Within this study, segmental bioimpedance measurements combined with standard 12-channel ECG were investigated for their ability to detect AAA in a mixed study population consisting of patients both with and without AAA as well as healthy controls. Due to the hyper dimensionality problem caused by the 225 output variables and limited sample size in this pilot study feature selection was performed and four different machine learning models were subsequently established. Model 1, a random forest applied to a filtered dataset and with selected predictors based on Boruta, showed satisfying performance in its ability to detect AAAs – no false negatives and one false positive in the test partition. Model 2 was the best performing model with 100% accuracy in the test sample. Model 3, a LASSO regression performed on a filtered dataset, had the worst performance with a specificity of 77.8% and a sensitivity of 66.7%. Model 4, the most conservative model, also showed reliable performance regarding the evaluation of a potential screening tool with 100% sensitivity in the test set. While solid regression models to approximate maximum AAA diameter could not be established, association analysis resulted in promising candidates for further exploration in larger cohorts such as the variance of the time interval of signal detection in the lower extremities.
The potential application of bioimpedance analysis for the detection of aneurysms has been previously discussed and repeatedly been praised as a non-invasive, low-cost technology.[19, 24, 25] Additionally, changes in body composition have been recently shown to be associated with abdominal aortic aneurysm growth after endovascular treatment.[26] The CombynECG uses segmental bioimpedance including a routine ECG that can also reliably assess body composition[20, 21] and grants insight to hemodynamic parameters such as the aortofemoral pulse wave velocity[27]. Based on these premises it appears very reasonable that the device can facilitate AAA diagnosis and provide a non-invasive screening tool.
The required training sample size for the application of machine learning algorithms is currently discussed and a matter of great interest in clinical research.[28] The sample size in the present study certainly constitutes a limitation and is prone to introduce bias during model development and might inflate accuracy in the test partitions.[29] Nevertheless, even the conservative conventional logistic regression, model 4, had good accuracy during prediction. Therefore, performance of the advanced machine learning models is unlikely to be driven by bias alone. Technical issues during the execution of CombynECG measurements were encountered that led to missing data points. A filtering workflow that included removing variables with missing data points for example led to the loss of more than 50% of all possible predictors. However, these errors are at least partially attributed to the learning curve associated with the use of new diagnostic devices. Furthermore, even the loss of possible predictors resulted in acceptable performance as evidenced with model 3, and even our most conservative approach in the present project, model 4, had a sensitivity of 100%. The best performing model was established by statistically imputing missing data points on a k-nearest neighbor graph, therefore creating an inherent bias. However, all four models showed predictive ability thereby cross validating our main finding: AAA detection via bioimpedance analysis is not only theoretically possible but is achievable with open market accessible devices.
In the present work we present a potential solution in the quest for radiation- and contrast-enhancement-free diagnostic tools in the form of a non-invasive cardiovascular screening device. The added benefit of the device is the absent inter-rater variation of ultrasound studies and its simultaneous assessment of other physiological features and pathologies such as body composition[20] or heart failure[21]. Ultimately, promising results were discovered with a device that was not specifically designed for the research question at hand. While residual confounding as a limitation of our data analysis certainly persists and analyses were conducted in a limited sample size, the level of evidence legitimizes further exploration in larger trials. With the renaissance of bioimpedance for diagnostics, devices could be engineered to optimize their potential for AAA detection and even size approximation, for example through the application of additional abdominal electrodes. While it cannot supplant imaging modalities for the planning of open or endovascular treatments, it might be able to facilitate simple and cost-effective screenings and subsequent surveillance.