Addressing the critical issue of Carbapenemase-Producing Enterobacteriaceae (CPE) within hospital infection prevention and control, this study applies an eXplainable AI (XAI) model to analyze patient outcomes using Electronic Medical Records (EMR) from an Irish hospital. The focus was a previously underexplored dataset from 2018 to 2022 involving variables such as diagnostic codes, ward transitions, patient demographics, and infection-related aspects. Transformer-based architectures, like TabTransformer and TabNet, were used alongside conventional machine learning models. The outcomes revealed the superior performance of Transformer-based models in effectively predicting clinical conditions, particularly concerning acquisition of CPE.
The study employed an XAI method to interpret the results, showing key risk factors for CPE such as ‘Area of Residence’, ‘Admission Ward’, and prior admissions. The ‘Ward PageRank’ network variable also stood out, shedding light on the value of exposure structure data. The transparent explainable AI framework used in the study highlighted the importance of a diverse set of clinical and network features in identifying risk factors and predicting CPE-related outcomes.
Further emphasizing the importance of infection prevention and control (IPC) efforts, the study identified multidrug-resistant organisms like CPE as a severe threat, predominantly affecting treatment efficacy within clinical environments. The world health organization has officially recognized healthcare-associated infections (HCAIs) as a potential patient safety issue. In Ireland, the steady growth in CPE cases over the past decade underscores the urgent need for advanced surveillance, screening strategies, and data-driven models to facilitate timely IPC interventions.
Creating AI technology that incorporates EMRs and applies machine learning to examine the relationship between transmission risk, patient characteristics, and their clinical outcomes, is part of the study’s focus. Such technology can potentially offer crucial insights supporting both risk stratification and resource planning in healthcare setups.
Transformer-based models, such as Med-BERT, BEHRT, and TabNet have shown substantial performance in patient outcome predictions using EMR. However, the use of these models in the case of CPE has been limited, especially within the Irish healthcare context. This study aims to fill that gap by using these models to predict CPE outcome, adding contact network features, and incorporating explainable AI tools for the provision of actionable insights.
The study detailed an elaborate process utilizing anonymized EMR from an Irish acute hospital to model CPE risk and outcomes. A mix of machine learning models, like TabNet, ResNet, and Transformer-based architectures, was used in conjunction with contact network-derived features. The framework ensured explainability by using multiple XAI techniques to decipher feature contributions and instill trust in model-driven IPC strategies.
Source: https://bmcmedinformdecismak.biomedcentral.com/articles/10.1186/s12911-025-03214-1