Explainable Spatio-Temporal GCNNs for Irregular Multivariate Time Series: Architecture and Application to ICU Patient Data

Óscar Escudero-Arnanz, Cristina Soguero-Ruiz, Antonio G. Marques · 2024

Read Explainable Spatio-Temporal GCNNs for Irregular Multivariate Time Series: Architecture and Application to ICU Patient Data by Óscar Escudero-Arnanz, Cristina Soguero-Ruiz, Antonio G. Marques (2024) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2411-01070.txt.

Description

Abstract In this paper, we present XST-GCNN (eXplainable Spatio-Temporal Graph Convolutional Neural Network), an innovative architecture designed for processing heterogeneous and irregular Multivariate Time Series (MTS) data. Our processing architecture captures both temporal and feature dependencies within a unified spatio-temporal pipeline by leveraging a GCNN that uses a spatio-temporal graph and aims at optimizing predictive accuracy and interpretability.

Catalog record

CREATORÓscar Escudero-Arnanz, Cristina Soguero-Ruiz, Antonio G. Marques
YEAR2024
COLLECTIONPAPERS
FORMATTXT
FILE SIZE92 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2411-01070

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Explainable Spatio-Temporal GCNNs for Irregular Multivariate Time Series: Architecture and Application to ICU Patient Data Óscar Escudero-Arnanz, Cristina Soguero-Ruiz, Antonio G. Marques Abstract In this paper, we present XST-GCNN (eXplainable Spatio-Temporal Graph Convolutional Neural Network), an innovative architecture designed for processing heterogeneous and irregular Multivariate Time Series (MTS) data. Our processing architecture captures both temporal and feature dependencies within a unified spatio-temporal pipeline by leveraging a GCNN that uses a spatio-temporal graph and aims at optimizing predictive accuracy and interpretability. For graph estimation, we propose several techniques, including a novel approach based on the (heterogeneous) Gower distance. Once the graphs are estimated, we propose two approaches for graph construction: one based on the Cartesian product that treats temporal instants homogeneously, and a spatio-temporal approach that considers different graphs per time step. Finally, we propose two GCNN architectures: a standard GCNN with a normalized adjacency matrix, and a higher-order polynomial GCNN. In addition to accuracy, we emphasize explainability by designing an inherently interpretable architecture and conducting a thorough interpretability analysis, identifying key feature-time combinations that drive the model’s predictions. We evaluate XST-GCNN using real-world Electronic Health Record data from the University Hospital of Fuenlabrada to predict Multidrug Resistance (MDR) in Intensive Care Unit patients, a critical healthcare challenge associated with high mortality and complex treatments. Our architecture outperforms traditional models, achieving a mean Receiver Operating Characteristic Area Under the Curve score of 81.03 ± 2.43 \textbf{81.03}\pm\textbf{2.43} . Additionally, the interpretability analysis provides actionable insights into clinical factors driving MDR predictions, enhancing model transparency and trust. This work sets a new benchmark for addressing complex inference tasks with heterogeneous and irregular MTS, offering a versatile and interpretable solution for real-world applications. I Introduction Graphs have become powerful tools in both Machine Learning (ML) and Signal Processing (SP) due to their capacity to model complex interactions and capture intrinsic relationships within structured data [ 1 ] . In real-world applications, data often originate from multiple domains and exhibit heterogeneous characteristics, presenting a significant challenge for graph analysis and estimation [ 2 , 1 ] . This heterogeneity spans various data types—including numerical, categorical, and textual, among others—collected at different temporal and spatial scales, further complicating traditional graph construction methods [ 2 ] . These methods frequently rely on domain-specific adjustments for each variable or data source, leading to inconsistent graph representations and limiting their generalizability [ 2 , 3 ] . To address the complexities described above, graph learning techniques have emerged as a powerful approach, allowing the inference of graph topologies directly from data, without imposing prior assumptions on the graph structure [ 4 , 5 ] . However, conventional methods that construct multiple graphs for different data characteristics introduce redundancy and computational inefficiencies [ 6 ] . This highlights the need for advanced models capable of estimating a unified graph that integrates relationships among heterogeneous variables while maintaining computational efficiency and interpretability [ 7 ] . Once a unified graph that captures the relationships within heterogeneous data is estimated, it becomes foundational for subsequent