(C) PLOS One
This story was originally published by PLOS One and is unaltered.
 . . . . . . . . . .



AI modeling for outbreak prediction: A graph-neural-network approach for identifying vancomycin-resistant enterococcus carriers [1]

['Gregor Donabauer', 'Department Of Infection Prevention', 'Infectious Diseases', 'University Medical Center Regensburg', 'Regensburg', 'Information Science', 'University Of Regensburg', 'Anca Rath', 'Aila Caplunik-Pratsch', 'Anja Eichner']

Date: 2025-05 

The isolation of affected patients and intensified infection control measures are used to prevent nosocomial transmission of vancomycin-resistant enterococci (VRE), but early detection of VRE carriers is needed. However, there are still no standard screening criteria for VRE, which poses a significant threat to patient safety. Our study aimed to develop and evaluate an artificial intelligence (AI)-based approach for identifying and predicting of at-risk patients who could assist infection prevention and control staff through a human-in-the-loop approach. We used data from 8,372 patients, combining more than 125,000 movements within our hospital with patient-related information to create time-dependent graph sequences and applied graph neural networks (GNNs) to classify patients as VRE carriers or noncarriers. Our model achieves a macro F1 score of 0.880 on the task (sensitivity of 0.808, specificity of 0.942). The parameters with the strongest impact on the prediction are the codes for clinical diagnosis (ICD) and operations/procedures (OPS), which are integrated as high-dimensional patient node features in our model. We demonstrate that modeling a “living” hospital with a GNN is a promising approach for the early detection of potential VRE carriers. This proves that AI-based tools combining heterogeneous information types can predict VRE carriage with high sensitivity and could therefore serve as a promising basis for future automated infection prevention control systems. Such systems could help enhance patient safety and proactively reduce nosocomial transmission events through targeted, cost-efficient interventions. Moreover, they could enable a more effective approach to managing antimicrobial resistance.

The rise of antibiotic-resistant bacteria, particularly vancomycin-resistant enterococci (VRE), poses a significant challenge in hospitals worldwide. VRE, which naturally reside in the human gut, can spread easily in healthcare settings, primarily through contact and contaminated surfaces. These bacteria are highly resilient, surviving for months in dry environments, making them difficult to control and eradicate. Certain patients, such as those with central venous catheters or Clostridioides difficile infections, are more susceptible to VRE colonization, which can lead to severe bloodstream infections. To combat this, hospitals need effective infection prevention and control strategies. However, detecting and isolating VRE carriers is challenging, especially when many carriers are asymptomatic and undetected. Current screening methods are limited, and there is no consensus on the best approach. In response, our research introduces a novel deep learning model that uses a heterogeneous graph neural network to predict undetected VRE carriers in hospitals. This model integrates various patient data and transmission pathways, offering a dynamic and comprehensive representation of the hospital environment. By embedding different spatial entities into the graph and using advanced temporal learning techniques, our approach aims to improve early detection of VRE carriers, thereby reducing the risk of further infections.

Data Availability: The test data set is a historical data set of our hospital and may not be published for data protection reasons. The implementations of our approach are publicly available via GitLab and can be accessed via the following link: https://git.uni-regensburg.de/dog21258/vre-outbreak-detection.git

Fig 1. The idea behind the learning model presented in this paper is to link as much information as possible that is available as a standard in hospital systems in a time-dependent graph neural network. The output is a classification: Would the patient be positive or negative in a possible test for vancomycin resistant enterococci?.

Going beyond these limitations, we introduce a deep learning model utilizing a heterogeneous GNN to predict undetected VRE carriers in hospitals. The model combines multidimensional patient data with possible transmission pathways, incorporating both direct and indirect interactions over time ( Fig 1 ). Building on prior research [ 41 , 42 ], our approach involves two key innovations: (1) representing and embedding various spatial entities—such as organizational units, wards, rooms, and beds—into a heterogeneous graph structure, and (2) employing advanced temporal graph learning techniques for a comprehensive temporal representation. In the following, we refer to the first stage of development as the “static model” and the second stage as the “dynamic model”.

