Spatio-Temporal Machine Learning Architectures for Clinical Multivariate Time Series: Toward Accurate and Explainable Prediction in ICUs

Abstract

In the era of data-driven medicine, the development of trustworthy Artificial Intelligence (AI) systems that can effectively handle heterogeneous and irregular clinical data has become essential. Among the most pressing applications is the early prediction of Multidrug Resistance (MDR) in Intensive Care Units, where accurate and interpretable decision-support tools can directly influence patient outcomes. This doctoral dissertation proposes a suite of machine learning and graphbased architectures tailored to analyze real-world, multivariate Electronic Health Records, with the primary objective of enabling early detection of MDR at Intensive Care Units while preserving clinical transparency. By integrating methodologies from Time Series (TS) modeling, graph signal processing, and Explainable Artificial Intelligence, the thesis offers a comprehensive framework that jointly optimizes predictive performance and explainability—two pillars of responsible AI in healthcare. The dissertation is structured around four core objectives. First, a clinically-informed preprocessing pipeline is designed to handle a large-scale, 16-year Electronic Health Record dataset from the University Hospital of Fuenlabrada. This step ensures patient synchronization, temporal consistency, and meaningful label generation based on microbiological cultures, thus establishing a robust foundation for temporal modeling. Second, a patient-to-patient similarity framework is introduced, leveraging dynamic time warping, time cluster kernel, and feature extraction to embed Multivariate Time Series (MTS) into interpretable similarity spaces. These embeddings support graph-based clustering and classification using classical machine learning models such as logistic regression, random forests, and support vector machines, achieving a competitive Receiver Operating Characteristic Area Under the Curve (ROC AUC) of up to 0.810. Unlike typical deep learning approaches, this lightweight method preserves explainability while delivering strong performance, enabling the identification of clinical patterns linked to MDR. Notably, non-MDR cases were associated with the presence of CF3, while MDR-positive patients showed increased exposure to mechanical ventilation and co-patient antibiotic usage over time, particularly from the PEN and AMG families. These insights demonstrate the framework’s ability to extract actionable information directly from raw MTS, supporting explainable and anticipatory decision-making in ICU settings. However, the method did not capture spatial dependencies or provide temporal explainability, both of which are addressed in subsequent contributions. Third, a novel explainable deep learning architecture—Explainable Spatio-Temporal Graph Convolutional Neural Network—is proposed to model spatio-temporal dependencies in MTS. This architecture integrates graph convolutional layers over temporally estimated adjacency matrices constructed using correlation, smoothness constraints, and a novel Heterogeneous Gower Distance. Two graph structures—Cartesian product graphs and spatio-temporal graphs—are evaluated. The model achieves robust predictive performance (ROC AUC 0.810 ± 0.024), while attention mechanisms enable intrinsic explainability by assigning variable-level importance scores across time. Attention analysis reveals clinically meaningful patterns, such as early exposure to antibiotics (e.g., CAR), organ dysfunction markers (e.g., renal and respiratory failure), and co-treatment with neighboring patients during the first 24 hours—consistently associated with increased MDR risk. These findings, validated across real and xvi Abstract synthetic datasets, confirm the model’s ability to preserve the structural and temporal complexity of clinical data while delivering interpretable insights. Fourth, the thesis addresses a critical gap in temporal explainability for MTS-to-TS tasks. Most traditional Explainable Artificial Intelligence methods provide static or global importance scores and fail to capture how feature relevance evolves over time. To overcome this limitation, three novel methods are introduced: (i) a pre-hoc approach based on causal conditional mutual information, (ii) an intrinsic Hadamard attention mechanism, and (iii) a post-hoc method called Irregular Time SHapley Additive Explanation (IT-SHAP). These methods offer fine-grained, temporally resolved interpretations of MDR risk trajectories, allowing clinicians to understand not only what the model predicts, but also when and why specific variables contribute. IT-SHAP uncovered consistent patterns aligned with clinical reasoning: early Staphylococcus and Pseudomonas cultures and multiorgan failure were strong MDR predictors, whereas insulin therapy and artificial nutrition were associated with non-MDR profiles. Additionally, the importance of bacterial cultures decreased over time, potentially reflecting treatment effects. Expert validation confirmed the practical utility of these methods for supporting explainable, timely, and actionable decision-making in critical care. Results demonstrate that temporal and relational modeling are critical for developing robust AI systems in clinical settings. By combining explainable MTS representations, adaptive graph estimation strategies, and time-resolved explainability, this research contributes to the growing field of trustworthy AI in healthcare. The proposed architectures empower clinicians to monitor the probability of MDR onset, identify patient-specific risk patterns, and intervene proactively, supported by transparent and data-driven evidence. Finally, this work identifies several promising directions for future research, including the use of topological data analysis for enhanced graph estimation, the integration of static patient variables into multimodal graph architectures, and the exploration of novel explainability methods for MTS-to-TS inference tasks. These advances aim to further personalize, stabilize, and clarify clinical predictions, reinforcing explainable graph modeling as a cornerstone in the development of trustworthy AI systems for critical care decision support.