Comparison of machine learning techniques in the early detection of abdominal aortic aneurysms from in-vivo photoplethysmography data

Abstract

Digitalisation and the generation of large amounts of data from examinations and medical interventions are making an ever-increasing amount of data available, of which only a fraction has been used to date. In this context, artificial intelligence and artificial neural networks play an emerging role generating new insights from these data. In particular, new models for description and understanding are generated, and new methods for early diagnosis make use of artificial intelligence. In this article, we consider the case of cardiovascular diseases that are the leading cause of death worldwide. Aortic aneurysms are particularly problematic and underdiagnosed: Existing diagnostic techniques are not suitable for effective screening at the family physician level. Several authors recently made attempts to develop a new, easy-to-use, non-invasive diagnostic method for the early detection of aortic aneurysms using machine or deep learning techniques. In a previous study, the authors achieved promising but still expandable results using machine learning techniques on non-invasive in-vivo data. Here, a variation is presented that increases the accuracy of the classification by 4%. We compare these results with those for other classifiers and classification approaches and an approach via a bidirectional recurrent neural network for diagnosing abdominal aortic aneurysms from the current literature. Applying a classification approach from the previous study of the authors provides accuracy values of more than 86% on an in-silico photoplethysmography dataset that served as a training set for the bidirectional recurrent neural network in the literature. The results are better the younger the virtual patients and the larger the aneurysms and are sufficiently good even with much smaller subsets of the datasets. The latter is important for the practical use of the methods on in-vivo data, where normally only smaller datasets are available. The bidirectional recurrent neural network was trained on in-silico data and the article in the literature suggested that it will work also on in-vivo data. Application of the neural network trained on in-silico data to our in-vivo data could not support this assumption because the sensitivity value received was lower than 50% (or 40%, respectively), depending on the choice of the dataset.