However, both studies failed to consider the detailed heterogeneous hospital environment and patient movement data, limiting nodes inside the graphs to a single type, such as either wards or patients. This limitation is critical, as VRE spread is influenced by surface contamination and patient interactions. In addition, Van Niekerk et al. rely on statistical models that are limited in their ability to learn nonlinear relationships in the data, while Gouareb et al. do not consider the temporal course of interactions and focus on MREs in general ignoring their different biological properties.

In the domain of epidemiology, where our work is located, graphs and graph neural networks (GNNs) have drawn particular interest for their capabilities in modeling epidemics and virus outbreaks [ 33 , 34 ]. Graphs, or networks, offer a unique way to model interactions between components in a system [ 35 ]. Generally, a graph can simply be interpreted as a collection of objects (i.e., nodes) alongside a set of interactions (i.e., edges) between pairs of these objects [ 36 ]. For example, studies have used graph data structures to investigate spatiotemporal settings for predicting the course of epidemic diseases such as influenza or COVID-19 [ 37 – 40 ]. These types of techniques have also been used for the automatic detection of outbreaks of VRE and multidrug-resistant Enterobacterales (MRE). For example, Van Niekerk et al. worked with a proprietary dataset to develop a spatiotemporal graph model with nodes as wards and patients as edges to detect VRE colonization at the ward level using a rule-based approach [ 41 ]. Furthermore, Gouareb et al. modeled temporally evolving information in homogeneous graphs to retrospectively classify patients at risk of MRE infection [ 42 ].

However, beyond analytical capabilities, the interaction between humans and AI is particularly important in medical contexts. Factors such as data privacy, security, and trust perception play a crucial role in the successful adoption of AI-driven solutions. Ensuring that AI tools are interpretable, transparent, and aligned with clinical decision-making processes is essential for their integration into real-world healthcare settings [ 28 , 29 ].

A major challenge to this approach – and generally in rapid infection prevention and control (IPC) - is the high number of unknown VRE carriers who can release bacteria into their environment [ 4 , 18 ] but are not subjected to special hygiene and isolation measures, facilitating further transmission. Currently, there is no consensus or sufficient evidence on the use of the VRE screening criterion as an evidence-based measure to prevent VRE spread in hospitals [ 19 , 20 ]. Additionally, cost-benefit analyses have shown that targeted screening is a more practical approach for VRE detection [ 21 ]. Therefore, given the complexity of transmission pathways and VRE patient identification, innovative approaches are needed for early predictive detection of potential VRE carriers to prevent subsequent colonization or infection of other patients.

Our current knowledge is that nosocomial VRE transmission occurs mainly through direct contact or through fomites, with healthcare workers’ hands being the primary vector. VRE are highly stable in the environment and can survive for months in dry conditions [ 8 – 14 ]. When surfaces are not properly decontaminated, they act as permanent reservoirs for VRE, affecting not only the initial patient but also subsequent patients in the same room or nearby [ 15 – 17 ]. This knowledge has led to the consideration that reduction of shared surfaces between VRE colonized and non-colonized patients – hence contact isolation – may reduce the number of nosocomial transmissions.

VRE are Gram-positive cocci found in the human gut. Infections - like bloodstream or urinary tract infections - commonly occur subsequent to colonization and lead to prolonged hospitalization, increase patient suffering and significantly worsen the prognosis of affected patients [ 4 ]. A European systematic review analyzing data over a 10-year period reported a mortality rate of 33.5% [ 5 , 6 ]. Certain patient characteristics such as bone marrow transplant, central venous catheter, bedsores, and concurrent Clostridioides difficile infection, increase the susceptibility of patients to VRE [ 7 ]. Therefore, the development of effective strategies to slow down the spread of VRE and prevent their nosocomial transmission is crucial to avoiding additional risks, particularly for these critically ill patients.

