPlanation (IT-SHAP) for post-hoc analysis.
• Implemented Hadamard Attention for intrinsic temporal dependency capture.
• Utilized Causal Conditional Mutual Information for pre-hoc feature selection.

Main Results:

• GRU with Hadamard Attention achieved high performance for MDR prediction (ROC-AUC=0.783).
• LSTM demonstrated superior performance for circulatory failure prediction (ROC-AUC=0.9970).
• IT-SHAP identified early antibiotic use and bacterial cultures as key risk factors, validated by clinicians.

Conclusions:

• The proposed framework provides temporal explainability in "MTS-to-TS" models.
• Clinicians can trace disease trajectories and understand variable contributions at each time step.
• Integration into EHR systems can improve early interventions, antimicrobial stewardship, and infection control.