Real-time prediction of atrial fibrillation in intensive care unit: a meta-learning approach.

Objective: Atrial fibrillation (AF) is a common arrhythmia affecting a large fraction of patients in intensive care units (ICUs). Predicting patients at risk of AF in the ICU is a challenging task and is not commonly practiced but could have clinical implications considering that AF is a proxy for poorer outcomes. Our goal is to develop a real-time AF prediction model using ICU numerical data to improve alarm quality, helping clinicians to identify at-risk patients and intervene before AF onset, thereby potentially improving clinical outcomes. Methods: We employed AmsterdamUMCdb, Europe's first openly accessible ICU dataset. The dataset includes static features, including demographics, as well as dynamic bedside monitoring data, like blood pressure and respiratory rate, along with detailed records of medication and fluid administration. To enhance generalizability across diverse ICU patients, we trained a long short-term memory (LSTM) model combined with a Model-Agnostic Meta-Learning (MAML) approach. Model performance was tested on an imbalanced test set, reflective of the real-world ratio of AF to non-AF patients. Direct external validation was performed using the MIMIC-IV database to test generalizability across different clinical settings. Results: The LSTM-MAML model achieved an AUC of 0.92, accuracy of 0.89, and precision of 0.29 on the internal validation set. In external validation with MIMIC-IV, it showed near-equivalent performance with an AUC of 0.89, accuracy of 0.85, and precision of 0.18. Conclusions: The real-time prediction model demonstrated predictive value for AF and has potential for future clinical benefits. However, for clinical implementation, further refinement is necessary to improve its performance in identifying patients at risk for AF.