Aims: Denoising autoencoders (DAEs) have the ability to understand complex patterns. The heart's electrical activity results from complex physiological mechanisms. Our aim is to use DAEs to learn the complex patterns found in body-surface ECG obtained from healthy individuals, and use this model to identify pathological activity.
Method: The database consists of body surface potential maps of 60 healthy individuals, each recorded using a 128-electrode device, and two groups of patients suffering from arrhythmogenic right ventricular dysplasia (ARVD) (n = 75) and idiopathic ventricular fibrillations (IVFs) (n = 45). For each diseased group, a cross-validation approach was used to select features relevant for discrimination from healthy individuals. Two DAEs were trained using healthy individuals with these selected features. For validation, a correspondence score was developed to assess the closeness of a test patient to the healthy population learned by the DAE: the input vector is partially masked to the DAE. The correspondence score is then defined as the root mean square error between the actual masked values and those predicted by the DAE.
Results: For both groups, 20 features were identified as significant. The median correspondence score for the healthy control group for the DAE trained with ARVD features is 0.66 [0.36-0.95] ([1st,3rd quartile]) versus 1.58 [0.52-2.02] for the ARVD group (p<0.05). For the healthy control group with IVF features the median correspondence score is 0.53 [0.26-0.82] versus 1.32 [0.42-2.17] for the IVF group (p<0.05). The use of the correspondence score for the discrimination between healthy individuals and those with pathology was assessed with ROC curve analysis, giving an area under the curve of 0.79 for ARVC and 0.78 for IVF, respectively.
Conclusion: By integrating dimensionality reduction techniques with DAE, we developed a new strategy for distinguishing healthy individuals from two specific pathologies: those with ARVD and those with IVF.