Given the global rise in antibiotic resistance, hospitals must develop effective strategies to combat nosocomial infections caused by multidrug-resistant organisms. According to the WHO Bacterial Priority Pathogen List [ 1 ], vancomycin-resistant enterococci (VRE) are a significant threat to patient safety and causing a growing burden of disease. In Europe. A rising number of reported enterococcal bloodstream infections (BSI; +57.3%) – and especially of VRE BSI (+80.0%) – was reported between 2017-2021 [ 2 ]. Moreover, in 2019, 200,000 deaths worldwide were attributable to or associated with VRE [ 3 ].

As the length of the training period of the dynamic model was 30 days, these data were selected to calculate the values discussed thus far and are presented in Table 1 . To evaluate how this number influences the model performance we varied it to 60 and 90 days in combination with all available patient characteristics. However, no significant difference was found in any of the evaluation metrics with 60 days of training (F1 score=0.884 (STD=0.033)) or 90 days of training (F1 score=0.882 (STD=0.026)) (30 days vs. 60 days: p F1 =0.841, p sensitivity =1.000, p specificity =0.841; 30 days vs. 90 days: p F1 =0.841, p sensitivity =0.690, p specificity =0.690), although the significance of the statistical test is severely limited by the low power.

When analyzing the impact of the features, it can be seen that including OPS codes (F1 score=0.881, STD=0.029) and ICD codes (F1 score=0.853, STD=0.026) has the highest positive impact on performance, resulting in a significantly improved F1 score in the dynamic model compared to having no features at all (no features vs. OPS codes: p F1 =0.008; no features vs. ICD codes: p F1 =0.008; no features vs. all: p F1 =0.008). S1 Table includes F1 scores; sensitivity; specificity; precision; recall; accuracy; negative predictive value; positive predictive value; standard deviations from cross-validation or time-curve shifts; p-values for all evaluated model setups and feature combinations. When comparing the dynamic model with all features pairwise against the dynamic model with ICD codes, the model with OPS codes, and the model with both ICD and OPS codes, no statistically significant differences were observed (all features vs. OPS codes: p F1 =1.000, power = 0.050; all features vs. ICD codes: p F1 =0.421, power = 0.273; all features vs. OPS and ICD codes: p F1 =0.841, power = 0.058). However, the statistical power of the analysis is limited by the small sample size (five), suggesting that these results should be interpreted with caution. Nonetheless, the data imply that any potential differences, if present, are likely to be small.

The results clearly demonstrated the superiority of the dynamic model as well as the positive impact of incorporating specific patient features such as OPS codes (codes for operations and procedures) and ICD codes (codes for diagnosis), which significantly improved prediction quality. When no patient information was used at all and only the movement and relation network represented by the graph was relied upon, the static model achieved an F1 score of 0.497 (standard deviation (STD) STD=0.064), while the dynamic model predictions resulted in a significantly greater F1 score of 0.670 (STD=0.019). In contrast, when all the available patient features were combined, the static model achieved an F1 score of 0.685 (STD=0.023) while the dynamic model resulted in a significant improvement of roughly 20% with a F1 score of 0.880 (STD=0.029), and the values for sensitivity and specificity also increased markedly. In general, significance tests showed that the dynamic model consistently achieved a significantly better F1 score than did the static model with the same set of features in all setups.

To enhance the expressiveness of the GNN, we assigned node features to the patient nodes and evaluated different feature combinations (no features, one of the feature groups, combination of features, all features) to determine their relevance. Table 1 presents the macro F1 scores, sensitivity, and specificity for both the static model (average results using five-fold cross-validation: a proportion of all patients is classified here) and the dynamic model (average results over five runs distributed equally throughout the year: all patients hospitalized during the test period are classified here). The scores in each row are reported based on results obtained using the individual features for training and testing listed in the “Features” column. Fig 4 provides a graphical comparison of the F1 scores between the static and dynamic models.

Fig 3. Patient labels over time: on each day, the respective values represent the absolute number of patients who fell in each of the four included categories. pos_total = all positive cases known on that day; pos_new = all new positive cases added to the total positive cases on that day; neg_total = all negative cases known on that day; unknown = not falling in either category that day.

