Postoperative atrial fibrillation (PAF) affects a large fraction of patients undergoing cardiac surgery. It is associated with increased postoperative mortality and morbidity, and also results in longer and more expensive hospital stays. While treatments such as prophylactic administration of beta-adrenergic blockers and amiodarone can reduce the incidence of PAF, existing tools to match patients to treatments that are appropriate for their risk need to be improved. Many high-risk patients may benefit from more aggressive treatment than they receive presently. Conversely, many low-risk patients are currently treated in excess of their actual risk of developing PAF.
The aim of our research is to develop novel computational markers that can be used to risk stratify patients undergoing cardiac surgery for PAF. Our research uses sophisticated signal processing and machine learning techniques to discover information in the electrocardiogram (ECG) signal related to autonomic and impulse conduction abnormalities. We describe how this information can be combined with existing metrics to identify high risk patients. In addition, we propose the creation of a large public dataset of ECG signals with detailed patient meta-data and outcomes for research by the broader scientific community on predicting PAF.
The specific aims of this proposal are: (1) To develop a public database with ECG signals and detailed patient metadata for research on predictive PAF. We will collect ECG data from over 900 patients undergoing cardiac surgery at the University of Michigan Medical Center during the first 12 months of this project with an expected PAF incidence of around 30%. We will share this data and other clinical information (i.e., demographics, comorbidities, laboratory reports, information related to the procedure, and the outcome of PAF) in a de- identified manner with the broader research community through PhysioNet;and (2) To develop and validate novel ECG-based risk markers for PAF. We will explore metrics based on the sympathovagal modulation of the heart, e.g., heart rate turbulence (HRT) and deceleration capacity (DC), which have shown recent promise in predicting ventricular arrhythmias. We will also develop new metrics based on the shape of the ECG to assess atrial myocardial instability. These approaches will be validated on the data collected in Aim 1, and will be used to develop models based on non-parametric machine learning techniques to accurately assess PAF risk.
Postoperative atrial fibrillation commonly occurs in patients undergoing cardiac surgery (e.g., coronary artery bypass graft or valve repair). It is associated with an increased risk of both mortality and morbidity, and increases the cost of hospitalization significantly. Successful management of patients undergoing cardiac surgery therefore relies on the ability of clinicians to estimate the risk of postoperative atrial fibrillation, and ideally, to anticipate and prevent this event. We propose to develop low-cost, and widely applicable methods that would improve the ability of physicians to identify patients at high risk of postoperative atrial fibrillation. They key concept for this proposal is that ECG signals contain valuable information that can help predict high or low risk cases using sophisticated techniques from signal processing and machine learning. This information can be used to develop clinical tools to improve management of patients undergoing cardiac surgery. In addition to developing these tools, we also propose the creation of a large public database that can help other researchers with an interest in predicting postoperative atrial fibrillation.