inference and prediction tasks. Graph Convolutional Neural Networks (GCNNs) have proven highly effective in this context, leveraging the graph structure to capture hidden patterns by iterating over nodes and aggregating information from their neighbors [ 8 ] . More recent developments, such as Spatio-Temporal Graph Neural Networks (ST-GNNs), combine the strengths of GCNNs with Recurrent Neural Networks (RNNs) to handle both spatial and temporal dependencies in dynamic data [ 9 , 3 , 10 ] . These models have demonstrated significant potential in solving complex problems involving heterogeneous and dynamic datasets [ 9 ] . In response to the challenges of heterogeneous temporal data, we propose the eXplainable Spatio-Temporal Graph Convolutional Neural Network (XST-GCNN). This architecture is designed to efficiently capture and integrate Spatio-Temporal (ST) relationships within irregular Multivariate Time Series (MTS). Unlike conventional models, XST-GCNN unifies discrete and continuous data types into a cohesive graph representation, allowing for the modeling of complex, domain-spanning interactions. This approach not only estimates a single graph that encapsulates both spatial and temporal dependencies but also provides an interpretable architecture, essential for clinical decision-making. By capturing both local and global relationships within medical data, XST-GCNN offers a robust solution for enhancing outcome prediction and supporting decisions in high-stakes environments such as healthcare. The effectiveness of the XST-GCNN architecture is demonstrated through a case study on real-world medical data from the Intensive Care Unit (ICU) of the University Hospital of Fuenlabrada (UHF), addressing the critical issue of Multidrug Resistance (MDR). Recognized by the World Health Organization as a growing global threat, MDR complicates infection treatments, increases mortality rates, and imposes significant pressures on healthcare systems [ 11 , 12 , 13 ] . It is important to highlight that MDR is a particularly alarming subset of Antimicrobial Resistance (AMR). The dataset, derived from Electronic Health Records (EHRs) of the ICU-UHF, was modeled as irregular MTS and included heterogeneous features such as real-valued and discrete variables. EHRs are frequently used to address a range of healthcare challenges, including early detection of sepsis, prediction of patient deterioration, and personalized treatment planning [ 14 ] . The inherent irregularity and variability in EHR data, recorded at varying intervals across different patients, exacerbate the limitations of traditional Time Series (TS) models, which struggle to provide accurate and interpretable predictions in clinical settings [ 3 , 15 ] . Graph-based models, on the other hand, excel in representing the intricate relationships within clinical data as networks, facilitating more comprehensive analysis and better clinical insights [ 4 ] . By capturing detailed, interconnected relationships within biomedical data, graph-based models enhance the integration and analysis of critical healthcare information, such as patient histories, diagnoses, and treatments [ 16 , 4 ] . When combined with advanced SP techniques, these models significantly improve the extraction of meaningful patterns from clinical data, boosting both interpretability and predictive accuracy [ 5 , 17 ] . GCNNs, in particular, have revolutionized medical informatics by leveraging graph structures to reveal hidden patterns and improve the interpretability of clinical data [ 18 ] . This is especially valuable for assessing patient conditions and predicting clinical outcomes. Despite their potential, current research in GCNNs often overlooks the unique challenges posed by irregular MTS, such as varying recording frequencies and the heterogeneous nature of clinical data. The XST-GCNN architecture proposed in this paper directly addresses these challenges (heterogeneity, irregular MTS, spatial and temporal dependencies, and the need for interpretability) by integrating spatial and temporal dependencies within heterogeneous clinical data, offering a unified, interpretable architecture that enhances clinical decision-making in the context of MDR. I-A Related Work The heterogeneity in real-world data has led to growing interest in architectures that support representation and learning in heterogeneous graphs [ 19 , 20 ] . Various approaches address this heterogeneity from different perspectives: [ 20 ] explores the integration of natural language and code snippets, focusing on structural and semantic heterogeneity across mixed domains, while [ 19 ] models recommendation systems using heterogeneous graphs defined by multiple nodes and relation types. In [ 21 ] , heterogeneity is examined as a combination of node and edge types within information networks, employing meta paths to interpret relational dependencies. A comprehensive review

Read the complete text