For training, validation, and testing of our classification approach all patients in the dataset were labeled VRE-positive or VRE-negative. Given that this labeling is based on microbiological test results which are unavailable for all patients in the dataset, some patients are also treated as unknown. Using our defined labeling schema, the distributions for the static model were 247 positive, 801 negative, and 7,328 unknown instances. The distributions for the dynamic model were 53,760 positive, 81,809 negative, and 708,064 unknown instances. The high numbers in the dynamic setting result from the fact that for this model the hospital graphs are generated on a daily basis, each containing all patients hospitalized on the respective day. The labels assigned to the patient nodes in the separate graphs are also determined on a daily basis. Fig 3 shows this distribution on a daily basis.

Fig 2. (A): Graph of the "uninhabited" hospital: Each of the 39 wards is represented by a node (blue), as are all 606 rooms (magenta), 770 bedplaces (brown) and 150 organizational units (org_unit) (green). The edges between the nodes show the links (room, on, ward), (room, on, org_unit) and (bed, in, room). (B): Graph of the "living" hospital: In addition to the nodes shown in (A), this graph contains the patients as a further type of node (red) and the new edges (patient, on, ward), (patient, in, room), (patient, in, bed), and (org_unit, visits, patient), but only for one day (here day 100).

To train our deep learning models and evaluate the approach, we utilized a historical dataset from one year (2019, deliberately chosen before the COVID-19 pandemic) from our tertiary care hospital. This dataset includes data from 8,372 patients (sex: 5,128 males, 3,244 females; age distribution: 0-10 years: 55, 11-20 years: 59, 21-30 years: 327, 31-40 years: 477, 41-50 years: 570, 51-60 years: 1,328, 61-70 years: 1,969, 71-80 years: 1,913, 81-90 years: 1,462, 91-100 years: 208, 101-110 years: 4), representing approximately one-third of the patients hospitalized in our facility during 2019. The dataset contains 125,947 interactions of these patients with the hospital environment, such as admissions, discharges, transfers between departments and rooms for examinations or treatments. The patients interacted with 39 departments, 606 rooms, 770 beds, and 150 organizational units. Fig 2 displays the heterogeneous graph of the hospital, which forms the basis of our models, once without patients (A) and once with the patient movement data for a single day (B).

Discussion

In this study, we successfully developed an AI-based method that uses hospital movement data within heterogeneous graph structures to predict the putative VRE status of patients. The findings from various studies [4,7,10–17] indicating that VRE transmission occurs through contact with VRE carriers and surface contamination inspired us to model hospitals as nodes and edges, thereby replicating transmission routes.

The simplest static model, excluding features such as OPS and ICD codes, showed good sensitivity but low specificity (sensitivity=0.848 (STD=0.099); specificity=0.401 (STD=0.119)). This means that too many cases are falsely classified as positive. By integrating additional features, this model became more precise and accurate (all patient features; sensitivity=0.624 (STD=0.041); specificity=0.788 (STD=0.027)). A fundamental issue with this model is the distortion of the data representation, as it models spatial dependencies that do not exist in reality. For instance, in our model, patients hospitalized at different times within the year are indirectly connected through shared rooms and wards, creating links across these nodes despite months of separation, thereby introducing noise into the data. This means that this model estimates VRE risk per ward/room averaged over the entire observation period.

The improvement from the static to the dynamic model through the integration of temporal information flow is based on enhanced representation of patient histories. The GNN can now learn that certain combinations of diagnoses or procedures occurring in specific temporal patterns lead to a change from VRE negative to VRE positive. Additionally, the time dependence of VRE outbreaks in wards was resolved, accounting for differences in VRE risk per ward/room over time. This is reflected in an average improvement of approximately 20% in all evaluated metrics (see S1 Table) compared to the static model in the all-features setup, which is substantial.

