Early detection of Multidrug Resistance using Multivariate Time Series analysis and interpretable patient-similarity...

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Clinical informatics

Background:

• Multidrug Resistance (MDR) is a significant global health threat with severe socioeconomic impacts.
• Existing prediction models lack interpretability, hindering clinical application.
• There is a need for explainable AI (XAI) in critical care for better patient management.

Purpose of the Study:

• To develop a novel interpretable Machine Learning (ML) approach for predicting Multidrug Resistance (MDR).
• To achieve accurate MDR inference using patient similarity representations.
• To enhance explainability for identifying key risk factors and patient subgroups.

Main Methods:

• Modeled patients as Multivariate Time Series (MTS) to capture clinical progression and interactions.
• Utilized MTS-based similarity metrics (e.g., Dynamic Time Warping, Time Cluster Kernel) as inputs for classification models (Logistic Regression, Random Forest, SVM).
• Employed graph-based methods (spectral clustering, t-SNE) on patient similarity networks for pattern extraction and subgroup identification.

Main Results:

• Achieved an 81% Receiver Operating Characteristic Area Under the Curve on an ICU dataset.
• Outperformed existing ML and deep learning models through graph-based patient similarity.
• Identified key MDR risk factors (e.g., antibiotic exposure, invasive procedures) and meaningful patient clusters.

Conclusions:

• Patient similarity representations and graph-based methods are effective for MDR prediction and interpretability.
• The interpretable ML approach enhances prediction accuracy and identifies critical risk factors.
• This framework enables early detection, targeted interventions, and improved patient stratification in critical care settings.