Interpretable large language models for early prediction of antimicrobial multidrug resistance

The growing burden of Antimicrobial Resistance (AMR) in Intensive Care Units (ICUs) poses a significant threat to global health, increasing patient mortality, morbidity, and healthcare costs. Early prediction of AMR is essential for timely intervention and effective treatment. This study proposes novel Large Language Model (LLM)-based architectures for the classification of AMR in ICU patients, using Electronic Health Records (EHRs) modeled as irregular Multivariate Time Series (MTS).

We evaluated the proposed LLM-based models using a dataset of 3,502 anonymized EHRs from ICU patients at the University Hospital of Fuenlabrada (Madrid, Spain), collected between 2004 and 2020. Their performance was compared to that of benchmark Deep Learning (DL) models, including Gated Recurrent Units, Long Short-Term Memory networks, Transformer, Mamba, and Graph Convolutional Neural Networks. All models were trained on irregular MTS data with predictive windows of 4, 7, and 14 days, and model interpretability was assessed using SHapley Additive exPlanations (SHAP) values.

The LLM-based models significantly outperformed baseline DL architectures. In the 7-day prediction window, InstructTime-LLM achieved the best performance with a ROC-AUC of 0.792 ± 0.009, while Blocks-LLM showed comparable results in the 14-day window. SHAP analysis further highlighted clinically consistent risk factors, such as catheter use, antibiotic exposure, and microbial cultures.

This study demonstrates the potential of the proposed LLM-based architectures in AMR classification, combining high predictive performance with model interpretability. These models offer a promising foundation for trustworthy, early-warning systems in critical care.