The key predictors of VRE status are diagnosis and procedure codes (OPS: sensitivity=0.753, specificity=0.937; ICD: sensitivity=0.820, specificity=0.933; OPS and ICD: sensitivity=0.796, specificity=0.958), which offer more extensive data than individual features such as bedsore or care level, which were identified as risk factors by Meschiari et al. [7], for example. Additionally, incorporating these individual characteristics does not yield significant improvements in model performance when ICD and OPS codes are already included, indicating that combining all available patient features does not necessarily result in the best performance. Given the high dimensional input data, the results also highlight the model’s ability to extract relationships from large datasets, demonstrating the efficacy of AI and neural networks. An analysis of the training period indicated that 30 days was sufficient for robust results (sensitivity=0.808, specificity=0.942), with additional days having minimal impact (60 days: sensitivity=0.811, specificity=0.946; 90 days: sensitivity=0.816, specificity=0.938).

Due to the lack of standardized methods for identifying patients at-risk of VRE infection, no established baselines are currently used in hospitals, which limits direct comparison with our findings. Comparing our setup to existing work from Van Niekerk et al. [41] and Gouareb et al. [42], who classified patients retrospectively after infections occurred, their approaches correspond to our static model. In contrast, our dynamic model approach not only classifies existing VRE carriers but also predicts patients with newly acquired VRE colonizations up to three days before to proven VRE diagnosis. Additionally, our heterogeneous graphs include all available spatial units — such as organizational units, wards, rooms, and even beds — and thus account for differences in node and edge types compared to homogeneous settings in related work, meeting the need for a detailed model of the environment to track VRE transmission chains [15–17]. In addition, we offer a comprehensive temporal representation by leveraging recent methods in temporal graph learning. This approach is much more comprehensive than the approaches used in these existing studies. When comparing the sensitivity and specificity scores from these two studies to our results, only Gouareb et al. [42] reported these metrics, whereas Van Niekerk et al. [41] focused on ward-level predictions and reported on the area under the curve (AUC) metric instead. Gouareb et al.’s best-performing setup demonstrated comparable performance (sensitivity = 0.806, specificity = 0.966) when addressing a broad range of multidrug-resistant bacteria. However, different bacteria have variable transmission pathways, patient populations at risk and pathogenicity, making the feasibility of this score insufficient for the evaluation of VRE in particular due to a lack of precision. While our experimental setup is different to this work (we focus on VRE as one type of multi-resistant bacteria and predict infections up to three days in advance) it places our findings in line with prior research and considers the specific insight into the characteristics of this particular pathogen.

Future work should focus on deploying our system in a “live” setting within the ICP workflow, which leads us to discuss practical implications and alignment with existing hospital processes. The question of what constitutes an ideal interface between a human and an AI application has attracted a lot of attention recently and the answer to that question clearly depends on each individual use case [43]. As outlined in the Methodology, a GPU with 6GB of VRAM is sufficient to retrain our proposed model daily (even with increasing graph complexity, e.g., a higher number of patients) to maintain an up-to-date predictor. The most challenging part of setting up the system is integrating existing hospital information systems, which house the relevant data, with the graph representation upon which our algorithm is based. This might require manual effort by technical and medical experts in individual environments. However, once the relevant data are matched, the remaining steps of model training and infection prediction are straightforward. After deployment, another challenge is making the system easily accessible to IPC personnel. The ideal solution may involve developing a user-friendly interface that bridges hygiene experts responsible for screening with the algorithm predicting at-risk patients. However, developing such an interface requires additional effort, such as identifying an easily accessible and user-friendly presentation of the information or ensuring that there is no misinterpretation of the algorithm’s screening recommendations. These can then serve as guidance for but not replace IPC team.

It is important for us to emphasize that our project, made possible only through the digitization of patient records, aims to give this documentation a broader purpose by actively utilizing the resulting dataset to improve diagnostics, therapy, and patient care. In the future, we hope to contribute to evidence-based data utilization for the benefit of patients and to support efforts in combating multidrug-resistant pathogens and antibiotic resistance.

[END]
---
[1] Url: https://journals.plos.org/digitalhealth/article?id=10.1371/journal.pdig.0000821

Published and (C) by PLOS One
Content appears here under this condition or license: Creative Commons - Attribution BY 4.0.

via Magical.Fish Gopher News Feeds:
gopher://magical.fish/1/feeds/news/plosone/