# Association between deep learning–based atrial fibrillation burden and in-hospital mortality

## TL;DR

High atrial fibrillation burden in ICU patients is linked to increased in-hospital mortality, and deep learning can help predict outcomes using routine heart monitoring data.

## Contribution

This study introduces AF burden as a dynamic, real-time predictor of mortality in critically ill patients using deep learning analysis of ECG data.

## Key findings

- High AF burden was associated with significantly higher in-hospital mortality (18.1% vs. 8.6%).
- AF burden was identified as an independent risk factor with an adjusted odds ratio of 1.63.
- Machine learning models showed AF burden contributes to mortality prediction with an area under the curve of 0.86.

## Abstract

Despite its clinical significance, research on atrial fibrillation (AF) burden as a dynamic, real-time predictor of adverse outcomes in patients with critical illness is lacking. This study examined the association between high AF burden and in-hospital mortality in critically ill patients, using intensive care unit (ICU) data from the Medical Information Mart for Intensive Care III (MIMIC-III; 2001–2012) and Yongin Severance Hospital (2021–2023). Electrocardiogram waveform data were analyzed using deep learning models to calculate AF burden. Adult ICU patients were included, with exclusion of those aged ≥90 years and those with an AF burden >0.9. AF burden was defined as the ratio of AF waveforms to total waveforms during ICU admission, with a high burden defined as ≥7.0%. Logistic regression and machine learning models were employed to assess the association between AF burden and in-hospital mortality, as well as to evaluate the contribution of AF burden to mortality prediction. From the MIMIC-III database, 7,734 patients were included: 5,734 (74.1%) had a low AF burden (median, 0.3%) and 2,000 (25.9%) had a high AF burden (median, 22.5%). High AF burden was associated with significantly higher in-hospital mortality (18.1% vs. 8.6%, P < 0.001) and was identified as an independent risk factor (adjusted odds ratio, 1.63; 95% confidence interval, 1.36–1.95; P < 0.001). Machine learning models demonstrated that AF burden is a significant contributor to mortality prediction, with an area under the curve of 0.86. AF burden may serve as a dynamic marker for real-time alerts of clinical deterioration and for risk stratification in critically ill patients.

When people become critically ill and are admitted to the intensive care unit, irregular heart rhythms such as atrial fibrillation are common and can be dangerous. In the past, most research has looked at atrial fibrillation simply as present or absent. However, this approach ignores how much time a patient actually spends in this rhythm. In our study, we measured the total “burden” of atrial fibrillation, meaning the percentage of time a patient’s heart was in this rhythm during their stay in the intensive care unit. We analyzed over 7,700 patients from a large public hospital database in the United States and confirmed our results using data from a Korean hospital. We found that patients with a high atrial fibrillation burden had a significantly higher risk of dying in the hospital. Using deep learning and machine learning methods, we also showed that atrial fibrillation burden was an important factor in predicting patient outcomes, alongside age, sepsis, and use of a ventilator. Because heart rhythm monitoring is already part of routine care in intensive care units, our approach could allow doctors to identify high-risk patients in real time, without extra cost or procedures, and potentially guide early interventions.