Neural networks as a tool to predict syncope risk in the Emergency Department

The optimal disposition approach to patients presenting to an Emergency Department (ED) with syncope is unclear. Many low-risk patients with syncope are unnecessarily admitted to hospital. This may increase risk associated with hospitalization (including medication errors and hospital-acquired infections), and to an excessive use of resources.1 On the other hand, patients inappropriately discharged from the ED may experience serious adverse events or even death that may have been preventable with hospital-based interventions.2-4 This can partially explain the high heterogeneity in practice and hospitalization rate observed from multiple reports obtained from different countries.4,5 Despite many attempts to optimize ED syncope management, such as the use of structured management pathways and ED syncope observation units, a rigorous and effective approach remains elusive.6,7 Several syncope prediction tools have been developed to guide clinician decision-making in the ED.8-12 However, none has proved superior to clinical practice.5,6,13 Artificial neural networks (ANNs) are complex non-linear models inspired by the working of biological neural networks (i.e. the central nervous system). They are used to estimate complex functions (i.e. non-linear) that require a large number of inputs. Artificial neural networks are presented as systems of layers (multilayer) composed of neurons (also called perceptrons) which exchange messages between each other by synapsis (weights). The synapses have numeric weights that can be tuned based on experience, making neural nets adaptive to inputs and capable of learning. 14 As one of the major problems in syncope risk stratification is that syncope itself can be the final common presentation of several conditions which are very heterogeneous in terms of prognosis, the absence of linearity in such a context could make the application of ANNs appealing.15 The use of ANNs has already shown promising results in emergency medicine. For example, ANNs have been developed to reduce computed tomography imaging for suspected craniocervical junction injury in major head trauma patients.16 Artificial neural networks have also been used to predict risk ofmyocardial infarction in patients with chest pain.17 The aim of our study was to investigate the effectiveness of ANNs as a short-term risk stratification tool for syncope patients in the ED.