Multimodal Interpretable Data-Driven Models for Early Prediction of Antimicrobial Multidrug Resistance Using Multivariate Time-Series

TL;DR

This work tackles early AMR prediction in the ICU using multimodal EHR data by developing interpretable DNNs that fuse static patient features with irregular multivariate time-series. The authors introduce three fusion-focused architectures—Joint Heterogeneous Fusioner, First Hidden State Initializer, and Late Fusion models—alongside feature selection and interpretable mechanisms (attention, time-slot perturbations, and dynamic masks). Through experiments on 3,158 ICU stays with a 14-day window, multimodal models achieve competitive ROC AUC (best around 0.84 under feature-selected conditions), with FS and interpretability analyses highlighting mechanical ventilation and AMR-neighbor counts as key drivers. The methodology promises practical ICU decision-support insights and is extensible to other EHR-based problems, with code available for replication.

Abstract

Electronic health records (EHR) is an inherently multimodal register of the patient's health status characterized by static data and multivariate time series (MTS). While MTS are a valuable tool for clinical prediction, their fusion with other data modalities can possibly result in more thorough insights and more accurate results. Deep neural networks (DNNs) have emerged as fundamental tools for identifying and defining underlying patterns in the healthcare domain. However, fundamental improvements in interpretability are needed for DNN models to be widely used in the clinical setting. In this study, we present an approach built on a collection of interpretable multimodal data-driven models that may anticipate and understand the emergence of antimicrobial multidrug resistance (AMR) germs in the intensive care unit (ICU) of the University Hospital of Fuenlabrada (Madrid, Spain). The profile and initial health status of the patient are modeled using static variables, while the evolution of the patient's health status during the ICU stay is modeled using several MTS, including mechanical ventilation and antibiotics intake. The multimodal DNNs models proposed in this paper include interpretable principles in addition to being effective at predicting AMR and providing an explainable prediction support system for AMR in the ICU. Furthermore, our proposed methodology based on multimodal models and interpretability schemes can be leveraged in additional clinical problems dealing with EHR data, broadening the impact and applicability of our results.

Figures (6)

• Figure 1: Graphical illustration of the workflow implemented. As illustrated in the left column, we begin by running a pre-processing stage to promote consistent and reliable results. Then, non-multimodal and multimodal models using all available features are trained (see Sec. \ref{['sec:NN']} and central column). We perform a knowledge extraction step for studying the most important features and time-slots, using two different FS schemes (see Sec. \ref{['subsec:FeatureSelection']} and right column) and different interpretable models (see Sec. \ref{['subsec:Interpretable_methods']} and right column). Once the most important variables are selected, we train models using the knowledge acquired using the FS and interpretable schemes. Finally, the models' performance and interpretability are evaluated using several figures of merit (see Sec. \ref{['sec:Experiments']}).
• Figure 2: High-level architecture of the FHSI. FHSI deals with static and time-varying inputs. The different blocks of the architecture are represented using different colors. The SE is represented in different green colors; the light green color blocks represent a first embedding mapping network, and the dark green color blocks represent the Variable Selection Network. The GRU block is represented in a light blue box, and the last non-linear dense layer is represented in a dark blue box.
• Figure 3: Matrix of features (in columns) and FS approaches (in rows, organized as classical and PFI techniques). The blue cells (darker ones) represent the selected features. Note that the NM results consider two different models: a MLP for dealing with static data and an RNN when considering MTS.
• Figure 4: Importance score heatmap using all the features as input representing the average of the $A_{i}$ matrices corresponding to the NLHA model. Columns represent features, while rows show time-slots of the MTS under study ('0' refers to the day of the ICU admission).
• Figure 5: Importance score heatmap representing the matrix ${\stackunder[1.2pt]{$$\mathbf A$$}{} }$ of the HAM model when using all the MTS variables. Columns represent features, while rows show time-slots of the MTS under study ('0' refers to the day of the ICU admission).