Multimodal interpretable data-driven models for early prediction of multidrug resistance using multivariate time...

Area of Science:

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

Background:

• Electronic Health Records (EHRs) contain valuable multimodal patient data, including static demographics and dynamic Multivariate Time Series (MTS).
• Integrating static and temporal data enhances clinical predictions, but Deep Neural Networks (DNNs) lack interpretability, hindering clinical adoption.
• Multidrug Resistance (MDR) in Intensive Care Units (ICUs) poses a significant challenge, necessitating accurate prediction and understanding.

Purpose of the Study:

• To develop interpretable multimodal DNN architectures for predicting and understanding the emergence of MDR in ICUs.
• To integrate static demographic data with temporal variables for a holistic patient health view.
• To enhance predictive performance and model interpretability for clinical decision support.

Main Methods:

• Proposed interpretable multimodal DNN architectures integrating static and temporal EHR data.
• Combined feature selection with attention mechanisms and post-hoc explainability tools.
• Evaluated model performance using Receiver Operating Characteristic Area Under the Curve (ROC AUC).

Main Results:

• Achieved a ROC AUC of 76.90 ± 3.10, significantly outperforming baseline models.
• The methodology effectively reduced feature redundancy and highlighted key risk factors for MDR.
• Demonstrated improved model accuracy and robustness through integrated feature selection and attention mechanisms.

Conclusions:

• The proposed framework provides a scalable and interpretable solution for MDR prediction in ICUs using EHR data.
• The methodology enhances predictive accuracy while offering crucial explanatory insights into risk factors.
• Supports timely, evidence-based interventions to improve patient outcomes in critical care settings.