Early diagnosis of aortic aneurysms based on the classification of transfer function parameters estimated from two photoplethysmographic signals. (2021)

• Record Type: Journal Article
• Authors: Hackstein, Urs; Krüger, Tobias; Mair, Alexander; Degünther, Charlotte; Krickl, Stefan; Schlensak, Christian; Bernhard, Stefan
• Is Part Of: Informatics in medicine unlocked. Volume 25(2021)
• Publication Date: 2021
• Subjects: Aortic aneurysm -- Clinical study -- ARMAX-models -- Parameter estimation -- Frequency response -- Classification; Medical informatics -- Periodicals; 610.285
• DOI: 10.1016/j.imu.2021.100652
• Languages: English
• ISSNs: 2352-9148

Abstract

Abstract: Background: Cardiovascular diseases are the leading cause of death worldwide. Particularly aortic aneurysms are problematic and underdiagnosed: Existing diagnostic techniques either lack sensitivity or require specialized expertise, are invasive and costly which hampers effective screening. The objective of this proof-of-concept study is to develop an easy-to-use, non-invasive diagnostic method to generate suspicious facts about the presence of aortic aneurysms at the family physician level. This study is based on previous in-silico results and reports on the first clinical study ever to validate attempts to diagnose aortic aneurysms using machine learning techniques. Method: Two feature selection approaches for a classification "control group vs. aneurysms in the thoracic or abdominal aorta" using Naive Bayes and K-Nearest-Neighbor-algorithms were applied to in-vivo data from a clinical study including 55 patients. The first attempt based on parameter estimation of AutoRegressive-MovingAverage (ARMAX)-models, the second on frequency response of the transfer function, both computed using two peripheral photoplethysmographic signals. Results: In both cases only around 60% overall accuracy was achieved, but as classifiers trained and tested with randomly permuted labels show less accuracy, an intrinsic effect of aneurysms could be verified. Compared to previous in-silico results, the lower accuracy can be explained by low signal-to-noise ratio of the sensors used in connection with the low peripheral perfusion of the highly morbid patients, variability and number of patients in the clinical study. Conclusion: Both approaches have shown basic classification quality in a proof-of-concept clinical setting, however to overcome remaining uncertainties, training of the classifier with a larger number of patients is necessary. Nevertheless, we assume, that the method will improve the interpretation of cardiovascular signals in the early diagnosis of aortic aneurysms in near future. (clinical study approved by Ethics Committee, University Hospital Tübingen (Germany) (No. 136/2018801)). Highlights: First clinical study to validate machine learning techniques to diagnose aneurysms. Can detect influence of aortic aneurysms on selected features. Use ARMA-coefficients or frequency response elements as features for classification. Determine possible improvements of presented method.